diff --git a/dnn/src/cuda/conv_bias/algo.cpp b/dnn/src/cuda/conv_bias/algo.cpp index a212ee4c..adf6a93c 100644 --- a/dnn/src/cuda/conv_bias/algo.cpp +++ b/dnn/src/cuda/conv_bias/algo.cpp @@ -90,6 +90,12 @@ ConvBiasForwardImpl::AlgoPack::AlgoPack() { for (auto&& algo : uint4_int4_nchw64_imma) { all_algos.push_back(&algo); } + for (auto&& algo : int4_int4_nhwc_imma) { + all_algos.push_back(&algo); + } + for (auto&& algo : uint4_int4_nhwc_imma) { + all_algos.push_back(&algo); + } #endif #endif fill_dp4a_algos(); @@ -247,6 +253,34 @@ void ConvBiasForwardImpl::AlgoPack::fill_imma_algos() { uint4_int4_nchw64_imma.emplace_back( AlgoParam{256, 128, 128, 64, 64, 128}); } + { + using AlgoParam = AlgoInt4Int4NHWCIMMAImplicitGemm::AlgoParam; + int4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 32, 64, 64, 32, 64, 32}); + int4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 32, 64, 64, 32, 64, 16}); + int4_int4_nhwc_imma.emplace_back(AlgoParam{128, 32, 64, 64, 32, 64, 8}); + int4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 64, 64, 64, 64, 64, 32}); + int4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 64, 64, 64, 64, 64, 16}); + int4_int4_nhwc_imma.emplace_back(AlgoParam{128, 64, 64, 64, 64, 64, 8}); + } + { + using AlgoParam = AlgoUInt4Int4NHWCIMMAImplicitGemm::AlgoParam; + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 32, 64, 64, 32, 64, 32}); + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 32, 64, 64, 32, 64, 16}); + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 32, 64, 64, 32, 64, 8}); + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 64, 64, 64, 64, 64, 32}); + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 64, 64, 64, 64, 64, 16}); + uint4_int4_nhwc_imma.emplace_back( + AlgoParam{128, 64, 64, 64, 64, 64, 8}); + } #endif } #endif diff --git a/dnn/src/cuda/conv_bias/algo.h b/dnn/src/cuda/conv_bias/algo.h index 869777b6..691298c6 100644 --- a/dnn/src/cuda/conv_bias/algo.h +++ b/dnn/src/cuda/conv_bias/algo.h @@ -63,6 +63,8 @@ public: CUDA_IMPLICIT_GEMM_IMMA_NCHW32_INT8, CUDA_IMPLICIT_GEMM_IMMA_NCHW64_INT4_INT4, CUDA_IMPLICIT_GEMM_IMMA_NCHW64_UINT4_INT4, + CUDA_IMPLICIT_GEMM_IMMA_NHWC_INT4_INT4, + CUDA_IMPLICIT_GEMM_IMMA_NHWC_UINT4_INT4, CUDA_BFLOAT16, CUDA_IMPLICIT_GEMM_SASS_NCHW4_DOTPROD_INT8, CUDA_IMPLICIT_GEMM_1X1_SASS_NCHW4_DOTPROD_INT8, @@ -896,6 +898,133 @@ private: void update_bias(const ExecArgs& args, void* updated_bias, void* reduce_filter_ptr, void* reduce_workspace) const; }; + +class ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase : public AlgoBase { +public: + struct AlgoParam { + int threadblock_m; + int threadblock_n; + int threadblock_k; + int warp_m; + int warp_n; + int warp_k; + int access_size; + }; + + AlgoInt4NHWCIMMAImplicitGemmBase(AlgoParam algo_param) + : m_algo_param(algo_param) {} + + AlgoAttribute attribute() const override { + return AlgoAttribute::REPRODUCIBLE; + } + const char* name() const override { return m_name.c_str(); } + std::string param() const override; + + bool is_available(const SizeArgs& args) const override; + void exec(const ExecArgs& args) const override; + + std::string to_string(AlgoParam algo_param); + +protected: + virtual DTypeEnum src_dtype() const = 0; + + // return filter_ptr, bias_ptr + virtual std::tuple prepare_filter_bias( + const ExecArgs& args) const = 0; + + // return alpha, beta, gamma, delta, theta + virtual std::tuple get_constants( + const ExecArgs& args) const = 0; + + virtual void do_exec(const ExecArgs& args, void* filter_ptr, void* bias_ptr, + void* z_ptr, convolution::ConvParam kern_param, + uint32_t nonlinear_mode, float alpha, float beta, + float gamma, float delta, float theta, + cudaStream_t stream) const = 0; + + void reorder_filter(const ExecArgs& args, int interleaved, + void* reordered_filter) const; + + std::string m_name; + AlgoParam m_algo_param; +}; + +class ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm final + : public AlgoInt4NHWCIMMAImplicitGemmBase { +public: + using Base = AlgoInt4NHWCIMMAImplicitGemmBase; + using AlgoParam = Base::AlgoParam; + + AlgoInt4Int4NHWCIMMAImplicitGemm(AlgoParam algo_param) : Base{algo_param} { + m_name = ConvBias::algo_name( + ssprintf("INT4_INT4_NHWC_IMMA_IMPLICIT_GEMM_%s", + to_string(m_algo_param).c_str()), + ConvBias::DirectParam{}); + } + + size_t get_workspace_in_bytes(const SizeArgs& args) const override; + size_t get_preprocess_workspace_in_bytes( + const SizeArgs& args) const override; + SmallVector deduce_preprocessed_filter_layout( + const SizeArgs& args) const override; + void exec_preprocess(const ExecArgs& args) const override; + + MEGDNN_DECL_ALGO_TYPE(CUDA_IMPLICIT_GEMM_IMMA_NHWC_INT4_INT4) + +private: + DTypeEnum src_dtype() const override { return DTypeEnum::QuantizedS4; } + + std::tuple prepare_filter_bias( + const ExecArgs& args) const override; + + std::tuple get_constants( + const ExecArgs& args) const override; + + void do_exec(const ExecArgs& args, void* filter_ptr, void* bias_ptr, + void* z_ptr, convolution::ConvParam kern_param, + uint32_t nonlinear_mode, float alpha, float beta, float gamma, + float delta, float theta, cudaStream_t stream) const override; +}; + +class ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm final + : public AlgoInt4NHWCIMMAImplicitGemmBase { +public: + using Base = AlgoInt4NHWCIMMAImplicitGemmBase; + using AlgoParam = Base::AlgoParam; + + AlgoUInt4Int4NHWCIMMAImplicitGemm(AlgoParam algo_param) : Base{algo_param} { + m_name = ConvBias::algo_name( + ssprintf("UINT4_INT4_NHWC_IMMA_IMPLICIT_GEMM_%s", + to_string(m_algo_param).c_str()), + ConvBias::DirectParam{}); + } + + size_t get_workspace_in_bytes(const SizeArgs& args) const override; + size_t get_preprocess_workspace_in_bytes( + const SizeArgs& args) const override; + SmallVector deduce_preprocessed_filter_layout( + const SizeArgs& args) const override; + void exec_preprocess(const ExecArgs& args) const override; + + MEGDNN_DECL_ALGO_TYPE(CUDA_IMPLICIT_GEMM_IMMA_NHWC_UINT4_INT4) + +private: + DTypeEnum src_dtype() const override { return DTypeEnum::Quantized4Asymm; } + + std::tuple prepare_filter_bias( + const ExecArgs& args) const override; + + std::tuple get_constants( + const ExecArgs& args) const override; + + void do_exec(const ExecArgs& args, void* filter_ptr, void* bias_ptr, + void* z_ptr, convolution::ConvParam kern_param, + uint32_t nonlinear_mode, float alpha, float beta, float gamma, + float delta, float theta, cudaStream_t stream) const override; + + void update_bias(const ExecArgs& args, void* updated_bias, + void* reduce_filter_ptr, void* reduce_workspace) const; +}; #endif class ConvBiasForwardImpl::AlgoBFloat16 final : public AlgoBase { @@ -955,6 +1084,8 @@ public: std::vector int8_nchw32_imma; std::vector int4_int4_nchw64_imma; std::vector uint4_int4_nchw64_imma; + std::vector int4_int4_nhwc_imma; + std::vector uint4_int4_nhwc_imma; #endif std::vector> gconv_refhold; AlgoBFloat16 bfloat16; diff --git a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu index 67de7770..e77a599a 100644 --- a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu +++ b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cu @@ -321,7 +321,8 @@ void megdnn::cuda::cutlass_wrapper:: ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp, \ cutlass::conv::threadblock:: \ ConvolutionFpropNCxHWxThreadblockSwizzle, \ - stage_, 4, aligned_, NeedLoadFromConstMem>; \ + stage_, 4, aligned_, NeedLoadFromConstMem, \ + cutlass::arch::OpMultiplyAdd>; \ typename Convolution::ConvolutionParameter conv_param( \ param.n, param.hi, param.wi, param.ci, param.co, param.fh, \ param.fw, param.ho, param.wo, param.ph, param.pw, param.sh, \ @@ -582,7 +583,8 @@ void megdnn::cuda::cutlass_wrapper:: ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp, \ cutlass::conv::threadblock:: \ ConvolutionFpropNCxHWxThreadblockSwizzle, \ - stages_, 4, aligned_, NeedLoadFromConstMem>; \ + stages_, 4, aligned_, NeedLoadFromConstMem, \ + cutlass::arch::OpMultiplyAdd>; \ typename Convolution::ConvolutionParameter conv_param( \ param.n, param.hi, param.wi, param.ci, param.co, param.fh, \ param.fw, param.ho, param.wo, param.ph, param.pw, param.sh, \ @@ -1037,4 +1039,262 @@ INST(true); INST(false); #undef INST +/* ====== cutlass kernel wrapper for int4 x int4 nchw64 layout ====== */ + +#if MEGDNN_TEGRA_X1 +template +void megdnn::cuda::cutlass_wrapper:: + do_conv_bias_int4_int4_implicit_gemm_imma_nhwc( + const int8_t* /* d_src */, const int8_t* /* d_filter */, + const int32_t* /* d_bias */, const int8_t* /* d_z */, + int8_t* /* d_dst */, int* /* workspace */, + const convolution::ConvParam& /* param */, + uint32_t /* nonlinear_mode */, float /* alpha */, + float /* beta */, float /* gamma */, float /* scale */, + const GemmCoord& /* threadblock_shape */, + const GemmCoord& /* warp_shape */, + const int32_t /* access_size */, cudaStream_t /* stream */) {} +#else +template +void megdnn::cuda::cutlass_wrapper:: + do_conv_bias_int4_int4_implicit_gemm_imma_nhwc( + const int8_t* d_src, const int8_t* d_filter, + const int32_t* d_bias, const int8_t* d_z, int8_t* d_dst, + int* workspace, const convolution::ConvParam& param, + uint32_t nonlinear_mode, float alpha, float beta, float gamma, + float scale, const GemmCoord& threadblock_shape, + const GemmCoord& warp_shape, const int32_t access_size, + cudaStream_t stream) { +#define DISPATCH_KERNEL_WITH_TILE_SHAPE(threadblock_m_, threadblock_n_, \ + threadblock_k_, warp_m_, warp_n_, \ + warp_k_, access_size_) \ + if (threadblock_shape.m() == threadblock_m_ && \ + threadblock_shape.n() == threadblock_n_ && \ + threadblock_shape.k() == threadblock_k_ && \ + warp_shape.m() == warp_m_ && warp_shape.n() == warp_n_ && \ + warp_shape.k() == warp_k_ && access_size == access_size_) { \ + using ThreadBlockShape = \ + cutlass::gemm::GemmShape; \ + using WarpShape = cutlass::gemm::GemmShape; \ + using InstructionShape = cutlass::gemm::GemmShape<8, 8, 32>; \ + using Convolution = cutlass::conv::device::Convolution< \ + cutlass::int4b_t, cutlass::layout::TensorNHWC, \ + cutlass::int4b_t, cutlass::layout::TensorNCxHWx, \ + ElementOutput, cutlass::layout::TensorNHWC, int32_t, \ + cutlass::layout::TensorNHWC, int32_t, \ + cutlass::conv::ConvType::kConvolution, \ + cutlass::arch::OpClassTensorOp, cutlass::arch::Sm75, \ + ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp, \ + cutlass::conv::threadblock:: \ + ConvolutionFpropNHWCThreadblockSwizzle, \ + 2, access_size_, access_size_, NeedLoadFromConstMem, \ + cutlass::arch::OpMultiplyAddSaturate, \ + cutlass::conv::ImplicitGemmMode::GEMM_TN>; \ + typename Convolution::ConvolutionParameter conv_param( \ + param.n, param.hi, param.wi, param.ci, param.co, param.fh, \ + param.fw, param.ho, param.wo, param.ph, param.pw, param.sh, \ + param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation); \ + return cutlass_convolution_wrapper( \ + reinterpret_cast(d_src), \ + reinterpret_cast(d_filter), d_bias, \ + reinterpret_cast(d_z), \ + reinterpret_cast(d_dst), workspace, \ + conv_param, epilogue, stream); \ + } +#define DISPATCH_KERNEL \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 32); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 16); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 8); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 32); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 16); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 8); \ + megdnn_assert(false, \ + "unsupported threadblock shape (%dx%dx%d) and warp shape " \ + "(%dx%dx%d) and access_size (%d)", \ + threadblock_shape.m(), threadblock_shape.n(), \ + threadblock_shape.k(), warp_shape.m(), warp_shape.n(), \ + warp_shape.k(), access_size); + using ElementOutput = cutlass::int4b_t; + using ElementAccumulator = int32_t; + using ElementBias = int32_t; + using ElementCompute = float; + using NonlineMode = megdnn::param_enumv::ConvBias::NonlineMode; + switch (nonlinear_mode) { + case NonlineMode::IDENTITY: { + using EpilogueOp = + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + ElementOutput, 8, ElementAccumulator, ElementBias, + ElementCompute>; + typename EpilogueOp::Params epilogue{alpha, beta, gamma}; + DISPATCH_KERNEL; + } + case NonlineMode::RELU: { + using EpilogueOp = cutlass::epilogue::thread:: + BiasAddLinearCombinationReluClamp< + ElementOutput, 8, ElementAccumulator, ElementBias, + ElementCompute>; + typename EpilogueOp::Params epilogue{alpha, beta, gamma, 0}; + DISPATCH_KERNEL; + } + case NonlineMode::H_SWISH: { + using EpilogueOp = cutlass::epilogue::thread:: + BiasAddLinearCombinationHSwishClamp< + ElementOutput, 8, ElementAccumulator, ElementBias, + ElementCompute>; + typename EpilogueOp::Params epilogue{alpha, beta, gamma, scale}; + DISPATCH_KERNEL; + } + default: + megdnn_assert(false, + "unsupported nonlinear mode for conv bias operator"); + } +#undef DISPATCH_KERNEL_WITH_TILE_SHAPE +#undef DISPATCH_KERNEL +} +#endif + +#define INST(need_load_from_const_mem) \ + template void megdnn::cuda::cutlass_wrapper:: \ + do_conv_bias_int4_int4_implicit_gemm_imma_nhwc< \ + need_load_from_const_mem>( \ + const int8_t* d_src, const int8_t* d_filter, \ + const int32_t* d_bias, const int8_t* d_z, int8_t* d_dst, \ + int* workspace, const convolution::ConvParam& param, \ + uint32_t nonlinear_mode, float alpha, float beta, \ + float gamma, float scale, \ + const GemmCoord& threadblock_shape, \ + const GemmCoord& warp_shape, const int32_t access_size, \ + cudaStream_t stream); +INST(true); +INST(false); +#undef INST + +#if MEGDNN_TEGRA_X1 +template +void megdnn::cuda::cutlass_wrapper:: + do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc( + const uint8_t* /* d_src */, const int8_t* /* d_filter */, + const int32_t* /* d_bias */, const uint8_t* /* d_z */, + uint8_t* /* d_dst */, int* /* workspace */, + const convolution::ConvParam& /* param */, + uint32_t /* nonlinear_mode */, float /* alpha */, + float /* beta */, float /* gamma */, float /* delta */, + float /* theta */, float /* scale */, + uint8_t /* src_zero_point */, + const GemmCoord& /* threadblock_shape */, + const GemmCoord& /* warp_shape */, + const int32_t /* access_size */, cudaStream_t /* stream */) {} +#else +template +void megdnn::cuda::cutlass_wrapper:: + do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc( + const uint8_t* d_src, const int8_t* d_filter, + const int32_t* d_bias, const uint8_t* d_z, uint8_t* d_dst, + int* workspace, const convolution::ConvParam& param, + uint32_t nonlinear_mode, float alpha, float beta, float gamma, + float delta, float theta, float /* scale */, + uint8_t src_zero_point, const GemmCoord& threadblock_shape, + const GemmCoord& warp_shape, const int32_t access_size, + cudaStream_t stream) { +#define DISPATCH_KERNEL_WITH_TILE_SHAPE(threadblock_m_, threadblock_n_, \ + threadblock_k_, warp_m_, warp_n_, \ + warp_k_, access_size_) \ + if (threadblock_shape.m() == threadblock_m_ && \ + threadblock_shape.n() == threadblock_n_ && \ + threadblock_shape.k() == threadblock_k_ && \ + warp_shape.m() == warp_m_ && warp_shape.n() == warp_n_ && \ + warp_shape.k() == warp_k_ && access_size == access_size_) { \ + using ThreadBlockShape = \ + cutlass::gemm::GemmShape; \ + using WarpShape = cutlass::gemm::GemmShape; \ + using InstructionShape = cutlass::gemm::GemmShape<8, 8, 32>; \ + using Convolution = cutlass::conv::device::Convolution< \ + cutlass::uint4b_t, cutlass::layout::TensorNHWC, \ + cutlass::int4b_t, cutlass::layout::TensorNCxHWx, \ + ElementOutput, cutlass::layout::TensorNHWC, int32_t, \ + cutlass::layout::TensorNHWC, int32_t, \ + cutlass::conv::ConvType::kConvolution, \ + cutlass::arch::OpClassTensorOp, cutlass::arch::Sm75, \ + ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp, \ + cutlass::conv::threadblock:: \ + ConvolutionFpropNHWCThreadblockSwizzle, \ + 2, access_size_, access_size_, NeedLoadFromConstMem, \ + cutlass::arch::OpMultiplyAddSaturate, \ + cutlass::conv::ImplicitGemmMode::GEMM_TN>; \ + typename Convolution::ConvolutionParameter conv_param( \ + param.n, param.hi, param.wi, param.ci, param.co, param.fh, \ + param.fw, param.ho, param.wo, param.ph, param.pw, param.sh, \ + param.sw, 1, 1, cutlass::conv::Mode::kCrossCorrelation); \ + return cutlass_convolution_wrapper( \ + reinterpret_cast(d_src), \ + reinterpret_cast(d_filter), d_bias, \ + reinterpret_cast(d_z), \ + reinterpret_cast(d_dst), workspace, \ + conv_param, epilogue, stream, {src_zero_point}); \ + } +#define DISPATCH_KERNEL \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 32); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 16); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 32, 64, 64, 32, 64, 8); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 32); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 16); \ + DISPATCH_KERNEL_WITH_TILE_SHAPE(128, 64, 64, 64, 64, 64, 8); \ + megdnn_assert(false, \ + "unsupported threadblock shape (%dx%dx%d) and warp shape " \ + "(%dx%dx%d) and access_size (%d)", \ + threadblock_shape.m(), threadblock_shape.n(), \ + threadblock_shape.k(), warp_shape.m(), warp_shape.n(), \ + warp_shape.k(), access_size); + using ElementOutput = cutlass::uint4b_t; + using ElementAccumulator = int32_t; + using ElementBias = int32_t; + using ElementCompute = float; + using NonlineMode = megdnn::param_enumv::ConvBias::NonlineMode; + switch (nonlinear_mode) { + case NonlineMode::IDENTITY: { + using EpilogueOp = + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + ElementOutput, 8, ElementAccumulator, ElementBias, + ElementCompute>; + typename EpilogueOp::Params epilogue{alpha, beta, gamma, + delta + theta}; + DISPATCH_KERNEL; + } + case NonlineMode::RELU: { + using EpilogueOp = cutlass::epilogue::thread:: + BiasAddLinearCombinationReluClamp< + ElementOutput, 8, ElementAccumulator, ElementBias, + ElementCompute>; + typename EpilogueOp::Params epilogue{alpha, beta, gamma, + 0, delta, theta}; + DISPATCH_KERNEL; + } + default: + megdnn_assert(false, + "unsupported nonlinear mode for conv bias operator"); + } +#undef DISPATCH_KERNEL_WITH_TILE_SHAPE +#undef DISPATCH_KERNEL +} +#endif + +#define INST(need_load_from_const_mem) \ + template void megdnn::cuda::cutlass_wrapper:: \ + do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc< \ + need_load_from_const_mem>( \ + const uint8_t* d_src, const int8_t* d_filter, \ + const int32_t* d_bias, const uint8_t* d_z, uint8_t* d_dst, \ + int* workspace, const convolution::ConvParam& param, \ + uint32_t nonlinear_mode, float alpha, float beta, \ + float gamma, float delta, float theta, float scale, \ + uint8_t src_zero_point, \ + const GemmCoord& threadblock_shape, \ + const GemmCoord& warp_shape, const int32_t access_size, \ + cudaStream_t stream); +INST(true); +INST(false); +#undef INST + // vim: syntax=cuda.doxygen diff --git a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cuh b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cuh index f2d7370d..0b695d33 100644 --- a/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cuh +++ b/dnn/src/cuda/conv_bias/cutlass_convolution_wrapper.cuh @@ -103,6 +103,25 @@ void do_conv_bias_int8_implicit_gemm_dp4a_ncdiv4hw4_nhwc( float scale, const GemmCoord& threadblock_shape, const GemmCoord& warp_shape, int stages, cudaStream_t stream); +template +void do_conv_bias_int4_int4_implicit_gemm_imma_nhwc( + const int8_t* d_src, const int8_t* d_filter, const int32_t* d_bias, + const int8_t* d_z, int8_t* d_dst, int* workspace, + const convolution::ConvParam& param, uint32_t nonlinear_mode, + float alpha, float beta, float gamma, float scale, + const GemmCoord& threadblock_shape, const GemmCoord& warp_shape, + const int32_t access_size, cudaStream_t stream); + +template +void do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc( + const uint8_t* d_src, const int8_t* d_filter, const int32_t* d_bias, + const uint8_t* d_z, uint8_t* d_dst, int* workspace, + const convolution::ConvParam& param, uint32_t nonlinear_mode, + float alpha, float beta, float gamma, float delta, float theta, + float scale, uint8_t src_zero_point, const GemmCoord& threadblock_shape, + const GemmCoord& warp_shape, const int32_t access_size, + cudaStream_t stream); + } // namespace cutlass_wrapper } // namespace cuda } // namespace megdnn diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_int4_int4_nhwc_imma.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_int4_int4_nhwc_imma.cpp new file mode 100644 index 00000000..5a72afc9 --- /dev/null +++ b/dnn/src/cuda/conv_bias/implicit_gemm_int4_int4_nhwc_imma.cpp @@ -0,0 +1,122 @@ +/** + * \file dnn/src/cuda/conv_bias/implicit_gemm_int4_int4_nhwc_imma.cpp + * MegEngine is Licensed under the Apache License, Version 2.0 (the "License") + * + * Copyright (c) 2014-2021 Megvii Inc. All rights reserved. + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or + * implied. + */ + +#include "./algo.h" +#include "src/cuda/conv_bias/cutlass_convolution_wrapper.cuh" + +using namespace megdnn; +using namespace cuda; +using namespace convolution; + +#if CUDA_VERSION >= 10020 +size_t +ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm::get_workspace_in_bytes( + const SizeArgs& args) const { + if (args.preprocessed_filter) { + return 0; + } else { + return args.filter_layout->span().dist_byte(); + } +} + +size_t ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm:: + get_preprocess_workspace_in_bytes(const SizeArgs& args) const { + return 0; +} + +SmallVector ConvBiasForwardImpl:: + AlgoInt4Int4NHWCIMMAImplicitGemm::deduce_preprocessed_filter_layout( + const SizeArgs& args) const { + return {args.filter_layout->collapse_contiguous()}; +} + +void ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm::exec_preprocess( + const ExecArgs& args) const { + megdnn_assert(args.preprocessed_filter->tensors.size() == 1); + void* filter_ptr = args.preprocessed_filter->tensors[0].raw_ptr; + reorder_filter(args, m_algo_param.access_size, filter_ptr); +} + +std::tuple +ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm::prepare_filter_bias( + const ExecArgs& args) const { + void* filter_ptr = nullptr; + if (args.preprocessed_filter) { + megdnn_assert(args.preprocessed_filter->tensors.size() == 1); + filter_ptr = args.preprocessed_filter->tensors[0].raw_ptr; + } else { + filter_ptr = reinterpret_cast(args.workspace.raw_ptr); + reorder_filter(args, m_algo_param.access_size, filter_ptr); + } + void* bias_ptr = args.bias_tensor->raw_ptr; + return {filter_ptr, bias_ptr}; +} + +std::tuple +ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm::get_constants( + const ExecArgs& args) const { + float src_scale = args.src_layout->dtype.param().scale, + filter_scale = + args.filter_layout->dtype.param().scale, + bias_scale = + args.bias_layout->dtype.param().scale, + dst_scale = args.dst_layout->dtype.param().scale; + + float alpha = src_scale * filter_scale / dst_scale, + beta = bias_scale / dst_scale, gamma = 0.f, delta = 0.f, theta = 0.f; + + if (args.z_layout->ndim > 0) { + float z_scale = args.z_layout->dtype.param().scale; + gamma = z_scale / dst_scale; + } + + return {alpha, beta, gamma, delta, theta}; +} + +void ConvBiasForwardImpl::AlgoInt4Int4NHWCIMMAImplicitGemm::do_exec( + const ExecArgs& args, void* filter_ptr, void* bias_ptr, void* z_ptr, + ConvParam kern_param, uint32_t nonlinear_mode, float alpha, float beta, + float gamma, float delta, float theta, cudaStream_t stream) const { + float dst_scale = args.dst_layout->dtype.param().scale; + + cutlass_wrapper::GemmCoord threadblock_shape{m_algo_param.threadblock_m, + m_algo_param.threadblock_n, + m_algo_param.threadblock_k}; + + cutlass_wrapper::GemmCoord warp_shape{ + m_algo_param.warp_m, m_algo_param.warp_n, m_algo_param.warp_k}; + + if (kern_param.fh == 1 && kern_param.fw == 1) { + cutlass_wrapper::do_conv_bias_int4_int4_implicit_gemm_imma_nhwc( + reinterpret_cast(args.src_tensor->raw_ptr), + reinterpret_cast(filter_ptr), + reinterpret_cast(bias_ptr), + reinterpret_cast(z_ptr), + reinterpret_cast(args.dst_tensor->raw_ptr), nullptr, + kern_param, nonlinear_mode, alpha, beta, gamma, dst_scale, + threadblock_shape, warp_shape, m_algo_param.access_size, + stream); + } else { + cutlass_wrapper::do_conv_bias_int4_int4_implicit_gemm_imma_nhwc( + reinterpret_cast(args.src_tensor->raw_ptr), + reinterpret_cast(filter_ptr), + reinterpret_cast(bias_ptr), + reinterpret_cast(z_ptr), + reinterpret_cast(args.dst_tensor->raw_ptr), nullptr, + kern_param, nonlinear_mode, alpha, beta, gamma, dst_scale, + threadblock_shape, warp_shape, m_algo_param.access_size, + stream); + } +} +#endif + +// vim: syntax=cpp.doxygen diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_int4_nchw64_imma_base.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_int4_nchw64_imma_base.cpp index 03335967..d99bf90a 100644 --- a/dnn/src/cuda/conv_bias/implicit_gemm_int4_nchw64_imma_base.cpp +++ b/dnn/src/cuda/conv_bias/implicit_gemm_int4_nchw64_imma_base.cpp @@ -60,9 +60,22 @@ bool ConvBiasForwardImpl::AlgoInt4NCHW64IMMAImplicitGemmBase::is_available( args.dst_layout->dtype.enumv() != src_dtype()) return false; + // uint4 do not support H_SWISH activition + if (src_dtype() == DTypeEnum::Quantized4Asymm && + param.nonlineMode == NonlineMode::H_SWISH) + return false; + if (!is_compute_capability_required(7, 5)) return false; + size_t fh = args.filter_layout->operator[](1), + fw = args.filter_layout->operator[](2); + + // param buffer size is 4K, use 3.4K to store precomputed offset + size_t kMaxFilterPixels = 848 / (2 * m_algo_param.warp_k / 64) - 2; + if (fh * fw > kMaxFilterPixels) + return false; + return true; } @@ -108,7 +121,7 @@ void ConvBiasForwardImpl::AlgoInt4NCHW64IMMAImplicitGemmBase::exec( std::string ConvBiasForwardImpl::AlgoInt4NCHW64IMMAImplicitGemmBase::to_string( AlgoParam algo_param) { - return ssprintf("%uX%uX%u_%uX%uX%u", algo_param.threadblock_m, + return ssprintf("%dX%dX%d_%dX%dX%d", algo_param.threadblock_m, algo_param.threadblock_n, algo_param.threadblock_k, algo_param.warp_m, algo_param.warp_n, algo_param.warp_k); } diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_int4_nhwc_imma_base.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_int4_nhwc_imma_base.cpp new file mode 100644 index 00000000..21aca925 --- /dev/null +++ b/dnn/src/cuda/conv_bias/implicit_gemm_int4_nhwc_imma_base.cpp @@ -0,0 +1,159 @@ +/** + * \file dnn/src/cuda/conv_bias/implicit_gemm_int4_nhwc_imma_base.cpp + * MegEngine is Licensed under the Apache License, Version 2.0 (the "License") + * + * Copyright (c) 2014-2021 Megvii Inc. All rights reserved. + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or + * implied. + */ + +#include "./algo.h" +#include "src/common/conv_bias.h" +#include "src/cuda/conv_bias/cutlass_convolution_wrapper.cuh" +#include "src/cuda/conv_bias/reduce_filter.cuh" +#include "src/cuda/convolution_helper/parameter.cuh" +#include "src/cuda/utils.h" + +using namespace megdnn; +using namespace cuda; +using namespace convolution; + +#if CUDA_VERSION >= 10020 +std::string ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase::param() + const { + std::string ret; + serialize_write_pod(m_algo_param, ret); + return ret; +} + +bool ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase::is_available( + const SizeArgs& args) const { + if (args.bias_layout->ndim <= 0) + return false; + + using Param = param::ConvBias; + using Format = Param::Format; + using Sparse = Param::Sparse; + using Mode = Param::Mode; + using NonlineMode = megdnn::param::ConvBias::NonlineMode; + + auto&& param = args.opr->param(); + + if (!check_bias_share_in_channel(*(args.bias_layout), param.format)) + return false; + + if (param.format != Format::NHWC || param.sparse != Sparse::DENSE || + param.mode != Mode::CROSS_CORRELATION) + return false; + + if (param.nonlineMode != NonlineMode::IDENTITY && + param.nonlineMode != NonlineMode::RELU && + param.nonlineMode != NonlineMode::H_SWISH) + return false; + + if (args.src_layout->dtype.enumv() != src_dtype() || + args.filter_layout->dtype.enumv() != DTypeEnum::QuantizedS4 || + args.bias_layout->dtype.enumv() != DTypeEnum::QuantizedS32 || + args.dst_layout->dtype.enumv() != src_dtype()) + return false; + + // uint4 do not support H_SWISH activition + if (src_dtype() == DTypeEnum::Quantized4Asymm && + param.nonlineMode == NonlineMode::H_SWISH) + return false; + + if (!is_compute_capability_required(7, 5)) + return false; + + size_t co = args.filter_layout->operator[](0), + ci = args.filter_layout->operator[](3), + fh = args.filter_layout->operator[](1), + fw = args.filter_layout->operator[](2); + + // param buffer size is 4K, use 3.4K to store precomputed offset + size_t kMaxFilterPixels = + 848 / (m_algo_param.warp_k / m_algo_param.access_size) - 1; + if (fh * fw > kMaxFilterPixels) + return false; + // co should be aligned with 8, and ci should be aligned with + // algo_param.access_size + if ((co % 8 != 0) || (ci % m_algo_param.access_size != 0)) + return false; + + return true; +} + +void ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase::exec( + const ExecArgs& args) const { + auto&& param = args.opr->param(); + auto&& fm = args.filter_meta; + size_t n = args.src_layout->operator[](0), + ci = args.src_layout->operator[](3), + hi = args.src_layout->operator[](1), + wi = args.src_layout->operator[](2); + size_t co = args.dst_layout->operator[](3), + ho = args.dst_layout->operator[](1), + wo = args.dst_layout->operator[](2); + UNPACK_CONV_PARAMETER(fm, param); + MARK_USED_VAR + + void* filter_ptr = nullptr; + void* bias_ptr = nullptr; + void* z_ptr = nullptr; + + std::tie(filter_ptr, bias_ptr) = prepare_filter_bias(args); + if (args.z_layout->ndim > 0) + z_ptr = args.z_tensor->raw_ptr; + + float alpha, beta, gamma, delta, theta; + std::tie(alpha, beta, gamma, delta, theta) = get_constants(args); + + ConvParam kern_param; + kern_param.n = n, kern_param.co = co, kern_param.ci = ci, + kern_param.hi = hi, kern_param.wi = wi, kern_param.ho = ho, + kern_param.wo = wo, kern_param.ph = ph, kern_param.pw = pw, + kern_param.sh = sh, kern_param.sw = sw, kern_param.fh = fh, + kern_param.fw = fw; + + uint32_t nonlinear_mode = static_cast(param.nonlineMode); + + cudaStream_t stream = cuda_stream(args.opr->handle()); + + do_exec(args, filter_ptr, bias_ptr, z_ptr, kern_param, nonlinear_mode, + alpha, beta, gamma, delta, theta, stream); +} + +std::string ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase::to_string( + AlgoParam algo_param) { + return ssprintf("%dX%dX%d_%dX%dX%d_%d", algo_param.threadblock_m, + algo_param.threadblock_n, algo_param.threadblock_k, + algo_param.warp_m, algo_param.warp_n, algo_param.warp_k, + algo_param.access_size); +} + +void ConvBiasForwardImpl::AlgoInt4NHWCIMMAImplicitGemmBase::reorder_filter( + const ExecArgs& args, const int iterleaved, + void* reordered_filter) const { + size_t co = args.filter_layout->operator[](0), + ci = args.filter_layout->operator[](3), + fh = args.filter_layout->operator[](1), + fw = args.filter_layout->operator[](2); + + // reformat grad from nhwc to ncxhwx + TensorLayout exec_src{{co, fh, fw, ci / iterleaved, (size_t)iterleaved / 2}, + dtype::Int8()}; + TensorLayout exec_dst{{co, ci / iterleaved, fh, fw, (size_t)iterleaved / 2}, + dtype::Int8()}; + + exec_src = exec_src.dimshuffle({0, 3, 1, 2, 4}); + + auto&& relayout = args.opr->handle()->create_operator(); + relayout->exec({args.filter_tensor->raw_ptr, exec_src}, + {reordered_filter, exec_dst}); +} +#endif + +// vim: syntax=cpp.doxygen diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw32_imma.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw32_imma.cpp index 03cb358b..b04b1a01 100644 --- a/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw32_imma.cpp +++ b/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw32_imma.cpp @@ -75,8 +75,9 @@ bool ConvBiasForwardImpl::AlgoInt8NCHW32IMMAImplicitGemm::is_available( // only support sm_75 or later, platform should have tensorcore int8 // support available &= is_compute_capability_required(7, 5); - // FIXME: too large filter size is not supported now - available &= fh * fw <= 49; + // FIXME: too large filter size is not supported now + size_t kMaxFilterPixels = 848 / (2 * m_algo_param.warp_k / 32) - 2; + available &= fh * fw <= kMaxFilterPixels; return available; } diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp index 3287b80a..ea40e918 100644 --- a/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp +++ b/dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp @@ -95,8 +95,10 @@ bool ConvBiasForwardImpl::AlgoInt8NCHW4DotProdImplicitGemm::is_available( // only support sm_61 or later, platform should have fast native int8 // support available &= is_compute_capability_required(6, 1); - // FIXME: too large filter size is not supported now - available &= fh * fw <= 49; + // FIXME: too large filter size is not supported now + size_t kMaxFilterPixels = 848 / (2 * m_algo_param.warp_k / 4) - 2; + available &= fh * fw <= kMaxFilterPixels; + ; return available; } diff --git a/dnn/src/cuda/conv_bias/implicit_gemm_uint4_int4_nhwc_imma.cpp b/dnn/src/cuda/conv_bias/implicit_gemm_uint4_int4_nhwc_imma.cpp new file mode 100644 index 00000000..ab62e1d9 --- /dev/null +++ b/dnn/src/cuda/conv_bias/implicit_gemm_uint4_int4_nhwc_imma.cpp @@ -0,0 +1,186 @@ +/** + * \file dnn/src/cuda/conv_bias/implicit_gemm_uint4_int4_nhwc_imma.cpp + * MegEngine is Licensed under the Apache License, Version 2.0 (the "License") + * + * Copyright (c) 2014-2021 Megvii Inc. All rights reserved. + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or + * implied. + */ + +#include "./algo.h" +#include "src/cuda/conv_bias/cutlass_convolution_wrapper.cuh" +#include "src/cuda/conv_bias/reduce_filter.cuh" +#include "src/cuda/utils.h" + +using namespace megdnn; +using namespace cuda; +using namespace convolution; + +#if CUDA_VERSION >= 10020 +size_t +ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::get_workspace_in_bytes( + const SizeArgs& args) const { + if (args.preprocessed_filter) { + return 0; + } else { + size_t ws_filter = args.filter_layout->span().dist_byte(), + ws_bias = args.bias_layout->span().dist_byte(), + ws_reduce_filter = get_preprocess_workspace_in_bytes(args); + return ws_filter + ws_bias + ws_reduce_filter; + } +} + +size_t ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm:: + get_preprocess_workspace_in_bytes(const SizeArgs& args) const { + size_t co = args.filter_layout->operator[](0), + ci = args.filter_layout->operator[](3), + fh = args.filter_layout->operator[](1), + fw = args.filter_layout->operator[](2); + size_t ws_size_reduce_filter = co * sizeof(int32_t); + size_t A = co, B = ci * fh * fw / 8, C = 1; + ws_size_reduce_filter += do_dispatch_reduce_workspace_in_bytes(A, B, C); + return ws_size_reduce_filter; +} + +SmallVector ConvBiasForwardImpl:: + AlgoUInt4Int4NHWCIMMAImplicitGemm::deduce_preprocessed_filter_layout( + const SizeArgs& args) const { + return {args.filter_layout->collapse_contiguous(), + args.bias_layout->collapse_contiguous()}; +} + +void ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::exec_preprocess( + const ExecArgs& args) const { + megdnn_assert(args.preprocessed_filter->tensors.size() == 2); + void* filter_ptr = args.preprocessed_filter->tensors[0].raw_ptr; + void* bias_ptr = args.preprocessed_filter->tensors[1].raw_ptr; + void* reduce_filter_ptr = reinterpret_cast(args.workspace.raw_ptr); + void* reduce_workspace = reinterpret_cast( + args.workspace.raw_ptr + args.bias_layout->span().dist_byte()); + reorder_filter(args, m_algo_param.access_size, filter_ptr); + update_bias(args, bias_ptr, reduce_filter_ptr, reduce_workspace); +} + +std::tuple +ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::prepare_filter_bias( + const ExecArgs& args) const { + void* filter_ptr = nullptr; + void* bias_ptr = nullptr; + if (args.preprocessed_filter) { + megdnn_assert(args.preprocessed_filter->tensors.size() == 2); + filter_ptr = args.preprocessed_filter->tensors[0].raw_ptr; + bias_ptr = args.preprocessed_filter->tensors[1].raw_ptr; + return {filter_ptr, bias_ptr}; + } else { + filter_ptr = reinterpret_cast(args.workspace.raw_ptr); + bias_ptr = + reinterpret_cast(args.workspace.raw_ptr + + args.filter_layout->span().dist_byte()); + void* reduce_filter_ptr = + reinterpret_cast(args.workspace.raw_ptr + + args.filter_layout->span().dist_byte() + + args.bias_layout->span().dist_byte()); + void* reduce_workspace = + reinterpret_cast(args.workspace.raw_ptr + + args.filter_layout->span().dist_byte() + + args.bias_layout->span().dist_byte() + + args.bias_layout->span().dist_byte()); + reorder_filter(args, m_algo_param.access_size, filter_ptr); + update_bias(args, bias_ptr, reduce_filter_ptr, reduce_workspace); + } + return {filter_ptr, bias_ptr}; +} + +std::tuple +ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::get_constants( + const ExecArgs& args) const { + float src_scale = + args.src_layout->dtype.param().scale, + filter_scale = + args.filter_layout->dtype.param().scale, + bias_scale = + args.bias_layout->dtype.param().scale, + dst_scale = + args.dst_layout->dtype.param().scale; + + uint8_t dst_zero = + args.dst_layout->dtype.param().zero_point; + float alpha = src_scale * filter_scale / dst_scale, + beta = bias_scale / dst_scale, gamma = 0.f, delta = 0.f, + theta = dst_zero; + + if (args.z_layout->ndim > 0) { + float z_scale = + args.z_layout->dtype.param().scale; + gamma = z_scale / dst_scale; + uint8_t z_zero = + args.z_layout->dtype.param().zero_point; + delta = -z_zero * gamma; + } + + return {alpha, beta, gamma, delta, theta}; +} + +void ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::do_exec( + const ExecArgs& args, void* filter_ptr, void* bias_ptr, void* z_ptr, + ConvParam kern_param, uint32_t nonlinear_mode, float alpha, float beta, + float gamma, float delta, float theta, cudaStream_t stream) const { + float dst_scale = + args.dst_layout->dtype.param().scale; + uint8_t src_zero = + args.src_layout->dtype.param().zero_point; + cutlass_wrapper::GemmCoord threadblock_shape{m_algo_param.threadblock_m, + m_algo_param.threadblock_n, + m_algo_param.threadblock_k}; + + cutlass_wrapper::GemmCoord warp_shape{ + m_algo_param.warp_m, m_algo_param.warp_n, m_algo_param.warp_k}; + if (kern_param.fh == 1 && kern_param.fw == 1) { + cutlass_wrapper::do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc( + reinterpret_cast(args.src_tensor->raw_ptr), + reinterpret_cast(filter_ptr), + reinterpret_cast(bias_ptr), + reinterpret_cast(z_ptr), + reinterpret_cast(args.dst_tensor->raw_ptr), nullptr, + kern_param, nonlinear_mode, alpha, beta, gamma, delta, theta, + dst_scale, src_zero, threadblock_shape, warp_shape, + m_algo_param.access_size, stream); + } else { + cutlass_wrapper::do_conv_bias_uint4_int4_implicit_gemm_imma_nhwc( + reinterpret_cast(args.src_tensor->raw_ptr), + reinterpret_cast(filter_ptr), + reinterpret_cast(bias_ptr), + reinterpret_cast(z_ptr), + reinterpret_cast(args.dst_tensor->raw_ptr), nullptr, + kern_param, nonlinear_mode, alpha, beta, gamma, delta, theta, + dst_scale, src_zero, threadblock_shape, warp_shape, + m_algo_param.access_size, stream); + } +} + +void ConvBiasForwardImpl::AlgoUInt4Int4NHWCIMMAImplicitGemm::update_bias( + const ExecArgs& args, void* updated_bias, void* reduce_filter_ptr, + void* reduce_workspace) const { + size_t co = args.filter_layout->operator[](0), + ci = args.filter_layout->operator[](3), + fh = args.filter_layout->operator[](1), + fw = args.filter_layout->operator[](2); + + auto&& stream = cuda_stream(args.opr->handle()); + + int src_zero_point = + args.src_tensor->layout.dtype.param() + .zero_point; + do_dispatch_reduce_filter_and_update_bias_4bit( + reinterpret_cast(args.filter_tensor->raw_ptr), + args.bias_tensor->compatible_ptr(), co, ci * fh * fw / 8, + reinterpret_cast(updated_bias), + reinterpret_cast(reduce_workspace), src_zero_point, + stream); +} +#endif + +// vim: syntax=cpp.doxygen diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..059b3287 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..ecab6b8c --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..aef4745a --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..bd482580 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..1a20121c --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..3f19c584 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..c66ac54e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..45590035 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..64a3e278 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..d0ead0c3 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..6994ef17 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..1d9e6c79 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..952bf9fe --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..83b6a72f --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..4d1c3231 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..d5d174c6 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..b39a3b38 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..1a6f8b69 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_1x1_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 5ed7c5ee..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<128, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..ef91b066 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..574fbba8 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..22ed7316 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..35db883a --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..33a25779 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..fc8339b4 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..b4934fb2 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 2052984d..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<256, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..d65d8dca --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_hswish_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<256, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationHSwishClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 26acda83..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<128, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..33aa6e47 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..99511641 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..5b023a5c --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..2a4de50e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..0586bbd0 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..b4c9d9ec --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..c5139b49 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 7bbe727c..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<256, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..77d958b1 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_identity_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<256, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 2367b398..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<128, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..52f4334d --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..c2d41fdc --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..331686a1 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..e0d8b1e8 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..e0917aa6 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..ba7aacb2 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..1bb9c179 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::int4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 3874374b..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::int4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<256, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< - cutlass::int4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..7fc1a3e3 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_s4_i8832fprop_relu_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<256, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::int4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..d0e906e7 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..80c9fc67 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..ba1875ff --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..4502520f --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..2ead7f5d --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..d0f6e113 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..94a82f82 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..0d17a91e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..967b0341 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..f780d1a0 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..8db5b1af --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..e4a28e3e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_1x1_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + false, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 755c024e..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<128, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationClamp< - cutlass::uint4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..db1db695 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..f6b7d1e7 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..1353ffd1 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..dbdc1bd6 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..a7f2d479 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..7c3a1ef3 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..d4e2e602 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 277b3308..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<256, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationClamp< - cutlass::uint4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..65b56c5e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_identity_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<256, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationClamp< + cutlass::uint4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index 5773bc47..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<128, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< - cutlass::uint4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..80f50629 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..0429ab43 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..6f9fbba8 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..e8f1001d --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x32x64_64x32x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 32, 64>, + cutlass::gemm::GemmShape<64, 32, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu new file mode 100644 index 00000000..35d1f344 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc16hw16" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<16>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 16, + 16, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu new file mode 100644 index 00000000..9ac7388e --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc32hw32" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<32>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu new file mode 100644 index 00000000..46e1bdea --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_128x64x64_64x64x64_2_nhwc_nc8hw8" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + cutlass::int4b_t, + cutlass::layout::TensorNCxHWx<8>, + cutlass::uint4b_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::layout::TensorNHWC, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<128, 64, 64>, + cutlass::gemm::GemmShape<64, 64, 64>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 8, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNHWCThreadblockSwizzle, + 2, + 8, + 8, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_TN>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu deleted file mode 100644 index d9f26575..00000000 --- a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64.cu +++ /dev/null @@ -1,55 +0,0 @@ - -#if !MEGDNN_TEGRA_X1 -// ignore warning of cutlass -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wunused-parameter" -#pragma GCC diagnostic ignored "-Wstrict-aliasing" -#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" - - -// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64" generated by cutlass generator -using Convolution = - typename cutlass::conv::device::Convolution< - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - cutlass::int4b_t, - cutlass::layout::TensorCxRSKx<64>, - cutlass::uint4b_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::layout::TensorNCxHWx<64>, - int32_t, - cutlass::conv::ConvType::kConvolution, - cutlass::arch::OpClassTensorOp, - cutlass::arch::Sm75, - cutlass::gemm::GemmShape<256, 128, 128>, - cutlass::gemm::GemmShape<64, 64, 128>, - cutlass::gemm::GemmShape<8, 8, 32>, - cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< - cutlass::uint4b_t, - 16, - int32_t, - int32_t, - float - >, - cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, - 2, - 32, - 32, - true, - cutlass::arch::OpMultiplyAddSaturate>; - - -template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( - const typename Convolution::ElementSrc* d_src, - const typename Convolution::ElementFilter* d_filter, - const typename Convolution::ElementBias* d_bias, - const typename Convolution::ElementDst* d_z, - typename Convolution::ElementDst* d_dst, - int* workspace, - typename Convolution::ConvolutionParameter const& conv_param, - typename Convolution::EpilogueOutputOp::Params const& epilogue, - cudaStream_t stream, - typename Convolution::ExtraParam extra_param); -#pragma GCC diagnostic pop -#endif diff --git a/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu new file mode 100644 index 00000000..55acfe28 --- /dev/null +++ b/dnn/src/cuda/conv_bias/int4/kimpl/cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64.cu @@ -0,0 +1,59 @@ + +#if !MEGDNN_TEGRA_X1 +// ignore warning of cutlass +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wunused-parameter" +#pragma GCC diagnostic ignored "-Wstrict-aliasing" +#include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" + + +// kernel instance "cutlass_tensorop_u4_i8832fprop_relu_u4_s4_256x128x128_64x64x128_2_nc64hw64_c64rsk64" generated by cutlass generator +using Convolution = + typename cutlass::conv::device::Convolution< + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + cutlass::int4b_t, + cutlass::layout::TensorCxRSKx<64>, + cutlass::uint4b_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::layout::TensorNCxHWx<64>, + int32_t, + cutlass::conv::ConvType::kConvolution, + cutlass::arch::OpClassTensorOp, + cutlass::arch::Sm75, + cutlass::gemm::GemmShape<256, 128, 128>, + cutlass::gemm::GemmShape<64, 64, 128>, + cutlass::gemm::GemmShape<8, 8, 32>, + cutlass::epilogue::thread::BiasAddLinearCombinationReluClamp< + cutlass::uint4b_t, + 16, + int32_t, + int32_t, + float + >, + cutlass::conv::threadblock::ConvolutionFpropNCxHWxThreadblockSwizzle, + 2, + 32, + 32, + true, + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; + + + +template void megdnn::cuda::cutlass_wrapper::cutlass_convolution_wrapper( + const typename Convolution::ElementSrc* d_src, + const typename Convolution::ElementFilter* d_filter, + const typename Convolution::ElementBias* d_bias, + const typename Convolution::ElementDst* d_z, + typename Convolution::ElementDst* d_dst, + int* workspace, + typename Convolution::ConvolutionParameter const& conv_param, + typename Convolution::EpilogueOutputOp::Params const& epilogue, + cudaStream_t stream, + typename Convolution::ExtraParam extra_param); + + +#pragma GCC diagnostic pop +#endif diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 828ad12a..0188b8c9 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 5fb1087f..634732ef 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 60ed11c6..10978a20 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index c6b29e41..e8ac6042 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index 57c1f91a..811cdfcd 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index f97f8c72..d33a30f0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index c9346d31..6692dfb0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 82895299..73e10efc 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 152a17be..550dabf0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index eab74ead..0ea16d92 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index c05e4e9d..6949e97f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index e5e41b33..7bd8c0a1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 35c48f8e..d9560c0e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 50b324d0..f49957b0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index 8e749a6c..08f98ebd 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index a81e5ae8..13c20c41 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 66bfdd0a..116d1981 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index 0e0f2263..772e9be8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index f06c69bd..b2f08485 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 7f43a2c2..995a362b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 95ef3609..5e1d157f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 1e2a14b8..7d758dd7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 7106d1f7..d0b56d97 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index ecf89ca6..1963beb0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index c1d06a19..cb6b9468 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index eeb92cda..f3db7a99 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index defd1f2d..f3dc2913 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 79f65f6c..2bcac258 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index cff3a8df..f40cfa4c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 0b56bcf5..ad941b12 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index e4d88dd9..161bbea5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index ea872272..e6f63ba8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 6ab659be..b3722c74 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 742ed499..2990d138 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 049b42c3..60d6b36b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 79b40aad..24f333e4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index 3072bdb4..4a7c4f82 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index 842ca304..695dbb8d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index d2c18579..ff4780c9 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index 831e696c..3e169cec 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 302935c5..d5ec49e3 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 26c6ada9..7c9903b5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 3b419be3..bc521eed 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index e99a6852..901fe816 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index a4fad1bf..70f635f0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index e3521828..92aaf901 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index fbf8c3b9..7218d5f3 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index 73cd3839..d37b1c08 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index b3d47b02..e79fca2b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 2f26acf8..16bc9200 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index 7b9219be..ff3abbd1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index e5b3eb1e..7b51944c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 2cd897a8..04846642 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 29e25395..bf95aad5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 386de077..8d147553 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 27bdaeb5..eb0ddf9a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index f3f73e94..61265f6b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 2bbee6fd..197018a8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu index ba873407..95017836 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu index 09082101..85115e8e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index 84b1ffba..f165fac4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu index 28a74223..7986b73a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu index fafc7a18..10dcd177 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 8e1835f8..2cf55999 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index a5f3721e..a1b665d1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu index 1401390e..bf5a59d7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nchw" generated by cutlass generator +// kernel instance "cutlass_simt_f32_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nchw" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 41d64cb9..c62d96d8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index a41c0294..6455430d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 566dc284..e63d982c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index 2467cc4e..696542f2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index c4d1697e..1af499ad 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 0514f278..f70e8d25 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index a86e72b1..de958f95 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index ae88eaf2..3a195832 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index ab6c8b00..7fd8b76f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 3ff004f9..4205a742 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index de94428e..431b43ca 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 86859c9a..3d6ada6f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 724f1026..77384ac6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 7b28b0d5..d30558d8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index 2f5bfcfc..3e32e6e5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index e5e7a144..d69a9e9e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 089cb208..c8678666 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 1bd0a97f..972816c7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 4a5a33fc..5117e5c1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index c20d5847..b4b5550a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 747219ed..c18385a5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index a7b322a5..9725a58a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index fe01d0b8..d2ab9fd3 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 1737acc2..91885956 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index e8f671ce..e0deccc2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index d4bcd815..6f7c1dac 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index a925dde6..d256e23a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index fe55edfe..649d27ed 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 9f5e15d5..8e5d4f44 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 03b3dd1e..799219ab 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index df33dc24..a0fe7fb2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 2cd0a451..cf2f55e0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index c92680aa..47634aee 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_s4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index eb5f8857..52a026ba 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 119ca59e..6c59c066 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index f06f8f31..e080b713 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ec713ff7..97dae192 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 9affaa3c..cbff4aaa 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 92afe974..7027e886 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index 9e331632..3b8ea536 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index 026d72a9..97b4ce07 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index f9311df3..71c0baca 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 65ae55c5..61fdf2bf 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index d9267ca2..065b4d97 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index cb957cd3..45e2ecd1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 53f215f1..66394d5f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index c865ecbc..7d0383cb 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index a6fcb3eb..c1104914 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 0008c79f..9216e8c1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 36d4ab6e..3faa0456 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 707e4e72..2657a4f1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 77b89b03..def60f3d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 287d21c2..8d37fb31 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index cd828293..f743166d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 06575697..5dfe6b54 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 27f290f8..65154c27 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index d3cfd4b3..0aeb15f7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 5da3327b..f02fd855 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 869fac22..0cd9d768 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index a8db4021..9d1b9824 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index c0c23080..95f37bff 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index 9271b3b3..3de086be 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 0e37f1c7..f346d1c1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index 8d5d9743..cec7576b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ad4aca7b..94062d99 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 27263d12..aa901c37 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index f1b452ea..a4581d29 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 859887dd..f6be5d3f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index a377250d..b2513493 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 7b259d59..56164c72 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 73a472e7..1a46c48e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 8e9ad219..5b40ac21 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 9a22a896..4fd9bb43 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 948d86fa..05117134 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index d03293a2..05b313d0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 54c059ad..5c2f505c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 0ae92d4f..8efd1931 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index abe5e521..79d51ce2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index c1b497ac..371cb6a7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index cfd0926b..8d9ef1af 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index bb989bf7..6acff4cc 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index c48fb744..35c8fcdf 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 68f9d8cc..b6cab90e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index 90a17d4d..818cdad0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index b4094622..e2e9b4a4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 05e3e2b6..0c4d8206 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ba2ef184..5f436fa1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index ed2e4f27..aa49f1be 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index e2141044..15bdd42a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 06e5b01f..79168d31 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 9b5a5b06..a196862c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 4fde39a8..e3d8de90 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 8b13b712..11bd4fd6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_1x1_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, false, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 8b8f3ac4..a1aebf28 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index c4a85c47..f1411857 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 971d5016..45394e91 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 5d5fb2c6..80fe01d4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index ee0c2a2e..3310ab82 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 26d1bd97..9afa9cca 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index 68585a3a..11ec8089 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index 6111f919..f98c6e3f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index 68794f6b..2a169bd1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 16e415f6..943e23c6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index 01b92bc1..ebe14204 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index d0620ca0..fd088482 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 5cdcce09..7da9f455 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index a633169a..d544a0dc 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 9e7ae5fd..901424c0 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 5897f989..90e777c8 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 9c856339..685a9d7e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 3277092d..73bf1d62 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index a1f282e9..c6e05459 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ba6e5045..588bbc12 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 812274a7..dce1e586 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index d37d0edb..fa043faa 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 837f1704..ec624a3a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index f7856b22..545b0478 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 341fd2e5..67b97d90 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index a318b38c..ba4e2f2d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index 8e017652..aa7116c4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index 79e3d032..3471b9fa 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index 00847daa..ceed408d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index a9aa737e..b7db1261 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index e3c88e3d..9fcc2f40 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ecaa42c0..94e3fc97 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 9a6c225d..76628d7f 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 9eab2b78..d0fda4ee 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 67eed08e..287ab3bc 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 125506b0..7f930475 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index c81ee566..6f884781 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index bbc3e062..1925a760 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 3b04adaa..b0948fd2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 54730939..75789416 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index a639f5bd..ea5b2fd7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 043533be..bb56f241 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index c1f70a43..2a0cf695 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 5ad99300..0d6b624b 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 6968d4fc..746dd299 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 2bfdea16..6619b613 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu index c458632a..824d2580 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu index f1023cbd..72d7da12 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu index 26af7b6c..bcc78062 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index baf6788e..b4cb06c1 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu index 8651eeaf..95002a67 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index ab66f277..8ea2592e 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu index 1adccb33..12b96460 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu index cfbc9dc7..8aa4d121 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu index 86073427..a8dfd9fb 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 624df6b1..45e4d7e4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu index 9ce8d067..1d12d500 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index dd78af30..2666fe1c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu index 7f2d171b..dbba1fc5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu index 408169dd..c177dd20 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_simt_s8_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nc32hw32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index c1f9f0cc..fce2a7b2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 42cce868..b8fe8657 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index ce2faaf9..e602b4fb 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index a8b5ef0f..9a11ca1d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index 0ba74f4c..3a416f6a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 61e4ebae..5e83d253 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index a01dc2b9..f5ca0a86 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 968403a5..d4b88bd6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 65b8157f..f83ef4f6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 0742f4c7..ab9dd016 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 9641af3e..eda36cc2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_hswish_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index ad8152a1..c59981f5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index e2612b5d..8df0ceaf 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 1b168e3d..ca939ce4 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index 13cca83e..b97524ee 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index d93d7191..29ec4813 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 031ad7e8..d440abb2 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 082e7f9c..bc064a26 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index c5c70528..e40529af 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 01bd0597..f57da562 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index d1643f9a..9891e9d5 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 40231a90..28823bb6 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_identity_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 11cd5722..25d4453a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 50e30b60..0ccecb64 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 8ae67566..cae3e75d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_128x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu index 73fed81f..a0d014f7 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_16x128x16_16x128x16_1_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu index cc8985f3..8518aabb 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_16x64x8_16x64x8_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 882ea15c..e2ca5566 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x128x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 10464380..a6e75395 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x32x32_32x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu index 768771c2..0ef0f83a 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_32x64x32_32x64x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index f09590a2..5acd525c 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x128x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index 57fbd2f3..8a70bd48 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x32x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu similarity index 92% rename from dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu rename to dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu index afbdbbbb..f0abfb3d 100644 --- a/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc.cu +++ b/dnn/src/cuda/conv_bias/int8/kimpl/cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_nhwc" generated by cutlass generator +// kernel instance "cutlass_simt_u4_ifprop_relu_s8_64x64x32_64x32x32_2_nc4hw4_c4rsk4_nhwc" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 8c2e425a..91ef9b82 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index db1c950c..4327a877 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index 99e63380..f81c5345 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 9b646f5a..8cd3d419 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index f2777718..6eb00a4c 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 52bf3538..de3b726c 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 6adce813..747240fc 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 94d14423..1a7151f3 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 9127e532..092869c9 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index 1089e31f..3418a6a6 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index 676de64e..9ed169ed 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index ac1ea8b0..86f0b034 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index be1cd355..f8e44905 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 58b209be..77815d13 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 0dd9d49a..10a95d79 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 7a9c681b..95ed7a79 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index 264516f3..45112372 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 5c7d5a81..e0e78d1f 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index f43d76a6..33795436 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 5981be60..1ab249ba 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 051fe0aa..f7e1bbb6 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 61c20863..8340a760 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 94c98d8a..1c00df1e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index 259fa2b4..a848d1a6 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index fdb3e383..7226dcb7 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index f499f162..3424de30 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index 4b5ce90c..8056afff 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 90% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 2ea46668..c7c0cb91 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 57a858e8..a8605228 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 24be9280..cf313aa2 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index 4861c0ba..ad862cfd 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 5f5efcc6..bf4044e5 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index a71bf73a..000c3420 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index bc85e8e9..94d728ef 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 7f6d69a5..61936ad4 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index bd99ad62..bdded445 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 1cde8955..f4add434 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index e9c07ff1..01200c9e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index e5792726..c00a1f2a 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 2f6b5d1a..53d97da9 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index f216358a..b9829d3c 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 8cd3ff76..112feff5 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_1x1_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, false, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index f1476e65..c190a1ed 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 36c0437d..c0b7c567 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index 954f44e8..95fd179e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 3bbe65d5..d33faf86 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index e9ab71c6..2f1a0e66 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 47ff8c28..0d78e2bb 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index f6a8fea0..3c128645 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index be9d5f00..f807bd34 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 7f9b454f..89604a7d 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index e49d4672..90271408 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index ed2411d3..a2942ae4 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 36bd53e1..e8685e45 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index bf0f2d37..eb9cf847 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index a5463e79..5bfa7ff9 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_hswish_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 2f7395fe..82765158 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 3f3d43b8..c0ae8376 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index cfc62fa6..2b66bb58 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 3a6c6eb1..aee8a5d1 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index 1599f6d7..67e37171 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 6aefb106..a9514f6c 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index 54a259e7..37984bc9 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 7b798cd2..70ed2f42 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 5119b75f..6a5c7d2e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index f07ee7af..98e4034e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index a86754dc..70708175 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 07d3a81d..84a88a6b 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index b32291d8..234db1ec 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index d5fd8891..46ed4028 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_identity_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index a0d811bf..07f2e0a7 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index d0e151da..ecf3cfc3 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu index 4f7a3e28..ba90c2af 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index b5dff9fb..e5208560 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x256x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu index 650b53b7..594b5e66 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index a6c1709c..2548d373 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_128x64x64_64x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu index be070f8b..2b3bcc3e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 0e0f5c4c..7e2c8596 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_256x128x64_64x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index b3bba802..5876b004 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_16x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu index fff4773f..d88ee03e 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_32x64x64_32x16x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu index 32455afb..13ab6aed 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 169f7179..e0b8a0fd 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x128x64_32x64x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu index 50bb8921..ea2343b9 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu similarity index 91% rename from dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu rename to dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu index 7c9302c9..8a5b7e8b 100644 --- a/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4.cu +++ b/dnn/src/cuda/conv_bias/int8_imma/kimpl/cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4.cu @@ -7,7 +7,7 @@ #include "src/cuda/conv_bias/implicit_gemm_conv_bias_cutlass_wrapper.cuinl" -// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_tensorop_s8_i8816fprop_relu_s8_64x64x64_32x32x64_2_nc32hw32_c32rsk32_nc4hw4" generated by cutlass generator using Convolution = typename cutlass::conv::device::Convolution< int8_t, @@ -37,7 +37,8 @@ using Convolution = 16, 16, true, - cutlass::arch::OpMultiplyAddSaturate>; + cutlass::arch::OpMultiplyAddSaturate, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/conv_bias/opr_impl.h b/dnn/src/cuda/conv_bias/opr_impl.h index 07eaefd1..43e612d2 100644 --- a/dnn/src/cuda/conv_bias/opr_impl.h +++ b/dnn/src/cuda/conv_bias/opr_impl.h @@ -67,6 +67,9 @@ public: class AlgoInt4NCHW64IMMAImplicitGemmBase; class AlgoInt4Int4NCHW64IMMAImplicitGemm; class AlgoUInt4Int4NCHW64IMMAImplicitGemm; + class AlgoInt4NHWCIMMAImplicitGemmBase; + class AlgoInt4Int4NHWCIMMAImplicitGemm; + class AlgoUInt4Int4NHWCIMMAImplicitGemm; class AlgoBFloat16; class AlgoPack; diff --git a/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu index 1d054ae4..63ca450e 100644 --- a/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu +++ b/dnn/src/cuda/convolution/backward_data/cutlass_deconvolution_wrapper.cu @@ -62,7 +62,7 @@ void megdnn::cuda::cutlass_wrapper::do_deconv_int8_implicit_gemm_dp4a_ncdiv4hw4( ThreadBlockShape, WarpShape, InstructionShape, EpilogueOp, \ cutlass::conv::threadblock:: \ ConvolutionDgradNCxHWxThreadblockSwizzle, \ - stage_, 4, aligned_>; \ + stage_, 4, aligned_, true, cutlass::arch::OpMultiplyAdd>; \ typename Deconvolution::ConvolutionParameter conv_param( \ param.n, param.hi, param.wi, param.ci, param.co, param.fh, \ param.fw, param.ho, param.wo, param.ph, param.pw, param.sh, \ diff --git a/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp index 10f0c70d..15380a34 100644 --- a/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp +++ b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw4_dp4a.cpp @@ -48,7 +48,8 @@ bool ConvolutionBackwardDataImpl::AlgoInt8NCHW4DotProdImplicitGemm:: // TODO: support dialtion available &= (fm.dilation[0] == 1 && fm.dilation[1] == 1); // FIXME: too large filter size is not supported now - available &= fm.spatial[0] * fm.spatial[1] <= 64; + available &= fm.spatial[0] * fm.spatial[1] <= + (uint32_t)(848 / (2 * m_algo_param.warp_k / 4) - 2); // only support sm_61 or later, platform should have fast native int8 // support available &= is_compute_capability_required(6, 1); diff --git a/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw_dp4a.cpp b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw_dp4a.cpp index a8f2fd9d..ccf466ab 100644 --- a/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw_dp4a.cpp +++ b/dnn/src/cuda/convolution/backward_data/implicit_gemm_int8_nchw_dp4a.cpp @@ -50,7 +50,7 @@ bool ConvolutionBackwardDataImpl::AlgoInt8NCHWDotProdImplicitGemm:: // TODO: support dialtion available &= (fm.dilation[0] == 1 && fm.dilation[1] == 1); // FIXME: too large filter size is not supported now - available &= fm.spatial[0] * fm.spatial[1] <= 64; + available &= fm.spatial[0] * fm.spatial[1] <= (848 / (2 * 8 / 4) - 2); // only support sm_61 or later, platform should have fast native int8 // support available &= is_compute_capability_required(6, 1); diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4.cu similarity index 92% rename from dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu rename to dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4.cu index 63d8b005..c0c8a8fa 100644 --- a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4.cu +++ b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4.cu @@ -7,7 +7,7 @@ #include "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x128x16_1_nc4hw4_k4rsc4" generated by cutlass generator using Deconvolution = typename cutlass::conv::device::Deconvolution< int8_t, @@ -36,7 +36,8 @@ using Deconvolution = 4, 8, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4.cu similarity index 92% rename from dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu rename to dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4.cu index e704b902..0eb7ddc3 100644 --- a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4.cu +++ b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4.cu @@ -7,7 +7,7 @@ #include "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x128x16_16x64x16_2_nc4hw4_k4rsc4" generated by cutlass generator using Deconvolution = typename cutlass::conv::device::Deconvolution< int8_t, @@ -36,7 +36,8 @@ using Deconvolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4.cu similarity index 92% rename from dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu rename to dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4.cu index a55ed79e..39a50fc1 100644 --- a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4.cu +++ b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4.cu @@ -7,7 +7,7 @@ #include "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_idgrad_identity_s8_16x64x8_16x64x8_2_nc4hw4_k4rsc4" generated by cutlass generator using Deconvolution = typename cutlass::conv::device::Deconvolution< int8_t, @@ -36,7 +36,8 @@ using Deconvolution = 4, 4, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4.cu similarity index 92% rename from dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu rename to dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4.cu index 0cc081ec..cfcbe8ef 100644 --- a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4.cu +++ b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4.cu @@ -7,7 +7,7 @@ #include "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_idgrad_identity_s8_32x128x32_32x64x32_2_nc4hw4_k4rsc4" generated by cutlass generator using Deconvolution = typename cutlass::conv::device::Deconvolution< int8_t, @@ -36,7 +36,8 @@ using Deconvolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4_k4rsc4.cu similarity index 92% rename from dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu rename to dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4_k4rsc4.cu index 5eaa5d46..1da05f64 100644 --- a/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4.cu +++ b/dnn/src/cuda/convolution/backward_data/int8/kimpl/cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4_k4rsc4.cu @@ -7,7 +7,7 @@ #include "src/cuda/convolution/backward_data/implicit_gemm_deconv_cutlass_wrapper.cuinl" -// kernel instance "cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4" generated by cutlass generator +// kernel instance "cutlass_simt_s8_idgrad_identity_s8_64x128x32_64x32x32_2_nc4hw4_k4rsc4" generated by cutlass generator using Deconvolution = typename cutlass::conv::device::Deconvolution< int8_t, @@ -36,7 +36,8 @@ using Deconvolution = 4, 16, true, - cutlass::arch::OpMultiplyAdd>; + cutlass::arch::OpMultiplyAdd, + cutlass::conv::ImplicitGemmMode::GEMM_NT>; diff --git a/dnn/test/cuda/conv_test_utils.cpp b/dnn/test/cuda/conv_test_utils.cpp index f6c4abe3..387e9705 100644 --- a/dnn/test/cuda/conv_test_utils.cpp +++ b/dnn/test/cuda/conv_test_utils.cpp @@ -100,6 +100,11 @@ std::vector get_det_first_bench_args(size_t batch) { std::vector args; args.emplace_back(BenchArgs{batch, 4, 736, 1280, 16, 3, 2}); args.emplace_back(BenchArgs{batch, 16, 384, 640, 16, 3, 1}); + args.emplace_back(BenchArgs{batch, 16, 384, 640, 32, 3, 2}); + args.emplace_back(BenchArgs{batch, 32, 184, 320, 32, 3, 1}); + args.emplace_back(BenchArgs{batch, 32, 384, 640, 64, 3, 2}); + args.emplace_back(BenchArgs{batch, 32, 184, 320, 32, 1, 1}); + args.emplace_back(BenchArgs{batch, 32, 384, 640, 64, 1, 2}); return args; }