Megvii Engine Team
a3ea1f153c
feat(mgb/opr): add fast profile and combined Execution strategy
GitOrigin-RevId: 843dc3a790
4 years ago
Megvii Engine Team
c82d88751a
fix(dnn/cuda): add cuda nchw int8 conv impl with nchw4 to fix cu111 compatibility
GitOrigin-RevId: 771968f9ac
4 years ago
Megvii Engine Team
652ec9f251
fix(mgb/dnn): fix backward computation of tqt
GitOrigin-RevId: 850d11a5ce
4 years ago
Megvii Engine Team
f2b42bf09e
chore(dotprod): add arm dotprod attribute for easy use
GitOrigin-RevId: 78c3e72218
4 years ago
Megvii Engine Team
c33a717314
feat(dnn): repalce is_reproducible with algo attribute in opencl, cpu, rocm and cuda
GitOrigin-RevId: 86dead0a11
4 years ago
Megvii Engine Team
97beae2fd8
fix(megdnn): fix megdnn benchmark testcase
GitOrigin-RevId: 726971474a
4 years ago
Megvii Engine Team
2de2222e46
feat(dnn/cuda): add cutlass batched gemv kernel for matmul operator
GitOrigin-RevId: 51702c4e79
4 years ago
Megvii Engine Team
973d2a0ac2
feat(dnn/cuda): add cutlass matmul using split k parallel
GitOrigin-RevId: 650209e35f
4 years ago
Megvii Engine Team
03c921f7c4
feat(dnn/cuda): add cutlass matmul impls
GitOrigin-RevId: 619c8c299c
4 years ago
Megvii Engine Team
5b62acfa01
feat(dnn/armv7): add new matmul strategy k8x8x4
GitOrigin-RevId: 0c6b7fa1b2
4 years ago
Megvii Engine Team
ad87f78a14
chore(imperative): refine tblgen for generating op name
GitOrigin-RevId: f47ceae726
4 years ago
Megvii Engine Team
9cc732f82d
fix(opencl): fix opencl search algo negative stride support
GitOrigin-RevId: 0642d1718d
4 years ago
Megvii Engine Team
cf27dd642c
fix(cuda): use cudnn8.0.4 as cu111 default libs
GitOrigin-RevId: 721ca73bae
4 years ago
Megvii Engine Team
649e4dd750
test(cuda): fix test for cu111
GitOrigin-RevId: 04fe5eb23f
4 years ago
Megvii Engine Team
c69359d00d
fix(dnn/cuda): disable cudnn conv_bias kernels for NCHW4_NCHW tensor format
due to illegal memory access errors caused by the kernels
GitOrigin-RevId: 584cedeeab
4 years ago
Megvii Engine Team
2e4b9a42f7
fix(mgb/gopt): fix folding conv dimshuffle opt pass
GitOrigin-RevId: 878b7de9de
4 years ago
Megvii Engine Team
0e3a6329ff
build(cuda): support cu111 build
GitOrigin-RevId: b3067ba4d5
4 years ago
Megvii Engine Team
e9db061e45
fix(mgb): fix compiling error for cuda-11.1
GitOrigin-RevId: f63e71afa7
4 years ago
Megvii Engine Team
cd7090acbb
fix(opencl): enable image on mali(cl2.1)
GitOrigin-RevId: 0c670fba80
4 years ago
Megvii Engine Team
c51a687cef
chore(mge): update copyright years
GitOrigin-RevId: 46104ac891
4 years ago
Megvii Engine Team
af42ce7e69
fix(megdnn): some fixes of execution policy
GitOrigin-RevId: 920f39bcb6
4 years ago
Megvii Engine Team
7afa422df4
refactor(megdnn): refactor sub opr setter
GitOrigin-RevId: 475afb9c10
4 years ago
Megvii Engine Team
821656aa4b
refactor(megdnn): refactor brute force algo in batched matmul
GitOrigin-RevId: 5c143ab3ac
4 years ago
Megvii Engine Team
08ff62deb6
refactor(megdnn): refactor batched matmul algo in conv bias
GitOrigin-RevId: 64fda611ff
4 years ago
Megvii Engine Team
8773926ef8
refactor(megdnn): refactor matmul algo in conv bias
GitOrigin-RevId: 932f7d6f81
4 years ago
Megvii Engine Team
e4b71bdf64
refactor(megdnn): remove unnessary 1x1 algo
GitOrigin-RevId: 809f9b2bba
4 years ago
Megvii Engine Team
44c8d2d16f
refactor(megdnn): refactor matmul algo in deformable conv
GitOrigin-RevId: 05291baf98
4 years ago
Megvii Engine Team
b04ad06f84
refactor(megdnn): refactor matmul algo in conv backward filter
GitOrigin-RevId: bb65777583
4 years ago
Megvii Engine Team
25089e520e
refactor(megdnn): refactor matmul algo in conv backward data
GitOrigin-RevId: 8de601df6e
4 years ago
Megvii Engine Team
0d720653ac
refactor(megdnn): add default algo for convolution forward
GitOrigin-RevId: a12a7d399a
4 years ago
Megvii Engine Team
659217acd2
refactor(megdnn): refactor bfloat16 convbias to recursive inteface
GitOrigin-RevId: 378194fb7f
4 years ago
Megvii Engine Team
4a1d52c9c6
refactor(megdnn): refactor bfloat16 matmul to recursive inteface
GitOrigin-RevId: 641c508aec
4 years ago
Megvii Engine Team
b8febaf91f
refactor(megdnn): refactor bfloat16 convolutionbackwardfilter to recursive inteface
GitOrigin-RevId: 37c08a5b8b
4 years ago
Megvii Engine Team
f14e0c17e7
feat(mgb): add recursive for fastrun and megdnn test
GitOrigin-RevId: 743846f645
4 years ago
Megvii Engine Team
85fa988348
refactor(dnn): add get_algorithm_from_desc interface
GitOrigin-RevId: 6d211ca167
4 years ago
Megvii Engine Team
2b8150ab52
fix(dnn): fix bazel build issue for cambricon platform
GitOrigin-RevId: a79e1b6f6a
4 years ago
Megvii Engine Team
0e8b81c20e
fix(dnn/opencl): fix elemwise negative stride support
GitOrigin-RevId: 506d7e6104
4 years ago
Megvii Engine Team
329306b031
fix(cmake/cuda): fix build at cuda `copy` env caused by b278a69e1
`copy` env vs `normal` env:
case a:`normal` env install CUDNN/TRT by package tool, eg, apt/yum, which
will put include file(cudnn.h/NvInfer.h) to /usr/include etc.
case b: self copy env after export CPATH with cudnn/trt include dir
case c: just copy libs, not export CPATH, only export
TRT_ROOT_DIR/CUDNN_ROOT_DIR
at case a and case b, BUILD system can find cudnn/trt headers, but
at case c, can not find. `FIX` it!
what`s more, recommend install CUDA env by case a and case c, NOT
recommend install by case b, which caused by CUDA SDK CPATH not
friendly to split c env, for example Android AOSP build env
GitOrigin-RevId: cb92123f9e
4 years ago
Megvii Engine Team
364afec033
chore(mge): update copyright years
GitOrigin-RevId: 3c0690bcc1
4 years ago
Megvii Engine Team
ae8b38f634
fix(cmake/whl): reduce wheel size
as some reason: some version of trt/cudnn need
-Wl,--whole-archive attribute to fix cuda issue,
target A(CMAKE OBJECT,eg megbrain/megdnn): -Wl,--whole-archive depend on trt/cudnn
target B(SHARED LIBS,eg libmegengine.so) --> A with `PUBLIC` depends
target C(SHARED LIBS,eg _impertive.so) --> B with `PUBLIC` depends,
then ld will force link fatbin section into C, which will be undesired
what`s more, attribute PUBLIC/PRIVATE do not take effect
to OBJECT library(megbrain/megdnn)
what`s more, megengine/megengine_export have to PUBLIC for
mgb/imperative target, as SRC/include depends
so we pull cudalib depend from megbrain/megdnn to
megengine/megengine_export on linux os, to finall
target at windows os, for example lar on windows
GitOrigin-RevId: b278a69e1c
4 years ago
Megvii Engine Team
3bda334798
fix(dnn/fallback): fix segmentfault caused by im2col/conv1x1 using
fallback naive matmul.
GitOrigin-RevId: 03ef904b11
4 years ago
Megvii Engine Team
87ff58f7fc
fix(megdnn): add algo for matmul/batchedmatrixmul of naive and opencl
GitOrigin-RevId: 2409b6ba16
4 years ago
Megvii Engine Team
a3caa5d3b7
fix(mgb(dnn)): fix convbias cudnnConvBiasActivation
GitOrigin-RevId: c0e44feffb
4 years ago
Megvii Engine Team
55042195d4
chore(winograd): add Convolutionv2 param
GitOrigin-RevId: 1a9e2ea340
4 years ago
Megvii Engine Team
409a877267
feat(dnn): add algo interface for rocm&fallback matmul and batched matrix mul
GitOrigin-RevId: dea03a0f7a
4 years ago
Megvii Engine Team
8f7f52ae4d
feat(jit): add memfwd in jit executor opr
GitOrigin-RevId: b58860bbe8
4 years ago
Megvii Engine Team
dfb2b2ce49
fix(dnn): change pooling window size smaller than padding constraint to log_error
GitOrigin-RevId: c3cda68f6d
4 years ago
Megvii Engine Team
d1fbec4fe2
feat(dnn/atlas): add atlas stub
GitOrigin-RevId: c63294378e
4 years ago
Megvii Engine Team
a85531dd0f
feat(mgb/opr): add tqt opr
GitOrigin-RevId: 49c62cd532
4 years ago
Megvii Engine Team
c3a4b2225d
feat(dnn/cuda): add cutlass impls for fused convolution reformat operation
GitOrigin-RevId: 02ef559c3f
4 years ago