fix(dnn/x86): fix matrix_mul quantized performance on vnni

GitOrigin-RevId: 4af6b8be60

dnn/src/x86/conv_bias/opr_impl.cpp

@@ -95,6 +95,14 @@ class ConvBiasImpl::AlgoPack : NonCopyableObj {
 
 public:
     AlgoPack() {
+        //! FIXME: preference to use mkldnn algo on VNNI devices
+        //! But now mkldnn algo preference issue with NCHW->NHWC->NCHW
+#if MEGDNN_X86_WITH_MKL_DNN
+        //! Create the mkldnn algo
+        all_algos.emplace_back(&mkldnn_conv_fp32);
+        all_algos.emplace_back(&mkldnn_matmul_qint8);
+        all_algos.emplace_back(&mkldnn_qint8);
+#endif
         all_algos.emplace_back(&stride1_direct_large_group);
         all_algos.emplace_back(&stride1_direct_small_group);
         all_algos.emplace_back(&stride2_direct_large_group);
@@ -105,14 +113,6 @@ public:
         all_algos.emplace_back(&avx2_stride2_chanwsie_qint8);
         all_algos.emplace_back(&matmul);
 
-        //! preference to use mkldnn algo on VNNI devices
-#if MEGDNN_X86_WITH_MKL_DNN
-        //! Create the mkldnn algo
-        all_algos.emplace_back(&mkldnn_conv_fp32);
-        all_algos.emplace_back(&mkldnn_matmul_qint8);
-        all_algos.emplace_back(&mkldnn_qint8);
-#endif
-
         static CpuOprDelegationStorage<> storage;
         auto matmul_opr = storage.get<MatrixMul>();
         auto&& matmul_algos =
@@ -172,15 +172,18 @@ bool ConvBiasImpl::is_matmul_quantized_prefer(
                 chanwise_avx2_stride2_qint8_usable_preferred(param) ||
                 direct_avx2_stride1_int8_usable_preferred(param) ||
                 direct_avx2_stride2_int8_usable_preferred(param);
-    }
 #if MEGDNN_X86_WITH_MKL_DNN
-    conv_direct_chanwise_mkldnn_usable =
-            conv_direct_chanwise_mkldnn_usable ||
-            mkldnn_qint8_usable_preferred(param) ||
-            mkldnn_matmul_qint8_usable_preferred(param);
+        conv_direct_chanwise_mkldnn_usable =
+                conv_direct_chanwise_mkldnn_usable ||
+                mkldnn_qint8_usable_preferred(param) ||
+                mkldnn_matmul_qint8_usable_preferred(param);
 #endif
+    }
 
-    return !conv_direct_chanwise_mkldnn_usable;
+    return !conv_direct_chanwise_mkldnn_usable ||
+           (is_supported(SIMDType::VNNI) &&
+            !chanwise_avx2_stride1_qint8_usable_preferred(param) &&
+            !chanwise_avx2_stride2_qint8_usable_preferred(param));
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/x86/matrix_mul/algos.cpp

@@ -6,7 +6,8 @@
  *
  * 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.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "src/x86/matrix_mul/algos.h"
@@ -17,7 +18,6 @@
 
 #include "src/x86/matrix_mul/f32/strategy.h"
 
-
 MIDOUT_DECL(megdnn_x86_matmul_kern)
 MIDOUT_DECL(megdnn_x86_matmul_kern_mk8_8x8)
 using namespace megdnn;
@@ -45,17 +45,16 @@ void f32_blas_kern(const MatrixMulImpl::KernParam& kern_param) {
 
 #if MEGDNN_X86_WITH_MKL && SUPPORT_MKL_PACKED_GEMM
 void f32_blas_kern_only_packA(const MatrixMulImpl::KernParam& kern_param,
-                                 const void* a_panel, const void* b_panel) {
-  MEGDNN_MARK_USED_VAR(b_panel);
+                              const void* a_panel, const void* b_panel) {
+    MEGDNN_MARK_USED_VAR(b_panel);
     auto m = kern_param.M, n = kern_param.N, k = kern_param.K;
     const auto Bptr = kern_param.B<dt_float32>();
     auto Cptr = kern_param.C<dt_float32>();
     auto Atrd = kern_param.LDA, Btrd = kern_param.LDB, Ctrd = kern_param.LDC;
     disable_denorm();
     cblas_sgemm_compute(CblasRowMajor, CblasPacked, CblasNoTrans, m, n, k,
-                        static_cast<const float*>(a_panel), Atrd,
-                        Bptr, Btrd, 0.0f, Cptr,
-                        Ctrd);
+                        static_cast<const float*>(a_panel), Atrd, Bptr, Btrd,
+                        0.0f, Cptr, Ctrd);
 }
 #endif
 
@@ -111,8 +110,9 @@ WorkspaceBundle MatrixMulImpl::AlgoF32MKLPackA::get_bundle(
     return {nullptr, {a_size, 0, 0}};
 }
 
-void MatrixMulImpl::AlgoF32MKLPackA::pack_A(const KernParam& kern_param, void* out,
-                                        size_t index, size_t stride) const {
+void MatrixMulImpl::AlgoF32MKLPackA::pack_A(const KernParam& kern_param,
+                                            void* out, size_t index,
+                                            size_t stride) const {
     MEGDNN_MARK_USED_VAR(stride);
     MEGDNN_MARK_USED_VAR(index);
     auto m = kern_param.M, n = kern_param.N, k = kern_param.K;
@@ -164,7 +164,7 @@ size_t get_kern_workspace(MatrixMulImpl::KernSizeParam kern_size_param) {
 
 bool MatrixMulImpl::AlgoInt8x8x32Vnni::usable(
         const KernSizeParam& kern_size_param) const {
-    return kern_size_param.A_type == kern_size_param.B_type &&
+    return kern_size_param.A_type.enumv() == kern_size_param.B_type.enumv() &&
            ((kern_size_param.A_type.enumv() == DTypeEnum::Int8 &&
              kern_size_param.C_type.enumv() == DTypeEnum::Int32) ||
             (kern_size_param.A_type.enumv() == DTypeEnum::QuantizedS8 &&
@@ -389,9 +389,10 @@ size_t MatrixMulImpl::AlgoInt8x8x32AVX2M2N4K16::get_workspace(
                    m, n, k, trans_a, trans_b, strategy, cacheline)
             .get_workspace_size();
 }
-MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL(AlgoInt8x8x32AVX2M2N4K16, megdnn_x86_matmul_kern,
-                                     8, x86::matmul::gemm_avx2_s8s8s32_2x4x16,
-                                    dt_int8, dt_int32);
+MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL(AlgoInt8x8x32AVX2M2N4K16,
+                                     megdnn_x86_matmul_kern, 8,
+                                     x86::matmul::gemm_avx2_s8s8s32_2x4x16,
+                                     dt_int8, dt_int32);
 
 /*************************AlgoInt8x8x32SSEM4N8K2********************/
 MatrixMulImpl::kern_t MatrixMulImpl::AlgoInt8x8x32SSEM4N8K2::get_kern(
@@ -426,9 +427,10 @@ size_t MatrixMulImpl::AlgoInt8x8x32SSEM4N8K2::get_workspace(
                    m, n, k, trans_a, trans_b, strategy, cacheline)
             .get_workspace_size();
 }
-MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL_DETAIL(
-        AlgoInt8x8x32SSEM4N8K2, megdnn_x86_matmul_kern, 9,
-        x86::matmul::gemm_sse_s8s8s32_4x8x2, dt_int8, dt_int32, dt_int16);
+MEGDNN_REG_GEMM_FUNC_FOR_IM2COL_IMPL_DETAIL(AlgoInt8x8x32SSEM4N8K2,
+                                            megdnn_x86_matmul_kern, 9,
+                                            x86::matmul::gemm_sse_s8s8s32_4x8x2,
+                                            dt_int8, dt_int32, dt_int16);
 
 /*************************AlgoF32MK8_8x8********************/
 MatrixMulImpl::kern_t MatrixMulImpl::AlgoF32MK8_8x8::get_kern(

dnn/src/x86/matrix_mul/opr_impl.cpp

@@ -6,7 +6,8 @@
  *
  * 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.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "src/x86/matrix_mul/opr_impl.h"
@@ -41,9 +42,6 @@ class MatrixMulImpl::AlgoPack : NonCopyableObj {
 public:
     AlgoPack() {
         if (is_supported(SIMDType::VNNI)) {
-#if MEGDNN_X86_WITH_MKL_DNN
-            all_algos.emplace_back(&algoint8x8x32mkldnn);
-#endif
 #if MEGDNN_X86_WITH_VNNI
             all_algos.emplace_back(&algoint8x8x32vnni);
 #endif