avx2 gemm now works for single thread

include/ck/tensor_operation/cpu/block/blockwise_gemm_avx2.hpp

@@ -13,21 +13,10 @@ namespace cpu {
 template <typename FloatA,
           typename FloatB,
           typename FloatC,
-          typename AccDataType,
+
           typename ABlockDesc,
           typename BBlockDesc,
-          typename CBlockDesc,
-
-          typename ABlockSliceLengths,
-          typename BBlockSliceLengths,
-          typename CBlockSliceLengths,
-
-          typename AThreadSliceLength,
-          typename BThreadSliceLength,
-
-          ck::index_t AThreadLoopOverDim, // thread slice loop over on block slice. 1d is enough for
-                                          // now
-          ck::index_t BThreadLoopOverDim,
+          typename CDesc,
 
           ck::index_t KPerBlock,
 
@@ -47,24 +36,14 @@ struct BlockwiseGemmAvx2_MxN
 
     static constexpr index_t nDimA = ABlockDesc::GetNumOfDimension();
     static constexpr index_t nDimB = BBlockDesc::GetNumOfDimension();
-    static constexpr index_t nDimC = CBlockDesc::GetNumOfDimension();
+    static constexpr index_t nDimC = CDesc::GetNumOfDimension();
     using IndexA                   = MultiIndex<nDimA>;
     using IndexB                   = MultiIndex<nDimB>;
     using IndexC                   = MultiIndex<nDimC>;
 
     using ACoord = decltype(make_tensor_coordinate(ABlockDesc{}, IndexA{}));
     using BCoord = decltype(make_tensor_coordinate(BBlockDesc{}, IndexB{}));
-    using CCoord = decltype(make_tensor_coordinate(CBlockDesc{}, IndexC{}));
-
-#if 0
-    constexpr BlockwiseGemmAvx2_MxN(const ABlockDesc & a_block_desc, const IndexA& a_thread_origin,
-                                    const BBlockDesc & b_block_desc, const IndexB& b_thread_origin)
-        : a_thread_coord_(make_tensor_coordinate(a_block_desc, a_thread_origin)),
-          b_thread_coord_(make_tensor_coordinate(b_block_desc, b_thread_origin)),
-    {
-
-    }
-#endif
+    using CCoord = decltype(make_tensor_coordinate(CDesc{}, IndexC{}));
 
     template <typename TensorDesc>
     constexpr auto GetLeadingElement(const TensorDesc& desc)
@@ -84,79 +63,175 @@ struct BlockwiseGemmAvx2_MxN
         }
     }
 
-    template <typename ABlockBuffer, typename BBlockBuffer, typename CBlockBuffer>
+    ck::index_t GetALeadingElement(const ABlockDesc& a_block_desc) const
+    {
+        return a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+    }
+
+    ck::index_t GetBLeadingElement(const BBlockDesc& b_block_desc) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            // K * N
+            return b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+        else
+        {
+            // N/8 * K * 8
+            return b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}] *
+                   b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+        }
+    }
+
+    ck::index_t GetCLeadingElement(const CDesc& c_desc) const
+    {
+        return c_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+    }
+
+    ck::index_t GetMPerBlock(const ABlockDesc& a_block_desc) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            // M * K
+            return a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+        }
+        else
+        {
+            // K * M
+            return a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+    }
+
+    ck::index_t GetKPerBlock(const ABlockDesc& a_block_desc) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            // M * K
+            return a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+        else
+        {
+            // K * M
+            return a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+        }
+    }
+
+    ck::index_t GetNPerBlock(const BBlockDesc& b_block_desc) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            // K * N
+            return b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+        else
+        {
+            // N/8 * K * 8
+            return b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}] *
+                   b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+        }
+    }
+
+    ck::index_t
+    GetABlockStartOffset(const ABlockDesc& a_block_desc, const index_t i_m, const index_t) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return i_m * a_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+        else
+        {
+            return i_m;
+        }
+    }
+
+    ck::index_t
+    GetBBlockStartOffset(const BBlockDesc& b_block_desc, const index_t, const index_t i_n) const
+    {
+        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
+                                  ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            // K * N
+            return i_n;
+        }
+        else
+        {
+            // N/8 * K * 8
+            return i_n * b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        }
+    }
+
+    ck::index_t
+    GetCBlockStartOffset(const CDesc& c_desc, const index_t i_m, const index_t i_n) const
+    {
+        return i_m * c_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}] + i_n;
+    }
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CBuffer>
     void Run(const ABlockDesc& a_block_desc,
              const ABlockBuffer& a_block_buf,
-             const IndexA& a_origin,
+             const IndexA& /* a_origin */,
 
              const BBlockDesc& b_block_desc,
              const BBlockBuffer& b_block_buf,
-             const IndexB& b_origin,
+             const IndexB& /* b_origin */,
+
+             const CDesc& c_desc,
+             CBuffer& c_buf,
+             const IndexC& /* c_origin */,
 
-             const CBlockDesc& c_block_desc,
-             CBlockBuffer& c_block_buf,
-             const IndexC& c_origin) const
+             bool is_accumulate_c = true) const
     {
+        auto lda = GetALeadingElement(a_block_desc) * sizeof(FloatA);
+        auto ldb = GetBLeadingElement(b_block_desc) * sizeof(FloatB);
+        auto ldc = GetCLeadingElement(c_desc) * sizeof(FloatC);
 
-        constexpr auto m_n_block_length =
-            ck::Sequence<ABlockSliceLengths::At(AThreadLoopOverDim),
-                         BBlockSliceLengths::At(BThreadLoopOverDim)>{};
-        constexpr auto m_n_thread_length =
-            ck::Sequence<AThreadSliceLength::At(AThreadLoopOverDim),
-                         BThreadSliceLength::At(BThreadLoopOverDim)>{};
+        // printf("lda:%d, ldb:%d, ldc:%d\n", lda, ldb, ldc);
 
-        constexpr auto m_n_access_length = m_n_block_length / m_n_thread_length;
+        const auto k_per_block  = GetKPerBlock(a_block_desc);
+        const auto m_per_block  = GetMPerBlock(a_block_desc);
+        const auto n_per_block  = GetNPerBlock(b_block_desc);
+        const auto m_per_thread = ThreadwiseGemm_Dispatch::ThreadMaxMr;
+        const auto n_per_thread = ThreadwiseGemm_Dispatch::ThreadMaxNr;
 
-        constexpr auto ordered_m_n_access_length =
-            container_reorder_given_new2old(m_n_access_length, ThreadMNAccessOrder{});
+        ck::cpu::ThreadwiseGemmParam param;
+        param.Kr          = k_per_block;
+        param.lda         = lda;
+        param.ldb         = ldb;
+        param.ldc         = ldc;
+        param.alpha       = 1.0f; // TODO
+        param.accmulate_c = is_accumulate_c ? 1 : 0;
+
+        if constexpr(std::is_same<ThreadMNAccessOrder, ck::Sequence<0, 1>>::value)
+        {
+            for(ck::index_t i_m = 0; i_m < m_per_block; i_m += m_per_thread)
+            {
+                auto current_mr = ck::math::min(m_per_block - i_m, m_per_thread);
+                param.p_a       = &a_block_buf.p_data_[GetABlockStartOffset(a_block_desc, i_m, 0)];
 
-        constexpr auto a_block_idx_zeros =
-            typename uniform_sequence_gen<nDimA, 0>::type{}; // starting point of the block
-        constexpr auto b_block_idx_zeros = typename uniform_sequence_gen<nDimB, 0>::type{};
+                // printf("YYYY: %d, i_m:%d, current_mr:%d, %d, %p\n",__LINE__, i_m, current_mr,
+                // GetABlockStartOffset(a_block_desc, i_m, 0), param.p_a);fflush(stdout);
 
-        constexpr auto lda = GetLeadingElement(a_block_desc) * sizeof(FloatA);
-        constexpr auto ldb = GetLeadingElement(b_block_desc) * sizeof(FloatB);
-        constexpr auto ldc = GetLeadingElement(c_block_desc) * sizeof(FloatC);
+                for(ck::index_t i_n = 0; i_n < n_per_block; i_n += n_per_thread)
+                {
+                    auto current_nr = ck::math::min(n_per_block - i_n, n_per_thread);
 
-        ck::cpu::ThreadwiseGemmParam param;
-        param.Kr    = KPerBlock;
-        param.lda   = lda;
-        param.ldb   = ldb;
-        param.ldc   = ldc;
-        param.alpha = 1.0f; // TODO
-
-        static_ford<decltype(ordered_m_n_access_length)>{}([&](auto ordered_idx) {
-            constexpr auto origin_m_n_idx = ordered_idx.ReorderGivenOld2New(ThreadMNAccessOrder{});
-
-            constexpr auto current_m_idx =
-                origin_m_n_idx.At(0) * AThreadSliceLength::At(AThreadLoopOverDim);
-            constexpr auto current_n_idx =
-                origin_m_n_idx.At(1) * BThreadSliceLength::At(BThreadLoopOverDim);
-
-            constexpr auto current_mr =
-                ck::math::min(m_n_block_length.At(0) - current_m_idx, m_n_thread_length.At(0));
-            constexpr auto current_nr =
-                ck::math::min(m_n_block_length.At(1) - current_n_idx, m_n_thread_length.At(1));
-
-            constexpr auto a_block_idx =
-                a_block_idx_zeros.Modify(AThreadLoopOverDim, current_m_idx);
-            constexpr auto a_block_coord =
-                make_tensor_coordinate(a_block_desc, to_multi_index(a_origin + a_block_idx));
-
-            constexpr auto b_block_idx =
-                b_block_idx_zeros.Modify(BThreadLoopOverDim, current_n_idx);
-            constexpr auto b_block_coord =
-                make_tensor_coordinate(b_block_desc, to_multi_index(b_origin + b_block_idx));
-
-            constexpr auto c_block_coord =
-                make_tensor_coordinate(c_block_desc, to_multi_index(c_origin + origin_m_n_idx));
-
-            param.p_a = &a_block_buf.p_data_[a_block_coord.GetOffset()];
-            param.p_b = &b_block_buf.p_data_[b_block_coord.GetOffset()];
-            param.p_c = &c_block_buf.p_data_[c_block_coord.GetOffset()];
-
-            ThreadwiseGemm_Dispatch::Run(&param, current_mr, current_nr);
-        });
+                    param.p_b = &b_block_buf.p_data_[GetBBlockStartOffset(b_block_desc, 0, i_n)];
+                    param.p_c = &c_buf.p_data_[GetCBlockStartOffset(c_desc, i_m, i_n)];
+
+                    // printf("YYYY: %d, i_n:%d, current_nr:%d, %d, %p, C:%d, %p\n",__LINE__, i_n,
+                    // current_nr, GetBBlockStartOffset(b_block_desc, 0, i_n), param.p_b,
+                    //        GetCBlockStartOffset(c_desc, i_m, i_n),
+                    //        param.p_c);fflush(stdout);
+
+                    ThreadwiseGemm_Dispatch::Run(&param, current_mr, current_nr);
+                }
+            }
+        }
     }
 };
 

include/ck/tensor_operation/cpu/device/convolution_forward_specialization_cpu.hpp

@@ -14,6 +14,19 @@ enum ConvolutionForwardSpecialization_t
     OddC,
 };
 
+enum ConvolutionForwardGemmKSpecialization_t
+{
+    DefaultGemmKLoop,
+    NHWC_GemmKLoopOverC, // not merge c*y*x, and c % k_per_block == 0
+};
+
+enum ConvolutionForwardBlockLoopOverSpecialization_t
+{
+    DefaultBlockLoopOver,
+    LoopOver_MNK,
+    LoopOver_MKN,
+};
+
 } // namespace device
 } // namespace cpu
 } // namespace tensor_operation

include/ck/tensor_operation/cpu/device/device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp

@@ -13,6 +13,8 @@
 #include "tensor_descriptor.hpp"
 #include "tensor_descriptor_helper.hpp"
 #include "gridwise_gemm_avx2.hpp"
+#include "threadwise_gemm_avx2.hpp"
+#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -23,20 +25,21 @@ namespace device {
 template <typename InDataType,
           typename WeiDataType,
           typename OutDataType,
-          typename AccDataType,
           typename InElementwiseOperation,
           typename WeiElementwiseOperation,
           typename OutElementwiseOperation,
           ConvolutionForwardSpecialization_t ConvForwardSpecialization,
+          ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
+          ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
           ck::index_t NumDimSpatial,
           ck::index_t MPerBlock, // block means data are designed to fit in cache (L1/L2/L3)
           ck::index_t NPerBlock,
           ck::index_t KPerBlock,
-          typename ThreadwiseGemm_Dispatch>
-
-//  bool IsGemmMPadded,
-//  bool IsGemmNPadded,
-//  bool IsGemmKPadded>
+          ck::index_t MPerThread,
+          ck::index_t NPerThread,
+          bool UseALocalBuffer,
+          bool UseBLocalBuffer,
+          bool UseCLocalBuffer>
 struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
     : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
 {
@@ -60,18 +63,89 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
     static constexpr auto I2 = Number<2>{};
     static constexpr auto I3 = Number<3>{};
 
+    static constexpr bool NonTemporalStore = false;
+
+    static constexpr auto GetBlockMNKAccessOrder()
+    {
+        if constexpr(BlockLoopOverSpecialization == DefaultBlockLoopOver ||
+                     BlockLoopOverSpecialization == LoopOver_MNK)
+            return ck::Sequence<0, 1, 2>{};
+        else if constexpr(BlockLoopOverSpecialization == LoopOver_MKN)
+            return ck::Sequence<0, 2, 1>{};
+    }
+
+    using BlockMNKAccessOrder = decltype(GetBlockMNKAccessOrder());
+
+    static constexpr auto GetThreadwiseGemm_Dispatch()
+    {
+        if constexpr(MPerThread == 4 && NPerThread == 24)
+        {
+            return ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<
+                InDataType,
+                WeiDataType,
+                OutDataType,
+                ck::tensor_layout::gemm::RowMajor,
+                ck::tensor_layout::gemm::ColumnMajor,
+                NonTemporalStore>{};
+        }
+        else if constexpr(MPerThread == 6 && NPerThread == 16)
+        {
+            return ck::cpu::ThreadwiseGemmAvx2_MxN_6x16_Dispatch<
+                InDataType,
+                WeiDataType,
+                OutDataType,
+                ck::tensor_layout::gemm::RowMajor,
+                ck::tensor_layout::gemm::ColumnMajor,
+                NonTemporalStore>{};
+        }
+        else
+        {
+            // static_assert(false, "invalid Mr/Nr");
+        }
+    }
+
+    using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
+
+    static constexpr auto GetInputBlockDescriptor()
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+    }
+
+    static constexpr auto GetWeightBlockDescriptor()
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(
+            math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
+            KPerBlock,
+            ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
+    }
+
+    static constexpr auto GetOutputBlockDescriptor()
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+    }
+
     static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
     {
+        ck::index_t gemm_n_padded =
+            math::integer_least_multiple(gemm_n, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
         const auto wei_gemm_n_k_grid_desc =
             make_naive_tensor_descriptor_packed(make_tuple(gemm_n, gemm_k));
 
-        const auto wei_gemm_n0_k_n1_grid_desc = transform_tensor_descriptor(
+        const auto wei_gemm_padn_k_grid_desc = transform_tensor_descriptor(
             wei_gemm_n_k_grid_desc,
-            ck::make_tuple(ck::make_unmerge_transform(
-                               ck::make_tuple(wei_gemm_n_k_grid_desc.GetLength(I0) /
-                                                  ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
-                                              ThreadwiseGemm_Dispatch::MatrixBMinVectorSize)),
-                           ck::make_pass_through_transform(wei_gemm_n_k_grid_desc.GetLength(I1))),
+            make_tuple(make_right_pad_transform(gemm_n, gemm_n_padded - gemm_n),
+                       make_pass_through_transform(gemm_k)),
+            ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
+            ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
+
+        const auto wei_gemm_n0_k_n1_grid_desc = transform_tensor_descriptor(
+            wei_gemm_padn_k_grid_desc,
+            ck::make_tuple(
+                ck::make_unmerge_transform(
+                    ck::make_tuple(wei_gemm_padn_k_grid_desc.GetLength(I0) /
+                                       ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
+                                   ThreadwiseGemm_Dispatch::MatrixBMinVectorSize)),
+                ck::make_pass_through_transform(wei_gemm_padn_k_grid_desc.GetLength(I1))),
             ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}),
             ck::make_tuple(ck::Sequence<0, 2>{}, ck::Sequence<1>{}));
 
@@ -409,6 +483,13 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                                    std::multiplies<ck::index_t>());
     }
 
+    static index_t GetGemmN(ck::index_t K)
+    {
+        // return ck::math::integer_least_multiple(K,
+        // ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
+        return K;
+    }
+
     static auto MakeABCGridDescriptor(ck::index_t N,
                                       ck::index_t K,
                                       ck::index_t C,
@@ -423,7 +504,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
         using namespace ck;
 
         const index_t GemmM = GetGemmM(N, output_spatial_lengths);
-        const index_t GemmN = K;
+        const index_t GemmN = GetGemmN(K);
         const index_t GemmK = GetGemmK(C, filter_spatial_lengths);
 
         // A:
@@ -474,13 +555,44 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
     using BGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
     using CGridDesc = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
 
-    static constexpr bool UseCLocalBuffer = true;
+    // static constexpr bool UseCLocalBuffer = false;
+
+    using AThreadwiseCopy =
+        ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC<
+            InDataType,
+            InDataType,
+            AGridDesc,
+            decltype(GetInputBlockDescriptor()),
+            InElementwiseOperation,
+            false,
+            ConvForwardSpecialization,
+            GemmKSpecialization>;
+
+    using BThreadwiseCopy =
+        ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC<
+            WeiDataType,
+            WeiDataType,
+            BGridDesc,
+            decltype(GetWeightBlockDescriptor()),
+            WeiElementwiseOperation,
+            false,
+            ConvForwardSpecialization,
+            GemmKSpecialization>;
+
+    using CThreadwiseCopy = ck::cpu::ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN<
+        OutDataType,
+        OutDataType,
+        CGridDesc,
+        decltype(GetOutputBlockDescriptor()),
+        OutElementwiseOperation,
+        !UseCLocalBuffer,
+        ConvForwardSpecialization,
+        GemmKSpecialization>;
 
     using GridwiseGemm =
         ck::cpu::GridwiseGemmAvx2_MxN<InDataType,              // InDataType,
                                       WeiDataType,             // WeiDataType,
                                       OutDataType,             // OutDataType,
-                                      AccDataType,             // AccDataType,
                                       AGridDesc,               // AGridDesc,
                                       BGridDesc,               // BGridDesc,
                                       CGridDesc,               // CGridDesc,
@@ -491,8 +603,13 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                                       NPerBlock,               // NPerBlock,
                                       KPerBlock,               // KPerBlock,
                                       ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
-                                      ck::Sequence<0, 1, 2>,   // BlockMNKAccessOrder,
+                                      AThreadwiseCopy,         // AThreadwiseCopy
+                                      BThreadwiseCopy,         // BThreadwiseCopy
+                                      CThreadwiseCopy,         // CThreadwiseCopy
+                                      BlockMNKAccessOrder,     // BlockMNKAccessOrder,
                                       ck::Sequence<0, 1>,      // ThreadMNAccessOrder
+                                      UseALocalBuffer,         // UseALocalBuffer
+                                      UseBLocalBuffer,         // UseBLocalBuffer
                                       UseCLocalBuffer          // UseCLocalBuffer
                                       >;
 
@@ -580,6 +697,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
 
+            memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
+
             const auto kernel = ck::cpu::kernel_gemm_avx_mxn<GridwiseGemm,
                                                              InDataType,
                                                              WeiDataType,
@@ -591,21 +710,24 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                                                              BElementwiseOperation,
                                                              CElementwiseOperation>;
 
-            float ave_time = launch_and_time_cpu_kernel(kernel,
-                                                        nrepeat,
-                                                        arg.p_a_grid_,
-                                                        arg.p_b_grid_,
-                                                        arg.p_c_grid_,
-                                                        arg.a_grid_desc_,
-                                                        arg.b_grid_desc_,
-                                                        arg.c_grid_desc_,
-                                                        arg.a_element_op_,
-                                                        arg.b_element_op_,
-                                                        arg.c_element_op_);
+            float ave_time = 0;
+
+            if(nrepeat != 1)
+                ave_time = launch_and_time_cpu_kernel(kernel,
+                                                      nrepeat,
+                                                      arg.p_a_grid_,
+                                                      arg.p_b_grid_,
+                                                      arg.p_c_grid_,
+                                                      arg.a_grid_desc_,
+                                                      arg.b_grid_desc_,
+                                                      arg.c_grid_desc_,
+                                                      arg.a_element_op_,
+                                                      arg.b_element_op_,
+                                                      arg.c_element_op_);
 
             // TODO: this is for benchmark purpose, so last time we clear c buffer and calculate the
             // result
-            memset(arg.p_c_grid_, 0, arg.a_grid_desc_.GetElementSpaceSize());
+            memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
                               arg.p_a_grid_,
@@ -659,6 +781,13 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             }
         }
 
+        if constexpr(GemmKSpecialization ==
+                     ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
+        {
+            if(!(arg.Conv_C_ % KPerBlock == 0))
+                return false;
+        }
+
         // Gridwise GEMM size
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
@@ -748,16 +877,25 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
     std::string GetTypeString() const override
     {
-        auto str = std::stringstream();
-
+        auto str                 = std::stringstream();
+        auto string_local_buffer = [](bool is_local_buffer) {
+            if(is_local_buffer)
+                return "L";
+            else
+                return "G";
+        };
         // clang-format off
         str << "DeviceConv" << std::to_string(NumDimSpatial) 
-            << "DFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
-            << "<"
-            << MPerBlock << ", "
-            << NPerBlock << ", "
-            << KPerBlock
-            << ">";
+            << "DFwdAvx2_NHWC_KYXC"
+            <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
+            <<"_KS"<< static_cast<int>(GemmKSpecialization)
+            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
+            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            << "_TT" << MPerThread << "x" << NPerThread 
+            << "_A" << string_local_buffer(UseALocalBuffer)
+            << "_B" << string_local_buffer(UseBLocalBuffer)
+            << "_C" << string_local_buffer(UseCLocalBuffer)
+            ;
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/cpu/grid/gridwise_gemm_avx2.hpp

@@ -7,7 +7,9 @@
 #include "tensor_descriptor_helper.hpp"
 #include "blockwise_gemm_avx2.hpp"
 #include "threadwise_tensor_slice_transfer_avx2.hpp"
+#include "threadwise_tensor_slice_transfer_avx2_specialization.hpp"
 #include "dynamic_buffer_cpu.hpp"
+#include <unistd.h>
 
 namespace ck {
 namespace cpu {
@@ -46,7 +48,6 @@ void kernel_gemm_avx_mxn(const FloatA* __restrict__ p_a_grid,
 template <typename FloatA,
           typename FloatB,
           typename FloatC,
-          typename AccDataType,
           typename AGridDesc,
           typename BGridDesc,
           typename CGridDesc,
@@ -57,334 +58,92 @@ template <typename FloatA,
           ck::index_t NPerBlock,
           ck::index_t KPerBlock,
           typename ThreadwiseGemm_Dispatch,
+          typename AThreadwiseCopy,
+          typename BThreadwiseCopy,
+          typename CThreadwiseCopy,
           typename BlockMNKAccessOrder, // how we accss gemm MNK to better fit in cache
           typename ThreadMNAccessOrder, // how we acces gemm MN to utilize micro kernel
+          bool UseALocalBuffer,
+          bool UseBLocalBuffer,
           bool UseCLocalBuffer // if true, will allocate a buffer and write to it in kernel, then
-                               // copy back to block buffer. if false, will write to C directly
+                               // copy back to block buffer (need CThreadwiseCopy).
+                               // if false, will write to C directly (no need CThreadwiseCopy)
           >
 struct GridwiseGemmAvx2_MxN
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
-    static constexpr auto I2 = Number<2>{};
-    static constexpr auto I3 = Number<3>{};
-    static constexpr auto I4 = Number<4>{};
-    static constexpr auto I5 = Number<5>{};
-    static constexpr auto I6 = Number<6>{};
-    static constexpr auto I7 = Number<7>{};
 
     // static constexpr auto Avx2RegisterVector = 8;   // 8 floats
     static constexpr index_t MemAlignmentByte = 32; // 256bit
 
-    static constexpr auto GetABlockDescriptor()
+    static auto GetABlockDescriptor(const ck::index_t m_per_blk, const ck::index_t k_per_blk)
     {
         if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
                                   ck::tensor_layout::gemm::RowMajor>::value)
         {
             // A : M, K
-            constexpr auto a_block_desc_m_k =
-                make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            auto a_block_desc_m_k =
+                make_naive_tensor_descriptor_packed(make_tuple(m_per_blk, k_per_blk));
             return a_block_desc_m_k;
         }
         else
         {
             // A : K, M
-            constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_packed(
-                make_tuple(KPerBlock,
+            auto a_block_desc_k_m = make_naive_tensor_descriptor_packed(
+                make_tuple(k_per_blk,
                            math::integer_least_multiple(
-                               MPerBlock, ThreadwiseGemm_Dispatch::MatrixAMinVectorSize)));
+                               m_per_blk, ThreadwiseGemm_Dispatch::MatrixAMinVectorSize)));
             return a_block_desc_k_m;
         }
     }
 
-    static constexpr auto GetBBlockDescriptor()
+    static auto GetBBlockDescriptor(const ck::index_t k_per_blk, const ck::index_t n_per_blk)
     {
+        // n_per_blk should be 8x
         if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
                                   ck::tensor_layout::gemm::RowMajor>::value)
         {
             // B : K, N
-            constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_packed(
-                make_tuple(KPerBlock,
-                           math::integer_least_multiple(
-                               NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize)));
+            auto b_block_desc_k_n =
+                make_naive_tensor_descriptor_packed(make_tuple(k_per_blk, n_per_blk));
             return b_block_desc_k_n;
         }
         else
         {
             // B : N/8, K, N8
-            constexpr auto b_block_desc_n0_k_n1 = make_naive_tensor_descriptor_packed(make_tuple(
-                math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
-                KPerBlock,
+            auto b_block_desc_n0_k_n1 = make_naive_tensor_descriptor_packed(make_tuple(
+                math::integer_divide_ceil(n_per_blk, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
+                k_per_blk,
                 ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
             return b_block_desc_n0_k_n1;
         }
     }
 
-    static constexpr auto GetABlockSliceLength()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // A : M, K
-            return ck::Sequence<MPerBlock, KPerBlock>{};
-        }
-        else
-        {
-            // A : K, M
-            return ck::Sequence<KPerBlock, MPerBlock>{};
-        }
-    }
-
-    static constexpr auto GetBBlockSliceLength()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return ck::Sequence<KPerBlock, NPerBlock>{};
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return ck::Sequence<NPerBlock / ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
-                                KPerBlock,
-                                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize>{};
-        }
-    }
-
-    static constexpr auto GetABlockDimAccessOrder() { return ck::Sequence<0, 1>{}; }
-
-    static constexpr auto GetBBlockDimAccessOrder()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return ck::Sequence<0, 1>{};
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return ck::Sequence<0, 1, 2>{};
-        }
-    }
-
-    static constexpr auto GetABlockMoveFwdStep()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // A : M, K
-            return ck::make_multi_index(0, KPerBlock);
-        }
-        else
-        {
-            // A : K, M
-            return ck::make_multi_index(KPerBlock, 0);
-        }
-    }
-
-    static constexpr auto GetBBlockMoveFwdStep()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return ck::make_multi_index(KPerBlock, 0);
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return ck::make_multi_index(0, KPerBlock, 0);
-        }
-    }
-
-#if 0
-    static constexpr auto GetAThreadDiscriptor()
-    {
-        if constexpr (std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout, ck::tensor_layout::gemm::RowMajor>::value){
-            // A : M, K
-            constexpr auto a_thread_desc_m_k = make_naive_tensor_descriptor_packed(make_tuple(ThreadwiseGemm_Dispatch::ThreadMaxMr, KPerBlock));
-            return a_thread_desc_m_k;
-        } else {
-            // A : K, M
-            constexpr auto a_thread_desc_k_m = make_naive_tensor_descriptor_packed(make_tuple(KPerBlock, ThreadwiseGemm_Dispatch::ThreadMaxMr));
-            return a_thread_desc_k_m;
-        }
-    }
-
-    static constexpr auto GetBThreadDescriptor()
-    {
-        if constexpr (std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout, ck::tensor_layout::gemm::RowMajor>::value){
-            // B : K, N
-            constexpr auto b_thread_desc_k_n = make_naive_tensor_descriptor_packed(make_tuple(KPerBlock, ThreadwiseGemm_Dispatch::ThreadMaxNr));
-            return b_thread_desc_k_n;
-        } else {
-            // B : N/8, K, N8
-            constexpr auto b_thread_desc_n_k_n8 = make_naive_tensor_descriptor_packed(make_tuple(math::integer_divide_ceil(ThreadwiseGemm_Dispatch::ThreadMaxNr, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize), KPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
-            return b_thread_desc_n_k_n8;
-        }
-    }
-#endif
-
-    static constexpr auto GetAThreadSliceLength()
+    static auto GetCBlockDescriptor(const ck::index_t m_per_blk, const ck::index_t n_per_blk)
     {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // A : M, K
-            return ck::Sequence<ThreadwiseGemm_Dispatch::ThreadMaxMr, KPerBlock>{};
-        }
-        else
-        {
-            // A : K, M
-            return ck::Sequence<KPerBlock, ThreadwiseGemm_Dispatch::ThreadMaxMr>{};
-        }
-    }
-
-    static constexpr auto GetBThreadSliceLength()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return ck::Sequence<KPerBlock, ThreadwiseGemm_Dispatch::ThreadMaxNr>{};
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return ck::Sequence<ThreadwiseGemm_Dispatch::ThreadMaxNr /
-                                    ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
-                                KPerBlock,
-                                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize>{};
-        }
+        return make_naive_tensor_descriptor_packed(make_tuple(m_per_blk, n_per_blk));
     }
 
-    static constexpr auto GetAThreadMoveFwdStep()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // A : M, K
-            return ck::make_multi_index(ThreadwiseGemm_Dispatch::ThreadMaxMr, 0);
-        }
-        else
-        {
-            // A : K, M
-            return ck::make_multi_index(0, ThreadwiseGemm_Dispatch::ThreadMaxMr);
-        }
-    }
-
-    static constexpr auto GetBThreadMoveFwdStep()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return ck::make_multi_index(0, ThreadwiseGemm_Dispatch::ThreadMaxNr);
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return ck::Sequence<ThreadwiseGemm_Dispatch::ThreadMaxNr /
-                                    ThreadwiseGemm_Dispatch::MatrixBMinVectorSize,
-                                0,
-                                0>{};
-        }
-    }
-
-    static constexpr ck::index_t GetAThreadLoopOverDim()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixALayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // A : M, K
-            return 0;
-        }
-        else
-        {
-            // A : K, M
-            return 1;
-        }
-    }
-
-    static constexpr ck::index_t GetBThreadLoopOverDim()
-    {
-        if constexpr(std::is_same<typename ThreadwiseGemm_Dispatch::MatrixBLayout,
-                                  ck::tensor_layout::gemm::RowMajor>::value)
-        {
-            // B : K, N
-            return 1;
-        }
-        else
-        {
-            // B : N/8, K, N88;
-            return 0;
-        }
-    }
-
-    static constexpr auto GetCBlockDescriptor()
-    {
-        if constexpr(UseCLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
-        }
-        else
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock)); // TODO:
-        }
-    }
-
-    static constexpr auto GetCBlockSliceLength() { return ck::Sequence<MPerBlock, NPerBlock>{}; }
-
     static constexpr bool CheckValidity(const AGridDesc& a_grid_desc,
                                         const BGridDesc& b_grid_desc,
                                         const CGridDesc& c_grid_desc)
     {
-#if 0
-        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
-        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
-        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
-
-        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
-             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
-             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
-            return false;
-
-        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
-            return false;
-
-        // check NumPrefetch
-        if constexpr(NumPrefetch == 1)
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        bool is_valid    = true;
+        const auto GemmN = c_grid_desc.GetLength(I1);
+        if constexpr(UseCLocalBuffer)
         {
-            // 1-stage prefetch always supported
-        }
-        else if constexpr(NumPrefetch == 2)
-        {
-            // 2-stage prefetch currently only support even number of K0 loop
-            // TODO: add support for odd number of K0 loop
-            if(!((K0 / K0PerBlock) % 2 == 0))
-            {
-                return false;
-            }
+            if(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value && NPerBlock < GemmN)
+                is_valid &= false;
         }
         else
         {
-            return false;
+            // TODO: need check c grid is simple transform?
+            if(GemmN % 8 != 0)
+                is_valid &= false;
         }
-
-        // check M01, N01
-        constexpr auto M1 = Number<MPerBlock>{};
-        constexpr auto N1 = Number<NPerBlock>{};
-
-        const auto M0 = M / M1;
-        const auto N0 = N / N1;
-
-        if(!(M0 % M01 == 0 && N0 % N01 == 0))
-            return false;
-#endif
-        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
-        return true;
+        return is_valid;
     }
 
     static void Run(const FloatA* __restrict__ p_a_grid,
@@ -397,178 +156,149 @@ struct GridwiseGemmAvx2_MxN
                     const BElementwiseOperation& b_element_op,
                     const CElementwiseOperation& c_element_op)
     {
-        ck::index_t m_per_block;
-        ck::index_t n_per_block;
-        ck::index_t k_per_block;
-
-        if constexpr(MPerBlock == 0 && NPerBlock == 0 && KPerBlock == 0) {}
-        else
-        {
-            m_per_block = MPerBlock;
-            n_per_block = NPerBlock;
-            k_per_block = KPerBlock;
-        }
-
-        const auto M = a_grid_desc.GetLength(I0);
-        const auto N = b_grid_desc.GetLength(I1);
-        const auto K = b_grid_desc.GetLength(I0);
-
-        const ck::index_t grid_m = math::integer_divide_ceil(M, m_per_block);
-        const ck::index_t grid_n = math::integer_divide_ceil(N, n_per_block);
-
-        const ck::index_t grid_size = grid_m * grid_n;
-
-        constexpr auto a_block_desc           = GetABlockDescriptor();
-        constexpr auto a_block_slice_length   = GetABlockSliceLength();
-        constexpr auto a_block_copy_dim       = decltype(a_block_slice_length)::Size();
-        constexpr auto a_dim_access_order     = GetABlockDimAccessOrder();
-        constexpr auto a_block_move_step      = GetABlockMoveFwdStep();
-        constexpr auto a_thread_slice_length  = GetAThreadSliceLength();
-        constexpr auto a_thread_loop_over_dim = GetAThreadLoopOverDim();
-
-        constexpr auto b_block_desc           = GetBBlockDescriptor();
-        constexpr auto b_block_slice_length   = GetBBlockSliceLength();
-        constexpr auto b_block_copy_dim       = decltype(b_block_slice_length)::Size();
-        constexpr auto b_dim_access_order     = GetBBlockDimAccessOrder();
-        constexpr auto b_block_move_step      = GetBBlockMoveFwdStep();
-        constexpr auto b_thread_slice_length  = GetBThreadSliceLength();
-        constexpr auto b_thread_loop_over_dim = GetBThreadLoopOverDim();
-
-        constexpr auto c_block_desc         = GetCBlockDescriptor();
-        constexpr auto c_block_slice_length = GetCBlockSliceLength();
-        constexpr auto c_block_move_step    = ck::make_multi_index(0, NPerBlock);
-
-        auto a_threadwise_copy = ck::cpu::ThreadwiseTensorSliceTransferAvx2<
-            FloatA,                               // SrcData
-            FloatA,                               // DstData
-            decltype(a_grid_desc),                // SrcDesc
-            decltype(a_block_desc),               // DstDesc
-            AElementwiseOperation,                // ElementwiseOperation
-            decltype(a_block_slice_length),       // SliceLengths
-            decltype(a_dim_access_order),         // DimAccessOrder
-            1,                                    // VectorDim
-            1,                                    // ScalarPerVector
-            ck::InMemoryDataOperationEnum_t::Set, // InMemoryDataOperationEnum_t
-            false,                                // SrcResetCoordinateAfterRun
-            true                                  // DstResetCoordinateAfterRun
-            >(a_grid_desc,
-              ck::make_zero_multi_index<a_block_copy_dim>(),
-              a_block_desc,
-              ck::make_zero_multi_index<a_block_copy_dim>(),
-              AElementwiseOperation{});
-
-        auto b_threadwise_copy = ck::cpu::ThreadwiseTensorSliceTransferAvx2<
-            FloatB,                               // SrcData
-            FloatB,                               // DstData
-            decltype(b_grid_desc),                // SrcDesc
-            decltype(b_block_desc),               // DstDesc
-            BElementwiseOperation,                // ElementwiseOperation
-            decltype(b_block_slice_length),       // SliceLengths
-            decltype(b_dim_access_order),         // DimAccessOrder
-            1,                                    // VectorDim
-            1,                                    // ScalarPerVector
-            ck::InMemoryDataOperationEnum_t::Set, // InMemoryDataOperationEnum_t
-            false,                                // SrcResetCoordinateAfterRun
-            true                                  // DstResetCoordinateAfterRun
-            >(b_grid_desc,
-              ck::make_zero_multi_index<b_block_copy_dim>(),
-              b_block_desc,
-              ck::make_zero_multi_index<b_block_copy_dim>(),
-              BElementwiseOperation{});
-
-        auto c_threadwise_copy = ck::cpu::ThreadwiseTensorSliceTransferAvx2<
-            FloatC,                               // SrcData
-            FloatC,                               // DstData
-            decltype(c_block_desc),               // SrcDesc
-            decltype(c_grid_desc),                // DstDesc
-            BElementwiseOperation,                // ElementwiseOperation
-            ck::Sequence<MPerBlock, NPerBlock>,   // SliceLengths
-            ck::Sequence<0, 1>,                   // DimAccessOrder
-            1,                                    // VectorDim
-            1,                                    // ScalarPerVector
-            ck::InMemoryDataOperationEnum_t::Set, // InMemoryDataOperationEnum_t
-            true,                                 // SrcResetCoordinateAfterRun
-            false                                 // DstResetCoordinateAfterRun
-            >(c_block_desc,
-              ck::make_zero_multi_index<2>(),
-              c_grid_desc,
-              ck::make_zero_multi_index<2>(),
-              CElementwiseOperation{});
-
-        DeviceAlignedMemCPU a_block_mem(MPerBlock * KPerBlock * sizeof(FloatA), MemAlignmentByte);
-        DeviceAlignedMemCPU b_block_mem(KPerBlock * NPerBlock * sizeof(FloatB), MemAlignmentByte);
-        DeviceAlignedMemCPU c_block_mem(MPerBlock * NPerBlock * sizeof(FloatC), MemAlignmentByte);
-
-        auto a_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
+        ck::index_t m_per_block = MPerBlock;
+        ck::index_t n_per_block = NPerBlock;
+        ck::index_t k_per_block = KPerBlock;
+
+        const auto GemmM = c_grid_desc.GetLength(I0);
+        const auto GemmN = c_grid_desc.GetLength(I1);
+        const auto GemmK = a_grid_desc.GetLength(I1);
+
+        constexpr auto a_block_copy_dim = AGridDesc::GetNumOfDimension();
+
+        constexpr auto b_block_copy_dim = BGridDesc::GetNumOfDimension();
+
+        auto a_threadwise_copy = AThreadwiseCopy(a_grid_desc,
+                                                 ck::make_zero_multi_index<a_block_copy_dim>(),
+                                                 GetABlockDescriptor(m_per_block, k_per_block),
+                                                 ck::make_zero_multi_index<a_block_copy_dim>(),
+                                                 AElementwiseOperation{});
+
+        auto b_threadwise_copy = BThreadwiseCopy(b_grid_desc,
+                                                 ck::make_zero_multi_index<b_block_copy_dim>(),
+                                                 GetBBlockDescriptor(k_per_block, n_per_block),
+                                                 ck::make_zero_multi_index<b_block_copy_dim>(),
+                                                 BElementwiseOperation{});
+
+        auto c_threadwise_copy = CThreadwiseCopy(GetCBlockDescriptor(m_per_block, n_per_block),
+                                                 ck::make_zero_multi_index<2>(),
+                                                 c_grid_desc,
+                                                 ck::make_zero_multi_index<2>(),
+                                                 CElementwiseOperation{});
+
+        DeviceAlignedMemCPU a_block_mem(m_per_block * k_per_block * sizeof(FloatA),
+                                        MemAlignmentByte);
+        DeviceAlignedMemCPU b_block_mem(k_per_block * n_per_block * sizeof(FloatB),
+                                        MemAlignmentByte);
+        DeviceAlignedMemCPU c_block_mem(m_per_block * n_per_block * sizeof(FloatC),
+                                        MemAlignmentByte);
+
+        auto a_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
             reinterpret_cast<const FloatA*>(p_a_grid), a_grid_desc.GetElementSpaceSize());
 
-        auto b_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
+        auto b_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
             reinterpret_cast<const FloatB*>(p_b_grid), b_grid_desc.GetElementSpaceSize());
 
-        auto c_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
+        auto c_grid_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
             reinterpret_cast<FloatC*>(p_c_grid), c_grid_desc.GetElementSpaceSize());
 
-        auto a_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
+        auto a_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
             reinterpret_cast<FloatA*>(a_block_mem.mpDeviceBuf),
             a_block_mem.mMemSize / sizeof(FloatA));
 
-        auto b_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
+        auto b_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
             reinterpret_cast<FloatB*>(b_block_mem.mpDeviceBuf),
             b_block_mem.mMemSize / sizeof(FloatB));
 
-        auto c_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum_t::Global>(
-            reinterpret_cast<FloatC*>(c_block_mem.mpDeviceBuf),
-            c_block_mem.mMemSize / sizeof(FloatC));
-
-        auto blockwise_gemm =
-            BlockwiseGemmAvx2_MxN<FloatA,                          // FloatA,
-                                  FloatB,                          // FloatB,
-                                  FloatC,                          // FloatC,
-                                  AccDataType,                     // AccDataType,
-                                  decltype(a_block_desc),          // ABlockDesc,
-                                  decltype(b_block_desc),          // BBlockDesc,
-                                  decltype(c_block_desc),          // CBlockDesc,
-                                  decltype(a_block_slice_length),  // ABlockSliceLengths,
-                                  decltype(b_block_slice_length),  // BBlockSliceLengths,
-                                  decltype(c_block_slice_length),  // CBlockSliceLengths,
-                                  decltype(a_thread_slice_length), // AThreadSliceLength,
-                                  decltype(b_thread_slice_length), // BThreadSliceLength,
-                                  a_thread_loop_over_dim,  // AThreadLoopOverDim,   // thread slice
-                                                           // loop over on block slice. 1d is enough
-                                                           // for now
-                                  b_thread_loop_over_dim,  // BThreadLoopOverDim,
-                                  KPerBlock,               // KPerBlock,
-                                  ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
-                                  ThreadMNAccessOrder>{};  // ThreadMNAccessOrder  // how we acces
-                                                           // gemm MN to utilize micro kernel>{};
+        auto c_block_buf = ck::cpu::make_dynamic_buffer<ck::AddressSpaceEnum::Global>(
+            UseCLocalBuffer ? reinterpret_cast<FloatC*>(c_block_mem.mpDeviceBuf)
+                            : reinterpret_cast<FloatC*>(p_c_grid),
+            UseCLocalBuffer ? c_block_mem.mMemSize / sizeof(FloatC)
+                            : c_grid_desc.GetElementSpaceSize());
+
+        auto blockwise_gemm = BlockwiseGemmAvx2_MxN<
+            FloatA,                                                  // FloatA,
+            FloatB,                                                  // FloatB,
+            FloatC,                                                  // FloatC,
+            decltype(GetABlockDescriptor(m_per_block, k_per_block)), // ABlockDesc,
+            decltype(GetBBlockDescriptor(k_per_block, n_per_block)), // BBlockDesc,
+            decltype(GetCBlockDescriptor(m_per_block, n_per_block)), // CBlockDesc,
+            KPerBlock,                                               // KPerBlock,
+            ThreadwiseGemm_Dispatch,                                 // ThreadwiseGemm_Dispatch,
+            ThreadMNAccessOrder>{}; // ThreadMNAccessOrder  // how we acces
+                                    // gemm MN to utilize micro kernel>{};
 
         // TODO: openmp aware ordering
+        //
         if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 1, 2>>::value)
         {
+            auto a_move_k_step       = ck::make_multi_index(0, k_per_block);
+            auto b_move_k_step       = ck::make_multi_index(0, k_per_block, 0);
+            const ck::index_t grid_m = math::integer_divide_ceil(GemmM, m_per_block);
+            const ck::index_t grid_n = math::integer_divide_ceil(GemmN, n_per_block);
+
+            const ck::index_t grid_size = grid_m * grid_n;
+// This version does not consider K panel re-usage. simple for openmp
 #pragma omp parallel for
             for(ck::index_t gid = 0; gid < grid_size; gid++)
             {
                 ck::index_t i_mc = (gid / grid_n) * m_per_block;
                 ck::index_t i_nc = (gid % grid_n) * n_per_block;
 
-                ck::index_t mc_size = ck::math::min(M - i_mc, m_per_block);
-                ck::index_t nc_size = ck::math::min(N - i_nc, n_per_block);
-
-                // pack_b
-                b_threadwise_copy.RunGeneric(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
-                b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_move_step);
-                if(i_nc == 0)
+                ck::index_t mc_size = ck::math::min(GemmM - i_mc, m_per_block);
+                ck::index_t nc_size =
+                    ck::math::min(GemmN - i_nc, n_per_block); // TODO: nc need be 8x
+                nc_size = math::integer_least_multiple(
+                    nc_size, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
+
+                a_threadwise_copy.SetSrcSliceOrigin(a_grid_desc, ck::make_multi_index(i_mc, 0));
+                b_threadwise_copy.SetSrcSliceOrigin(
+                    b_grid_desc,
+                    ck::make_multi_index(math::integer_divide_ceil(
+                                             i_nc, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
+                                         0,
+                                         0));
+
+                auto c_block_desc =
+                    UseCLocalBuffer ? GetCBlockDescriptor(mc_size, nc_size) : c_grid_desc;
+                if constexpr(UseCLocalBuffer)
+                {
+                    c_threadwise_copy.SetDstSliceOrigin(c_grid_desc,
+                                                        ck::make_multi_index(i_mc, i_nc));
+                }
+                else
                 {
-                    // pack_a
-                    a_threadwise_copy.RunGeneric(
-                        a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
-                    a_threadwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_move_step);
+                    c_threadwise_copy.SetSrcSliceOrigin(c_block_desc,
+                                                        ck::make_multi_index(i_mc, i_nc));
+                    c_threadwise_copy.Run(c_block_desc, c_block_buf, c_grid_desc, c_grid_buf);
                 }
 
-                for(ck::index_t i_kc = 0; i_kc < K; i_kc += k_per_block)
+                for(ck::index_t i_kc = 0; i_kc < GemmK; i_kc += k_per_block)
                 {
-                    ck::index_t kc_size = ck::math::min(K - i_kc, k_per_block);
+                    ck::index_t kc_size = ck::math::min(GemmK - i_kc, k_per_block);
 
+                    auto a_block_desc = GetABlockDescriptor(mc_size, kc_size);
+                    auto b_block_desc = GetBBlockDescriptor(kc_size, nc_size);
+
+                    // printf("==> i_m:%d, i_n:%d, i_k:%d, mc:%d, nc:%d, kc:%d(%d, %d)\n", i_mc,
+                    // i_nc, i_kc, mc_size, nc_size, kc_size, KPerBlock, GemmK); fflush(stdout);
+
+                    a_threadwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
+
+                    b_threadwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
+
+                    // for(auto i_elem = 0; i_elem < (mc_size * kc_size) ; i_elem++){
+                    //    printf("A ==> %3d : %f(0x%08x)\n", i_elem,
+                    //    (reinterpret_cast<float*>(a_block_buf.p_data_))[i_elem],
+                    //    (reinterpret_cast<uint32_t*>(a_block_buf.p_data_))[i_elem]);
+                    //}
+
+                    // for(auto i_elem = 0; i_elem < (kc_size * nc_size) ; i_elem++){
+                    //     printf("B ==> %3d : %f(0x%08x)\n", i_elem,
+                    //     (reinterpret_cast<float*>(b_block_buf.p_data_))[i_elem],
+                    //     (reinterpret_cast<uint32_t*>(b_block_buf.p_data_))[i_elem]);
+                    // }
+                    // printf("[%d] 2222 \n",__LINE__);
                     blockwise_gemm.Run(a_block_desc,
                                        a_block_buf,
                                        make_zero_multi_index<a_block_copy_dim>(),
@@ -577,14 +307,108 @@ struct GridwiseGemmAvx2_MxN
                                        make_zero_multi_index<b_block_copy_dim>(),
                                        c_block_desc,
                                        c_block_buf,
-                                       make_zero_multi_index<2>());
+                                       make_zero_multi_index<2>(),
+                                       i_kc != 0);
+
+                    // printf("[%d] 2222 \n",__LINE__);
+                    if((i_kc + k_per_block) < GemmK)
+                    {
+                        a_threadwise_copy.MoveSrcSliceWindow(a_grid_desc, a_move_k_step);
+                        b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc, b_move_k_step);
+                    }
+
+                    // printf("[%d] 2222 \n",__LINE__);
+
+                    // for(auto i_elem = 0; i_elem < (10) ; i_elem++){
+                    //     printf("C ==> %3d : %f(0x%08x)\n", i_elem,
+                    //     (reinterpret_cast<float*>(c_block_buf.p_data_))[i_elem],
+                    //     (reinterpret_cast<uint32_t*>(c_block_buf.p_data_))[i_elem]);
+                    // }
                 }
 
+                // for(auto i_elem = 0; i_elem < (c_block_mem.mMemSize / sizeof(FloatC)) ;
+                // i_elem++){
+                //     printf("C ==> %3d : %f(0x%08x)\n", i_elem,
+                //     (reinterpret_cast<float*>(c_block_buf.p_data_))[i_elem],
+                //     (reinterpret_cast<uint32_t*>(c_block_buf.p_data_))[i_elem]);
+                // }
                 if constexpr(UseCLocalBuffer)
+                    c_threadwise_copy.Run(c_block_desc, c_block_buf, c_grid_desc, c_grid_buf);
+            }
+        }
+        else if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value)
+        {
+            auto a_move_k_step = ck::make_multi_index(0, k_per_block);
+            auto b_move_k_step = ck::make_multi_index(
+                math::integer_divide_ceil(n_per_block,
+                                          ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
+                0,
+                0);
+
+// only parallel in gemm m dim
+#pragma omp parallel for
+            for(ck::index_t i_mc = 0; i_mc < GemmM; i_mc += m_per_block)
+            {
+                ck::index_t mc_size = ck::math::min(GemmM - i_mc, m_per_block);
+                a_threadwise_copy.SetSrcSliceOrigin(a_grid_desc, ck::make_multi_index(i_mc, 0));
+                for(ck::index_t i_kc = 0; i_kc < GemmK; i_kc += k_per_block)
                 {
-                    c_threadwise_copy.RunGeneric(
-                        c_block_desc, c_block_buf, c_grid_desc, c_grid_buf);
-                    c_threadwise_copy.MoveDstSliceWindow(c_grid_desc, c_block_move_step);
+                    ck::index_t kc_size = ck::math::min(GemmK - i_kc, k_per_block);
+
+                    auto a_block_desc = GetABlockDescriptor(mc_size, kc_size);
+                    a_threadwise_copy.Run(a_grid_desc, a_grid_buf, a_block_desc, a_block_buf);
+
+                    b_threadwise_copy.SetSrcSliceOrigin(b_grid_desc,
+                                                        ck::make_multi_index(0, i_kc, 0));
+
+                    // TODO: if use local C buffer, then this nc loop need to loop only once
+                    for(ck::index_t i_nc = 0; i_nc < GemmN; i_nc += n_per_block)
+                    {
+                        ck::index_t nc_size =
+                            ck::math::min(GemmN - i_nc, n_per_block); // TODO: nc need be 8x
+                        nc_size = math::integer_least_multiple(
+                            nc_size, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize);
+
+                        auto b_block_desc = GetBBlockDescriptor(kc_size, nc_size);
+
+                        b_threadwise_copy.Run(b_grid_desc, b_grid_buf, b_block_desc, b_block_buf);
+
+                        auto c_block_desc =
+                            UseCLocalBuffer ? GetCBlockDescriptor(mc_size, nc_size) : c_grid_desc;
+
+                        if constexpr(!UseCLocalBuffer)
+                        {
+                            c_threadwise_copy.SetSrcSliceOrigin(c_block_desc,
+                                                                ck::make_multi_index(i_mc, i_nc));
+                            c_threadwise_copy.Run(
+                                c_block_desc, c_block_buf, c_grid_desc, c_grid_buf);
+                        }
+
+                        blockwise_gemm.Run(a_block_desc,
+                                           a_block_buf,
+                                           make_zero_multi_index<a_block_copy_dim>(),
+                                           b_block_desc,
+                                           b_block_buf,
+                                           make_zero_multi_index<b_block_copy_dim>(),
+                                           c_block_desc,
+                                           c_block_buf,
+                                           make_zero_multi_index<2>(),
+                                           i_kc != 0);
+
+                        if((i_nc + n_per_block) < GemmN)
+                            b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc, b_move_k_step);
+
+                        if constexpr(UseCLocalBuffer)
+                        {
+                            c_threadwise_copy.SetDstSliceOrigin(c_grid_desc,
+                                                                ck::make_multi_index(i_mc, i_nc));
+                            c_threadwise_copy.Run(
+                                c_block_desc, c_block_buf, c_grid_desc, c_grid_buf);
+                        }
+                    }
+
+                    if((i_kc + k_per_block) < GemmK)
+                        a_threadwise_copy.MoveSrcSliceWindow(a_grid_desc, a_move_k_step);
                 }
             }
         }

include/ck/tensor_operation/cpu/thread/threadwise_gemm_avx2.hpp

@@ -7,7 +7,7 @@
 #include "common_header.hpp"
 #include "../../gpu/device/tensor_layout.hpp"
 #include "math.hpp"
-#include "threadwise_param.hpp"
+#include "threadwise_gemm_param.hpp"
 
 namespace ck {
 namespace cpu {
@@ -294,6 +294,9 @@ struct ThreadwiseGemmAvx2_MxN_6x16
             ".if (m_Mr > 4)\n lea  (%%rdx, %%rdi, 1), %%r8 \n .endif\n"
             ".if (m_Mr > 5)\n lea  (%%r8,  %%rdi, 1), %%r9 \n .endif\n"
 
+            "mov    60(%[m_param]),  %%edi\n"    // accmulate_c
+            "test   %%edi,    %%edi\n"
+            "je     L_GemmAvx2_MxN_6x16_Store_C%=\n"
             "                               vaddps  (%%rax),    %%ymm0,  %%ymm0 \n"
             ".if               (m_Nr > 8)\n vaddps  32(%%rax),  %%ymm1,  %%ymm1 \n .endif\n"
             ".if (m_Mr > 1)              \n vaddps  (%%rbx),    %%ymm2,  %%ymm2 \n .endif\n"
@@ -307,6 +310,7 @@ struct ThreadwiseGemmAvx2_MxN_6x16
             ".if (m_Mr > 5)              \n vaddps  (%%r9),     %%ymm10, %%ymm10\n .endif\n"
             ".if (m_Mr > 5) && (m_Nr > 8)\n vaddps  32(%%r9),   %%ymm11, %%ymm11\n .endif\n"
 
+        "L_GemmAvx2_MxN_6x16_Store_C%=:\n"
             ".if m_NTStore == 0\n"
             "                               vmovups %%ymm0,     (%%rax)  \n"
             ".if               (m_Nr > 8)\n vmovups %%ymm1,     32(%%rax)\n .endif\n"
@@ -424,18 +428,33 @@ struct ThreadwiseGemmAvx2_MxN_6x16
         };
 
         // clang-format off
-                                        ymm0  = _mm256_loadu_ps(p_c + 0 * ldc + 0 * 8);
-        if constexpr (          Nr > 8) ymm1  = _mm256_loadu_ps(p_c + 0 * ldc + 1 * 8);
-        if constexpr (Mr > 1          ) ymm2  = _mm256_loadu_ps(p_c + 1 * ldc + 0 * 8);
-        if constexpr (Mr > 1 && Nr > 8) ymm3  = _mm256_loadu_ps(p_c + 1 * ldc + 1 * 8);
-        if constexpr (Mr > 2          ) ymm4  = _mm256_loadu_ps(p_c + 2 * ldc + 0 * 8);
-        if constexpr (Mr > 2 && Nr > 8) ymm5  = _mm256_loadu_ps(p_c + 2 * ldc + 1 * 8);
-        if constexpr (Mr > 3          ) ymm6  = _mm256_loadu_ps(p_c + 3 * ldc + 0 * 8);
-        if constexpr (Mr > 3 && Nr > 8) ymm7  = _mm256_loadu_ps(p_c + 3 * ldc + 1 * 8);
-        if constexpr (Mr > 4          ) ymm8  = _mm256_loadu_ps(p_c + 4 * ldc + 0 * 8);
-        if constexpr (Mr > 4 && Nr > 8) ymm9  = _mm256_loadu_ps(p_c + 4 * ldc + 1 * 8);
-        if constexpr (Mr > 5          ) ymm10 = _mm256_loadu_ps(p_c + 5 * ldc + 0 * 8);
-        if constexpr (Mr > 5 && Nr > 8) ymm11 = _mm256_loadu_ps(p_c + 5 * ldc + 1 * 8);
+        if(param->accmulate_c){
+                                            ymm0  = _mm256_loadu_ps(p_c + 0 * ldc + 0 * 8);
+            if constexpr (          Nr > 8) ymm1  = _mm256_loadu_ps(p_c + 0 * ldc + 1 * 8);
+            if constexpr (Mr > 1          ) ymm2  = _mm256_loadu_ps(p_c + 1 * ldc + 0 * 8);
+            if constexpr (Mr > 1 && Nr > 8) ymm3  = _mm256_loadu_ps(p_c + 1 * ldc + 1 * 8);
+            if constexpr (Mr > 2          ) ymm4  = _mm256_loadu_ps(p_c + 2 * ldc + 0 * 8);
+            if constexpr (Mr > 2 && Nr > 8) ymm5  = _mm256_loadu_ps(p_c + 2 * ldc + 1 * 8);
+            if constexpr (Mr > 3          ) ymm6  = _mm256_loadu_ps(p_c + 3 * ldc + 0 * 8);
+            if constexpr (Mr > 3 && Nr > 8) ymm7  = _mm256_loadu_ps(p_c + 3 * ldc + 1 * 8);
+            if constexpr (Mr > 4          ) ymm8  = _mm256_loadu_ps(p_c + 4 * ldc + 0 * 8);
+            if constexpr (Mr > 4 && Nr > 8) ymm9  = _mm256_loadu_ps(p_c + 4 * ldc + 1 * 8);
+            if constexpr (Mr > 5          ) ymm10 = _mm256_loadu_ps(p_c + 5 * ldc + 0 * 8);
+            if constexpr (Mr > 5 && Nr > 8) ymm11 = _mm256_loadu_ps(p_c + 5 * ldc + 1 * 8);
+        } else {
+                                            ymm0  = _mm256_xor_ps(ymm0,   ymm0);
+            if constexpr (          Nr > 8) ymm1  = _mm256_xor_ps(ymm1,   ymm1);
+            if constexpr (Mr > 1          ) ymm2  = _mm256_xor_ps(ymm2,   ymm2);
+            if constexpr (Mr > 1 && Nr > 8) ymm3  = _mm256_xor_ps(ymm3,   ymm3);
+            if constexpr (Mr > 2          ) ymm4  = _mm256_xor_ps(ymm4,   ymm4);
+            if constexpr (Mr > 2 && Nr > 8) ymm5  = _mm256_xor_ps(ymm5,   ymm5);
+            if constexpr (Mr > 3          ) ymm6  = _mm256_xor_ps(ymm6,   ymm6);
+            if constexpr (Mr > 3 && Nr > 8) ymm7  = _mm256_xor_ps(ymm7,   ymm7);
+            if constexpr (Mr > 4          ) ymm8  = _mm256_xor_ps(ymm8,   ymm8);
+            if constexpr (Mr > 4 && Nr > 8) ymm9  = _mm256_xor_ps(ymm9,   ymm9);
+            if constexpr (Mr > 5          ) ymm10 = _mm256_xor_ps(ymm10,  ymm10);
+            if constexpr (Mr > 5 && Nr > 8) ymm11 = _mm256_xor_ps(ymm11,  ymm11);
+        }
 
         while (Kr > 4){
             #pragma unroll
@@ -532,6 +551,7 @@ struct ThreadwiseGemmAvx2_MxN_6x16
         }
 
         if constexpr (NonTemporalStore) {
+                                            _mm256_stream_ps(p_c + 0 * ldc + 0 * 8, ymm1);
             if constexpr (          Nr > 8) _mm256_stream_ps(p_c + 0 * ldc + 1 * 8, ymm1);
             if constexpr (Mr > 1          ) _mm256_stream_ps(p_c + 1 * ldc + 0 * 8, ymm2);
             if constexpr (Mr > 1 && Nr > 8) _mm256_stream_ps(p_c + 1 * ldc + 1 * 8, ymm3);
@@ -830,19 +850,23 @@ struct ThreadwiseGemmAvx2_MxN_4x24
             ".if (m_Mr > 2)\n lea  (%%rbx, %%rdi, 1), %%rcx\n .endif\n"
             ".if (m_Mr > 3)\n lea  (%%rcx, %%rdi, 1), %%rdx\n .endif\n"
 
-            // "                               vaddps  (%%rax),    %%ymm0,  %%ymm0 \n"
-            // ".if               (m_Nr > 8)\n vaddps  32(%%rax),  %%ymm1,  %%ymm1 \n .endif\n"
-            // ".if               (m_Nr >16)\n vaddps  64(%%rax),  %%ymm2,  %%ymm2 \n .endif\n"
-            // ".if (m_Mr > 1)              \n vaddps  (%%rbx),    %%ymm3,  %%ymm3 \n .endif\n"
-            // ".if (m_Mr > 1) && (m_Nr > 8)\n vaddps  32(%%rbx),  %%ymm4,  %%ymm4 \n .endif\n"
-            // ".if (m_Mr > 1) && (m_Nr >16)\n vaddps  64(%%rbx),  %%ymm5,  %%ymm5 \n .endif\n"
-            // ".if (m_Mr > 2)              \n vaddps  (%%rcx),    %%ymm6,  %%ymm6 \n .endif\n"
-            // ".if (m_Mr > 2) && (m_Nr > 8)\n vaddps  32(%%rcx),  %%ymm7,  %%ymm7 \n .endif\n"
-            // ".if (m_Mr > 2) && (m_Nr >16)\n vaddps  64(%%rcx),  %%ymm8,  %%ymm8 \n .endif\n"
-            // ".if (m_Mr > 3)              \n vaddps  (%%rdx),    %%ymm9,  %%ymm9 \n .endif\n"
-            // ".if (m_Mr > 3) && (m_Nr > 8)\n vaddps  32(%%rdx),  %%ymm10, %%ymm10\n .endif\n"
-            // ".if (m_Mr > 3) && (m_Nr >16)\n vaddps  64(%%rdx),  %%ymm11, %%ymm11\n .endif\n"
-
+            "mov    60(%[m_param]),  %%edi\n"    // accmulate_c
+            "test   %%edi,    %%edi\n"
+            "je     L_GemmAvx2_MxN_4x24_Store_C%=\n"
+            "                               vaddps  (%%rax),    %%ymm0,  %%ymm0 \n"
+            ".if               (m_Nr > 8)\n vaddps  32(%%rax),  %%ymm1,  %%ymm1 \n .endif\n"
+            ".if               (m_Nr >16)\n vaddps  64(%%rax),  %%ymm2,  %%ymm2 \n .endif\n"
+            ".if (m_Mr > 1)              \n vaddps  (%%rbx),    %%ymm3,  %%ymm3 \n .endif\n"
+            ".if (m_Mr > 1) && (m_Nr > 8)\n vaddps  32(%%rbx),  %%ymm4,  %%ymm4 \n .endif\n"
+            ".if (m_Mr > 1) && (m_Nr >16)\n vaddps  64(%%rbx),  %%ymm5,  %%ymm5 \n .endif\n"
+            ".if (m_Mr > 2)              \n vaddps  (%%rcx),    %%ymm6,  %%ymm6 \n .endif\n"
+            ".if (m_Mr > 2) && (m_Nr > 8)\n vaddps  32(%%rcx),  %%ymm7,  %%ymm7 \n .endif\n"
+            ".if (m_Mr > 2) && (m_Nr >16)\n vaddps  64(%%rcx),  %%ymm8,  %%ymm8 \n .endif\n"
+            ".if (m_Mr > 3)              \n vaddps  (%%rdx),    %%ymm9,  %%ymm9 \n .endif\n"
+            ".if (m_Mr > 3) && (m_Nr > 8)\n vaddps  32(%%rdx),  %%ymm10, %%ymm10\n .endif\n"
+            ".if (m_Mr > 3) && (m_Nr >16)\n vaddps  64(%%rdx),  %%ymm11, %%ymm11\n .endif\n"
+
+        "L_GemmAvx2_MxN_4x24_Store_C%=:\n"
             ".if m_NTStore == 0\n"
             "                               vmovups %%ymm0,     (%%rax)  \n"
             ".if               (m_Nr > 8)\n vmovups %%ymm1,     32(%%rax)\n .endif\n"
@@ -960,18 +984,33 @@ struct ThreadwiseGemmAvx2_MxN_4x24
         };
 
         // clang-format off
-                                        ymm0  = _mm256_loadu_ps(p_c + 0 * ldc + 0 * 8);
-        if constexpr (          Nr > 8) ymm1  = _mm256_loadu_ps(p_c + 0 * ldc + 1 * 8);
-        if constexpr (          Nr >16) ymm2  = _mm256_loadu_ps(p_c + 0 * ldc + 2 * 8);
-        if constexpr (Mr > 1          ) ymm3  = _mm256_loadu_ps(p_c + 1 * ldc + 0 * 8);
-        if constexpr (Mr > 1 && Nr > 8) ymm4  = _mm256_loadu_ps(p_c + 1 * ldc + 1 * 8);
-        if constexpr (Mr > 1 && Nr >16) ymm5  = _mm256_loadu_ps(p_c + 1 * ldc + 2 * 8);
-        if constexpr (Mr > 2          ) ymm6  = _mm256_loadu_ps(p_c + 2 * ldc + 0 * 8);
-        if constexpr (Mr > 2 && Nr > 8) ymm7  = _mm256_loadu_ps(p_c + 2 * ldc + 1 * 8);
-        if constexpr (Mr > 2 && Nr >16) ymm8  = _mm256_loadu_ps(p_c + 2 * ldc + 2 * 8);
-        if constexpr (Mr > 3          ) ymm9  = _mm256_loadu_ps(p_c + 3 * ldc + 0 * 8);
-        if constexpr (Mr > 3 && Nr > 8) ymm10 = _mm256_loadu_ps(p_c + 3 * ldc + 1 * 8);
-        if constexpr (Mr > 3 && Nr >16) ymm11 = _mm256_loadu_ps(p_c + 3 * ldc + 2 * 8);
+        if(param->accmulate_c) {
+                                            ymm0  = _mm256_loadu_ps(p_c + 0 * ldc + 0 * 8);
+            if constexpr (          Nr > 8) ymm1  = _mm256_loadu_ps(p_c + 0 * ldc + 1 * 8);
+            if constexpr (          Nr >16) ymm2  = _mm256_loadu_ps(p_c + 0 * ldc + 2 * 8);
+            if constexpr (Mr > 1          ) ymm3  = _mm256_loadu_ps(p_c + 1 * ldc + 0 * 8);
+            if constexpr (Mr > 1 && Nr > 8) ymm4  = _mm256_loadu_ps(p_c + 1 * ldc + 1 * 8);
+            if constexpr (Mr > 1 && Nr >16) ymm5  = _mm256_loadu_ps(p_c + 1 * ldc + 2 * 8);
+            if constexpr (Mr > 2          ) ymm6  = _mm256_loadu_ps(p_c + 2 * ldc + 0 * 8);
+            if constexpr (Mr > 2 && Nr > 8) ymm7  = _mm256_loadu_ps(p_c + 2 * ldc + 1 * 8);
+            if constexpr (Mr > 2 && Nr >16) ymm8  = _mm256_loadu_ps(p_c + 2 * ldc + 2 * 8);
+            if constexpr (Mr > 3          ) ymm9  = _mm256_loadu_ps(p_c + 3 * ldc + 0 * 8);
+            if constexpr (Mr > 3 && Nr > 8) ymm10 = _mm256_loadu_ps(p_c + 3 * ldc + 1 * 8);
+            if constexpr (Mr > 3 && Nr >16) ymm11 = _mm256_loadu_ps(p_c + 3 * ldc + 2 * 8);
+        } else {
+                                            ymm0  = _mm256_xor_ps(ymm0,   ymm0);
+            if constexpr (          Nr > 8) ymm1  = _mm256_xor_ps(ymm1,   ymm1);
+            if constexpr (          Nr >16) ymm2  = _mm256_xor_ps(ymm2,   ymm2);
+            if constexpr (Mr > 1          ) ymm3  = _mm256_xor_ps(ymm3,   ymm3);
+            if constexpr (Mr > 1 && Nr > 8) ymm4  = _mm256_xor_ps(ymm4,   ymm4);
+            if constexpr (Mr > 1 && Nr >16) ymm5  = _mm256_xor_ps(ymm5,   ymm5);
+            if constexpr (Mr > 2          ) ymm6  = _mm256_xor_ps(ymm6,   ymm6);
+            if constexpr (Mr > 2 && Nr > 8) ymm7  = _mm256_xor_ps(ymm7,   ymm7);
+            if constexpr (Mr > 2 && Nr >16) ymm8  = _mm256_xor_ps(ymm8,   ymm8);
+            if constexpr (Mr > 3          ) ymm9  = _mm256_xor_ps(ymm9,   ymm9);
+            if constexpr (Mr > 3 && Nr > 8) ymm10 = _mm256_xor_ps(ymm10,  ymm10);
+            if constexpr (Mr > 3 && Nr >16) ymm11 = _mm256_xor_ps(ymm11,  ymm11);
+        }
 
         while (Kr > 4){
             #pragma unroll
@@ -1221,33 +1260,36 @@ struct ThreadwiseGemmAvx2_MxN_6x16_Dispatch
 
     static constexpr pThreadwiseGemmAvx2Run dispatch_table[6][2] = {
         {
-            ThreadwiseGemm_6x16_t::Run,
-            ThreadwiseGemm_6x8_t::Run,
+            ThreadwiseGemm_1x8_t::Run,
+            ThreadwiseGemm_1x16_t::Run,
         },
         {
-            ThreadwiseGemm_5x16_t::Run,
-            ThreadwiseGemm_5x8_t::Run,
+            ThreadwiseGemm_2x8_t::Run,
+            ThreadwiseGemm_2x16_t::Run,
         },
         {
-            ThreadwiseGemm_4x16_t::Run,
-            ThreadwiseGemm_4x8_t::Run,
+            ThreadwiseGemm_3x8_t::Run,
+            ThreadwiseGemm_3x16_t::Run,
         },
         {
-            ThreadwiseGemm_3x16_t::Run,
-            ThreadwiseGemm_3x8_t::Run,
+            ThreadwiseGemm_4x8_t::Run,
+            ThreadwiseGemm_4x16_t::Run,
         },
         {
-            ThreadwiseGemm_2x16_t::Run,
-            ThreadwiseGemm_2x8_t::Run,
+            ThreadwiseGemm_5x8_t::Run,
+            ThreadwiseGemm_5x16_t::Run,
         },
         {
-            ThreadwiseGemm_1x16_t::Run,
-            ThreadwiseGemm_1x8_t::Run,
+            ThreadwiseGemm_6x8_t::Run,
+            ThreadwiseGemm_6x16_t::Run,
         },
     };
 
     static void Run(ThreadwiseGemmParam* param, index_t mr, index_t nr)
     {
+        index_t im = mr - 1;
+        index_t in = (nr >> 3) - 1;
+        assert(im >= 0 && im <= 5 && in >= 0 && in <= 1);
         return dispatch_table[mr][nr](param);
     }
 };
@@ -1371,30 +1413,33 @@ struct ThreadwiseGemmAvx2_MxN_4x24_Dispatch
 
     static constexpr pThreadwiseGemmAvx2Run dispatch_table[4][3] = {
         {
-            ThreadwiseGemm_4x24_t::Run,
-            ThreadwiseGemm_4x16_t::Run,
-            ThreadwiseGemm_4x8_t::Run,
+            ThreadwiseGemm_1x8_t::Run,
+            ThreadwiseGemm_1x16_t::Run,
+            ThreadwiseGemm_1x24_t::Run,
         },
         {
-            ThreadwiseGemm_3x24_t::Run,
-            ThreadwiseGemm_3x16_t::Run,
-            ThreadwiseGemm_3x8_t::Run,
+            ThreadwiseGemm_2x8_t::Run,
+            ThreadwiseGemm_2x16_t::Run,
+            ThreadwiseGemm_2x24_t::Run,
         },
         {
-            ThreadwiseGemm_2x24_t::Run,
-            ThreadwiseGemm_2x16_t::Run,
-            ThreadwiseGemm_2x8_t::Run,
+            ThreadwiseGemm_3x8_t::Run,
+            ThreadwiseGemm_3x16_t::Run,
+            ThreadwiseGemm_3x24_t::Run,
         },
         {
-            ThreadwiseGemm_1x24_t::Run,
-            ThreadwiseGemm_1x16_t::Run,
-            ThreadwiseGemm_1x8_t::Run,
+            ThreadwiseGemm_4x8_t::Run,
+            ThreadwiseGemm_4x16_t::Run,
+            ThreadwiseGemm_4x24_t::Run,
         },
     };
 
     static void Run(ThreadwiseGemmParam* param, index_t mr, index_t nr)
     {
-        return dispatch_table[mr][nr](param);
+        index_t im = mr - 1;
+        index_t in = (nr >> 3) - 1;
+        assert(im >= 0 && im <= 3 && in >= 0 && in <= 2);
+        return dispatch_table[im][in](param);
     }
 };
 

include/ck/tensor_operation/cpu/thread/threadwise_param.hpp → include/ck/tensor_operation/cpu/thread/threadwise_gemm_param.hpp

-#ifndef CK_THREADWISE_PARAM_HPP
-#define CK_THREADWISE_PARAM_HPP
+#ifndef CK_THREADWISE_GEMM_PARAM_HPP
+#define CK_THREADWISE_GEMM_PARAM_HPP
 
 #include "common_header.hpp"
 #include "math.hpp"
@@ -17,7 +17,7 @@ struct ThreadwiseGemmParam
     uint64_t ldb; // in unit of byte
     uint64_t ldc; // in unit of byte
     float alpha;
-    uint32_t _pack0;
+    int accmulate_c; // if 1, need load C and add into current fma. if 0, direct store out c result
 } __attribute__((packed));
 
 } // namespace cpu

include/ck/tensor_operation/cpu/thread/threadwise_tensor_slice_transfer_avx2.hpp

@@ -53,6 +53,42 @@ struct ThreadwiseTensorSliceTransferAvx2
     {
         static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
                       "wrong! cannot evenly divide");
+
+        int N  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+        int Hi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+        int Wi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+        int C  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<3>{}];
+
+        int Ho = src_desc.GetTransforms()[Number<9>{}].low_lengths_[Number<1>{}];
+        int Wo = src_desc.GetTransforms()[Number<9>{}].low_lengths_[Number<2>{}];
+
+        int Fy = src_desc.GetTransforms()[Number<10>{}].low_lengths_[Number<0>{}];
+        int Fx = src_desc.GetTransforms()[Number<10>{}].low_lengths_[Number<1>{}];
+
+        int Dy = src_desc.GetTransforms()[Number<6>{}].coefficients_[Number<0>{}];
+        int Sy = src_desc.GetTransforms()[Number<6>{}].coefficients_[Number<1>{}];
+        int Dx = src_desc.GetTransforms()[Number<7>{}].coefficients_[Number<0>{}];
+        int Sx = src_desc.GetTransforms()[Number<7>{}].coefficients_[Number<1>{}];
+
+        int Py = src_desc.GetTransforms()[Number<2>{}].left_pad_length_;
+        int Px = src_desc.GetTransforms()[Number<3>{}].left_pad_length_;
+
+        printf("N:%d, Hi:%d, Wi:%d, C:%d, Ho:%d, Wo:%d, Fy:%d, Fx:%d, Dy:%d, Sy:%d, Dx:%d, Sx:%d, "
+               "Py:%d, Px:%d\n",
+               N,
+               Hi,
+               Wi,
+               C,
+               Ho,
+               Wo,
+               Fy,
+               Fx,
+               Dy,
+               Sy,
+               Dx,
+               Sx,
+               Py,
+               Px);
     }
 
     void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
@@ -87,6 +123,10 @@ struct ThreadwiseTensorSliceTransferAvx2
 
         // std::cout<<"num_access:"<<num_access<<std::endl;
 
+        std::cout << "src hidden:" << SrcDesc::GetNumOfHiddenDimension() << std::endl;
+        std::cout << "dst hidden:" << DstDesc::GetNumOfHiddenDimension() << std::endl;
+
+#if 0
         static_for<0, num_access, 1>{}([&](auto idx_1d) {
             using src_vector_type = ck::cpu::vector_type_maker_t<SrcData, ScalarPerVector>;
             using src_vector_t    = typename src_vector_type::type;
@@ -148,6 +188,75 @@ struct ThreadwiseTensorSliceTransferAvx2
                     dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
             }
         });
+#endif
+        const auto src_slice_idx_zeros = typename uniform_sequence_gen<nDim, 0>::type{};
+        const auto src_slice_step      = make_tensor_coordinate_step(
+            src_desc, to_multi_index(src_slice_idx_zeros.Modify(Number<nDim - 1>{}, Number<1>{})));
+
+        const auto dst_slice_idx_zeros = typename uniform_sequence_gen<nDim, 0>::type{};
+        const auto dst_slice_step      = make_tensor_coordinate_step(
+            dst_desc, to_multi_index(dst_slice_idx_zeros.Modify(Number<nDim - 1>{}, Number<1>{})));
+
+        for(auto idx_id = 0; idx_id < num_access; idx_id++)
+        {
+            using src_vector_type = ck::cpu::vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = ck::cpu::vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            printf("[%s] ", is_src_valid ? "y" : "n");
+            print_multi_index(src_coord_.GetIndex());
+            printf("----");
+            // print_multi_index(src_coord_.GetHiddenIndex());
+
+            // printf(":%d", src_coord_.GetOffset());
+            // printf("\n");
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            // static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+            //     element_op_(dst_vector_container.template AsType<DstData>()(i),
+            //                 src_vector_container.template AsType<SrcData>()[i]);
+            // });
+            element_op_(dst_vector_container.template AsType<dst_vector_t>(),
+                        src_vector_container.template AsType<src_vector_t>());
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            printf(" -> ");
+            print_multi_index(dst_coord_.GetIndex());
+            // printf(":%d", dst_coord_.GetOffset());
+
+            // printf(", src:0x%x, dst:0x%x",
+            // *reinterpret_cast<uint32_t*>(&src_vector_container.template AsType<src_vector_t>()),
+            //                               *reinterpret_cast<uint32_t*>(&dst_vector_container.template
+            //                               AsType<dst_vector_t>()));
+            printf("\n");
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>());
+
+            // move coordinate
+            if(idx_id != num_access - 1)
+            {
+                // constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(src_desc, src_coord_, src_slice_step);
+                move_tensor_coordinate(dst_desc, dst_coord_, dst_slice_step);
+            }
+        }
 
         // move coordinate back to slice origin (or not)
         if constexpr(SrcResetCoordinateAfterRun)

include/ck/tensor_operation/cpu/thread/threadwise_tensor_slice_transfer_avx2_specialization.hpp

+#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_AVX2_SPECIALIZED_HPP
+#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_AVX2_SPECIALIZED_HPP
+
+#include "common_header.hpp"
+#include "data_type_cpu.hpp"
+#include "../../gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "tensor_space_filling_curve.hpp"
+#include "dynamic_buffer_cpu.hpp"
+#include <immintrin.h>
+#include "convolution_forward_specialization_cpu.hpp"
+#include <immintrin.h>
+
+namespace ck {
+namespace cpu {
+
+namespace avx2_util {
+
+inline void memcpy32_avx2(void* dst, const void* src, const ck::index_t n)
+{
+    // 16-8-4-2-1 pattern
+    ck::index_t i_n    = n;
+    float* p_dst       = reinterpret_cast<float*>(dst);
+    const float* p_src = reinterpret_cast<const float*>(src);
+    while(i_n >= 16)
+    {
+        _mm256_storeu_ps(p_dst + 0, _mm256_loadu_ps(p_src + 0));
+        _mm256_storeu_ps(p_dst + 8, _mm256_loadu_ps(p_src + 8));
+        p_dst += 16;
+        p_src += 16;
+        i_n -= 16;
+    }
+    if(i_n & 8)
+    {
+        _mm256_storeu_ps(p_dst, _mm256_loadu_ps(p_src));
+        p_dst += 8;
+        p_src += 8;
+    }
+    if(i_n & 4)
+    {
+        _mm_storeu_ps(p_dst, _mm_loadu_ps(p_src));
+        p_dst += 4;
+        p_src += 4;
+    }
+    if(i_n & 2)
+    {
+        _mm_storeu_si64(p_dst, _mm_loadu_si64(p_src));
+        p_dst += 2;
+        p_src += 2;
+    }
+    if(i_n & 1)
+    {
+        *p_dst = *p_src;
+    }
+}
+
+inline void memset32_avx2(void* dst, const int32_t value, const ck::index_t n)
+{
+    // 16-8-4-2-1 pattern
+    ck::index_t i_n = n;
+    float* p_dst    = reinterpret_cast<float*>(dst);
+    __m256 ymm      = _mm256_set1_ps(*reinterpret_cast<const float*>(&value));
+    __m128 xmm      = _mm_set1_ps(*reinterpret_cast<const float*>(&value));
+    while(i_n >= 16)
+    {
+        _mm256_storeu_ps(p_dst + 0, ymm);
+        _mm256_storeu_ps(p_dst + 8, ymm);
+        p_dst += 16;
+        i_n -= 16;
+    }
+    if(i_n & 8)
+    {
+        _mm256_storeu_ps(p_dst, ymm);
+        p_dst += 8;
+    }
+    if(i_n & 4)
+    {
+        _mm_storeu_ps(p_dst, xmm);
+        p_dst += 4;
+    }
+    if(i_n & 2)
+    {
+        _mm_storeu_si64(p_dst, xmm);
+        p_dst += 2;
+    }
+    if(i_n & 1)
+    {
+        *p_dst = *reinterpret_cast<const float*>(&value);
+    }
+}
+
+inline void
+transpose8x8_avx2(void* dst, ck::index_t stride_dst, const void* src, ck::index_t stride_src)
+{
+    // TODO: use vinsertf128 for better port usage. vpermf128 is slow
+    __m256 r0, r1, r2, r3, r4, r5, r6, r7;
+    __m256 t0, t1, t2, t3, t4, t5, t6, t7;
+
+    float* p_dst       = reinterpret_cast<float*>(dst);
+    const float* p_src = reinterpret_cast<const float*>(src);
+
+    r0 = _mm256_loadu_ps(p_src + 0 * stride_src);
+    r1 = _mm256_loadu_ps(p_src + 1 * stride_src);
+    r2 = _mm256_loadu_ps(p_src + 2 * stride_src);
+    r3 = _mm256_loadu_ps(p_src + 3 * stride_src);
+    r4 = _mm256_loadu_ps(p_src + 4 * stride_src);
+    r5 = _mm256_loadu_ps(p_src + 5 * stride_src);
+    r6 = _mm256_loadu_ps(p_src + 6 * stride_src);
+    r7 = _mm256_loadu_ps(p_src + 7 * stride_src);
+
+    t0 = _mm256_unpacklo_ps(r0, r1);
+    t1 = _mm256_unpackhi_ps(r0, r1);
+    t2 = _mm256_unpacklo_ps(r2, r3);
+    t3 = _mm256_unpackhi_ps(r2, r3);
+    t4 = _mm256_unpacklo_ps(r4, r5);
+    t5 = _mm256_unpackhi_ps(r4, r5);
+    t6 = _mm256_unpacklo_ps(r6, r7);
+    t7 = _mm256_unpackhi_ps(r6, r7);
+
+    r0 = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(1, 0, 1, 0));
+    r1 = _mm256_shuffle_ps(t0, t2, _MM_SHUFFLE(3, 2, 3, 2));
+    r2 = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(1, 0, 1, 0));
+    r3 = _mm256_shuffle_ps(t1, t3, _MM_SHUFFLE(3, 2, 3, 2));
+    r4 = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(1, 0, 1, 0));
+    r5 = _mm256_shuffle_ps(t4, t6, _MM_SHUFFLE(3, 2, 3, 2));
+    r6 = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(1, 0, 1, 0));
+    r7 = _mm256_shuffle_ps(t5, t7, _MM_SHUFFLE(3, 2, 3, 2));
+
+    t0 = _mm256_permute2f128_ps(r0, r4, 0x20);
+    t1 = _mm256_permute2f128_ps(r1, r5, 0x20);
+    t2 = _mm256_permute2f128_ps(r2, r6, 0x20);
+    t3 = _mm256_permute2f128_ps(r3, r7, 0x20);
+    t4 = _mm256_permute2f128_ps(r0, r4, 0x31);
+    t5 = _mm256_permute2f128_ps(r1, r5, 0x31);
+    t6 = _mm256_permute2f128_ps(r2, r6, 0x31);
+    t7 = _mm256_permute2f128_ps(r3, r7, 0x31);
+
+    _mm256_storeu_ps(p_dst + 0 * stride_dst, t0);
+    _mm256_storeu_ps(p_dst + 1 * stride_dst, t1);
+    _mm256_storeu_ps(p_dst + 2 * stride_dst, t2);
+    _mm256_storeu_ps(p_dst + 3 * stride_dst, t3);
+    _mm256_storeu_ps(p_dst + 4 * stride_dst, t4);
+    _mm256_storeu_ps(p_dst + 5 * stride_dst, t5);
+    _mm256_storeu_ps(p_dst + 6 * stride_dst, t6);
+    _mm256_storeu_ps(p_dst + 7 * stride_dst, t7);
+}
+
+} // namespace avx2_util
+
+using ConvolutionForwardSpecialization_t =
+    ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t;
+using ConvolutionForwardGemmKSpecialization_t =
+    ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t;
+
+// assume input -> a matrix
+// assume input -> MC * KC
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          bool BypassTransfer,
+          ConvolutionForwardSpecialization_t ConvForwardSpecialization,
+          ConvolutionForwardGemmKSpecialization_t GemmKSpecialization>
+struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC
+{
+    static constexpr ck::index_t nDim = SrcDesc::GetNumOfDimension();
+    using Index                       = MultiIndex<nDim>;
+
+    constexpr ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_In_NHWC(
+        const SrcDesc& src_desc,
+        const Index&,
+        const DstDesc&,
+        const Index&,
+        const ElementwiseOperation& element_op)
+        : element_op_(element_op)
+    {
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
+        {
+            N  = 1;
+            Hi = 1;
+            Wi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}]; // gemm_m
+            C  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}]; // gemm_k
+
+            Ho = 1;
+            Wo = Wi;
+
+            Fy = 1;
+            Fx = 1;
+
+            Dy = 1;
+            Sy = 1;
+            Dx = 1;
+            Sx = 1;
+
+            Py = 0;
+            Px = 0;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization_t::Filter1x1Pad0)
+        {
+            N  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+            Hi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+            Wi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+            C  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<3>{}];
+
+            Ho = src_desc.GetTransforms()[Number<2>{}].GetUpperLengths()[Number<0>{}];
+            Wo = src_desc.GetTransforms()[Number<3>{}].GetUpperLengths()[Number<0>{}];
+
+            Fy = 1;
+            Fx = 1;
+
+            Dy = 1;
+            Sy = src_desc.GetTransforms()[Number<2>{}].coefficients_[Number<0>{}];
+            Dx = 1;
+            Sx = src_desc.GetTransforms()[Number<3>{}].coefficients_[Number<0>{}];
+
+            Py = 0;
+            Px = 0;
+        }
+        else
+        {
+            N  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+            Hi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+            Wi = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+            C  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<3>{}];
+
+            Ho = src_desc.GetTransforms()[Number<9>{}].low_lengths_[Number<1>{}];
+            Wo = src_desc.GetTransforms()[Number<9>{}].low_lengths_[Number<2>{}];
+
+            Fy = src_desc.GetTransforms()[Number<10>{}].low_lengths_[Number<0>{}];
+            Fx = src_desc.GetTransforms()[Number<10>{}].low_lengths_[Number<1>{}];
+
+            Dy = src_desc.GetTransforms()[Number<6>{}].coefficients_[Number<0>{}];
+            Sy = src_desc.GetTransforms()[Number<6>{}].coefficients_[Number<1>{}];
+            Dx = src_desc.GetTransforms()[Number<7>{}].coefficients_[Number<0>{}];
+            Sx = src_desc.GetTransforms()[Number<7>{}].coefficients_[Number<1>{}];
+
+            Py = src_desc.GetTransforms()[Number<2>{}].left_pad_length_;
+            Px = src_desc.GetTransforms()[Number<3>{}].left_pad_length_;
+        }
+
+        // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+        // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+        input_offset_acc_wi        = Sx * C;
+        input_offset_ovf_wi_acc_hi = Sy * Wi * C - Wo * Sx * C;
+        input_offset_ovf_hi_acc_n  = Hi * Wi * C - Ho * Sy * Wi * C;
+
+        // input_offset_acc_c       = 1;
+        input_offset_ovf_c_acc_x = Dx * C - C;
+        input_offset_ovf_x_acc_y = Dy * Wi * C - Fx * Dx * C;
+
+        src_offset = -Py * Wi * C - Px * C;
+
+        i_n  = 0;
+        i_c  = 0;
+        i_hi = -Py;
+        i_wi = -Px;
+        i_ho = 0;
+        i_wo = 0;
+        i_y  = 0;
+        i_x  = 0;
+
+        i_gemm_k = 0;
+
+#if 0
+        printf("N:%d, Hi:%d, Wi:%d, C:%d, Ho:%d, Wo:%d, Fy:%d, Fx:%d, Dy:%d, Sy:%d, Dx:%d, Sx:%d, "
+               "Py:%d, Px:%d\n",
+               N,
+               Hi,
+               Wi,
+               C,
+               Ho,
+               Wo,
+               Fy,
+               Fx,
+               Dy,
+               Sy,
+               Dx,
+               Sx,
+               Py,
+               Px);
+#endif
+    }
+
+    void SetSrcSliceOrigin(const SrcDesc&, const Index& src_slice_origin_idx)
+    {
+        ck::index_t idx_m = src_slice_origin_idx[Number<0>{}];
+        ck::index_t idx_k = src_slice_origin_idx[Number<1>{}];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
+        {
+            i_wi       = idx_m;
+            i_c        = idx_k;
+            src_offset = i_wi * C + i_c;
+
+            // printf("src_offset:%d, i_wi:%d, i_c:%d\n", src_offset, i_wi, i_c);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization_t::Filter1x1Pad0)
+        {
+            i_wo = idx_m % Wo;
+            i_ho = (idx_m / Wo) % Ho;
+            i_n  = (idx_m / Wo) / Ho;
+
+            // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+            // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+            i_c = idx_k;
+            i_x = 0;
+            i_y = 0;
+
+            i_hi = i_ho * Sy;
+            i_wi = i_wo * Sx;
+
+            src_offset = i_n * Hi * Wi * C + i_hi * Wi * C + i_wi * C + i_c;
+
+            i_gemm_k = idx_k;
+        }
+        else
+        {
+            i_wo = idx_m % Wo;
+            i_ho = (idx_m / Wo) % Ho;
+            i_n  = (idx_m / Wo) / Ho;
+
+            // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+            // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+            if(idx_k == 0)
+            {
+                i_c = 0;
+                i_x = 0;
+                i_y = 0;
+
+                i_hi = i_ho * Sy - Py;
+                i_wi = i_wo * Sx - Px;
+            }
+            else
+            {
+                i_c = idx_k % C;
+                i_x = (idx_k / C) % Fx;
+                i_y = (idx_k / C) / Fx;
+
+                i_hi = i_ho * Sy + i_y * Dy - Py;
+                i_wi = i_wo * Sx + i_x * Dx - Px;
+            }
+
+            src_offset = i_n * Hi * Wi * C + i_hi * Wi * C + i_wi * C + i_c;
+
+            i_gemm_k = idx_k;
+
+            // printf("[%d] i_wo:%d, i_ho:%d, i_wi:%d, i_hi:%d, src_offset:%d\n",
+            //            __LINE__, i_wo, i_ho, i_wi, i_hi, src_offset);
+        }
+    }
+
+    void SetDstSliceOrigin(const DstDesc&, const Index&) {}
+
+    template <typename SrcBuffer, typename DstBuffer>
+    void Run(const SrcDesc& src_desc,
+             const SrcBuffer& src_buf,
+             const DstDesc& dst_desc,
+             DstBuffer& dst_buf)
+    {
+        if constexpr(BypassTransfer)
+        {
+            float* p_src    = reinterpret_cast<float*>(src_buf.p_data_) + src_offset;
+            dst_buf.p_data_ = p_src;
+        }
+        else
+        {
+            const ck::index_t m_per_block =
+                dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+            const ck::index_t k_per_block =
+                dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+
+            const float* p_src = reinterpret_cast<const float*>(src_buf.p_data_) + src_offset;
+            float* p_dst       = reinterpret_cast<float*>(dst_buf.p_data_);
+
+            // printf("src offset:%d, k_per_block:%d, m_per_block:%d\n", src_offset, k_per_block,
+            // m_per_block);
+
+            if constexpr(ConvForwardSpecialization ==
+                         ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
+            {
+                ck::index_t i_m_itr = m_per_block;
+                // standard 8-4-2-1 pattern
+                while(i_m_itr >= 8)
+                {
+                    avx2_util::memcpy32_avx2(p_dst + 0 * k_per_block, p_src + 0 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 1 * k_per_block, p_src + 1 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 2 * k_per_block, p_src + 2 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 3 * k_per_block, p_src + 3 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 4 * k_per_block, p_src + 4 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 5 * k_per_block, p_src + 5 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 6 * k_per_block, p_src + 6 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 7 * k_per_block, p_src + 7 * C, k_per_block);
+
+                    i_m_itr -= 8;
+                    p_dst += 8 * k_per_block;
+                    p_src += 8 * C;
+                }
+                if(i_m_itr & 4)
+                {
+                    avx2_util::memcpy32_avx2(p_dst + 0 * k_per_block, p_src + 0 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 1 * k_per_block, p_src + 1 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 2 * k_per_block, p_src + 2 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 3 * k_per_block, p_src + 3 * C, k_per_block);
+
+                    p_dst += 4 * k_per_block;
+                    p_src += 4 * C;
+                }
+
+                if(i_m_itr & 2)
+                {
+                    avx2_util::memcpy32_avx2(p_dst + 0 * k_per_block, p_src + 0 * C, k_per_block);
+                    avx2_util::memcpy32_avx2(p_dst + 1 * k_per_block, p_src + 1 * C, k_per_block);
+
+                    p_dst += 2 * k_per_block;
+                    p_src += 2 * C;
+                }
+
+                if(i_m_itr & 1)
+                {
+                    avx2_util::memcpy32_avx2(p_dst + 0 * k_per_block, p_src + 0 * C, k_per_block);
+                }
+            }
+            else if constexpr(ConvForwardSpecialization ==
+                              ConvolutionForwardSpecialization_t::Filter1x1Pad0)
+            {
+                ck::index_t i_m_itr  = m_per_block;
+                ck::index_t i_wo_itr = i_wo;
+                ck::index_t i_ho_itr = i_ho;
+                while(i_m_itr > 0)
+                {
+                    avx2_util::memcpy32_avx2(p_dst, p_src, k_per_block);
+                    p_dst += k_per_block;
+                    i_wo_itr++;
+                    p_src += input_offset_acc_wi;
+                    if(i_wo_itr >= Wo)
+                    {
+                        i_wo_itr = 0;
+                        i_ho_itr++;
+                        p_src += input_offset_ovf_wi_acc_hi;
+                    }
+                    if(i_ho_itr >= Ho)
+                    {
+                        i_ho_itr = 0;
+                        // i_n++;
+                        p_src += input_offset_ovf_hi_acc_n;
+                    }
+
+                    i_m_itr--;
+                }
+            }
+            else
+            {
+                // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+                // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+                if constexpr(GemmKSpecialization ==
+                             ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
+                {
+                    // c % k_per_block == 0, so every time k_per_block here is the same
+                    ck::index_t i_m_itr  = m_per_block;
+                    ck::index_t i_wo_itr = i_wo;
+                    ck::index_t i_ho_itr = i_ho;
+                    ck::index_t i_wi_itr = i_wi;
+                    ck::index_t i_hi_itr = i_hi;
+
+                    // printf("[%d] i_m_itr:%d, i_wo_itr:%d, i_ho_itr:%d, i_wi_itr:%d, i_hi_itr:%d,
+                    // src_offset:%d, input_offset_acc_wi:%d,
+                    // input_offset_ovf_wi_acc_hi:%d,input_offset_ovf_hi_acc_n:%d, %p(%p)\n",
+                    //         __LINE__, i_m_itr, i_wo_itr, i_ho_itr, i_wi_itr, i_hi_itr,
+                    //         src_offset, input_offset_acc_wi, input_offset_ovf_wi_acc_hi,
+                    //         input_offset_ovf_hi_acc_n, src_buf.p_data_, p_src);
+
+                    // printf("%p %p %p, %d, %x, %p\n",src_buf.p_data_, reinterpret_cast<const
+                    // float*>(src_buf.p_data_) + 1, reinterpret_cast<const float*>(src_buf.p_data_)
+                    // + ck::index_t(-1),
+                    //     sizeof(src_offset), *reinterpret_cast<uint32_t*>(&src_offset),
+                    //     reinterpret_cast<const float*>(src_buf.p_data_) + (-1088));
+
+                    while(i_m_itr > 0)
+                    {
+                        // printf("[%d] i_m_itr:%d, i_wo_itr:%d, i_ho_itr:%d, i_wi_itr:%d,
+                        // i_hi_itr:%d, src_offset:%d -> %p\n",
+                        //    __LINE__, i_m_itr, i_wo_itr, i_ho_itr, i_wi_itr, i_hi_itr, src_offset,
+                        //    p_src);
+
+                        if((*reinterpret_cast<uint32_t*>(&i_hi_itr) < Hi) &&
+                           (*reinterpret_cast<uint32_t*>(&i_wi_itr) < Wi))
+                            avx2_util::memcpy32_avx2(p_dst, p_src, k_per_block);
+                        else
+                            avx2_util::memset32_avx2(p_dst, 0, k_per_block);
+
+                        p_dst += k_per_block;
+
+                        i_wo_itr++;
+                        i_wi_itr += Sx;
+                        p_src += input_offset_acc_wi;
+                        if(i_wo_itr >= Wo)
+                        {
+                            i_wo_itr = 0;
+                            i_wi_itr -= Wo * Sx;
+                            i_ho_itr++;
+                            i_hi_itr += Sy;
+                            p_src += input_offset_ovf_wi_acc_hi;
+                        }
+
+                        if(i_ho_itr >= Ho)
+                        {
+                            i_ho_itr = 0;
+                            i_hi_itr -= Ho * Sy;
+                            // i_n++;
+                            p_src += input_offset_ovf_hi_acc_n;
+                        }
+
+                        i_m_itr--;
+                    }
+
+                    // printf("[%d]   \n", __LINE__);
+                }
+                else
+                {
+                    ck::index_t i_m_itr  = m_per_block;
+                    ck::index_t i_wo_itr = i_wo;
+                    ck::index_t i_ho_itr = i_ho;
+                    ck::index_t i_wi_itr = i_wi;
+                    ck::index_t i_hi_itr = i_hi;
+                    // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+                    // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+                    while(i_m_itr > 0)
+                    {
+                        /*** go along Gemm K ***/
+                        const float* p_src_k   = p_src;
+                        float* p_dst_k         = p_dst;
+                        ck::index_t i_wi_itr_k = i_wi_itr;
+                        ck::index_t i_hi_itr_k = i_hi_itr;
+                        ck::index_t i_c_itr_k  = i_c;
+                        ck::index_t i_y_itr_k  = i_y;
+                        ck::index_t i_x_itr_k  = i_x;
+
+                        ck::index_t i_k_itr = k_per_block;
+                        while(i_k_itr > 0)
+                        {
+                            ck::index_t current_k_block = ck::math::min(C - i_c_itr_k, k_per_block);
+
+                            if((*reinterpret_cast<uint32_t*>(&i_hi_itr_k) < Hi) &&
+                               (*reinterpret_cast<uint32_t*>(&i_wi_itr_k) < Wi))
+                                avx2_util::memcpy32_avx2(p_dst_k, p_src_k, current_k_block);
+                            else
+                                avx2_util::memset32_avx2(p_dst_k, 0, current_k_block);
+
+                            p_dst_k += current_k_block;
+                            p_src_k += current_k_block;
+
+                            i_c_itr_k += current_k_block;
+                            if(i_c_itr_k >= C)
+                            {
+                                i_c_itr_k = 0;
+                                i_x_itr_k++;
+                                i_wi_itr_k += Dx;
+                                p_src_k += input_offset_ovf_c_acc_x;
+                            }
+                            if(i_x_itr_k >= Fx)
+                            {
+                                i_x_itr_k = 0;
+                                i_y_itr_k++;
+                                i_hi_itr_k += Dy;
+                                p_src_k += input_offset_ovf_x_acc_y;
+                            }
+
+                            i_k_itr -= current_k_block;
+                        }
+                        /***  go along Gemm K ***/
+
+                        p_dst += k_per_block;
+
+                        i_wo_itr++;
+                        i_wi_itr += Sx;
+                        p_src += input_offset_acc_wi;
+                        if(i_wo_itr >= Wo)
+                        {
+                            i_wo_itr = 0;
+                            i_wi_itr -= Wo * Sx;
+                            i_ho_itr++;
+                            i_hi_itr += Sy;
+                            p_src += input_offset_ovf_wi_acc_hi;
+                        }
+
+                        if(i_ho_itr >= Ho)
+                        {
+                            i_ho_itr = 0;
+                            i_hi_itr -= Ho * Sy;
+                            // i_n++;
+                            p_src += input_offset_ovf_hi_acc_n;
+                        }
+
+                        i_m_itr--;
+                    }
+                }
+            }
+        }
+    }
+
+    void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& src_slice_origin_step_idx)
+    {
+        ck::index_t move_k = src_slice_origin_step_idx[Number<1>{}];
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
+        {
+            // printf("  => move_k:%d, src offset:%d\n", move_k, src_offset);
+            i_c += move_k;
+            src_offset += move_k;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization_t::Filter1x1Pad0)
+        {
+            i_c += move_k;
+            src_offset += move_k;
+        }
+        else
+        {
+            if constexpr(GemmKSpecialization ==
+                         ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC)
+            {
+                // c % k_per_block == 0, so every time k_per_block here is the same
+                // ihi = iho * s_stride_h + iy * s_dilation_h - s_pad_h
+                // iwi = iwo * s_stride_w + ix * s_dilation_w - s_pad_w
+                // printf("222222 C:%d, src_offset:%d, i_c:%d, i_x:%d\n", C, src_offset, i_c, i_x);
+                // fflush(stdout);
+
+                // TODO: branch seems weird
+
+                i_c += move_k;
+                src_offset += move_k;
+
+                // printf("3333[%d]  src_offset:%d\n", __LINE__, src_offset);
+
+                if(i_c >= C)
+                {
+                    i_c = 0;
+                    i_x++;
+                    i_wi += Dx;
+                    src_offset += Dx * C - C;
+                    // printf("3333[%d]  src_offset:%d\n", __LINE__, src_offset);
+                }
+                if(i_x >= Fx)
+                {
+                    i_x = 0;
+                    i_y++;
+                    i_wi = i_wi - Fx * Dx;
+                    i_hi += Dy;
+
+                    src_offset += Dy * Wi * C - Fx * Dx * C;
+                    // printf("3333[%d]  src_offset:%d\n", __LINE__, src_offset);
+                }
+
+                // printf("inp move:%d, i_c:%d, i_hi:%d, i_wi:%d src_offset:%d\n", move_k, i_c,
+                // i_hi, i_wi, src_offset); fflush(stdout);
+            }
+            else
+            {
+                i_gemm_k += move_k;
+
+                i_c = i_gemm_k % C;
+                i_x = (i_gemm_k / C) % Fx;
+                i_y = (i_gemm_k / C) / Fx;
+
+                i_hi = i_ho * Sy + i_y * Dy - Py;
+                i_wi = i_wo * Sx + i_x * Dx - Px;
+
+                src_offset = i_n * Hi * Wi * C + i_hi * Wi * C + i_wi * C + i_c;
+            }
+        }
+    }
+
+    void MoveDstSliceWindow(const DstDesc&, const Index&) {}
+
+    private:
+    const ElementwiseOperation element_op_;
+
+    ck::index_t i_n;
+    ck::index_t i_c;
+    ck::index_t i_hi;
+    ck::index_t i_wi;
+    ck::index_t i_ho;
+    ck::index_t i_wo;
+    ck::index_t i_y;
+    ck::index_t i_x;
+    ck::index_t i_gemm_k;
+
+    ck::index_t N;
+    // ck::index_t K;
+    ck::index_t C;
+    ck::index_t Hi;
+    ck::index_t Wi;
+    ck::index_t Ho;
+    ck::index_t Wo;
+
+    ck::index_t Sy;
+    ck::index_t Sx;
+
+    ck::index_t Dy;
+    ck::index_t Dx;
+
+    ck::index_t Py;
+    ck::index_t Px;
+
+    ck::index_t Fy;
+    ck::index_t Fx;
+
+    intptr_t input_offset_acc_wi;
+    intptr_t input_offset_ovf_wi_acc_hi;
+    intptr_t input_offset_ovf_hi_acc_n;
+
+    // intptr_t input_offset_acc_c;
+    intptr_t input_offset_ovf_c_acc_x;
+    intptr_t input_offset_ovf_x_acc_y;
+
+    intptr_t src_offset; // keep this as pointer type in case we have negative offset
+};
+
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          bool BypassTransfer,
+          ConvolutionForwardSpecialization_t ConvForwardSpecialization,
+          ConvolutionForwardGemmKSpecialization_t GemmKSpecialization>
+struct ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC
+{
+    static constexpr ck::index_t nDim = SrcDesc::GetNumOfDimension();
+    using Index                       = MultiIndex<nDim>;
+
+    // using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    // using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    constexpr ThreadwiseTensorSliceTransferAvx2Specialization_ConvFwd_Wei_NHWC(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const ElementwiseOperation& element_op)
+        : element_op_(element_op)
+    {
+        GemmN1 = src_desc.GetTransforms()[Number<3>{}].GetUpperLengths()[Number<1>{}];
+        GemmN  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+        GemmK  = src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+    }
+
+    void SetSrcSliceOrigin(const SrcDesc&, const Index& src_slice_origin_idx)
+    {
+        ck::index_t idx_n0 = src_slice_origin_idx[Number<0>{}];
+        ck::index_t idx_k  = src_slice_origin_idx[Number<1>{}];
+        ck::index_t idx_n1 = src_slice_origin_idx[Number<2>{}];
+
+        i_gemm_n = idx_n0 * GemmN1 + idx_n1;
+        // i_gemm_k = idx_k;
+
+        src_offset = idx_n0 * GemmK * GemmN1 + idx_k + idx_n1 * GemmN1; // Note we transpose here
+
+        // printf("xxxx  i_gemm_n:%d, i_gemm_k:%d, src_offset:%d\n", i_gemm_n, i_gemm_k,
+        // src_offset);
+    }
+
+    void SetDstSliceOrigin(const DstDesc&, const Index&) {}
+
+    template <typename SrcBuffer, typename DstBuffer>
+    void Run(const SrcDesc&, const SrcBuffer& src_buf, const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        if constexpr(BypassTransfer)
+        {
+            // TODO: weight NHWC not support this
+        }
+        else
+        {
+            const ck::index_t n_per_block =
+                dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}] *
+                dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+            const ck::index_t k_per_block =
+                dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+
+            // printf(" >>>> %d, %d, %d -> %d(%dx%d), %d\n", GemmN, GemmK, GemmN1, n_per_block,
+            //     dst_desc.GetTransforms()[Number<0>{}]
+            //         .GetUpperLengths()[Number<0>{}],
+            //         dst_desc.GetTransforms()[Number<0>{}]
+            //         .GetUpperLengths()[Number<2>{}],
+            // k_per_block);
+
+            const float* p_src = reinterpret_cast<const float*>(src_buf.p_data_) + src_offset;
+            float* p_dst       = reinterpret_cast<float*>(dst_buf.p_data_);
+
+            // n * k -> n0 * k * n1, n1 = 8, n0 = n/8
+            for(index_t i_n_itr = 0; i_n_itr < n_per_block; i_n_itr += 8)
+            {
+                ck::index_t current_n_8 = ck::math::min(GemmN - (i_n_itr + i_gemm_n), 8);
+                ck::index_t i_k_itr     = k_per_block;
+                if(current_n_8 == 8)
+                {
+                    const float* p_src_k = p_src;
+                    float* p_dst_k       = p_dst;
+                    while(i_k_itr >= 8)
+                    {
+                        avx2_util::transpose8x8_avx2(p_dst_k, 8, p_src_k, GemmK);
+                        p_dst_k += 8 * 8;
+                        p_src_k += 8;
+                        i_k_itr -= 8;
+                    }
+                    if(i_k_itr & 4)
+                    {
+                        p_dst_k[0 * 8 + 0] = p_src_k[0 * GemmK + 0];
+                        p_dst_k[0 * 8 + 1] = p_src_k[1 * GemmK + 0];
+                        p_dst_k[0 * 8 + 2] = p_src_k[2 * GemmK + 0];
+                        p_dst_k[0 * 8 + 3] = p_src_k[3 * GemmK + 0];
+                        p_dst_k[0 * 8 + 4] = p_src_k[4 * GemmK + 0];
+                        p_dst_k[0 * 8 + 5] = p_src_k[5 * GemmK + 0];
+                        p_dst_k[0 * 8 + 6] = p_src_k[6 * GemmK + 0];
+                        p_dst_k[0 * 8 + 7] = p_src_k[7 * GemmK + 0];
+
+                        p_dst_k[1 * 8 + 0] = p_src_k[0 * GemmK + 1];
+                        p_dst_k[1 * 8 + 1] = p_src_k[1 * GemmK + 1];
+                        p_dst_k[1 * 8 + 2] = p_src_k[2 * GemmK + 1];
+                        p_dst_k[1 * 8 + 3] = p_src_k[3 * GemmK + 1];
+                        p_dst_k[1 * 8 + 4] = p_src_k[4 * GemmK + 1];
+                        p_dst_k[1 * 8 + 5] = p_src_k[5 * GemmK + 1];
+                        p_dst_k[1 * 8 + 6] = p_src_k[6 * GemmK + 1];
+                        p_dst_k[1 * 8 + 7] = p_src_k[7 * GemmK + 1];
+
+                        p_dst_k[2 * 8 + 0] = p_src_k[0 * GemmK + 2];
+                        p_dst_k[2 * 8 + 1] = p_src_k[1 * GemmK + 2];
+                        p_dst_k[2 * 8 + 2] = p_src_k[2 * GemmK + 2];
+                        p_dst_k[2 * 8 + 3] = p_src_k[3 * GemmK + 2];
+                        p_dst_k[2 * 8 + 4] = p_src_k[4 * GemmK + 2];
+                        p_dst_k[2 * 8 + 5] = p_src_k[5 * GemmK + 2];
+                        p_dst_k[2 * 8 + 6] = p_src_k[6 * GemmK + 2];
+                        p_dst_k[2 * 8 + 7] = p_src_k[7 * GemmK + 2];
+
+                        p_dst_k[3 * 8 + 0] = p_src_k[0 * GemmK + 3];
+                        p_dst_k[3 * 8 + 1] = p_src_k[1 * GemmK + 3];
+                        p_dst_k[3 * 8 + 2] = p_src_k[2 * GemmK + 3];
+                        p_dst_k[3 * 8 + 3] = p_src_k[3 * GemmK + 3];
+                        p_dst_k[3 * 8 + 4] = p_src_k[4 * GemmK + 3];
+                        p_dst_k[3 * 8 + 5] = p_src_k[5 * GemmK + 3];
+                        p_dst_k[3 * 8 + 6] = p_src_k[6 * GemmK + 3];
+                        p_dst_k[3 * 8 + 7] = p_src_k[7 * GemmK + 3];
+
+                        p_dst_k += 4 * 8;
+                        p_src_k += 4;
+                    }
+                    if(i_k_itr & 2)
+                    {
+                        p_dst_k[0 * 8 + 0] = p_src_k[0 * GemmK + 0];
+                        p_dst_k[0 * 8 + 1] = p_src_k[1 * GemmK + 0];
+                        p_dst_k[0 * 8 + 2] = p_src_k[2 * GemmK + 0];
+                        p_dst_k[0 * 8 + 3] = p_src_k[3 * GemmK + 0];
+                        p_dst_k[0 * 8 + 4] = p_src_k[4 * GemmK + 0];
+                        p_dst_k[0 * 8 + 5] = p_src_k[5 * GemmK + 0];
+                        p_dst_k[0 * 8 + 6] = p_src_k[6 * GemmK + 0];
+                        p_dst_k[0 * 8 + 7] = p_src_k[7 * GemmK + 0];
+
+                        p_dst_k[1 * 8 + 0] = p_src_k[0 * GemmK + 1];
+                        p_dst_k[1 * 8 + 1] = p_src_k[1 * GemmK + 1];
+                        p_dst_k[1 * 8 + 2] = p_src_k[2 * GemmK + 1];
+                        p_dst_k[1 * 8 + 3] = p_src_k[3 * GemmK + 1];
+                        p_dst_k[1 * 8 + 4] = p_src_k[4 * GemmK + 1];
+                        p_dst_k[1 * 8 + 5] = p_src_k[5 * GemmK + 1];
+                        p_dst_k[1 * 8 + 6] = p_src_k[6 * GemmK + 1];
+                        p_dst_k[1 * 8 + 7] = p_src_k[7 * GemmK + 1];
+
+                        p_dst_k += 2 * 8;
+                        p_src_k += 2;
+                    }
+                    if(i_k_itr & 1)
+                    {
+                        p_dst_k[0 * 8 + 0] = p_src_k[0 * GemmK + 0];
+                        p_dst_k[0 * 8 + 1] = p_src_k[1 * GemmK + 0];
+                        p_dst_k[0 * 8 + 2] = p_src_k[2 * GemmK + 0];
+                        p_dst_k[0 * 8 + 3] = p_src_k[3 * GemmK + 0];
+                        p_dst_k[0 * 8 + 4] = p_src_k[4 * GemmK + 0];
+                        p_dst_k[0 * 8 + 5] = p_src_k[5 * GemmK + 0];
+                        p_dst_k[0 * 8 + 6] = p_src_k[6 * GemmK + 0];
+                        p_dst_k[0 * 8 + 7] = p_src_k[7 * GemmK + 0];
+                    }
+                }
+                else
+                {
+                    const float* p_src_k = p_src;
+                    float* p_dst_k       = p_dst;
+
+                    for(index_t i_sub_n = 0; i_sub_n < 8; i_sub_n++)
+                    {
+                        for(index_t i_sub_k = 0; i_sub_k < k_per_block; i_sub_k++)
+                        {
+                            ck::index_t i_current_n_itr = i_n_itr + i_sub_n + i_gemm_n;
+
+                            float v =
+                                i_current_n_itr < GemmN ? p_src_k[i_sub_n * GemmK + i_sub_k] : .0f;
+
+                            p_dst_k[i_sub_k * 8 + i_sub_n] = v;
+                        }
+                    }
+                }
+
+                p_dst += 8 * k_per_block;
+                p_src += 8 * GemmK;
+            }
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& src_slice_origin_step_idx)
+    {
+        ck::index_t move_k  = src_slice_origin_step_idx[Number<1>{}];
+        ck::index_t move_n0 = src_slice_origin_step_idx[Number<0>{}];
+
+        // i_gemm_k += move_k;
+
+        // printf("wei move:%d\n", move_k); fflush(stdout);
+
+        src_offset += move_k + move_n0 * GemmK * GemmN1;
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    void MoveDstSliceWindow(const DstDesc&, const Index&) {}
+
+    private:
+    const ElementwiseOperation element_op_;
+
+    ck::index_t i_gemm_n;
+    // ck::index_t i_gemm_k;
+
+    // ck::index_t GemmN0;
+    ck::index_t GemmN1;
+    ck::index_t GemmN;
+    ck::index_t GemmK;
+
+    intptr_t src_offset;
+};
+
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          bool BypassTransfer,
+          ConvolutionForwardSpecialization_t ConvForwardSpecialization,
+          ConvolutionForwardGemmKSpecialization_t GemmKSpecialization>
+struct ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN
+{
+    static constexpr ck::index_t nDim = SrcDesc::GetNumOfDimension();
+    using Index                       = MultiIndex<nDim>;
+
+    constexpr ThreadwiseTensorSliceTransferAvx2Specialization_MatC_Store_MxN(
+        const SrcDesc& src_desc,
+        const Index&,
+        const DstDesc& dst_desc,
+        const Index&,
+        const ElementwiseOperation& element_op)
+        : element_op_(element_op)
+    {
+        DstGemmM = dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<0>{}];
+        DstGemmN = dst_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+
+        src_offset = 0;
+        dst_offset = 0;
+    }
+
+    void SetSrcSliceOrigin(const SrcDesc&, const Index& src_slice_origin_idx)
+    {
+        if constexpr(BypassTransfer)
+        {
+            auto i_src_gemm_m = src_slice_origin_idx[Number<0>{}];
+            auto i_src_gemm_n = src_slice_origin_idx[Number<1>{}];
+
+            src_offset = i_src_gemm_m * DstGemmN + i_src_gemm_n;
+        }
+    }
+
+    void SetDstSliceOrigin(const DstDesc&, const Index& dst_slice_origin_idx)
+    {
+        i_dst_gemm_m = dst_slice_origin_idx[Number<0>{}];
+        i_dst_gemm_n = dst_slice_origin_idx[Number<1>{}];
+
+        dst_offset = i_dst_gemm_m * DstGemmN + i_dst_gemm_n;
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    void
+    Run(const SrcDesc& src_desc, SrcBuffer& src_buf, const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        if constexpr(BypassTransfer)
+        {
+            src_buf.p_data_ = reinterpret_cast<float*>(dst_buf.p_data_) + src_offset;
+        }
+        else
+        {
+            const ck::index_t m_per_block =
+                src_desc.GetTransforms()[Number<0>{}]
+                    .GetUpperLengths()[Number<0>{}]; // must be multiple of 8
+            const ck::index_t n_per_block =
+                src_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}];
+
+            const ck::index_t current_n = ck::math::min(DstGemmN - i_dst_gemm_n, n_per_block);
+
+            const float* p_src = reinterpret_cast<float*>(src_buf.p_data_) + src_offset;
+            float* p_dst       = reinterpret_cast<float*>(dst_buf.p_data_) + dst_offset;
+
+            ck::index_t i_m_itr = m_per_block;
+
+            // printf("xxxx %d, current_n:%d, DstGemmN:%d, n_per_block:%d\n",__LINE__, current_n,
+            // DstGemmN, n_per_block);fflush(stdout);
+
+            // standard 8-4-2-1 pattern
+            while(i_m_itr >= 8)
+            {
+                avx2_util::memcpy32_avx2(p_dst + 0 * DstGemmN, p_src + 0 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 1 * DstGemmN, p_src + 1 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 2 * DstGemmN, p_src + 2 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 3 * DstGemmN, p_src + 3 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 4 * DstGemmN, p_src + 4 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 5 * DstGemmN, p_src + 5 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 6 * DstGemmN, p_src + 6 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 7 * DstGemmN, p_src + 7 * n_per_block, current_n);
+
+                i_m_itr -= 8;
+                p_dst += 8 * DstGemmN;
+                p_src += 8 * n_per_block;
+            }
+
+            if(i_m_itr & 4)
+            {
+                avx2_util::memcpy32_avx2(p_dst + 0 * DstGemmN, p_src + 0 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 1 * DstGemmN, p_src + 1 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 2 * DstGemmN, p_src + 2 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 3 * DstGemmN, p_src + 3 * n_per_block, current_n);
+
+                p_dst += 4 * DstGemmN;
+                p_src += 4 * n_per_block;
+            }
+
+            if(i_m_itr & 2)
+            {
+                avx2_util::memcpy32_avx2(p_dst + 0 * DstGemmN, p_src + 0 * n_per_block, current_n);
+                avx2_util::memcpy32_avx2(p_dst + 1 * DstGemmN, p_src + 1 * n_per_block, current_n);
+
+                p_dst += 2 * DstGemmN;
+                p_src += 2 * n_per_block;
+            }
+
+            if(i_m_itr & 1)
+            {
+                avx2_util::memcpy32_avx2(p_dst + 0 * DstGemmN, p_src + 0 * n_per_block, current_n);
+            }
+
+            // printf("xxxx %d\n",__LINE__);fflush(stdout);
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    void MoveSrcSliceWindow(const SrcDesc&, const Index&) {}
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    void MoveDstSliceWindow(const DstDesc&, const Index&) {}
+
+    private:
+    const ElementwiseOperation element_op_;
+
+    ck::index_t i_dst_gemm_m;
+    ck::index_t i_dst_gemm_n;
+
+    ck::index_t DstGemmM;
+    ck::index_t DstGemmN;
+
+    intptr_t src_offset;
+    intptr_t dst_offset;
+};
+
+} // namespace cpu
+} // namespace ck
+
+#endif

library/include/ck/library/host_tensor/device.hpp

@@ -121,7 +121,11 @@ template <typename... Args, typename F>
 float launch_and_time_cpu_kernel(F kernel, int nrepeat, Args... args)
 {
     WallTimer timer;
-    kernel(args...);
+
+    int nwarmup = 3;
+
+    for(int i = 0; i < nwarmup; i++)
+        kernel(args...);
 
     timer.Start();
     for(int i = 0; i < nrepeat; i++)

library/src/tensor_operation_instance/cpu/conv2d_fwd/device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_instance.cpp

@@ -19,7 +19,7 @@ using InLayout                         = ck::tensor_layout::gemm::RowMajor;    /
 using WeiLayout                        = ck::tensor_layout::gemm::ColumnMajor; // KYXC
 static constexpr bool NonTemporalStore = false;
 
-using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough;
+using PT = ck::tensor_operation::cpu::element_wise::PassThrough;
 using ThreadwiseGemmAvx2_MxN_4x24_Dispatch =
     ck::cpu::ThreadwiseGemmAvx2_MxN_4x24_Dispatch<InType,
                                                   WeiType,
@@ -37,53 +37,37 @@ static constexpr auto ConvFwd1x1P0 =
 static constexpr auto ConvFwd1x1S1P0 =
     ck::tensor_operation::cpu::device::ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0;
 
+static constexpr auto DefaultGemmKLoop =
+    ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::DefaultGemmKLoop;
+static constexpr auto GemmKLoopOverC =
+    ck::tensor_operation::cpu::device::ConvolutionForwardGemmKSpecialization_t::NHWC_GemmKLoopOverC;
+
+static constexpr auto LoopOver_MNK = ck::tensor_operation::cpu::device::LoopOver_MNK;
+static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver_MKN;
+
+// clang-format off
+#define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(a_elem_op, b_elem_op, c_elem_op, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf) \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0,   GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0,   DefaultGemmKLoop, LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0,   GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1P0,   DefaultGemmKLoop, LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, a_local_buf, b_local_buf, c_local_buf>
+
+// clang-format on
+
 using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_instances = std::tuple<
-    //#################################################################|InDataType|WeiDataType|OutDataType|AccDataType|InElementwiseOp|WeiElementwiseOp|OutElementwiseOp|ConvForwardSp|NumDimSpatial|MPerBlock|NPerBlock|KPerBlock|ThreadwiseGemm_Dispatch
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
-        float,
-        float,
-        float,
-        float,
-        PassThrough,
-        PassThrough,
-        PassThrough,
-        ConvFwdDefault,
-        2,
-        256,
-        128,
-        64,
-        ThreadwiseGemmAvx2_MxN_4x24_Dispatch>,
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
-        float,
-        float,
-        float,
-        float,
-        PassThrough,
-        PassThrough,
-        PassThrough,
-        ConvFwdDefault,
-        2,
-        512,
-        256,
-        128,
-        ThreadwiseGemmAvx2_MxN_4x24_Dispatch>,
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
-        float,
-        float,
-        float,
-        float,
-        PassThrough,
-        PassThrough,
-        PassThrough,
-        ConvFwdDefault,
-        2,
-        1024,
-        144,
-        128,
-        ThreadwiseGemmAvx2_MxN_4x24_Dispatch>>;
+    // clang-format off
+    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  256, 120,  64,  4, 24, true, true, false),
+    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  512, 144, 128,  4, 24, true, true, false),
+    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  512, 240, 128,  4, 24, true, true, false),
+    // DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  768, 192, 128,  4, 24, true, true, false),
+    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  768, 288, 128,  4, 24, true, true, false)>;
+// clang-format on
 
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(
-    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances)
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
 {
     ck::tensor_operation::device::add_device_operation_instances(
         instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_instances{});

test/convnd_fwd_cpu/conv2d_fwd_cpu.cpp

-#include "config.hpp"
-#include "device.hpp"
-#include "host_tensor.hpp"
-#include "host_tensor_generator.hpp"
-#include "tensor_layout.hpp"
-#include "device_tensor.hpp"
-#include "device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp"
-#include "element_wise_operation_cpu.hpp"
-#include "reference_conv_fwd.hpp"
-#include "element_wise_operation_cpu.hpp"
-#include "dynamic_buffer_cpu.hpp"
-
-#define AVX2_DATA_ALIGNMENT 32
-
-namespace ck {
-namespace tensor_operation {
-namespace cpu {
-namespace device {
-namespace device_conv2d_fwd_avx2_instance {
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(
-    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
-
-} // namespace device_conv2d_fwd_avx2_instance
-} // namespace device
-} // namespace cpu
-} // namespace tensor_operation
-} // namespace ck
-
-using InElementOp  = ck::tensor_operation::cpu::element_wise::PassThrough;
-using WeiElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
-using OutElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
-
-template <typename T>
-static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
-{
-    float max_diff = 1e-6;
-
-    for(int i = 0; i < ref.mData.size(); ++i)
-    {
-        float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
-        if(max_diff < diff)
-        {
-            return false;
-        }
-    }
-
-    return true;
-}
-
-int main(int argc, char* argv[])
-{
-    int data_type   = 0;
-    int init_method = 0;
-
-    // Conv shape
-    ck::index_t N               = 128;
-    ck::index_t K               = 256;
-    ck::index_t C               = 192;
-    ck::index_t Y               = 3;
-    ck::index_t X               = 3;
-    ck::index_t Hi              = 71;
-    ck::index_t Wi              = 71;
-    ck::index_t conv_stride_h   = 2;
-    ck::index_t conv_stride_w   = 2;
-    ck::index_t conv_dilation_h = 1;
-    ck::index_t conv_dilation_w = 1;
-    ck::index_t in_left_pad_h   = 1;
-    ck::index_t in_left_pad_w   = 1;
-    ck::index_t in_right_pad_h  = 1;
-    ck::index_t in_right_pad_w  = 1;
-
-    if(argc == 1)
-    {
-        data_type   = 1;
-        init_method = 1;
-    }
-    else if(argc == 3)
-    {
-        data_type   = std::stoi(argv[1]);
-        init_method = std::stoi(argv[2]);
-    }
-    else if(argc == 18)
-    {
-        data_type   = std::stoi(argv[1]);
-        init_method = std::stoi(argv[2]);
-
-        N               = std::stoi(argv[3]);
-        K               = std::stoi(argv[4]);
-        C               = std::stoi(argv[5]);
-        Y               = std::stoi(argv[6]);
-        X               = std::stoi(argv[7]);
-        Hi              = std::stoi(argv[8]);
-        Wi              = std::stoi(argv[9]);
-        conv_stride_h   = std::stoi(argv[10]);
-        conv_stride_w   = std::stoi(argv[11]);
-        conv_dilation_h = std::stoi(argv[12]);
-        conv_dilation_w = std::stoi(argv[13]);
-        in_left_pad_h   = std::stoi(argv[14]);
-        in_left_pad_w   = std::stoi(argv[15]);
-        in_right_pad_h  = std::stoi(argv[16]);
-        in_right_pad_w  = std::stoi(argv[17]);
-    }
-    else
-    {
-        printf("arg1: data type (0=fp32, 1=fp16)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
-               "RightPx\n");
-        exit(1);
-    }
-
-    auto Run = [&](auto input_type, auto wei_type, auto out_type, auto acc_type) {
-        using InDataType  = decltype(input_type);
-        using WeiDataType = decltype(wei_type);
-        using OutDataType = decltype(out_type);
-        using AccDataType = decltype(acc_type);
-
-        using ReferenceConvBwdInstance =
-            ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
-                                                             WeiDataType,
-                                                             OutDataType,
-                                                             AccDataType,
-                                                             InElementOp,
-                                                             WeiElementOp,
-                                                             OutElementOp>;
-
-        const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
-        const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
-
-        const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
-        const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
-
-        const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
-        const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
-        const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
-        const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
-        const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
-        const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
-        const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
-
-        auto f_host_tensor_descriptor = [](std::size_t N_,
-                                           std::size_t C_,
-                                           std::size_t H_,
-                                           std::size_t W_) {
-            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H_, W_}),
-                                        std::vector<std::size_t>({C_ * H_ * W_, 1, W_ * C_, C_}));
-        };
-
-        Tensor<OutDataType> out_n_ho_wo_k(f_host_tensor_descriptor(N, K, Ho, Wo));
-        Tensor<WeiDataType> wei_k_y_x_c(f_host_tensor_descriptor(K, C, Y, X));
-        Tensor<InDataType> in_n_hi_wi_c_host_result(f_host_tensor_descriptor(N, C, Hi, Wi));
-        Tensor<InDataType> in_n_hi_wi_c_device_result(f_host_tensor_descriptor(N, C, Hi, Wi));
-
-        std::cout << "in (N, C, Hi, Wi): " << in_n_hi_wi_c_host_result.mDesc << std::endl;
-        std::cout << "wei(K, C,  Y,  X): " << wei_k_y_x_c.mDesc << std::endl;
-        std::cout << "out(N, K, Ho, Wo): " << out_n_ho_wo_k.mDesc << std::endl;
-
-        switch(init_method)
-        {
-        case 0: break;
-        case 1:
-            out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
-            wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
-            break;
-        case 2:
-            out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
-            wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
-            break;
-        default:
-            out_n_ho_wo_k.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
-            wei_k_y_x_c.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
-        }
-
-        DeviceAlignedMemCPU in_device_buf(sizeof(InDataType) *
-                                              in_n_hi_wi_c_device_result.mDesc.GetElementSpace(),
-                                          AVX2_DATA_ALIGNMENT);
-        DeviceAlignedMemCPU wei_device_buf(
-            sizeof(WeiDataType) * wei_k_y_x_c.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT);
-        DeviceAlignedMemCPU out_device_buf(
-            sizeof(OutDataType) * out_n_ho_wo_k.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT);
-
-        out_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
-        wei_device_buf.ToDevice(wei_k_y_x_c.mData.data());
-        // reset input to zero
-        in_n_hi_wi_c_device_result.GenerateTensorValue(GeneratorTensor_1<InDataType>{0});
-        in_device_buf.ToDevice(in_n_hi_wi_c_device_result.mData.data());
-
-        // get host result
-        {
-            auto ref_conv    = ReferenceConvFwdInstance{};
-            auto ref_invoker = ref_conv.MakeInvoker();
-
-            auto ref_argument = ref_conv.MakeArgument(in_n_hi_wi_c_host_result,
-                                                      wei_k_y_x_c,
-                                                      out_n_ho_wo_k,
-                                                      conv_filter_strides,
-                                                      conv_filter_dilations,
-                                                      input_left_pads,
-                                                      input_right_pads,
-                                                      InElementOp{},
-                                                      WeiElementOp{},
-                                                      OutElementOp{});
-            ref_invoker.Run(ref_argument);
-        }
-
-        using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough;
-        using DeviceConvFwdNoOpPtr =
-            ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>;
-
-        // add device Conv instances
-        std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
-
-        if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
-                     ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
-                     ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
-        {
-            ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
-                add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(conv_ptrs);
-        }
-
-        if(conv_ptrs.size() <= 0)
-        {
-            throw std::runtime_error("wrong! no device Conv instance found");
-        }
-
-        // profile device Conv instances
-        bool success = true;
-        for(auto& conv_ptr : conv_ptrs)
-        {
-            auto argument_ptr = conv_ptr->MakeArgumentPointer(
-                static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
-                static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
-                static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
-                N,
-                K,
-                C,
-                input_spatial_lengths,
-                filter_spatial_lengths,
-                output_spatial_lengths,
-                conv_filter_strides,
-                conv_filter_dilations,
-                input_left_pads,
-                input_right_pads,
-                InElementOp{},
-                WeiElementOp{},
-                OutElementOp{});
-
-            if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
-            {
-                auto invoker_ptr = conv_ptr->MakeInvokerPointer();
-                invoker_ptr->Run(argument_ptr.get(), 1);
-
-                in_device_buf.FromDevice(in_n_hi_wi_c_device_result.mData.data());
-
-                if(!check_out(in_n_hi_wi_c_host_result, in_n_hi_wi_c_device_result))
-                {
-                    std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
-                    success = false;
-                }
-                else
-                {
-                    std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
-                }
-            }
-            else
-            {
-                std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
-            }
-        }
-
-        if(success)
-        {
-            std::cout << "test conv2d fwd cpu : Pass" << std::endl;
-            return 0;
-        }
-        else
-        {
-            std::cout << "test conv2d fwd cpu: Fail " << std::endl;
-            return -1;
-        }
-    };
-
-    if(data_type == 0)
-    {
-        return Run(F32(), F32(), F32(), F32());
-    }
-    else if(data_type == 1)
-    {
-        return Run(F16(), F16(), F16(), F32());
-    }
-    else
-    {
-        return 1;
-    }
-}
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "tensor_layout.hpp"
+#include "device_tensor.hpp"
+#include "device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp"
+#include "element_wise_operation_cpu.hpp"
+#include "reference_conv_fwd.hpp"
+#include "element_wise_operation_cpu.hpp"
+#include "dynamic_buffer_cpu.hpp"
+#include <omp.h>
+
+#define AVX2_DATA_ALIGNMENT 32
+using F32 = float;
+using F16 = ck::half_t;
+
+namespace ck {
+namespace tensor_operation {
+namespace cpu {
+namespace device {
+namespace device_conv2d_fwd_avx2_instance {
+
+using PassThrough = ck::tensor_operation::cpu::element_wise::PassThrough;
+
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+
+} // namespace device_conv2d_fwd_avx2_instance
+} // namespace device
+} // namespace cpu
+} // namespace tensor_operation
+} // namespace ck
+
+using InElementOp  = ck::tensor_operation::cpu::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::cpu::element_wise::PassThrough;
+
+template <typename T>
+static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
+{
+    int error_count = 0;
+    float max_diff  = 1e-6;
+
+    for(int i = 0; i < ref.mData.size(); ++i)
+    {
+        float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
+        if(max_diff < diff)
+        {
+            error_count++;
+            printf("idx:%3d, ref:%f, res:%f (diff:%f)\n",
+                   i,
+                   double(ref.mData[i]),
+                   double(result.mData[i]),
+                   diff);
+        }
+    }
+
+    return error_count == 0;
+}
+
+float calculate_gflops() {}
+
+int main(int argc, char* argv[])
+{
+    int data_type   = 0;
+    int init_method = 0;
+
+    // Conv shape
+    ck::index_t N               = 2;
+    ck::index_t K               = 256;
+    ck::index_t C               = 192;
+    ck::index_t Y               = 3;
+    ck::index_t X               = 3;
+    ck::index_t Hi              = 71;
+    ck::index_t Wi              = 71;
+    ck::index_t conv_stride_h   = 1;
+    ck::index_t conv_stride_w   = 1;
+    ck::index_t conv_dilation_h = 1;
+    ck::index_t conv_dilation_w = 1;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    if(argc == 1)
+    {
+        data_type   = 0;
+        init_method = 1;
+    }
+    else if(argc == 3)
+    {
+        data_type   = std::stoi(argv[1]);
+        init_method = std::stoi(argv[2]);
+    }
+    else if(argc == 18)
+    {
+        data_type   = std::stoi(argv[1]);
+        init_method = std::stoi(argv[2]);
+
+        N               = std::stoi(argv[3]);
+        K               = std::stoi(argv[4]);
+        C               = std::stoi(argv[5]);
+        Y               = std::stoi(argv[6]);
+        X               = std::stoi(argv[7]);
+        Hi              = std::stoi(argv[8]);
+        Wi              = std::stoi(argv[9]);
+        conv_stride_h   = std::stoi(argv[10]);
+        conv_stride_w   = std::stoi(argv[11]);
+        conv_dilation_h = std::stoi(argv[12]);
+        conv_dilation_w = std::stoi(argv[13]);
+        in_left_pad_h   = std::stoi(argv[14]);
+        in_left_pad_w   = std::stoi(argv[15]);
+        in_right_pad_h  = std::stoi(argv[16]);
+        in_right_pad_w  = std::stoi(argv[17]);
+    }
+    else
+    {
+        printf("arg1: data type (0=fp32, 1=fp16)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3 to 17: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    auto Run = [&](auto input_type, auto wei_type, auto out_type) {
+        using InDataType  = decltype(input_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        using ReferenceConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
+                                                                                      WeiDataType,
+                                                                                      OutDataType,
+                                                                                      InElementOp,
+                                                                                      WeiElementOp,
+                                                                                      OutElementOp>;
+
+        const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+        const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+        const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+        const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+        const std::vector<ck::index_t> input_spatial_lengths{{Hi, Wi}};
+        const std::vector<ck::index_t> filter_spatial_lengths{{Y, X}};
+        const std::vector<ck::index_t> output_spatial_lengths{{Ho, Wo}};
+        const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
+        const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
+        const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
+        const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
+
+        auto f_host_tensor_descriptor = [](std::size_t N_,
+                                           std::size_t C_,
+                                           std::size_t H_,
+                                           std::size_t W_) {
+            return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H_, W_}),
+                                        std::vector<std::size_t>({C_ * H_ * W_, 1, W_ * C_, C_}));
+        };
+
+        Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
+        Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
+        Tensor<OutDataType> out_n_k_ho_wo_host_result(f_host_tensor_descriptor(N, K, Ho, Wo));
+        Tensor<OutDataType> out_n_k_ho_wo_device_result(f_host_tensor_descriptor(N, K, Ho, Wo));
+
+        std::cout << "in (N, C, Hi, Wi): " << in_n_c_hi_wi.mDesc << std::endl;
+        std::cout << "wei(K, C,  Y,  X): " << wei_k_c_y_x.mDesc << std::endl;
+        std::cout << "out(N, K, Ho, Wo): " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+        std::cout << "LPad(H, W):" << in_left_pad_h << "," << in_left_pad_w
+                  << ", RPad(H, W):" << in_right_pad_h << "," << in_right_pad_w
+                  << ", Stride(H, W):" << conv_stride_h << ", " << conv_stride_w
+                  << ", Dilation(H, W):" << conv_dilation_h << ", " << conv_dilation_w
+                  << ", Threads:" << omp_get_max_threads() << std::endl;
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+
+            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+            // in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+            // wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
+            break;
+        case 2:
+            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{});
+            break;
+        case 3:
+
+#define PACK_32(v24, v16, v8, v0) \
+    (((v24 & 0xff) << 24) | ((v16 & 0xff) << 16) | ((v8 & 0xff) << 8) | ((v0 & 0xff) << 0))
+
+            for(auto i_n = 0; i_n < N; i_n++)
+            {
+                for(auto i_c = 0; i_c < C; i_c++)
+                {
+                    for(auto i_hi = 0; i_hi < Hi; i_hi++)
+                    {
+                        for(auto i_wi = 0; i_wi < Wi; i_wi++)
+                        {
+                            uint32_t v                         = PACK_32(i_n, i_c, i_hi, i_wi);
+                            in_n_c_hi_wi(i_n, i_c, i_hi, i_wi) = *reinterpret_cast<float*>(&v);
+                        }
+                    }
+                }
+            }
+
+            for(auto i_k = 0; i_k < K; i_k++)
+            {
+                for(auto i_c = 0; i_c < C; i_c++)
+                {
+                    for(auto i_y = 0; i_y < Y; i_y++)
+                    {
+                        for(auto i_x = 0; i_x < X; i_x++)
+                        {
+                            uint32_t v                      = PACK_32(i_k, i_c, i_y, i_x);
+                            wei_k_c_y_x(i_k, i_c, i_y, i_x) = *reinterpret_cast<float*>(&v);
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0, 1});
+            wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-1, 1});
+        }
+
+        DeviceAlignedMemCPU in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace(),
+                                          AVX2_DATA_ALIGNMENT);
+        DeviceAlignedMemCPU wei_device_buf(
+            sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace(), AVX2_DATA_ALIGNMENT);
+        DeviceAlignedMemCPU out_device_buf(sizeof(OutDataType) *
+                                               out_n_k_ho_wo_host_result.mDesc.GetElementSpace(),
+                                           AVX2_DATA_ALIGNMENT);
+
+        in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+        wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+
+        // get host result
+        {
+            auto ref_conv    = ReferenceConvFwdInstance{};
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                      wei_k_c_y_x,
+                                                      out_n_k_ho_wo_host_result,
+                                                      conv_filter_strides,
+                                                      conv_filter_dilations,
+                                                      input_left_pads,
+                                                      input_right_pads,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+            ref_invoker.Run(ref_argument);
+        }
+
+        using PassThrough          = ck::tensor_operation::cpu::element_wise::PassThrough;
+        using DeviceConvFwdNoOpPtr = ck::tensor_operation::cpu::device::
+            DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>;
+
+        // add device Conv instances
+        std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
+
+        if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
+                     ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
+                     ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
+        {
+            ck::tensor_operation::cpu::device::device_conv2d_fwd_avx2_instance::
+                add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk(conv_ptrs);
+        }
+
+        if(conv_ptrs.size() <= 0)
+        {
+            throw std::runtime_error("wrong! no device Conv instance found");
+        }
+
+        // profile device Conv instances
+        bool success                    = true;
+        double fastest_kernel_time      = std::numeric_limits<double>::max();
+        std::string fastest_kernel_name = "";
+        double fastest_kernel_gflops    = 0;
+        for(auto& conv_ptr : conv_ptrs)
+        {
+            auto argument_ptr = conv_ptr->MakeArgumentPointer(
+                static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                N,
+                K,
+                C,
+                input_spatial_lengths,
+                filter_spatial_lengths,
+                output_spatial_lengths,
+                conv_filter_strides,
+                conv_filter_dilations,
+                input_left_pads,
+                input_right_pads,
+                InElementOp{},
+                WeiElementOp{},
+                OutElementOp{});
+
+            if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                auto invoker_ptr = conv_ptr->MakeInvokerPointer();
+                double time      = invoker_ptr->Run(argument_ptr.get(), 10);
+
+                double total_flop = static_cast<double>(2) * N * C * Ho * Wo * K * Y * X;
+
+                double gflops = (total_flop * 1e-6) / time;
+
+                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+                if(!check_out(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result))
+                {
+                    std::cout << "Fail Info: " << conv_ptr->GetTypeString() << std::endl;
+                    success = false;
+                }
+                else
+                {
+                    std::cout << "Pass Info: " << conv_ptr->GetTypeString() << ", Time:" << time
+                              << "ms, Gflops:" << gflops << std::endl;
+
+                    if(time < fastest_kernel_time)
+                    {
+                        fastest_kernel_time   = time;
+                        fastest_kernel_name   = conv_ptr->GetTypeString();
+                        fastest_kernel_gflops = gflops;
+                    }
+                }
+            }
+            else
+            {
+                std::cout << "Not support Info: " << conv_ptr->GetTypeString() << std::endl;
+            }
+        }
+
+        if(fastest_kernel_time != std::numeric_limits<double>::max())
+        {
+            std::cout << "  fastest:" << fastest_kernel_name << ", time:" << fastest_kernel_time
+                      << "ms, Gflops:" << fastest_kernel_gflops << std::endl;
+        }
+        return 0;
+        // if(success)
+        // {
+        //     std::cout << "test conv2d fwd cpu : Pass" << std::endl;
+        //     return 0;
+        // }
+        // else
+        // {
+        //     std::cout << "test conv2d fwd cpu: Fail " << std::endl;
+        //     return -1;
+        // }
+    };
+
+    if(data_type == 0)
+    {
+        return Run(F32(), F32(), F32());
+    }
+    else
+    {
+        return 1;
+    }
+}

test/cpu_threadwise_transfer/cpu_threadwise_transfer.cpp

@@ -226,6 +226,8 @@ int main(int argc, char** argv)
     static constexpr ck::index_t nDim =
         ck::remove_reference_t<decltype(input_desc)>::GetNumOfDimension();
 
+    input_desc.Print();
+
     auto threadwise_transfer = threadwise_transfer_t{input_desc,
                                                      ck::make_zero_multi_index<nDim>(),
                                                      input_cblock_desc,

test/cpu_ukernel/cpu_gemm_uk.cpp

@@ -313,14 +313,15 @@ void test_ukernel(ukenrel_t uk,
         float* private_c = mat_c + tid * m * n;
 
         ck::cpu::ThreadwiseGemmParam param;
-        param.p_a   = mat_a;
-        param.p_b   = mat_b;
-        param.p_c   = private_c;
-        param.Kr    = k;
-        param.lda   = (std::is_same<Row, ALayout>::value ? k : m) * sizeof(FloatA);
-        param.ldb   = (std::is_same<Row, BLayout>::value ? n : k * 8) * sizeof(FloatB);
-        param.ldc   = n * sizeof(float);
-        param.alpha = alpha;
+        param.p_a         = mat_a;
+        param.p_b         = mat_b;
+        param.p_c         = private_c;
+        param.Kr          = k;
+        param.lda         = (std::is_same<Row, ALayout>::value ? k : m) * sizeof(FloatA);
+        param.ldb         = (std::is_same<Row, BLayout>::value ? n : k * 8) * sizeof(FloatB);
+        param.ldc         = n * sizeof(float);
+        param.alpha       = alpha;
+        param.accmulate_c = 0;
 
         memset(private_c, 0, m * n * sizeof(float));