refactor device instance to use less template, more dynamic tunable

example/cpu_01_conv2d_fwd/cpu_conv2d_fwd.cpp

@@ -21,7 +21,7 @@
 #define TEST_LAYOUT_NHWC_KYXC_NHWK 0
 #define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1
 #define TEST_LAYOUT_NHWC_YXCK_NHWK 2
-#define TEST_LAYOUT TEST_LAYOUT_NHWC_YXCK_NHWK
+#define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXCK8_NHWK
 
 using F32 = float;
 using F16 = ck::half_t;

example/cpu_02_conv2d_fwd_bias_relu_add/cpu_conv2d_fwd_bias_relu_add.cpp

@@ -17,7 +17,7 @@
 #define TEST_LAYOUT_NHWC_KYXC_NHWK 0
 #define TEST_LAYOUT_NHWC_KYXCK8_NHWK 1
 #define TEST_LAYOUT_NHWC_YXCK_NHWK 2
-#define TEST_LAYOUT TEST_LAYOUT_NHWC_YXCK_NHWK
+#define TEST_LAYOUT TEST_LAYOUT_NHWC_KYXCK8_NHWK
 
 using F32 = float;
 using F16 = ck::half_t;

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

@@ -18,8 +18,6 @@ template <typename FloatA,
           typename BBlockDesc,
           typename CDesc,
 
-          ck::index_t KPerBlock,
-
           typename ThreadwiseGemm_Dispatch,
           typename ThreadMNAccessOrder // how we acces gemm MN to utilize micro kernel
           >
@@ -83,8 +81,11 @@ struct BlockwiseGemmAvx2_MxN
         else
         {
             // N/8 * K * 8
-            return b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<1>{}] *
-                   b_block_desc.GetTransforms()[Number<0>{}].GetUpperLengths()[Number<2>{}];
+            // return b_block_desc.GetTransforms()[Number<BBlockDesc::GetNumOfTransform() -
+            // 1>{}].GetUpperLengths()[Number<1>{}] *
+            //       b_block_desc.GetTransforms()[Number<BBlockDesc::GetNumOfTransform() -
+            //       1>{}].GetUpperLengths()[Number<2>{}];
+            return b_block_desc.GetLength(Number<1>{}) * b_block_desc.GetLength(Number<2>{});
         }
     }
 

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

@@ -3,6 +3,7 @@
 
 #include <iostream>
 #include "device_base_cpu.hpp"
+#include "convolution_forward_specialization_cpu.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -77,6 +78,24 @@ using DeviceConvFwdBiasActivationAddPtr =
                                                    WeiElementwiseOperation,
                                                    OutElementwiseOperation>>;
 
+struct DeviceConvFwdDynamicTunable
+{
+    ck::index_t m_per_block;
+    ck::index_t n_per_block;
+    ck::index_t k_per_block;
+
+    // ck::index_t m_per_thread;
+    // ck::index_t n_per_thread;
+
+    // bool use_a_local_buffer;
+    // bool use_b_local_buffer;
+    // bool use_c_local_buffer;
+
+    // ConvolutionForwardSpecialization_t  forward_spec;
+    // ConvolutionForwardGemmKSpecialization_t gemm_k_spec;
+    ConvolutionForwardBlockLoopOverSpecialization_t loop_over_spec;
+};
+
 } // namespace device
 } // namespace cpu
 } // namespace tensor_operation

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

@@ -30,11 +30,7 @@ template <typename InDataType,
           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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -65,17 +61,12 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -106,45 +97,6 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
     using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
 
-    static constexpr auto GetInputBlockDescriptor()
-    {
-        if constexpr(UseALocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
-        }
-        else
-        {
-            return AGridDesc{};
-        }
-    }
-
-    static constexpr auto GetWeightBlockDescriptor()
-    {
-        if constexpr(UseBLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(
-                math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
-                KPerBlock,
-                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
-        }
-        else
-        {
-            return BGridDesc{};
-        }
-    }
-
-    static constexpr auto GetOutputBlockDescriptor()
-    {
-        if constexpr(UseCLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
-        }
-        else
-        {
-            return CGridDesc{};
-        }
-    }
-
     static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
     {
         ck::index_t gemm_n_padded =
@@ -576,6 +528,48 @@ 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 auto GetInputBlockDescriptor()
+    {
+        if constexpr(UseALocalBuffer)
+        {
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return AGridDesc{};
+        }
+    }
+
+    static constexpr auto GetWeightBlockDescriptor()
+    {
+        if constexpr(UseBLocalBuffer)
+        {
+            // return make_naive_tensor_descriptor_packed(make_tuple(
+            //     math::integer_divide_ceil(NPerBlock,
+            //     ThreadwiseGemm_Dispatch::MatrixBMinVectorSize), KPerBlock,
+            //     ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0, 0));
+        }
+        else
+        {
+            return BGridDesc{};
+        }
+    }
+
+    static constexpr auto GetOutputBlockDescriptor()
+    {
+        if constexpr(UseCLocalBuffer)
+        {
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return CGridDesc{};
+        }
+    }
+
     // static constexpr bool UseCLocalBuffer = false;
 
     using AThreadwiseCopy =
@@ -620,20 +614,18 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                                       AElementwiseOperation,   // AElementwiseOperation,
                                       BElementwiseOperation,   // BElementwiseOperation,
                                       CElementwiseOperation,   // CElementwiseOperation,
-                                      MPerBlock,               // MPerBlock,
-                                      NPerBlock,               // NPerBlock,
-                                      KPerBlock,               // KPerBlock,
                                       ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
                                       AThreadwiseCopy,         // AThreadwiseCopy
                                       BThreadwiseCopy,         // BThreadwiseCopy
                                       CThreadwiseCopy,         // CThreadwiseCopy
-                                      BlockMNKAccessOrder,     // BlockMNKAccessOrder,
                                       ck::Sequence<0, 1>,      // ThreadMNAccessOrder
                                       UseALocalBuffer,         // UseALocalBuffer
                                       UseBLocalBuffer,         // UseBLocalBuffer
                                       UseCLocalBuffer          // UseCLocalBuffer
                                       >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -711,11 +703,15 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
     {
         using Argument = DeviceOp::Argument;
 
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
+
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -738,6 +734,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -753,6 +750,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -780,7 +778,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -811,7 +809,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -825,7 +823,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -868,7 +866,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -908,7 +906,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -925,8 +923,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
             << "DFwdAvx2_NHWC_KYXC"
             <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
             <<"_KS"<< static_cast<int>(GemmKSpecialization)
-            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
-            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_TT" << MPerThread << "x" << NPerThread 
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)

include/ck/tensor_operation/cpu/device/device_convnd_fwd_avx2_nhwc_kyxck8_nhwk.hpp

@@ -30,11 +30,8 @@ template <typename InDataType,
           typename OutElementwiseOperation,
           ConvolutionForwardSpecialization_t ConvForwardSpecialization,
           ConvolutionForwardGemmKSpecialization_t GemmKSpecialization,
-          ConvolutionForwardBlockLoopOverSpecialization_t BlockLoopOverSpecialization,
+          // 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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -65,17 +62,12 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -518,7 +510,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
     {
         if constexpr(UseALocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
         }
         else
         {
@@ -530,10 +523,11 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
     {
         if constexpr(UseBLocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(
-                math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
-                KPerBlock,
-                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
+            // return make_naive_tensor_descriptor_packed(make_tuple(
+            //     math::integer_divide_ceil(NPerBlock,
+            //     ThreadwiseGemm_Dispatch::MatrixBMinVectorSize), KPerBlock,
+            //     ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0, 0));
         }
         else
         {
@@ -545,7 +539,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
     {
         if constexpr(UseCLocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
         }
         else
         {
@@ -597,20 +592,18 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
                                       AElementwiseOperation,   // AElementwiseOperation,
                                       BElementwiseOperation,   // BElementwiseOperation,
                                       CElementwiseOperation,   // CElementwiseOperation,
-                                      MPerBlock,               // MPerBlock,
-                                      NPerBlock,               // NPerBlock,
-                                      KPerBlock,               // KPerBlock,
                                       ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
                                       AThreadwiseCopy,         // AThreadwiseCopy
                                       BThreadwiseCopy,         // BThreadwiseCopy
                                       CThreadwiseCopy,         // CThreadwiseCopy
-                                      BlockMNKAccessOrder,     // BlockMNKAccessOrder,
                                       ck::Sequence<0, 1>,      // ThreadMNAccessOrder
                                       UseALocalBuffer,         // UseALocalBuffer
                                       UseBLocalBuffer,         // UseBLocalBuffer
                                       UseCLocalBuffer          // UseCLocalBuffer
                                       >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -687,12 +680,15 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
     struct Invoker : public BaseInvoker
     {
         using Argument = DeviceOp::Argument;
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
 
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -715,6 +711,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -730,6 +727,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -757,7 +755,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -788,7 +786,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -805,7 +803,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -848,7 +846,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -888,7 +886,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -905,8 +903,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K
             << "DFwdAvx2_NHWC_KYXCK8"
             <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
             <<"_KS"<< static_cast<int>(GemmKSpecialization)
-            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
-            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_TT" << MPerThread << "x" << NPerThread 
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)

include/ck/tensor_operation/cpu/device/device_convnd_fwd_avx2_nhwc_yxck_nhwk.hpp

@@ -29,11 +29,7 @@ template <typename InDataType,
           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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -64,17 +60,12 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -514,7 +505,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
     {
         if constexpr(UseALocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
         }
         else
         {
@@ -526,7 +518,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
     {
         if constexpr(UseBLocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(KPerBlock, NPerBlock));
+            // return make_naive_tensor_descriptor_packed(make_tuple(KPerBlock, NPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
         }
         else
         {
@@ -538,7 +531,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
     {
         if constexpr(UseCLocalBuffer)
         {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+            // return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
         }
         else
         {
@@ -590,20 +584,18 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
                                       AElementwiseOperation,   // AElementwiseOperation,
                                       BElementwiseOperation,   // BElementwiseOperation,
                                       CElementwiseOperation,   // CElementwiseOperation,
-                                      MPerBlock,               // MPerBlock,
-                                      NPerBlock,               // NPerBlock,
-                                      KPerBlock,               // KPerBlock,
                                       ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
                                       AThreadwiseCopy,         // AThreadwiseCopy
                                       BThreadwiseCopy,         // BThreadwiseCopy
                                       CThreadwiseCopy,         // CThreadwiseCopy
-                                      BlockMNKAccessOrder,     // BlockMNKAccessOrder,
                                       ck::Sequence<0, 1>,      // ThreadMNAccessOrder
                                       UseALocalBuffer,         // UseALocalBuffer
                                       UseBLocalBuffer,         // UseBLocalBuffer
                                       UseCLocalBuffer          // UseCLocalBuffer
                                       >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -680,12 +672,15 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
     struct Invoker : public BaseInvoker
     {
         using Argument = DeviceOp::Argument;
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
 
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -708,6 +703,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -723,6 +719,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -750,7 +747,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -781,7 +778,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -801,7 +798,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -844,7 +841,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -884,7 +881,7 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -901,9 +898,8 @@ struct DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K
             << "DFwdAvx2_NHWC_YXCK"
             <<"_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 
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)
             << "_C" << string_local_buffer(UseCLocalBuffer)

include/ck/tensor_operation/cpu/device/device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk.hpp

@@ -32,11 +32,7 @@ template <typename InDataType,
           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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -73,17 +69,12 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -114,45 +105,6 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
 
     using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
 
-    static constexpr auto GetInputBlockDescriptor()
-    {
-        if constexpr(UseALocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
-        }
-        else
-        {
-            return AGridDesc{};
-        }
-    }
-
-    static constexpr auto GetWeightBlockDescriptor()
-    {
-        if constexpr(UseBLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(
-                math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
-                KPerBlock,
-                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
-        }
-        else
-        {
-            return BGridDesc{};
-        }
-    }
-
-    static constexpr auto GetOutputBlockDescriptor()
-    {
-        if constexpr(UseCLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
-        }
-        else
-        {
-            return CGridDesc{};
-        }
-    }
-
     static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
     {
         ck::index_t gemm_n_padded =
@@ -598,6 +550,42 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
     using C0GridDesc = remove_cvref_t<decltype(MakeBiasTensorDescriptor(1, 1))>;
     using C1GridDesc = CGridDesc;
 
+    static constexpr auto GetInputBlockDescriptor()
+    {
+        if constexpr(UseALocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return AGridDesc{};
+        }
+    }
+
+    static constexpr auto GetWeightBlockDescriptor()
+    {
+        if constexpr(UseBLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0, 0));
+        }
+        else
+        {
+            return BGridDesc{};
+        }
+    }
+
+    static constexpr auto GetOutputBlockDescriptor()
+    {
+        if constexpr(UseCLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return CGridDesc{};
+        }
+    }
+
     // static constexpr bool UseCLocalBuffer = false;
 
     using AThreadwiseCopy =
@@ -650,20 +638,18 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
         AElementwiseOperation,   // AElementwiseOperation,
         BElementwiseOperation,   // BElementwiseOperation,
         CElementwiseOperation,   // CElementwiseOperation,
-        MPerBlock,               // MPerBlock,
-        NPerBlock,               // NPerBlock,
-        KPerBlock,               // KPerBlock,
         ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
         AThreadwiseCopy,         // AThreadwiseCopy
         BThreadwiseCopy,         // BThreadwiseCopy
         CThreadwiseCopy,         // CThreadwiseCopy
-        BlockMNKAccessOrder,     // BlockMNKAccessOrder,
         ck::Sequence<0, 1>,      // ThreadMNAccessOrder
         UseALocalBuffer,         // UseALocalBuffer
         UseBLocalBuffer,         // UseBLocalBuffer
         UseCLocalBuffer          // UseCLocalBuffer
         >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -755,11 +741,15 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
     {
         using Argument = DeviceOp::Argument;
 
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
+
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -787,6 +777,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -806,6 +797,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -837,7 +829,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -868,7 +860,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -882,7 +874,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -929,7 +921,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -973,7 +965,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -990,8 +982,8 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Outpu
             << "DFwd_BAA_Avx2_NHWC_KYXC"
             <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
             <<"_KS"<< static_cast<int>(GemmKSpecialization)
-            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
-            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_TT" << MPerThread << "x" << NPerThread 
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)

include/ck/tensor_operation/cpu/device/device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk.hpp

@@ -32,11 +32,7 @@ template <typename InDataType,
           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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -73,17 +69,12 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -114,45 +105,6 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
 
     using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
 
-    static constexpr auto GetInputBlockDescriptor()
-    {
-        if constexpr(UseALocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
-        }
-        else
-        {
-            return AGridDesc{};
-        }
-    }
-
-    static constexpr auto GetWeightBlockDescriptor()
-    {
-        if constexpr(UseBLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(
-                math::integer_divide_ceil(NPerBlock, ThreadwiseGemm_Dispatch::MatrixBMinVectorSize),
-                KPerBlock,
-                ThreadwiseGemm_Dispatch::MatrixBMinVectorSize));
-        }
-        else
-        {
-            return BGridDesc{};
-        }
-    }
-
-    static constexpr auto GetOutputBlockDescriptor()
-    {
-        if constexpr(UseCLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
-        }
-        else
-        {
-            return CGridDesc{};
-        }
-    }
-
     static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
     {
         return make_naive_tensor_descriptor_packed(make_tuple(gemm_n / 8, gemm_k, 8));
@@ -575,6 +527,42 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
     using C0GridDesc = remove_cvref_t<decltype(MakeBiasTensorDescriptor(1, 1))>;
     using C1GridDesc = CGridDesc;
 
+    static constexpr auto GetInputBlockDescriptor()
+    {
+        if constexpr(UseALocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return AGridDesc{};
+        }
+    }
+
+    static constexpr auto GetWeightBlockDescriptor()
+    {
+        if constexpr(UseBLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0, 0));
+        }
+        else
+        {
+            return BGridDesc{};
+        }
+    }
+
+    static constexpr auto GetOutputBlockDescriptor()
+    {
+        if constexpr(UseCLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return CGridDesc{};
+        }
+    }
+
     // static constexpr bool UseCLocalBuffer = false;
 
     using AThreadwiseCopy =
@@ -627,20 +615,18 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
         AElementwiseOperation,   // AElementwiseOperation,
         BElementwiseOperation,   // BElementwiseOperation,
         CElementwiseOperation,   // CElementwiseOperation,
-        MPerBlock,               // MPerBlock,
-        NPerBlock,               // NPerBlock,
-        KPerBlock,               // KPerBlock,
         ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
         AThreadwiseCopy,         // AThreadwiseCopy
         BThreadwiseCopy,         // BThreadwiseCopy
         CThreadwiseCopy,         // CThreadwiseCopy
-        BlockMNKAccessOrder,     // BlockMNKAccessOrder,
         ck::Sequence<0, 1>,      // ThreadMNAccessOrder
         UseALocalBuffer,         // UseALocalBuffer
         UseBLocalBuffer,         // UseBLocalBuffer
         UseCLocalBuffer          // UseCLocalBuffer
         >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -732,11 +718,15 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
     {
         using Argument = DeviceOp::Argument;
 
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
+
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -764,6 +754,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -783,6 +774,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -814,7 +806,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -845,7 +837,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -862,7 +854,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -909,7 +901,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -953,7 +945,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -970,8 +962,8 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Ou
             << "DFwd_BAA_Avx2_NHWC_KYXCK8"
             <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
             <<"_KS"<< static_cast<int>(GemmKSpecialization)
-            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
-            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_TT" << MPerThread << "x" << NPerThread 
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)

include/ck/tensor_operation/cpu/device/device_convnd_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk.hpp

@@ -31,11 +31,7 @@ template <typename InDataType,
           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,
           ck::index_t MPerThread,
           ck::index_t NPerThread,
           bool UseALocalBuffer,
@@ -72,17 +68,12 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
 
     static constexpr bool NonTemporalStore = false;
 
-    static constexpr auto GetBlockMNKAccessOrder()
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K(
+        const DeviceConvFwdDynamicTunable& dtune)
+        : gridwise_gemm(dtune)
     {
-        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)
@@ -111,42 +102,6 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
 
     using ThreadwiseGemm_Dispatch = decltype(GetThreadwiseGemm_Dispatch());
 
-    static constexpr auto GetInputBlockDescriptor()
-    {
-        if constexpr(UseALocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, KPerBlock));
-        }
-        else
-        {
-            return AGridDesc{};
-        }
-    }
-
-    static constexpr auto GetWeightBlockDescriptor()
-    {
-        if constexpr(UseBLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(KPerBlock, NPerBlock));
-        }
-        else
-        {
-            return BGridDesc{};
-        }
-    }
-
-    static constexpr auto GetOutputBlockDescriptor()
-    {
-        if constexpr(UseCLocalBuffer)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(MPerBlock, NPerBlock));
-        }
-        else
-        {
-            return CGridDesc{};
-        }
-    }
-
     static auto GetWeightTensorDescriptor(ck::index_t gemm_k, ck::index_t gemm_n)
     {
         return make_naive_tensor_descriptor_packed(make_tuple(gemm_k, gemm_n));
@@ -568,6 +523,42 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
     using C0GridDesc = remove_cvref_t<decltype(MakeBiasTensorDescriptor(1, 1))>;
     using C1GridDesc = CGridDesc;
 
+    static constexpr auto GetInputBlockDescriptor()
+    {
+        if constexpr(UseALocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return AGridDesc{};
+        }
+    }
+
+    static constexpr auto GetWeightBlockDescriptor()
+    {
+        if constexpr(UseBLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return BGridDesc{};
+        }
+    }
+
+    static constexpr auto GetOutputBlockDescriptor()
+    {
+        if constexpr(UseCLocalBuffer)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(0, 0));
+        }
+        else
+        {
+            return CGridDesc{};
+        }
+    }
+
     // static constexpr bool UseCLocalBuffer = false;
 
     using AThreadwiseCopy =
@@ -620,20 +611,18 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
         AElementwiseOperation,   // AElementwiseOperation,
         BElementwiseOperation,   // BElementwiseOperation,
         CElementwiseOperation,   // CElementwiseOperation,
-        MPerBlock,               // MPerBlock,
-        NPerBlock,               // NPerBlock,
-        KPerBlock,               // KPerBlock,
         ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
         AThreadwiseCopy,         // AThreadwiseCopy
         BThreadwiseCopy,         // BThreadwiseCopy
         CThreadwiseCopy,         // CThreadwiseCopy
-        BlockMNKAccessOrder,     // BlockMNKAccessOrder,
         ck::Sequence<0, 1>,      // ThreadMNAccessOrder
         UseALocalBuffer,         // UseALocalBuffer
         UseBLocalBuffer,         // UseBLocalBuffer
         UseCLocalBuffer          // UseCLocalBuffer
         >;
 
+    GridwiseGemm gridwise_gemm;
+
     // Argument
     struct Argument : public BaseArgument
     {
@@ -725,11 +714,15 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
     {
         using Argument = DeviceOp::Argument;
 
+        GridwiseGemm gridwise_gemm;
+
+        Invoker(const GridwiseGemm& gridwise_gemm_) : gridwise_gemm(gridwise_gemm_) {}
+
         float Run(const Argument& arg,
                   const StreamConfig& stream_config = StreamConfig{},
                   int nrepeat                       = 1)
         {
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
+            if(!gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_))
             {
                 throw std::runtime_error("wrong! GridwiseGemmAvx2_MxN has invalid setting");
             }
@@ -757,6 +750,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
             if(nrepeat != 1)
                 ave_time = launch_and_time_cpu_kernel(kernel,
                                                       nrepeat,
+                                                      gridwise_gemm,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
@@ -776,6 +770,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
             memset(arg.p_c_grid_, 0, arg.c_grid_desc_.GetElementSpaceSize());
 
             launch_cpu_kernel(kernel,
+                              gridwise_gemm,
                               arg.p_a_grid_,
                               arg.p_b_grid_,
                               arg.p_c_grid_,
@@ -807,7 +802,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
         return true;
     }
 
-    static bool IsSupportedArgument(const Argument& arg)
+    bool IsSupportedArgument(const Argument& arg)
     {
         if constexpr(ConvForwardSpecialization ==
                      ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
@@ -838,7 +833,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
                      ConvForwardSpecialization !=
                          ConvolutionForwardSpecialization_t::Filter1x1Stride1Pad0)
         {
-            if(!(arg.Conv_C_ % KPerBlock == 0))
+            if(!(arg.Conv_C_ % gridwise_gemm.dynamic_tunable.k_per_block == 0))
                 return false;
         }
 
@@ -858,7 +853,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
         }
 
         // Gridwise GEMM size
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
+        return gridwise_gemm.CheckValidity(arg.a_grid_desc_, arg.b_grid_desc_, arg.c_grid_desc_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -905,7 +900,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
                         out_element_op};
     }
 
-    static auto MakeInvoker() { return Invoker{}; }
+    auto MakeInvoker() { return Invoker{gridwise_gemm}; }
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in_grid,
@@ -949,7 +944,7 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
     {
-        return std::make_unique<Invoker>(Invoker{});
+        return std::make_unique<Invoker>(Invoker{gridwise_gemm});
     }
 
     std::string GetTypeString() const override
@@ -966,8 +961,8 @@ struct DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Outpu
             << "DFwd_BAA_Avx2_NHWC_YXCK"
             <<"_FS"<< static_cast<int>(ConvForwardSpecialization)
             <<"_KS"<< static_cast<int>(GemmKSpecialization)
-            <<"_BS"<< static_cast<int>(BlockLoopOverSpecialization)
-            << "_BT" << MPerBlock << "x" << NPerBlock << "x" << KPerBlock
+            <<"_BS"<< static_cast<int>(gridwise_gemm.dynamic_tunable.loop_over_spec)
+            << "_BT" << gridwise_gemm.dynamic_tunable.m_per_block << "x" << gridwise_gemm.dynamic_tunable.n_per_block << "x" << gridwise_gemm.dynamic_tunable.k_per_block
             << "_TT" << MPerThread << "x" << NPerThread 
             << "_A" << string_local_buffer(UseALocalBuffer)
             << "_B" << string_local_buffer(UseBLocalBuffer)

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

@@ -28,7 +28,8 @@ template <typename GridwiseGemm,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
-void kernel_gemm_avx_mxn(const FloatA* __restrict__ p_a_grid,
+void kernel_gemm_avx_mxn(const GridwiseGemm& gridwise_gemm,
+                         const FloatA* __restrict__ p_a_grid,
                          const FloatB* __restrict__ p_b_grid,
                          FloatC* __restrict__ p_c_grid,
                          const AGridDesc& a_grid_desc,
@@ -38,7 +39,7 @@ void kernel_gemm_avx_mxn(const FloatA* __restrict__ p_a_grid,
                          const BElementwiseOperation& b_element_op,
                          const CElementwiseOperation& c_element_op)
 {
-    GridwiseGemm::Run(p_a_grid,
+    gridwise_gemm.Run(p_a_grid,
                       p_b_grid,
                       p_c_grid,
                       a_grid_desc,
@@ -58,14 +59,10 @@ template <typename FloatA,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
-          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,
           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,
@@ -75,12 +72,19 @@ template <typename FloatA,
           >
 struct GridwiseGemmAvx2_MxN
 {
+    ck::tensor_operation::cpu::device::DeviceConvFwdDynamicTunable dynamic_tunable;
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
     // static constexpr auto Avx2RegisterVector = 8;   // 8 floats
     static constexpr index_t MemAlignmentByte = 32; // 256bit
 
+    GridwiseGemmAvx2_MxN(
+        const ck::tensor_operation::cpu::device::DeviceConvFwdDynamicTunable dynamic_tunable_)
+        : dynamic_tunable(dynamic_tunable_)
+    {
+    }
+
     static auto GetABlockDescriptor(const ck::index_t m_per_blk,
                                     const ck::index_t k_per_blk,
                                     const AGridDesc& a_grid_desc)
@@ -238,7 +242,7 @@ struct GridwiseGemmAvx2_MxN
         return ck::make_multi_index(i_m, i_n);
     }
 
-    static constexpr bool CheckValidity(const AGridDesc& a_grid_desc,
+    bool CheckValidity(const AGridDesc& a_grid_desc,
                        const BGridDesc& b_grid_desc,
                        const CGridDesc& c_grid_desc)
     {
@@ -247,7 +251,12 @@ struct GridwiseGemmAvx2_MxN
         const auto GemmN = c_grid_desc.GetLength(I1);
         if constexpr(UseCLocalBuffer)
         {
-            if(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value && NPerBlock < GemmN)
+            // if(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value &&
+            // dynamic_tunable.gemm_n_per_block < GemmN)
+            if(dynamic_tunable.loop_over_spec ==
+                   ck::tensor_operation::cpu::device::
+                       ConvolutionForwardBlockLoopOverSpecialization_t::LoopOver_MKN &&
+               dynamic_tunable.n_per_block < GemmN)
                 is_valid &= false;
         }
         else
@@ -259,7 +268,7 @@ struct GridwiseGemmAvx2_MxN
         return is_valid;
     }
 
-    static void Run(const FloatA* __restrict__ p_a_grid,
+    void Run(const FloatA* __restrict__ p_a_grid,
              const FloatB* __restrict__ p_b_grid,
              FloatC* __restrict__ p_c_grid,
              const AGridDesc& a_grid_desc,
@@ -267,11 +276,11 @@ struct GridwiseGemmAvx2_MxN
              const CGridDesc& c_grid_desc,
              const AElementwiseOperation& a_element_op,
              const BElementwiseOperation& b_element_op,
-                    const CElementwiseOperation& c_element_op)
+             const CElementwiseOperation& c_element_op) const
     {
-        ck::index_t m_per_block = MPerBlock;
-        ck::index_t n_per_block = NPerBlock;
-        ck::index_t k_per_block = KPerBlock;
+        ck::index_t m_per_block = dynamic_tunable.m_per_block;
+        ck::index_t n_per_block = dynamic_tunable.n_per_block;
+        ck::index_t k_per_block = dynamic_tunable.k_per_block;
 
         const auto GemmM = c_grid_desc.GetLength(I0);
         const auto GemmN = c_grid_desc.GetLength(I1);
@@ -297,7 +306,6 @@ struct GridwiseGemmAvx2_MxN
             decltype(GetABlockDescriptor(m_per_block, k_per_block, a_grid_desc)), // ABlockDesc,
             decltype(GetBBlockDescriptor(k_per_block, n_per_block, b_grid_desc)), // BBlockDesc,
             decltype(GetCBlockDescriptor(m_per_block, n_per_block, c_grid_desc)), // CBlockDesc,
-            KPerBlock,                                                            // KPerBlock,
             ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
             ThreadMNAccessOrder>{};  // ThreadMNAccessOrder  // how we acces
                                      // gemm MN to utilize micro kernel>{};
@@ -323,7 +331,9 @@ struct GridwiseGemmAvx2_MxN
 
         // TODO: openmp aware ordering
         //
-        if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 1, 2>>::value)
+        if(dynamic_tunable.loop_over_spec ==
+           ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
+               LoopOver_MNK)
         {
             auto a_move_k_step = GetAIndex(0, k_per_block);
             auto b_move_k_step = GetBIndex(k_per_block, 0);
@@ -467,7 +477,9 @@ struct GridwiseGemmAvx2_MxN
                 }
             }
         }
-        else if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value)
+        else if(dynamic_tunable.loop_over_spec ==
+                ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
+                    LoopOver_MKN)
         {
             auto a_move_k_step = GetAIndex(0, k_per_block);
             auto b_move_k_step = GetBIndex(0, n_per_block);

include/ck/tensor_operation/cpu/grid/gridwise_gemm_bias_activation_add_avx2.hpp

@@ -32,7 +32,8 @@ template <typename GridwiseGemm,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>
-void kernel_gemm_bias_activation_add_avx_mxn(const FloatA* __restrict__ p_a_grid,
+void kernel_gemm_bias_activation_add_avx_mxn(const GridwiseGemm& gridwise_gemm,
+                                             const FloatA* __restrict__ p_a_grid,
                                              const FloatB* __restrict__ p_b_grid,
                                              FloatC* __restrict__ p_c_grid,
                                              const FloatC0* __restrict__ p_c0_grid,
@@ -46,7 +47,7 @@ void kernel_gemm_bias_activation_add_avx_mxn(const FloatA* __restrict__ p_a_grid
                                              const BElementwiseOperation& b_element_op,
                                              const CElementwiseOperation& c_element_op)
 {
-    GridwiseGemm::Run(p_a_grid,
+    gridwise_gemm.Run(p_a_grid,
                       p_b_grid,
                       p_c_grid,
                       p_c0_grid,
@@ -74,14 +75,10 @@ template <typename FloatA,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
-          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,
           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,
@@ -91,12 +88,19 @@ template <typename FloatA,
           >
 struct GridwiseGemmBiasActivationAddAvx2_MxN
 {
+    ck::tensor_operation::cpu::device::DeviceConvFwdDynamicTunable dynamic_tunable;
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
     // static constexpr auto Avx2RegisterVector = 8;   // 8 floats
     static constexpr index_t MemAlignmentByte = 32; // 256bit
 
+    GridwiseGemmBiasActivationAddAvx2_MxN(
+        const ck::tensor_operation::cpu::device::DeviceConvFwdDynamicTunable dynamic_tunable_)
+        : dynamic_tunable(dynamic_tunable_)
+    {
+    }
+
     static auto GetABlockDescriptor(const ck::index_t m_per_blk,
                                     const ck::index_t k_per_blk,
                                     const AGridDesc& a_grid_desc)
@@ -254,7 +258,7 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
         return ck::make_multi_index(i_m, i_n);
     }
 
-    static constexpr bool CheckValidity(const AGridDesc& a_grid_desc,
+    bool CheckValidity(const AGridDesc& a_grid_desc,
                        const BGridDesc& b_grid_desc,
                        const CGridDesc& c_grid_desc)
     {
@@ -263,7 +267,11 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
         const auto GemmN = c_grid_desc.GetLength(I1);
         if constexpr(UseCLocalBuffer)
         {
-            if(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value && NPerBlock < GemmN)
+            if(dynamic_tunable.loop_over_spec ==
+                   ck::tensor_operation::cpu::device::
+                       ConvolutionForwardBlockLoopOverSpecialization_t::LoopOver_MKN &&
+               dynamic_tunable.n_per_block < GemmN)
+
                 is_valid &= false;
         }
         else
@@ -275,7 +283,7 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
         return is_valid;
     }
 
-    static void Run(const FloatA* __restrict__ p_a_grid,
+    void Run(const FloatA* __restrict__ p_a_grid,
              const FloatB* __restrict__ p_b_grid,
              FloatC* __restrict__ p_c_grid,
              const FloatC0* __restrict__ p_c0_grid,
@@ -287,11 +295,11 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
              const C1GridDesc& c1_grid_desc,
              const AElementwiseOperation& a_element_op,
              const BElementwiseOperation& b_element_op,
-                    const CElementwiseOperation& c_element_op)
+             const CElementwiseOperation& c_element_op) const
     {
-        ck::index_t m_per_block = MPerBlock;
-        ck::index_t n_per_block = NPerBlock;
-        ck::index_t k_per_block = KPerBlock;
+        ck::index_t m_per_block = dynamic_tunable.m_per_block;
+        ck::index_t n_per_block = dynamic_tunable.n_per_block;
+        ck::index_t k_per_block = dynamic_tunable.k_per_block;
 
         const auto GemmM = c_grid_desc.GetLength(I0);
         const auto GemmN = c_grid_desc.GetLength(I1);
@@ -323,7 +331,6 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
             decltype(GetABlockDescriptor(m_per_block, k_per_block, a_grid_desc)), // ABlockDesc,
             decltype(GetBBlockDescriptor(k_per_block, n_per_block, b_grid_desc)), // BBlockDesc,
             decltype(GetCBlockDescriptor(m_per_block, n_per_block, c_grid_desc)), // CBlockDesc,
-            KPerBlock,                                                            // KPerBlock,
             ThreadwiseGemm_Dispatch, // ThreadwiseGemm_Dispatch,
             ThreadMNAccessOrder>{};  // ThreadMNAccessOrder  // how we acces
                                      // gemm MN to utilize micro kernel>{};
@@ -349,7 +356,10 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
 
         // TODO: openmp aware ordering
         //
-        if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 1, 2>>::value)
+
+        if(dynamic_tunable.loop_over_spec ==
+           ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
+               LoopOver_MNK)
         {
             auto a_move_k_step = GetAIndex(0, k_per_block);
             auto b_move_k_step = GetBIndex(k_per_block, 0);
@@ -505,7 +515,9 @@ struct GridwiseGemmBiasActivationAddAvx2_MxN
                 }
             }
         }
-        else if constexpr(std::is_same<BlockMNKAccessOrder, ck::Sequence<0, 2, 1>>::value)
+        else if(dynamic_tunable.loop_over_spec ==
+                ck::tensor_operation::cpu::device::ConvolutionForwardBlockLoopOverSpecialization_t::
+                    LoopOver_MKN)
         {
             auto a_move_k_step = GetAIndex(0, k_per_block);
             auto b_move_k_step = GetBIndex(0, n_per_block);

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

 #include <stdlib.h>
+#include <utility>
 #include "convolution_forward_specialization_cpu.hpp"
 #include "config.hpp"
 #include "device_convnd_fwd_avx2_nhwc_kyxc_nhwk.hpp"
@@ -48,18 +49,24 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // 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, 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, true,  true,  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, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  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, true,  true,  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, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, true,  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, true,  true,  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, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  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, true,  true,  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, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, true,  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  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    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, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    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, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    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, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    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, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    \
+    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  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    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, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    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, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    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, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    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, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+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,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, false),
@@ -70,11 +77,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, true),
@@ -85,12 +98,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,   24,  24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,   32,  24, 256, 4, 24, false),
@@ -117,10 +135,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, false),
@@ -131,11 +156,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_relu_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, true),
@@ -146,12 +177,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_relu_instances = std::tu
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,   24, 24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,   32, 24, 256, 4, 24, false),
@@ -178,48 +214,9 @@ using device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_relu_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
-
-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{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_local_c_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_local_c_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_mt_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxc_nhwk_f32_mt_relu_instances{});
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_avx2_instance

library/src/tensor_operation_instance/cpu/conv2d_fwd/device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_instance.cpp

 #include <stdlib.h>
+#include <utility>
 #include "config.hpp"
 #include "convolution_forward_specialization_cpu.hpp"
 #include "device_convnd_fwd_avx2_nhwc_kyxck8_nhwk.hpp"
@@ -41,20 +42,25 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // clang-format off
 #define DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_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, c_local_buf) \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  false, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
     \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_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, true,  false, c_local_buf>
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, false),
@@ -65,11 +71,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, true),
@@ -80,12 +92,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances = std::tuple
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,   24,  24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,   32,  24, 256, 4, 24, false),
@@ -104,7 +121,6 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  120,  32, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  120,  64, 128, 6, 16, false),
 
-    // DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128,  64,  6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  128, 256, 128, 6, 16, true),
 
@@ -113,10 +129,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, false),
@@ -127,11 +150,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_relu_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, true),
@@ -142,12 +171,17 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_relu_instances = std::
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,   24,  24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,   32,  24, 256, 4, 24, false),
@@ -170,7 +204,6 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances = std::tuple
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  120,  32, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  120,  64, 128, 6, 16, false),
 
-    // DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, PT,  256, 128,  64,  6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  128, 256, 128, 6, 16, true),
 
@@ -179,49 +212,9 @@ using device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances = std::tuple
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_local_c_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_local_c_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_mt_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_kyxck8_nhwk_f32_mt_relu_instances{});
+            DEVICE_CONV2D_FWD_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_avx2_instance

library/src/tensor_operation_instance/cpu/conv2d_fwd/device_conv2d_fwd_avx2_nhwc_yxck_nhwk_instance.cpp

 #include <stdlib.h>
+#include <utility>
 #include "config.hpp"
 #include "convolution_forward_specialization_cpu.hpp"
 #include "device_convnd_fwd_avx2_nhwc_yxck_nhwk.hpp"
@@ -40,20 +41,24 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // clang-format off
 #define DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_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, c_local_buf) \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  false, c_local_buf>, \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
     \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  true,  c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf>, \
-    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_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, true,  false, c_local_buf>
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk(std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, false),
@@ -64,11 +69,17 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_local_c(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  256, 128, 128, 6, 16, true),
@@ -79,12 +90,17 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_mt(
+    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,   24,  24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,   32,  24, 256, 4, 24, false),
@@ -112,10 +128,17 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, PT, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, false),
@@ -126,11 +149,17 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_relu_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, false),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_local_c_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  256, 128,  64, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  256, 128, 128, 6, 16, true),
@@ -141,12 +170,17 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_relu_instances = std::tu
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_relu_instances = std::tuple<
+void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_mt_relu(
+    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,   24,  24, 256, 4, 24, false),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,   32,  24, 256, 4, 24, false),
@@ -178,48 +212,9 @@ using device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_relu_instances = std::tuple<
 
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  768, 320, 128, 6, 16, true),
             DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu,  896, 352, 128, 6, 16, true),
-    DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128,  6, 16, true)>;
-// clang-format on
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk(std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_local_c(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_mt(
-    std::vector<DeviceConvFwdPtr<PT, PT, PT>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_local_c_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_local_c_relu_instances{});
-}
-
-void add_device_conv2d_fwd_avx2_nhwc_yxck_nhwk_mt_relu(
-    std::vector<DeviceConvFwdPtr<PT, PT, Relu>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_avx2_nhwc_yxck_nhwk_f32_mt_relu_instances{});
+            DEVICE_CONV2D_FWD_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, Relu, 1024, 416, 128, 6, 16, true)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_avx2_instance

library/src/tensor_operation_instance/cpu/conv2d_fwd_bias_activation_add/device_conv2d_bias_activation_add_avx2_nhwc_kyxc_nhwk_instance.cpp

 #include <stdlib.h>
+#include <utility>
 #include "convolution_forward_specialization_cpu.hpp"
 #include "config.hpp"
 #include "device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk.hpp"
@@ -41,18 +42,25 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // clang-format off
 #define DEVICE_CONV2D_FWD_BAA_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, c_local_buf, bias_along_m) \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, true , c_local_buf, bias_along_m>, \
-\
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, true , c_local_buf, bias_along_m>
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, false, false),
@@ -63,11 +71,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_instances =
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, false, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_local_c_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, true, false),
@@ -78,12 +92,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_local_c_inst
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  24,  24, 256,  4, 24, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  32,  24, 256,  4, 24, false, false),
@@ -110,29 +129,9 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_mt_instances
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_local_c(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances,
-        device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_mt(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxc_nhwk_f32_mt_instances{});
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXC_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_bias_activation_add_avx2_instance

library/src/tensor_operation_instance/cpu/conv2d_fwd_bias_activation_add/device_conv2d_bias_activation_add_avx2_nhwc_kyxck8_nhwk_instance.cpp

 #include <stdlib.h>
+#include <utility>
 #include "config.hpp"
 #include "convolution_forward_specialization_cpu.hpp"
 #include "device_convnd_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk.hpp"
@@ -41,21 +42,25 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // clang-format off
 #define DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_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, c_local_buf, bias_along_m) \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , false, c_local_buf, bias_along_m>, \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
     \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , 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, true , false, c_local_buf, bias_along_m>
-
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_K_Y_X_C_K8_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, false, false),
@@ -66,12 +71,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_instances
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, false, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances =
-    std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_local_c(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, true, false),
@@ -82,12 +92,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_local_c_in
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_mt(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  24,  24, 256,  4, 24, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  32,  24, 256,  4, 24, false, false),
@@ -114,29 +129,9 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_mt_instanc
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_local_c(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances,
-        device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_mt(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_kyxck8_nhwk_f32_mt_instances{});
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_KYXCK8_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_bias_activation_add_avx2_instance

library/src/tensor_operation_instance/cpu/conv2d_fwd_bias_activation_add/device_conv2d_bias_activation_add_avx2_nhwc_yxck_nhwk_instance.cpp

 #include <stdlib.h>
+#include <utility>
 #include "config.hpp"
 #include "convolution_forward_specialization_cpu.hpp"
 #include "device_convnd_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk.hpp"
@@ -40,21 +41,25 @@ static constexpr auto LoopOver_MKN = ck::tensor_operation::cpu::device::LoopOver
 
 // clang-format off
 #define DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_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, c_local_buf, bias_along_m) \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MNK, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , false, c_local_buf, bias_along_m>, \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MNK}), \
     \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , true , c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float , float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , LoopOver_MKN, 2, m_per_block, n_per_block, k_per_block, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>, \
-    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , 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, true , false, c_local_buf, bias_along_m>
-
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  true,  c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwd1x1S1P0, GemmKLoopOverC  , 2, m_per_thread, n_per_thread, false, false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN}), \
+    DeviceConvNDFwdBiasActivationAddAvx2_Input_N_Hi_Wi_C_Weight_Y_X_C_K_Output_N_Ho_Wo_K<float , float , float, float, float, a_elem_op, b_elem_op, c_elem_op, ConvFwdDefault, DefaultGemmKLoop, 2, m_per_thread, n_per_thread, true,  false, c_local_buf, bias_along_m>({m_per_block, n_per_block, k_per_block, LoopOver_MKN})
 // clang-format on
 
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, false, false),
@@ -65,11 +70,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_instances =
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, false, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, false, false)
+            // clang-format on
+            ));
+}
 
-// use this in single thread, but gemm_n is not multiple of 8
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_local_c_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_local_c(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  256, 128,  64,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  256, 128, 128,  6, 16, true, false),
@@ -80,12 +91,17 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_local_c_inst
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
+}
 
-// use this in multi thread environment (need local C buffer to avoid cache coherence, although some
-// time no local c is better...)
-using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_mt_instances = std::tuple<
+void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_mt(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
+{
+    ck::tensor_operation::device::add_device_operation_instances(
+        instances,
+        std::make_tuple(
             // clang-format off
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  24,  24, 256,  4, 24, false, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  32,  24, 256,  4, 24, false, false),
@@ -112,29 +128,9 @@ using device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_mt_instances
 
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  768, 320, 128,  6, 16, true, false),
             DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd,  896, 352, 128,  6, 16, true, false),
-    DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)>;
-// clang-format on
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_local_c(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances,
-        device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_local_c_instances{});
-}
-
-void add_device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_mt(
-    std::vector<DeviceConvFwdBiasActivationAddPtr<PT, PT, AddReluAdd>>& instances)
-{
-    ck::tensor_operation::device::add_device_operation_instances(
-        instances, device_conv2d_fwd_bias_activation_add_avx2_nhwc_yxck_nhwk_f32_mt_instances{});
+            DEVICE_CONV2D_FWD_BAA_AVX2_NHWC_YXCK_NHWK_F32(PT, PT, AddReluAdd, 1024, 416, 128,  6, 16, true, false)
+            // clang-format on
+            ));
 }
 
 } // namespace device_conv2d_fwd_bias_activation_add_avx2_instance