add nchw atomic , nhwc and nhwc atomic method for backward weight (#30)

* add add new algorithm from v4r4r2

* program once issue

* add split k functiion

* redefine code

* add a matrix unmerge

* add b matrix unmerge k0

* trans a and b to gridegemm

* nhwc init

* no hacks and vector load

* add hacks

* modify some parameter

* fix tuning prometer for fp32

* fix tuning prometer for fp16

* start change gridwise k split

* init ok

* revome a b matrix k0mk1 desc in grid

* carewrite lculate gridsize

* add kbatch to CalculateBottomIndex

* remove some unused funtion

* add clear data function before call kernel

* out hacks

* in hacks

* rename device convolution file and function name

* modify kBatch value

* fix some tuning code

* start from v4r4 nhwc

* nhwc atomic is able to run

* just for fp32

* enable nchw atomic

* tweak

* tweak

* re-arrange gridwise gemm hot loop for wrw

* add wrw v4r5

* v4r4r5 fp16

* v4r4r4 fp16

* v4r4r2 fp16

* V4R4R4XDLNHWC fp16

* V4R4R2XDLATOMICNCHW fp16

* adjust for fp16

* input gridsize

* change kbatch to gridsize

* testing wrw

* clean up

* k_batch to gridsize

* fix bug

* wrw v4r4r4 kbatch change to gride size

* wrw v4r4r2 kbatch change to gride size

* after merge , change gridwise gemm v2r4

* change MakeCBlockClusterAdaptor

* other method use new gridwise gemm

* clean up

* chapad method nge to make_right_pad_transform

* kbatch out from transform function

* clean up and fix bug

* fix bug

* using function type reduce template parameters

* using auto replace define fuction type

* clean up
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: Jing Zhang <jizhan@amd.com>

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmK = N * Ho * Wo
+// GemmN = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = C * Y * X;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = Y * X * C;
+    const auto GemmN      = K;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: output tensor
+    const auto out_gemmktotal_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = Y * X * C;
+    const auto GemmN  = K;
+    const auto GemmK  = N * Ho * Wo;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: output tensor
+    const auto out_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(out_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: out
+// B: in
+// C: wei
+// GemmM = K
+// GemmN = Y * X * C
+// GemmKTotal = N * Ho * Wo
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = Y * X * C;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: input tensor
+    const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+        in_n_hi_wi_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r4.hpp

+#ifndef CK_GRIDWISE_GEMM_XDLOPS_V2R4_HPP
+#define CK_GRIDWISE_GEMM_XDLOPS_V2R4_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_gemm_xdlops.hpp"
+#include "blockwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_tensor_slice_set.hpp"
+
+namespace ck {
+
+#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CM0N0M1N1M2M3M4N2GridDesc,
+          typename CBlockClusterAdaptor>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4(const FloatAB* __restrict__ p_a_grid,
+                                const FloatAB* __restrict__ p_b_grid,
+                                FloatC* __restrict__ p_c_grid,
+                                const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
+                                const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
+                                const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_b_k0_m_k1_grid_desc,
+                      b_b_k0_n_k1_grid_desc,
+                      c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                      c_block_cluster_adaptor);
+}
+#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CM0N0M1N1M2M3M4N2GridDesc,
+          typename CBlockClusterAdaptor>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4(const FloatAB* __restrict__ p_a_grid,
+                                const FloatAB* __restrict__ p_b_grid,
+                                FloatC* __restrict__ p_c_grid,
+                                const void CONSTANT* p_a_b_k0_m_k1_grid_desc,
+                                const void CONSTANT* p_b_b_k0_n_k1_grid_desc,
+                                const void CONSTANT* p_c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                const void CONSTANT* p_c_block_cluster_adaptor)
+{
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    const auto a_b_k0_m_k1_grid_desc = *reinterpret_cast<const ABK0MK1GridDesc*>(
+        cast_pointer_to_generic_address_space(p_a_b_k0_m_k1_grid_desc));
+    const auto b_b_k0_n_k1_grid_desc = *reinterpret_cast<const BBK0NK1GridDesc*>(
+        cast_pointer_to_generic_address_space(p_b_b_k0_n_k1_grid_desc));
+    const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
+        *reinterpret_cast<const CM0N0M1N1M2M3M4N2GridDesc*>(
+            cast_pointer_to_generic_address_space(p_c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc));
+    const auto c_block_cluster_adaptor = *reinterpret_cast<const CBlockClusterAdaptor*>(
+        cast_pointer_to_generic_address_space(p_c_block_cluster_adaptor));
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_b_k0_m_k1_grid_desc,
+                      b_b_k0_n_k1_grid_desc,
+                      c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                      c_block_cluster_adaptor);
+}
+#endif
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CMNGridDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          typename AGridStepHacks,
+          typename BGridStepHacks,
+          typename CGridStepHacks,
+          typename AGridMoveSliceWindowStepHacks,
+          typename BGridMoveSliceWindowStepHacks,
+          bool CAccessOrderMRepeatNRepeat,
+          bool ABlockLdsExtraM,
+          bool BBlockLdsExtraN>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                  const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % KPerBlock == 0))
+            return false;
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CMNGridDesc& c_m_n_grid_desc, index_t KBatch)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock) * KBatch;
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto kbatch_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(KBatch),
+                           make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
+
+        const auto c_blockid_to_kbatch_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(KBatch, M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_blockid_to_kbatch_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(kbatch_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  c_blockid_to_kbatch_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return c_blockid_to_kbatch_m0_n0_block_cluster_adaptor;
+    }
+
+    using CM0N0M1N1M2M3M4N2GridDesc = decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{}));
+    using CBlockClusterAdaptor      = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               FloatAB* __restrict__ p_shared_block,
+                               const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                               const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                               const CM0N0M1N1M2M3M4N2GridDesc& c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t k_batch_id = block_work_idx[I0];
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<KPerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<KPerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                               Number<MPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<KPerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<KPerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<KPerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                               Number<NPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<KPerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum_t::Set,
+                                            Sequence<1, KPerBlock, MPerBlock, K1>,
+                                            ABlockTransferThreadSliceLengths_K0_M_K1,
+                                            ABlockTransferThreadClusterLengths_K0_M_K1,
+                                            ABlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(a_b_k0_m_k1_grid_desc),
+                                            decltype(a_b_k0_m_k1_block_desc),
+                                            ABlockTransferSrcAccessOrder,
+                                            Sequence<0, 2, 1, 3>,
+                                            ABlockTransferSrcVectorDim,
+                                            3,
+                                            ABlockTransferSrcScalarPerVector,
+                                            ABlockTransferDstScalarPerVector_K1,
+                                            1,
+                                            1,
+                                            AThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                            InMemoryDataOperationEnum_t::Set,
+                                            Sequence<1, KPerBlock, NPerBlock, K1>,
+                                            BBlockTransferThreadSliceLengths_K0_N_K1,
+                                            BBlockTransferThreadClusterLengths_K0_N_K1,
+                                            BBlockTransferThreadClusterArrangeOrder,
+                                            FloatAB,
+                                            FloatAB,
+                                            decltype(b_b_k0_n_k1_grid_desc),
+                                            decltype(b_b_k0_n_k1_block_desc),
+                                            BBlockTransferSrcAccessOrder,
+                                            Sequence<0, 2, 1, 3>,
+                                            BBlockTransferSrcVectorDim,
+                                            3,
+                                            BBlockTransferSrcScalarPerVector,
+                                            BBlockTransferDstScalarPerVector_K1,
+                                            1,
+                                            1,
+                                            BThreadTransferSrcResetCoordinateAfterRun,
+                                            true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[KPerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = p_shared_block;
+        FloatAB* p_b_block = p_shared_block + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock, 0, 0);
+
+        // hack to control index calculation when iterating over A and B matrix for threadwise copy
+        constexpr auto a_k0_m_k1_grid_step_hacks = AGridStepHacks{};
+        constexpr auto b_k0_n_k1_grid_step_hacks = BGridStepHacks{};
+
+        // hack to control index calculation when move slice window for A and B matrix for
+        // threadwise copy
+        constexpr auto a_k0_m_k1_grid_move_slice_window_step_hack = AGridMoveSliceWindowStepHacks{};
+        constexpr auto b_k0_n_k1_grid_move_slice_window_step_hack = BGridMoveSliceWindowStepHacks{};
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf, a_k0_m_k1_grid_step_hacks);
+            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf, b_k0_n_k1_grid_step_hacks);
+
+            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+        }
+
+        // main body
+        index_t k_block_data_begin = 0;
+
+        do
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc,
+                                                a_block_slice_copy_step,
+                                                a_k0_m_k1_grid_move_slice_window_step_hack);
+            b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc,
+                                                b_block_slice_copy_step,
+                                                b_k0_n_k1_grid_move_slice_window_step_hack);
+
+            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf, a_k0_m_k1_grid_step_hacks);
+
+            block_sync_lds();
+
+            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf, b_k0_n_k1_grid_step_hacks);
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            block_sync_lds();
+
+            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+
+            k_block_data_begin += KPerBlock;
+        } while(k_block_data_begin < (K0 - KPerBlock));
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCM0N0M1N1M2M3M4N2BlockDescriptor();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(
+                    Number<M0>{}, Number<N0>{}, I1, I1, Number<M2>{}, I1, Number<M4>{}, I1));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks = CGridStepHacks{};
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+
+                    c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2])};
+
+            c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                              c_grid_buf,
+                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks);
+        }
+    }
+}; // namespace ck
+
+} // namespace ck
+#endif

host/driver_offline/include/device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk.hpp

@@ -3,6 +3,7 @@
 #include "host_tensor.hpp"
 #include "transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp"
 #include "driver_gemm_xdlops_v2r3.hpp"
+#include "debug.hpp"
 
 template <typename TInWei,
           typename TAcc,
@@ -275,8 +276,8 @@ void device_convolution_backward_data_implicit_gemm_v4r1_xdlops_nhwc_kyxc_nhwk(
               wei_gemmk0_gemmm_gemmk1_grid_desc,
               out_gemmk0_gemmn_gemmk1_grid_desc,
               in_gemmm_gemmn_grid_desc,
-              debug_driver_gemm_xdlops_v2r3::M01,
-              debug_driver_gemm_xdlops_v2r3::N01,
+              debug::debug_driver_gemm_xdlops_v2r3::M01,
+              debug::debug_driver_gemm_xdlops_v2r3::N01,
               wei_gemmk0_gemmm_gemmk1_grid_step_hacks,
               out_gemmk0_gemmn_gemmk1_grid_step_hacks,
               in_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,

host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw.hpp

+#include <unistd.h>
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp"
+#include "driver_gemm_xdlops_v2r4.hpp"
+
+template <typename TIn,
+          typename TWei,
+          typename TAcc,
+          typename TOut,
+          typename InLengths,
+          typename WeiLengths,
+          typename OutLengths,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          typename GridSizeType>
+void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw(
+    const InLengths& in_n_c_hi_wi_lengths,
+    const WeiLengths& wei_k_c_y_x_lengths,
+    const OutLengths& out_n_k_ho_wo_lengths,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const Tensor<TIn>& in_n_c_hi_wi,
+    Tensor<TWei>& wei_k_c_y_x,
+    const Tensor<TOut>& out_n_k_ho_wo,
+    GridSizeType desired_grid_size,
+    ck::index_t nrepeat)
+{
+    using namespace ck;
+
+    std::cout << __func__ << std::endl;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    DeviceMem in_n_c_hi_wi_device_buf(sizeof(TIn) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_k_c_y_x_device_buf(sizeof(TWei) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
+
+    in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    out_n_k_ho_wo_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
+
+    const auto in_n_c_hi_wi_desc  = make_naive_tensor_descriptor_packed(in_n_c_hi_wi_lengths);
+    const auto wei_k_c_y_x_desc   = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
+    const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
+
+#if 1
+    // [M, N, K0, K1] = [128, 128, 4, 8] for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerWave = 32;
+    constexpr index_t GemmNPerWave = 32;
+    constexpr index_t GemmK1       = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmB_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 2, 8>;
+    using GemmABlockTransferThreadClusterLengths_GemmB_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 64, 1>;
+    // using vector load 4, so config's wo*ho  must be a multiple of 4
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmK1 = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmB_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 2, 8>;
+    using GemmBBlockTransferThreadClusterLengths_GemmB_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 64, 1>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 1;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#endif
+
+    const auto N = in_n_c_hi_wi_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_desc.GetLength(I1);
+
+    const auto Ho = out_n_k_ho_wo_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_desc.GetLength(I3);
+
+    const auto GemmM      = K;
+    const auto GemmN      = Y * X * C;
+    const auto GemmKTotal = N * Ho * Wo;
+
+    const auto GemmK = GemmKTotal / GemmK1;
+
+    const auto GridMN        = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
+    const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
+    const index_t BatchLen   = std::ceil(GemmK * 1.0 / (GemmKPerBlock * GemmKBatch));
+    const index_t GemmK0     = BatchLen * GemmKPerBlock;
+    const index_t GemmKPad   = GemmKBatch * GemmK0 * GemmK1;
+
+    std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
+              << " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
+              << std::endl;
+    const auto descs =
+        transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
+            wei_k_c_y_x_desc,
+            in_n_c_hi_wi_desc,
+            out_n_k_ho_wo_desc,
+            conv_strides,
+            conv_dilations,
+            in_left_pads,
+            in_right_pads,
+            Number<GemmK1>{},
+            GemmKBatch,
+            GemmKPad);
+
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
+    const auto in_gemmk0_gemmn_gemmk1_grid_desc  = descs[I1];
+    const auto wei_gemmm_gemmn_grid_desc         = descs[I2];
+
+    // HACK: hacks that control index calculation when iterating over A, B, C matrix
+    constexpr auto out_gemmk0_gemmm_gemmk1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 1, 0, 0, 0, 0>{},   // 0+: GemmB
+                              Sequence<0, 0, 1, 0, 0, 0, 0>{},   // 1+: GemmK0
+                              Sequence<0, 0, 0, 0, 0, 0, 0>{},   // 2+: GemmM
+                              Sequence<0, 0, 1, 0, 0, 0, 0>{}),  // 3+: GemmK1
+                   make_tuple(Sequence<0, 0, 2, 0, 0, 0, 0>{},   // 0-: GemB
+                              Sequence<0, 0, 2, 0, 0, 0, 0>{},   // 1-: GemmK0
+                              Sequence<0, 0, 0, 0, 0, 0, 0>{},   // 2-: GemmM
+                              Sequence<0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
+
+    constexpr auto in_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 0+: GemmB
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 1+: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 2+: GemmN
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}),  // 3+: GemmK1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 0-: GemmB
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 1-: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 2-: GemmN
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
+
+    constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    constexpr auto out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 1, 0, 0, 0, 0>{};
+
+    constexpr auto in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
+
+    const auto driver_gemm_xdlops =
+        driver_gemm_xdlops_v2r4<BlockSize,
+                                TIn,
+                                TAcc,
+                                TWei,
+                                InMemoryDataOperationEnum_t::AtomicAdd,
+                                decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
+                                decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
+                                decltype(wei_gemmm_gemmn_grid_desc),
+                                GemmMPerBlock,
+                                GemmNPerBlock,
+                                GemmKPerBlock,
+                                GemmMPerWave,
+                                GemmNPerWave,
+                                GemmK1,
+                                MRepeat,
+                                NRepeat,
+                                GemmABlockTransferThreadSliceLengths_GemmB_GemmK0_GemmM_GemmK1,
+                                GemmABlockTransferThreadClusterLengths_GemmB_GemmK0_GemmM_GemmK1,
+                                Sequence<0, 2, 1, 3>,
+                                Sequence<0, 2, 1, 3>,
+                                3,
+                                GemmABlockTransferSrcScalarPerVector_GemmK1,
+                                GemmABlockTransferDstScalarPerVector_GemmK1,
+                                false, // don't move back src coordinate after threadwise copy
+                                GemmBBlockTransferThreadSliceLengths_GemmB_GemmK0_GemmN_GemmK1,
+                                GemmBBlockTransferThreadClusterLengths_GemmB_GemmK0_GemmN_GemmK1,
+                                Sequence<0, 2, 1, 3>,
+                                Sequence<0, 2, 1, 3>,
+                                3,
+                                GemmBBlockTransferSrcScalarPerVector_GemmN,
+                                GemmBBlockTransferDstScalarPerVector_GemmK1,
+                                false, // don't move back src coordinate after threadwise copy
+                                Sequence<3, 0, 1, 2, 7, 5, 4, 6>,
+                                7,
+                                GemmCThreadTransferDstScalarPerVector,
+                                decltype(out_gemmk0_gemmm_gemmk1_grid_step_hacks),
+                                decltype(in_gemmk0_gemmn_gemmk1_grid_step_hacks),
+                                decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
+                                decltype(out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
+                                decltype(in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
+                                false,
+                                true,
+                                true>;
+
+    for(index_t i = 0; i < 5; ++i)
+    {
+        float ave_time =
+            driver_gemm_xdlops(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
+                               static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
+                               static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
+                               out_gemmk0_gemmm_gemmk1_grid_desc,
+                               in_gemmk0_gemmn_gemmk1_grid_desc,
+                               wei_gemmm_gemmn_grid_desc,
+                               debug::debug_driver_gemm_xdlops_v2r3::M01,
+                               debug::debug_driver_gemm_xdlops_v2r3::N01,
+                               out_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                               in_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                               wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                               out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                               in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                               nrepeat);
+
+        float perf = static_cast<float>(calculate_convolution_flops(
+                         in_n_c_hi_wi_desc, wei_k_c_y_x_desc, out_n_k_ho_wo_desc)) /
+                     (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+        std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
+    }
+
+    wei_k_c_y_x_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    driver_gemm_xdlops(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
+                       static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
+                       static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
+                       out_gemmk0_gemmm_gemmk1_grid_desc,
+                       in_gemmk0_gemmn_gemmk1_grid_desc,
+                       wei_gemmm_gemmn_grid_desc,
+                       debug::debug_driver_gemm_xdlops_v2r3::M01,
+                       debug::debug_driver_gemm_xdlops_v2r3::N01,
+                       out_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                       in_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                       wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                       out_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                       in_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                       0);
+    // copy result back to host
+    wei_k_c_y_x_device_buf.FromDevice(wei_k_c_y_x.mData.data());
+}

host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp

@@ -4,7 +4,8 @@
 #include "transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp"
 #include "driver_gemm_xdlops_v2r3.hpp"
 
-template <typename TInWei,
+template <typename TIn,
+          typename TWei,
           typename TAcc,
           typename TOut,
           typename InLengths,
@@ -22,8 +23,8 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
     const ConvDilations& conv_dilations,
     const InLeftPads& in_left_pads,
     const InRightPads& in_right_pads,
-    const Tensor<TInWei>& in_n_c_hi_wi,
-    Tensor<TInWei>& wei_k_c_y_x,
+    const Tensor<TIn>& in_n_c_hi_wi,
+    Tensor<TWei>& wei_k_c_y_x,
     const Tensor<TOut>& out_n_k_ho_wo,
     ck::index_t nrepeat)
 {
@@ -35,8 +36,8 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
     constexpr auto I1 = Number<1>{};
     constexpr auto I2 = Number<2>{};
 
-    DeviceMem in_n_c_hi_wi_device_buf(sizeof(TInWei) * in_n_c_hi_wi.mDesc.GetElementSpace());
-    DeviceMem wei_k_c_y_x_device_buf(sizeof(TInWei) * wei_k_c_y_x.mDesc.GetElementSpace());
+    DeviceMem in_n_c_hi_wi_device_buf(sizeof(TIn) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_k_c_y_x_device_buf(sizeof(TWei) * wei_k_c_y_x.mDesc.GetElementSpace());
     DeviceMem out_n_k_ho_wo_device_buf(sizeof(TOut) * out_n_k_ho_wo.mDesc.GetElementSpace());
 
     in_n_c_hi_wi_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
@@ -47,7 +48,7 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
     const auto wei_k_c_y_x_desc   = make_naive_tensor_descriptor_packed(wei_k_c_y_x_lengths);
     const auto out_n_k_ho_wo_desc = make_naive_tensor_descriptor_packed(out_n_k_ho_wo_lengths);
 
-#if 1
+#if 0
     // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
     constexpr index_t BlockSize = 256;
 
@@ -164,9 +165,9 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
     {
         float ave_time = driver_gemm_xdlops_v2r3<
             BlockSize,
-            TInWei,
+            TIn,
             TAcc,
-            TOut,
+            TWei,
             InMemoryDataOperationEnum_t::Set,
             decltype(out_gemmk0_gemmm_gemmk1_grid_desc),
             decltype(in_gemmk0_gemmn_gemmk1_grid_desc),
@@ -207,8 +208,8 @@ void device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nk
             true,  // ABlockLdsExtraM
             true   // BBlockLdsExtraN
             >(static_cast<TOut*>(out_n_k_ho_wo_device_buf.GetDeviceBuffer()),
-              static_cast<TInWei*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
-              static_cast<TInWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
+              static_cast<TIn*>(in_n_c_hi_wi_device_buf.GetDeviceBuffer()),
+              static_cast<TWei*>(wei_k_c_y_x_device_buf.GetDeviceBuffer()),
               out_gemmk0_gemmm_gemmk1_grid_desc,
               in_gemmk0_gemmn_gemmk1_grid_desc,
               wei_gemmm_gemmn_grid_desc,

host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk.hpp

+#include <unistd.h>
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp"
+#include "driver_gemm_xdlops_v2r4.hpp"
+
+template <typename TIn,
+          typename TWei,
+          typename TAcc,
+          typename TOut,
+          typename InLengths,
+          typename WeiLengths,
+          typename OutLengths,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          typename GridSizeType>
+void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk(
+    const InLengths& in_n_hi_wi_c_lengths,
+    const WeiLengths& wei_k_y_x_c_lengths,
+    const OutLengths& out_n_ho_wo_k_lengths,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const Tensor<TIn>& in_n_hi_wi_c,
+    Tensor<TWei>& wei_k_y_x_c,
+    const Tensor<TOut>& out_n_ho_wo_k,
+    GridSizeType desired_grid_size,
+    ck::index_t nrepeat)
+{
+    using namespace ck;
+
+    std::cout << __func__ << std::endl;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
+    DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
+    DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
+
+    in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
+    wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
+    out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
+
+    const auto in_n_hi_wi_c_desc  = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
+    const auto wei_k_y_x_c_desc   = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
+    const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
+
+#if 0
+    // [M, N, K0, K1] = [128, 256, 4, 4] for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 256;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 4;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 4;
+#elif 1
+    // [M, N, K0, K1] = [128, 128, 4, 4] for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 4;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#endif
+
+    const auto N = in_n_hi_wi_c_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_desc.GetLength(I3);
+
+    const auto Ho = out_n_ho_wo_k_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_desc.GetLength(I2);
+
+    const auto GemmM      = Y * X * C;
+    const auto GemmN      = K;
+    const auto GemmKTotal = N * Ho * Wo;
+
+    const auto GemmK = GemmKTotal / GemmK1;
+
+    const auto GridMN        = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
+    const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
+    const index_t BatchLen   = std::ceil(GemmK * 1.0 / (GemmKPerBlock * GemmKBatch));
+    const index_t GemmK0     = BatchLen * GemmKPerBlock;
+    const index_t GemmKPad   = GemmKBatch * GemmK0 * GemmK1;
+
+    std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
+              << " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
+              << std::endl;
+
+    const auto descs =
+        transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
+            in_n_hi_wi_c_desc,
+            wei_k_y_x_c_desc,
+            out_n_ho_wo_k_desc,
+            conv_strides,
+            conv_dilations,
+            in_left_pads,
+            in_right_pads,
+            Number<GemmK1>{},
+            GemmKBatch,
+            GemmKPad);
+
+    const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc  = descs[I0];
+    const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
+    const auto wei_gemmm_gemmn_grid_desc                    = descs[I2];
+
+    // HACK: hacks that control index calculation when iterating over A, B, C matrix
+    constexpr auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 0+: GemmKBatch
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 1+: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 2+: GemmM
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}),  // 3+: GemmK1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 0-: GemmKBatch
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 1-: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 2-: GemmM
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 3-: GemmK1
+
+    constexpr auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: GemmN
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: GemmK1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 0-: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: GemmN
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
+
+    constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    constexpr auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
+
+    constexpr auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0>{};
+
+    const auto driver_gemm_xdlops = driver_gemm_xdlops_v2r4<
+        BlockSize,
+        TIn,
+        TAcc,
+        TWei,
+        InMemoryDataOperationEnum_t::AtomicAdd,
+        decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc),
+        decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc),
+        decltype(wei_gemmm_gemmn_grid_desc),
+        GemmMPerBlock,
+        GemmNPerBlock,
+        GemmKPerBlock,
+        GemmMPerXDL,
+        GemmNPerXDL,
+        GemmK1,
+        MRepeat,
+        NRepeat,
+        GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
+        GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
+        Sequence<0, 1, 2, 3>,
+        Sequence<0, 1, 2, 3>,
+        2,
+        GemmABlockTransferSrcScalarPerVector_GemmM,
+        GemmABlockTransferDstScalarPerVector_GemmK1,
+        false, // don't move back src coordinate after threadwise copy
+        GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
+        GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
+        Sequence<0, 1, 2, 3>,
+        Sequence<0, 1, 2, 3>,
+        2,
+        GemmBBlockTransferSrcScalarPerVector_GemmN,
+        GemmBBlockTransferDstScalarPerVector_GemmK1,
+        false, // don't move back src coordinate after threadwise copy
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        6,
+        GemmCThreadTransferDstScalarPerVector,
+        decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks),
+        decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks),
+        decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
+        decltype(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
+        decltype(out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
+        false, // CAccessOrderMRepeatNRepeat
+        true,
+        true>;
+
+    for(index_t i = 0; i < 5; ++i)
+    {
+        float ave_time =
+            driver_gemm_xdlops(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
+                               static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
+                               static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
+                               in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                               out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                               wei_gemmm_gemmn_grid_desc,
+                               debug::debug_driver_gemm_xdlops_v2r3::M01,
+                               debug::debug_driver_gemm_xdlops_v2r3::N01,
+                               in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                               out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                               wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                               in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                               out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                               nrepeat);
+
+        {
+
+            float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
+                         (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+            std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
+                      << std::endl;
+        }
+    }
+
+    wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
+    driver_gemm_xdlops(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
+                       static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
+                       static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
+                       in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                       out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                       wei_gemmm_gemmn_grid_desc,
+                       debug::debug_driver_gemm_xdlops_v2r3::M01,
+                       debug::debug_driver_gemm_xdlops_v2r3::N01,
+                       in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                       out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                       wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                       in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                       out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                       0);
+    // copy result back to host
+    wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
+}

host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp

+#include <unistd.h>
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp"
+#include "driver_gemm_xdlops_v2r3.hpp"
+#include "debug.hpp"
+
+template <typename TIn,
+          typename TWei,
+          typename TAcc,
+          typename TOut,
+          typename InLengths,
+          typename WeiLengths,
+          typename OutLengths,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads>
+void device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk(
+    const InLengths& in_n_hi_wi_c_lengths,
+    const WeiLengths& wei_k_y_x_c_lengths,
+    const OutLengths& out_n_ho_wo_k_lengths,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const Tensor<TIn>& in_n_hi_wi_c,
+    Tensor<TWei>& wei_k_y_x_c,
+    const Tensor<TOut>& out_n_ho_wo_k,
+    ck::index_t nrepeat)
+{
+    using namespace ck;
+
+    std::cout << __func__ << std::endl;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
+    DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
+    DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
+
+    in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
+    wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
+    out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
+
+    const auto in_n_hi_wi_c_desc  = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
+    const auto wei_k_y_x_c_desc   = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
+    const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
+
+#if 0
+    // [M, N, K0, K1] = [256, 128, 4, 4] for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 256;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1       = 4;
+
+    constexpr index_t MRepeat = 4;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 4, 4>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 64, 1>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM = 2;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 2, 4>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 64, 1>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN = 2;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [128, 128, 4, 4] for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 4, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 4, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 4;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+
+#elif 0
+    // [M, N, K0, K1] = [128, 128, 4, 8] for fp16
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 4, 4>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 4;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 4, 4>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 4;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#endif
+
+    const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
+        in_n_hi_wi_c_desc,
+        wei_k_y_x_c_desc,
+        out_n_ho_wo_k_desc,
+        conv_strides,
+        conv_dilations,
+        in_left_pads,
+        in_right_pads,
+        Number<GemmK1>{});
+
+    const auto in_gemmk0_gemmm_gemmk1_grid_desc  = descs[I0];
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc = descs[I1];
+    const auto wei_gemmm_gemmn_grid_desc         = descs[I2];
+
+    // HACK: hacks that control index calculation when iterating over A, B, C matrix
+    constexpr auto in_gemmk0_gemmm_gemmk1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0>{},   // 1+: GemmM
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0>{}),  // 2+: GemmK1
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{},   // 0-: GemmK0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0>{},   // 1-: GemmM
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0>{})); // 2-: GemmK1
+
+    constexpr auto out_gemmk0_gemmn_gemmk1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: GemmN
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: GemmK1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0-: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: GemmN
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
+
+    constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    constexpr auto in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0>{};
+
+    constexpr auto out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0>{};
+
+    for(index_t i = 0; i < 5; ++i)
+    {
+        float ave_time = driver_gemm_xdlops_v2r3<
+            BlockSize,
+            TIn,
+            TAcc,
+            TWei,
+            InMemoryDataOperationEnum_t::Set,
+            decltype(in_gemmk0_gemmm_gemmk1_grid_desc),
+            decltype(out_gemmk0_gemmn_gemmk1_grid_desc),
+            decltype(wei_gemmm_gemmn_grid_desc),
+            GemmMPerBlock,
+            GemmNPerBlock,
+            GemmKPerBlock,
+            GemmMPerXDL,
+            GemmNPerXDL,
+            GemmK1,
+            MRepeat,
+            NRepeat,
+            GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
+            GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
+            Sequence<0, 2, 1>,
+            Sequence<0, 2, 1>,
+            1,
+            GemmABlockTransferSrcScalarPerVector_GemmM,
+            GemmABlockTransferDstScalarPerVector_GemmK1,
+            false, // don't move back src coordinate after threadwise copy
+            GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
+            GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
+            Sequence<0, 2, 1>,
+            Sequence<0, 2, 1>,
+            1,
+            GemmBBlockTransferSrcScalarPerVector_GemmN,
+            GemmBBlockTransferDstScalarPerVector_GemmK1,
+            false, // don't move back src coordinate after threadwise copy
+            Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+            7,
+            GemmCThreadTransferDstScalarPerVector,
+            decltype(in_gemmk0_gemmm_gemmk1_grid_step_hacks),
+            decltype(out_gemmk0_gemmn_gemmk1_grid_step_hacks),
+            decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
+            decltype(in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
+            decltype(out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
+            false, // CAccessOrderMRepeatNRepeat
+            true,
+            true>(static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
+                  static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
+                  static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
+                  in_gemmk0_gemmm_gemmk1_grid_desc,
+                  out_gemmk0_gemmn_gemmk1_grid_desc,
+                  wei_gemmm_gemmn_grid_desc,
+                  debug::debug_driver_gemm_xdlops_v2r3::M01,
+                  debug::debug_driver_gemm_xdlops_v2r3::N01,
+                  in_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                  out_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                  wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                  in_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                  out_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                  nrepeat);
+
+        {
+            const auto N = out_n_ho_wo_k_lengths[I0];
+            const auto K = out_n_ho_wo_k_lengths[I3];
+            const auto C = wei_k_y_x_c_lengths[I3];
+
+            const auto Ho = out_n_ho_wo_k_lengths[I1];
+            const auto Wo = out_n_ho_wo_k_lengths[I2];
+
+            const auto Y = wei_k_y_x_c_lengths[I1];
+            const auto X = wei_k_y_x_c_lengths[I2];
+
+            float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
+                         (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+            std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
+                      << std::endl;
+        }
+    }
+
+    // copy result back to host
+    wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
+}

host/driver_offline/include/device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk.hpp

+#include <unistd.h>
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp"
+#include "driver_gemm_xdlops_v2r4.hpp"
+
+template <typename TIn,
+          typename TWei,
+          typename TAcc,
+          typename TOut,
+          typename InLengths,
+          typename WeiLengths,
+          typename OutLengths,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          typename GridSizeType>
+void device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk(
+    const InLengths& in_n_hi_wi_c_lengths,
+    const WeiLengths& wei_k_y_x_c_lengths,
+    const OutLengths& out_n_ho_wo_k_lengths,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    const Tensor<TIn>& in_n_hi_wi_c,
+    Tensor<TWei>& wei_k_y_x_c,
+    const Tensor<TOut>& out_n_ho_wo_k,
+    GridSizeType desired_grid_size,
+    ck::index_t nrepeat)
+{
+    using namespace ck;
+
+    std::cout << __func__ << std::endl;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    DeviceMem in_n_hi_wi_c_device_buf(sizeof(TIn) * in_n_hi_wi_c.mDesc.GetElementSpace());
+    DeviceMem wei_k_y_x_c_device_buf(sizeof(TWei) * wei_k_y_x_c.mDesc.GetElementSpace());
+    DeviceMem out_n_ho_wo_k_device_buf(sizeof(TOut) * out_n_ho_wo_k.mDesc.GetElementSpace());
+
+    in_n_hi_wi_c_device_buf.ToDevice(in_n_hi_wi_c.mData.data());
+    wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
+    out_n_ho_wo_k_device_buf.ToDevice(out_n_ho_wo_k.mData.data());
+
+    const auto in_n_hi_wi_c_desc  = make_naive_tensor_descriptor_packed(in_n_hi_wi_c_lengths);
+    const auto wei_k_y_x_c_desc   = make_naive_tensor_descriptor_packed(wei_k_y_x_c_lengths);
+    const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(out_n_ho_wo_k_lengths);
+
+#if 0
+    // [M, N, K0, K1] = [256, 128, 4, 4], C 128, for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 256;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 4;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 4;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 0
+    // [M, N, K0, K1] = [128, 256, 4, 4], C 128, for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 256;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 4;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 0
+    // [M, N, K0, K1] = [128, 128, 4, 4], C 64, for fp32
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 4;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 4;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 4, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 32, 2>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 4;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [256, 128, 4, 8], C 128, for fp16
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 256;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 4;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 16, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 16, 4>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 16, 4>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [128, 128, 4, 8], C 64, for fp16
+    constexpr index_t BlockSize = 256;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 16, 4>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 16, 4>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 0
+    // [M, N, K0, K1] = [128, 64, 4, 8], C 64, for fp16
+    constexpr index_t BlockSize = 128;
+
+    constexpr index_t GemmMPerBlock = 128;
+    constexpr index_t GemmNPerBlock = 64;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 16, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [64, 128, 4, 8], C 64, for fp16
+    constexpr index_t BlockSize = 128;
+
+    constexpr index_t GemmMPerBlock = 64;
+    constexpr index_t GemmNPerBlock = 128;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 2;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 16, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#elif 1
+    // [M, N, K0, K1] = [64, 64, 4, 8], C 32, for fp16
+    constexpr index_t BlockSize = 128;
+
+    constexpr index_t GemmMPerBlock = 64;
+    constexpr index_t GemmNPerBlock = 64;
+    constexpr index_t GemmKPerBlock = 4;
+
+    constexpr index_t GemmMPerXDL = 32;
+    constexpr index_t GemmNPerXDL = 32;
+    constexpr index_t GemmK1      = 8;
+
+    constexpr index_t MRepeat = 2;
+    constexpr index_t NRepeat = 1;
+
+    using GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmABlockTransferSrcScalarPerVector_GemmM  = 8;
+    constexpr index_t GemmABlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    using GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1   = Sequence<1, 1, 8, 2>;
+    using GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1 = Sequence<1, 4, 8, 4>;
+
+    constexpr index_t GemmBBlockTransferSrcScalarPerVector_GemmN  = 8;
+    constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 2;
+
+    constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
+#endif
+
+    const auto N = in_n_hi_wi_c_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_desc.GetLength(I3);
+
+    const auto Ho = out_n_ho_wo_k_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_desc.GetLength(I2);
+
+    const auto GemmM      = K;
+    const auto GemmN      = Y * X * C;
+    const auto GemmKTotal = N * Ho * Wo;
+
+    const auto GemmK = GemmKTotal / GemmK1;
+
+    const auto GridMN        = GemmM * GemmN / (GemmMPerBlock * GemmNPerBlock);
+    const index_t GemmKBatch = std::max(desired_grid_size / GridMN, 1);
+    const index_t BatchLen   = std::ceil(GemmK * 1.0 / (GemmKPerBlock * GemmKBatch));
+    const index_t GemmK0     = BatchLen * GemmKPerBlock;
+    const index_t GemmKPad   = GemmKBatch * GemmK0 * GemmK1;
+
+    std::cout << "GemmKTotal: " << GemmKTotal << " GrideSizeMN: " << GridMN
+              << " GemmKBatch: " << GemmKBatch << " GemmK0: " << GemmK0 << " gemmKPad: " << GemmKPad
+              << std::endl;
+
+    const auto descs = transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
+        in_n_hi_wi_c_desc,
+        wei_k_y_x_c_desc,
+        out_n_ho_wo_k_desc,
+        conv_strides,
+        conv_dilations,
+        in_left_pads,
+        in_right_pads,
+        Number<GemmK1>{},
+        GemmKBatch,
+        GemmKPad);
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = descs[I0];
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc  = descs[I1];
+    const auto wei_gemmm_gemmn_grid_desc                    = descs[I2];
+
+    // HACK: hacks that control index calculation when iterating over A, B, C matrix
+    constexpr auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1+: GemmN
+                              Sequence<0, 0, 0, 0, 0>{}),  // 2+: GemmK1
+                   make_tuple(Sequence<0, 0, 0, 0, 0>{},   // 0+: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 0-: GemmK0
+                              Sequence<0, 0, 0, 0, 0>{},   // 1-: GemmN
+                              Sequence<0, 0, 0, 0, 0>{})); // 2-: GemmK1
+
+    constexpr auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks = make_tuple(
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 0+: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{},   // 0+: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{},   // 1+: GemmM
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0>{}),  // 2+: GemmK1
+        make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 0-: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{},   // 0-: GemmK0
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{},   // 1-: GemmM
+                   Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0>{})); // 2-: GemmK1
+
+    constexpr auto wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
+        make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0+: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1+: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2+: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3+: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4+: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5+: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6+: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}),  // 7+: N2
+                   make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 0-: M0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 1-: N0
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 2-: M1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 3-: N1
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 4-: M2
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 5-: M3
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{},   // 6-: M4
+                              Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
+
+    constexpr auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0>{};
+
+    constexpr auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks =
+        Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 0, 0, 0, 0>{};
+
+    const auto driver_gemm_xdlops = driver_gemm_xdlops_v2r4<
+        BlockSize,
+        TIn,
+        TAcc,
+        TWei,
+        InMemoryDataOperationEnum_t::AtomicAdd,
+        decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc),
+        decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc),
+        decltype(wei_gemmm_gemmn_grid_desc),
+        GemmMPerBlock,
+        GemmNPerBlock,
+        GemmKPerBlock,
+        GemmMPerXDL,
+        GemmNPerXDL,
+        GemmK1,
+        MRepeat,
+        NRepeat,
+        GemmABlockTransferThreadSliceLengths_GemmK0_GemmM_GemmK1,
+        GemmABlockTransferThreadClusterLengths_GemmK0_GemmM_GemmK1,
+        Sequence<0, 1, 2, 3>,
+        Sequence<0, 1, 2, 3>,
+        2,
+        GemmABlockTransferSrcScalarPerVector_GemmM,
+        GemmABlockTransferDstScalarPerVector_GemmK1,
+        false, // don't move back src coordinate after threadwise copy
+        GemmBBlockTransferThreadSliceLengths_GemmK0_GemmN_GemmK1,
+        GemmBBlockTransferThreadClusterLengths_GemmK0_GemmN_GemmK1,
+        Sequence<0, 1, 2, 3>,
+        Sequence<0, 1, 3, 2>,
+        2,
+        GemmBBlockTransferSrcScalarPerVector_GemmN,
+        GemmBBlockTransferDstScalarPerVector_GemmK1,
+        false, // don't move back src coordinate after threadwise copy
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        7,
+        GemmCThreadTransferDstScalarPerVector,
+        decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks),
+        decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks),
+        decltype(wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks),
+        decltype(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks),
+        decltype(in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks),
+        false, // CAccessOrderMRepeatNRepeat
+        true,
+        true>;
+
+    // timing
+    for(index_t i = 0; i < 5; ++i)
+    {
+        float ave_time =
+            driver_gemm_xdlops(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
+                               static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
+                               static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
+                               out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                               wei_gemmm_gemmn_grid_desc,
+                               debug::debug_driver_gemm_xdlops_v2r3::M01,
+                               debug::debug_driver_gemm_xdlops_v2r3::N01,
+                               out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                               wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                               out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                               nrepeat);
+
+        {
+            float perf = static_cast<float>((std::size_t(2) * N * K * Ho * Wo * C * Y * X)) /
+                         (std::size_t(1000) * 1000 * 1000) / ave_time;
+
+            std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s"
+                      << std::endl;
+        }
+    }
+
+    // verification
+    wei_k_y_x_c_device_buf.ToDevice(wei_k_y_x_c.mData.data());
+    driver_gemm_xdlops(static_cast<TOut*>(out_n_ho_wo_k_device_buf.GetDeviceBuffer()),
+                       static_cast<TIn*>(in_n_hi_wi_c_device_buf.GetDeviceBuffer()),
+                       static_cast<TWei*>(wei_k_y_x_c_device_buf.GetDeviceBuffer()),
+                       out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                       in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                       wei_gemmm_gemmn_grid_desc,
+                       debug::debug_driver_gemm_xdlops_v2r3::M01,
+                       debug::debug_driver_gemm_xdlops_v2r3::N01,
+                       out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_step_hacks,
+                       in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_step_hacks,
+                       wei_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks,
+                       out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_move_slice_window_step_hacks,
+                       in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_move_slice_window_step_hacks,
+                       0);
+    // copy result back to host
+    wei_k_y_x_c_device_buf.FromDevice(wei_k_y_x_c.mData.data());
+}

host/driver_offline/include/driver_gemm_xdlops_v2r4.hpp

+#ifndef DRIVER_GEMM_XDLOPS_V2R4
+#define DRIVER_GEMM_XDLOPS_V2R4
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r4.hpp"
+
+template <ck::index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          ck::InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CMNGridDesc,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t K1,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadSliceLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          typename BBlockTransferThreadSliceLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          typename CThreadTransferSrcDstAccessOrder,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          typename AGridStepHacks,
+          typename BGridStepHacks,
+          typename CGridStepHacks,
+          typename AGridMoveSliceWindowStepHacks,
+          typename BGridMoveSliceWindowStepHacks,
+          bool CAccessOrderMRepeatNRepeat,
+          bool ABlockLdsAddExtraM,
+          bool BBlockLdsAddExtraN>
+__host__ float driver_gemm_xdlops_v2r4(const FloatAB* p_a_grid,
+                                       const FloatAB* p_b_grid,
+                                       FloatC* p_c_grid,
+                                       const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                                       const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                                       const CMNGridDesc& c_m_n_grid_desc,
+                                       ck::index_t M01,
+                                       ck::index_t N01,
+                                       AGridStepHacks,
+                                       BGridStepHacks,
+                                       CGridStepHacks,
+                                       AGridMoveSliceWindowStepHacks,
+                                       BGridMoveSliceWindowStepHacks,
+                                       ck::index_t nrepeat)
+
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    using GridwiseGemm =
+        GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4<BlockSize,
+                                                  FloatAB,
+                                                  FloatAcc,
+                                                  FloatC,
+                                                  CGlobalMemoryDataOperation,
+                                                  ABK0MK1GridDesc,
+                                                  BBK0NK1GridDesc,
+                                                  CMNGridDesc,
+                                                  MPerBlock,
+                                                  NPerBlock,
+                                                  KPerBlock,
+                                                  MPerXDL,
+                                                  NPerXDL,
+                                                  K1,
+                                                  MRepeat,
+                                                  NRepeat,
+                                                  ABlockTransferThreadSliceLengths_K0_M_K1,
+                                                  ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                  ABlockTransferThreadClusterArrangeOrder,
+                                                  ABlockTransferSrcAccessOrder,
+                                                  ABlockTransferSrcVectorDim,
+                                                  ABlockTransferSrcScalarPerVector,
+                                                  ABlockTransferDstScalarPerVector_K1,
+                                                  AThreadTransferSrcResetCoordinateAfterRun,
+                                                  BBlockTransferThreadSliceLengths_K0_N_K1,
+                                                  BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                  BBlockTransferThreadClusterArrangeOrder,
+                                                  BBlockTransferSrcAccessOrder,
+                                                  BBlockTransferSrcVectorDim,
+                                                  BBlockTransferSrcScalarPerVector,
+                                                  BBlockTransferDstScalarPerVector_K1,
+                                                  BThreadTransferSrcResetCoordinateAfterRun,
+                                                  CThreadTransferSrcDstAccessOrder,
+                                                  CThreadTransferSrcDstVectorDim,
+                                                  CThreadTransferDstScalarPerVector,
+                                                  AGridStepHacks,
+                                                  BGridStepHacks,
+                                                  CGridStepHacks,
+                                                  AGridMoveSliceWindowStepHacks,
+                                                  BGridMoveSliceWindowStepHacks,
+                                                  CAccessOrderMRepeatNRepeat,
+                                                  ABlockLdsAddExtraM,
+                                                  BBlockLdsAddExtraN>;
+
+    {
+        std::cout << "a_b_k0_m_k1_grid_desc{" << a_b_k0_m_k1_grid_desc.GetLength(I0) << ", "
+                  << a_b_k0_m_k1_grid_desc.GetLength(I1) << ", "
+                  << a_b_k0_m_k1_grid_desc.GetLength(I2) << ", "
+                  << a_b_k0_m_k1_grid_desc.GetLength(I3) << "}" << std::endl;
+
+        std::cout << "b_b_k0_n_k1_grid_desc{" << b_b_k0_n_k1_grid_desc.GetLength(I0) << ", "
+                  << b_b_k0_n_k1_grid_desc.GetLength(I1) << ", "
+                  << b_b_k0_n_k1_grid_desc.GetLength(I2) << ", "
+                  << b_b_k0_n_k1_grid_desc.GetLength(I3) << "}" << std::endl;
+
+        std::cout << "c_m_n_grid_desc{ " << c_m_n_grid_desc.GetLength(I0) << ", "
+                  << c_m_n_grid_desc.GetLength(I1) << "}" << std::endl;
+    }
+
+    if(!GridwiseGemm::CheckValidity(
+           a_b_k0_m_k1_grid_desc, b_b_k0_n_k1_grid_desc, c_m_n_grid_desc, M01, N01))
+    {
+        throw std::runtime_error(
+            "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r4 has invalid setting");
+    }
+
+    const auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc =
+        GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_m_n_grid_desc);
+
+    using CM0N0M1N1M2M3M4N2GridDesc = decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
+
+    const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+    const auto c_block_cluster_adaptor =
+        GridwiseGemm::MakeCBlockClusterAdaptor(c_m_n_grid_desc, M01, N01, KBatch);
+
+    using CBlockClusterAdaptor = decltype(c_block_cluster_adaptor);
+
+    const index_t grid_size = GridwiseGemm::CalculateGridSize(c_m_n_grid_desc, KBatch);
+    {
+        std::cout << "gridSize : " << grid_size << std::endl;
+    }
+
+    const auto kernel = kernel_gemm_xdlops_v2r4<GridwiseGemm,
+                                                FloatAB,
+                                                FloatC,
+                                                remove_reference_t<ABK0MK1GridDesc>,
+                                                remove_reference_t<BBK0NK1GridDesc>,
+                                                remove_reference_t<CM0N0M1N1M2M3M4N2GridDesc>,
+                                                remove_reference_t<CBlockClusterAdaptor>>;
+
+#if CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE
+    float ave_time = launch_and_time_kernel(kernel,
+                                            nrepeat,
+                                            dim3(grid_size),
+                                            dim3(BlockSize),
+                                            0,
+                                            p_a_grid,
+                                            p_b_grid,
+                                            p_c_grid,
+                                            a_b_k0_m_k1_grid_desc,
+                                            b_b_k0_n_k1_grid_desc,
+                                            c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                            c_block_cluster_adaptor);
+
+#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
+    DeviceMem a_b_k0_m_k1_grid_desc_dev_buf(sizeof(ABK0MK1GridDesc));
+    DeviceMem b_b_k0_n_k1_grid_desc_dev_buf(sizeof(BBK0NK1GridDesc));
+    DeviceMem c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf(sizeof(CM0N0M1N1M2M3M4N2GridDesc));
+    DeviceMem c_block_cluster_adaptor_dev_buf(sizeof(CBlockClusterAdaptor));
+
+    a_b_k0_m_k1_grid_desc_dev_buf.ToDevice(&a_b_k0_m_k1_grid_desc);
+    b_b_k0_n_k1_grid_desc_dev_buf.ToDevice(&b_b_k0_n_k1_grid_desc);
+    c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.ToDevice(&c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc);
+    c_block_cluster_adaptor_dev_buf.ToDevice(&c_block_cluster_adaptor);
+
+    float ave_time = launch_and_time_kernel(
+        kernel,
+        nrepeat,
+        dim3(grid_size),
+        dim3(BlockSize),
+        0,
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        cast_pointer_to_constant_address_space(a_b_k0_m_k1_grid_desc_dev_buf.GetDeviceBuffer()),
+        cast_pointer_to_constant_address_space(b_b_k0_n_k1_grid_desc_dev_buf.GetDeviceBuffer()),
+        cast_pointer_to_constant_address_space(
+            c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc_dev_buf.GetDeviceBuffer()),
+        cast_pointer_to_constant_address_space(c_block_cluster_adaptor_dev_buf.GetDeviceBuffer()));
+#endif
+    return ave_time;
+}
+#endif

host/driver_offline/src/conv_wrw_driver_offline.cpp

@@ -14,13 +14,25 @@
 #include "host_conv_bwd_weight.hpp"
 #include "device_tensor.hpp"
 #include "device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp"
+#include "device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp"
+#include "device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw.hpp"
+#include "device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk.hpp"
+#include "device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk.hpp"
 
 #define USE_DYNAMIC_MODE 1
-#define USE_CONV_WRW_V4R4R2_XDL_NCHW 1
+#define USE_CONV_WRW_V4R4R2_XDL_NCHW 0
+#define USE_CONV_WRW_V4R4R4_XDL_NHWC 0
+#define USE_CONV_WRW_V4R4R2_XDL_ATOMIC_NCHW 0
+#define USE_CONV_WRW_V4R4R4_XDL_ATOMIC_NHWC 0
+#define USE_CONV_WRW_V4R4R5_XDL_ATOMIC_NHWC 1
 
 enum ConvBackwardWeightAlgo
 {
-    V4R4R2XDLNCHW,
+    V4R4R2XDLNCHW,       // 0
+    V4R4R4XDLNHWC,       // 1
+    V4R4R2XDLATOMICNCHW, // 2
+    V4R4R4XDLATOMICNHWC, // 3
+    V4R4R5XDLATOMICNHWC, // 4
 };
 
 int main(int argc, char* argv[])
@@ -37,10 +49,11 @@ int main(int argc, char* argv[])
 
 #if USE_DYNAMIC_MODE
     // dynamic mode
-    if(argc != 22)
+    if(argc != 23)
     {
         printf("arg1 to 6: layout, algo, do_verification, init_method, do_log, nrepeat\n");
         printf("rest: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx\n");
+        printf("additional: desired_grid_size\n");
         exit(1);
     }
 
@@ -68,6 +81,8 @@ int main(int argc, char* argv[])
     const index_t in_right_pad_h  = std::stoi(argv[20]);
     const index_t in_right_pad_w  = std::stoi(argv[21]);
 
+    const index_t desired_grid_size = std::stoi(argv[22]);
+
     const index_t YEff = (Y - 1) * conv_dilation_h + 1;
     const index_t XEff = (X - 1) * conv_dilation_w + 1;
 
@@ -114,16 +129,19 @@ int main(int argc, char* argv[])
 
 #if 0
     using in_data_t  = float;
+    using wei_data_t = float;
     using acc_data_t = float;
     using out_data_t = float;
 #elif 1
     using in_data_t   = half_t;
-    using acc_data_t  = float;
     using out_data_t  = half_t;
+    using acc_data_t  = float;
+    using wei_data_t  = float;
 #elif 1
     using in_data_t  = int8_t;
-    using acc_data_t = int32_t;
     using out_data_t = int8_t;
+    using acc_data_t = int32_t;
+    using wei_data_t = int8_t;
 #endif
 
     std::vector<std::size_t> in_lengths_host(4), wei_lengths_host(4), out_lengths_host(4);
@@ -164,8 +182,8 @@ int main(int argc, char* argv[])
     }
 
     Tensor<in_data_t> in(in_lengths_host);
-    Tensor<in_data_t> wei_device(wei_lengths_host);
-    Tensor<out_data_t> wei_host(wei_lengths_host);
+    Tensor<wei_data_t> wei_device(wei_lengths_host);
+    Tensor<wei_data_t> wei_host(wei_lengths_host);
     Tensor<out_data_t> out(out_lengths_host);
 
     std::cout << "layout: " << layout << std::endl;
@@ -231,6 +249,26 @@ int main(int argc, char* argv[])
                           in_right_pads_dev);
     };
 
+    auto f_make_for_device_nhwc = [&]() {
+        const auto in_lengths_dev     = make_tuple(N, Hi, Wi, C);
+        const auto wei_lengths_dev    = make_tuple(K, Y, X, C);
+        const auto out_lengths_dev    = make_tuple(N, Ho, Wo, K);
+        const auto conv_strides_dev   = make_tuple(conv_stride_h, conv_stride_w);
+        const auto conv_dilations_dev = make_tuple(conv_dilation_h, conv_dilation_w);
+        const auto in_left_pads_dev   = make_tuple(in_left_pad_h, in_left_pad_w);
+        const auto in_right_pads_dev  = make_tuple(in_right_pad_h, in_right_pad_w);
+
+        return make_tuple(in_lengths_dev,
+                          wei_lengths_dev,
+                          out_lengths_dev,
+                          conv_strides_dev,
+                          conv_dilations_dev,
+                          in_left_pads_dev,
+                          in_right_pads_dev);
+    };
+
+    // set zero to wei_device
+    wei_device.GenerateTensorValue(GeneratorTensor_0{}, num_thread);
 #if USE_CONV_WRW_V4R4R2_XDL_NCHW
     if(algo == ConvBackwardWeightAlgo::V4R4R2XDLNCHW)
     {
@@ -242,6 +280,35 @@ int main(int argc, char* argv[])
         const auto tmp = f_make_for_device_nchw();
 
         device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw<in_data_t,
+                                                                                      wei_data_t,
+                                                                                      acc_data_t,
+                                                                                      out_data_t>(
+            tmp[I0],
+            tmp[I1],
+            tmp[I2],
+            tmp[I3],
+            tmp[I4],
+            tmp[I5],
+            tmp[I6],
+            in,
+            wei_device,
+            out,
+            nrepeat);
+    }
+#endif
+
+#if USE_CONV_WRW_V4R4R4_XDL_NHWC
+    if(algo == ConvBackwardWeightAlgo::V4R4R4XDLNHWC)
+    {
+        if(layout != ConvTensorLayout::NHWC)
+        {
+            throw std::runtime_error("wrong! layout");
+        }
+
+        const auto tmp = f_make_for_device_nhwc();
+
+        device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk<in_data_t,
+                                                                                      wei_data_t,
                                                                                       acc_data_t,
                                                                                       out_data_t>(
             tmp[I0],
@@ -258,6 +325,93 @@ int main(int argc, char* argv[])
     }
 #endif
 
+#if USE_CONV_WRW_V4R4R2_XDL_ATOMIC_NCHW
+    if(algo == ConvBackwardWeightAlgo::V4R4R2XDLATOMICNCHW)
+    {
+        if(layout != ConvTensorLayout::NCHW)
+        {
+            throw std::runtime_error("wrong! layout");
+        }
+
+        const auto tmp = f_make_for_device_nchw();
+
+        device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_atomic_nchw_kcyx_nkhw<
+            in_data_t,
+            wei_data_t,
+            acc_data_t,
+            out_data_t>(tmp[I0],
+                        tmp[I1],
+                        tmp[I2],
+                        tmp[I3],
+                        tmp[I4],
+                        tmp[I5],
+                        tmp[I6],
+                        in,
+                        wei_device,
+                        out,
+                        desired_grid_size,
+                        nrepeat);
+    }
+#endif
+
+#if USE_CONV_WRW_V4R4R4_XDL_ATOMIC_NHWC
+    if(algo == ConvBackwardWeightAlgo::V4R4R4XDLATOMICNHWC)
+    {
+        if(layout != ConvTensorLayout::NHWC)
+        {
+            throw std::runtime_error("wrong! layout");
+        }
+
+        const auto tmp = f_make_for_device_nhwc();
+
+        device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk<
+            in_data_t,
+            wei_data_t,
+            acc_data_t,
+            out_data_t>(tmp[I0],
+                        tmp[I1],
+                        tmp[I2],
+                        tmp[I3],
+                        tmp[I4],
+                        tmp[I5],
+                        tmp[I6],
+                        in,
+                        wei_device,
+                        out,
+                        desired_grid_size,
+                        nrepeat);
+    }
+#endif
+
+#if USE_CONV_WRW_V4R4R5_XDL_ATOMIC_NHWC
+    if(algo == ConvBackwardWeightAlgo::V4R4R5XDLATOMICNHWC)
+    {
+        if(layout != ConvTensorLayout::NHWC)
+        {
+            throw std::runtime_error("wrong! layout");
+        }
+
+        const auto tmp = f_make_for_device_nhwc();
+
+        device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk<
+            in_data_t,
+            wei_data_t,
+            acc_data_t,
+            out_data_t>(tmp[I0],
+                        tmp[I1],
+                        tmp[I2],
+                        tmp[I3],
+                        tmp[I4],
+                        tmp[I5],
+                        tmp[I6],
+                        in,
+                        wei_device,
+                        out,
+                        desired_grid_size,
+                        nrepeat);
+    }
+#endif
+
     if(do_verification)
     {
         host_direct_convolution_backward_weights(out,

host/host_tensor/include/device.hpp

@@ -2,6 +2,7 @@
 #define DEVICE_HPP
 
 #include <memory>
+#include <functional>
 #include <thread>
 #include <chrono>
 #include "hip/hip_runtime.h"
@@ -80,5 +81,4 @@ float launch_and_time_kernel(
 
     return timer.GetElapsedTime() / nrepeat;
 }
-
 #endif

host/host_tensor/include/host_tensor_generator.hpp

@@ -15,6 +15,17 @@ struct GeneratorTensor_1
     }
 };
 
+struct GeneratorTensor_0
+{
+    int value = 0;
+
+    template <typename... Is>
+    float operator()(Is...)
+    {
+        return value;
+    }
+};
+
 struct GeneratorTensor_2
 {
     int min_value = 0;