merge develop to attn-train-develop-qloop

include/ck/host_utility/io.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/host_utility/kernel_launch.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -20,6 +20,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
 #if CK_TIME_KERNEL
     if(stream_config.time_kernel_)
     {
+#if DEBUG_LOG
         printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
                __func__,
                grid_dim.x,
@@ -29,15 +30,15 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
                block_dim.y,
                block_dim.z);
 
-        const int nrepeat = 10;
-
         printf("Warm up 1 time\n");
-
+#endif
         // warm up
         kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
 
+        const int nrepeat = 10;
+#if DEBUG_LOG
         printf("Start running %d times...\n", nrepeat);
-
+#endif
         hipEvent_t start, stop;
 
         hip_check_error(hipEventCreate(&start));

include/ck/host_utility/stream_utility.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+#include "ck/stream_config.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+static inline int getAvailableComputeUnitCount(const StreamConfig& stream_config)
+{
+    constexpr int MAX_MASK_DWORDS = 64;
+
+    // assume at most 64*32 = 2048 CUs
+    uint32_t cuMask[MAX_MASK_DWORDS];
+
+    for(int i = 0; i < MAX_MASK_DWORDS; i++)
+        cuMask[i] = 0;
+
+    auto countSetBits = [](uint32_t dword) {
+        int count = 0;
+
+        while(dword != 0)
+        {
+            if(dword & 0x1)
+                count++;
+
+            dword = dword >> 1;
+        };
+
+        return (count);
+    };
+
+    hip_check_error(hipExtStreamGetCUMask(stream_config.stream_id_, MAX_MASK_DWORDS, &cuMask[0]));
+
+    int ret = 0;
+
+    for(int i = 0; i < MAX_MASK_DWORDS; i++)
+        ret += countSetBits(cuMask[i]);
+
+    return (ret);
+};

include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
-#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// Number of GEMMs = YTilde * XTilde
-// GemmM = C
-// GemmN = N * HTildeSlice * WTildeSlice
-// GemmK = K * YDotSlice * XDotSlice
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads,
-          index_t IYTildeValue,
-          index_t IXTildeValue,
-          index_t GemmK1Value>
-__host__ __device__ constexpr auto
-transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk(
-    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
-    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
-    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
-    const ConvStrides& conv_strides,
-    const ConvDilations& conv_dilations,
-    const InLeftPads& in_left_pads,
-    const InRightPads& in_right_pads,
-    Number<IYTildeValue>,
-    Number<IXTildeValue>,
-    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>{};
-    constexpr auto IYTilde = Number<IYTildeValue>{};
-    constexpr auto IXTilde = Number<IXTildeValue>{};
-
-    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 GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
-    const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
-
-    const auto YTilde = ConvStrideH / GcdStrideDilationH;
-    const auto XTilde = ConvStrideW / GcdStrideDilationW;
-
-    const auto YDot = math::integer_divide_ceil(Y, YTilde);
-    const auto XDot = math::integer_divide_ceil(X, XTilde);
-
-    const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
-    const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
-
-    // only work on HTilde and WTilde that contribute to non-padding area of input tensor
-    const auto IHTildeSliceBegin = math::integer_divide_floor(
-        math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
-    const auto IWTildeSliceBegin = math::integer_divide_floor(
-        math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
-
-    const auto IHTildeSliceEnd =
-        math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
-    const auto IWTildeSliceEnd =
-        math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
-
-    const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
-    const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
-
-    // GemmK is different for each GEMM
-    const auto YDotSlice = math::integer_divide_ceil(Y - IYTilde, YTilde);
-    const auto XDotSlice = math::integer_divide_ceil(X - IXTilde, XTilde);
-
-    const auto K1 = GemmK1;
-    const auto K0 = K / K1;
-
-    // weight tensor
-    const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
-        wei_k_y_x_c_grid_desc,
-        make_tuple(make_pass_through_transform(K),
-                   make_embed_transform(make_tuple(YDot, YTilde),
-                                        make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
-                   make_embed_transform(make_tuple(XDot, XTilde),
-                                        make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
-                   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 wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
-        transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
-                                    make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
-                                               make_slice_transform(YDot, I0, YDotSlice),
-                                               make_slice_transform(XDot, I0, XDotSlice),
-                                               make_freeze_transform(IYTilde),
-                                               make_freeze_transform(IXTilde),
-                                               make_pass_through_transform(C)),
-                                    make_tuple(Sequence<0>{},
-                                               Sequence<1>{},
-                                               Sequence<3>{},
-                                               Sequence<2>{},
-                                               Sequence<4>{},
-                                               Sequence<5>{}),
-                                    make_tuple(Sequence<0, 1>{},
-                                               Sequence<2>{},
-                                               Sequence<3>{},
-                                               Sequence<>{},
-                                               Sequence<>{},
-                                               Sequence<4>{}));
-
-#if 1
-    const auto wei_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
-                   make_pass_through_transform(C),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#else
-    const auto wei_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
-                   make_pass_through_transform(C),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<0, 2, 3>{}, Sequence<4>{}, Sequence<1>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#endif
-
-    // output tensor
-    // this add padding check
-    const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
-        out_n_ho_wo_k_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_pad_transform(Ho, I0, I0),
-                   make_pad_transform(Wo, I0, I0),
-                   make_pass_through_transform(K)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
-
-    const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
-        out_n_hop_wop_k_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_embed_transform(make_tuple(YDot, HTilde),
-                                        make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
-                   make_embed_transform(make_tuple(XDot, WTilde),
-                                        make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
-                   make_pass_through_transform(K)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
-
-    const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
-        transform_tensor_descriptor(
-            out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
-            make_tuple(make_pass_through_transform(N),
-                       make_slice_transform(YDot, I0, YDotSlice),
-                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                       make_slice_transform(XDot, I0, XDotSlice),
-                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                       make_unmerge_transform(make_tuple(K0, K1))),
-            make_tuple(Sequence<0>{},
-                       Sequence<1>{},
-                       Sequence<2>{},
-                       Sequence<3>{},
-                       Sequence<4>{},
-                       Sequence<5>{}),
-            make_tuple(Sequence<0>{},
-                       Sequence<1>{},
-                       Sequence<2>{},
-                       Sequence<3>{},
-                       Sequence<4>{},
-                       Sequence<5, 6>{}));
-
-#if 1
-    const auto out_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
-                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#else
-    const auto out_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
-                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<5, 1, 3>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#endif
-
-    // 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_ytilde_htilde_xtilde_wtilde_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(YTilde, HTilde),
-                                        make_tuple(ConvDilationH, ConvStrideH)),
-                   make_embed_transform(make_tuple(XTilde, WTilde),
-                                        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_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
-        in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_freeze_transform(IYTilde),
-                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                   make_freeze_transform(IXTilde),
-                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                   make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{},
-                   Sequence<1>{},
-                   Sequence<2>{},
-                   Sequence<3>{},
-                   Sequence<4>{},
-                   Sequence<5>{}),
-        make_tuple(Sequence<0>{},
-                   Sequence<>{},
-                   Sequence<1>{},
-                   Sequence<>{},
-                   Sequence<2>{},
-                   Sequence<3>{}));
-
-    const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
-        in_n_htildeslice_wtildeslice_c_grid_desc,
-        make_tuple(make_pass_through_transform(C),
-                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice))),
-        make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
-                      out_gemmk0_gemmn_gemmk1_grid_desc,
-                      in_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
-#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// A: out
-// B: wei
-// C: in
-// Number of GEMMs = YTilde * XTilde
-// GemmM = N * HTildeSlice * WTildeSlice
-// GemmN = C
-// GemmK = K * YDotSlice * XDotSlice
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads,
-          typename IYTilde,
-          typename IXTilde,
-          index_t GemmK1Value>
-__host__ __device__ constexpr auto
-transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
-    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
-    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
-    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
-    const ConvStrides& conv_strides,
-    const ConvDilations& conv_dilations,
-    const InLeftPads& in_left_pads,
-    const InRightPads& in_right_pads,
-    IYTilde i_ytilde,
-    IXTilde i_xtilde,
-    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 GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
-    const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
-
-    const auto YTilde = ConvStrideH / GcdStrideDilationH;
-    const auto XTilde = ConvStrideW / GcdStrideDilationW;
-
-    const auto YDot = math::integer_divide_ceil(Y, YTilde);
-    const auto XDot = math::integer_divide_ceil(X, XTilde);
-
-    const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
-    const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
-
-    // only work on HTilde and WTilde that contribute to non-padding area of input tensor
-    const auto IHTildeSliceBegin = math::integer_divide_floor(
-        math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
-    const auto IWTildeSliceBegin = math::integer_divide_floor(
-        math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
-
-    const auto IHTildeSliceEnd =
-        math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
-    const auto IWTildeSliceEnd =
-        math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
-
-    const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
-    const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
-
-    // GemmK is different for each GEMM
-    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
-    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
-
-    const auto K1 = GemmK1;
-    const auto K0 = K / K1;
-
-    // A: output tensor
-    // this add padding check
-    const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
-        out_n_ho_wo_k_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_pad_transform(Ho, I0, I0),
-                   make_pad_transform(Wo, I0, I0),
-                   make_pass_through_transform(K)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
-
-    const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
-        out_n_hop_wop_k_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_embed_transform(make_tuple(YDot, HTilde),
-                                        make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
-                   make_embed_transform(make_tuple(XDot, WTilde),
-                                        make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
-                   make_pass_through_transform(K)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
-
-    const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
-        transform_tensor_descriptor(
-            out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
-            make_tuple(make_pass_through_transform(N),
-                       make_slice_transform(YDot, I0, YDotSlice),
-                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                       make_slice_transform(XDot, I0, XDotSlice),
-                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                       make_unmerge_transform(make_tuple(K0, K1))),
-            make_tuple(Sequence<0>{},
-                       Sequence<1>{},
-                       Sequence<2>{},
-                       Sequence<3>{},
-                       Sequence<4>{},
-                       Sequence<5>{}),
-            make_tuple(Sequence<0>{},
-                       Sequence<1>{},
-                       Sequence<2>{},
-                       Sequence<3>{},
-                       Sequence<4>{},
-                       Sequence<5, 6>{}));
-
-#if 1
-    const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
-                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#else
-    const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-        out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
-                   make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<5, 1, 3>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#endif
-
-    // B: weight tensor
-    const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
-        wei_k_y_x_c_grid_desc,
-        make_tuple(make_pass_through_transform(K),
-                   make_embed_transform(make_tuple(YDot, YTilde),
-                                        make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
-                   make_embed_transform(make_tuple(XDot, XTilde),
-                                        make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
-                   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 wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
-        transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
-                                    make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
-                                               make_slice_transform(YDot, I0, YDotSlice),
-                                               make_slice_transform(XDot, I0, XDotSlice),
-                                               make_freeze_transform(i_ytilde),
-                                               make_freeze_transform(i_xtilde),
-                                               make_pass_through_transform(C)),
-                                    make_tuple(Sequence<0>{},
-                                               Sequence<1>{},
-                                               Sequence<3>{},
-                                               Sequence<2>{},
-                                               Sequence<4>{},
-                                               Sequence<5>{}),
-                                    make_tuple(Sequence<0, 1>{},
-                                               Sequence<2>{},
-                                               Sequence<3>{},
-                                               Sequence<>{},
-                                               Sequence<>{},
-                                               Sequence<4>{}));
-
-#if 1
-    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
-                   make_pass_through_transform(C),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#else
-    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-        wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(K0, YDotSlice, XDotSlice)),
-                   make_pass_through_transform(C),
-                   make_pass_through_transform(K1)),
-        make_tuple(Sequence<0, 2, 3>{}, Sequence<4>{}, Sequence<1>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-#endif
-
-    // C: 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_ytilde_htilde_xtilde_wtilde_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(YTilde, HTilde),
-                                        make_tuple(ConvDilationH, ConvStrideH)),
-                   make_embed_transform(make_tuple(XTilde, WTilde),
-                                        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_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
-        in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_freeze_transform(i_ytilde),
-                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                   make_freeze_transform(i_xtilde),
-                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                   make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{},
-                   Sequence<1>{},
-                   Sequence<2>{},
-                   Sequence<3>{},
-                   Sequence<4>{},
-                   Sequence<5>{}),
-        make_tuple(Sequence<0>{},
-                   Sequence<>{},
-                   Sequence<1>{},
-                   Sequence<>{},
-                   Sequence<2>{},
-                   Sequence<3>{}));
-
-    const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
-        in_n_htildeslice_wtildeslice_c_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                   make_pass_through_transform(C)),
-        make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
-                      wei_gemmk0_gemmn_gemmk1_grid_desc,
-                      in_gemmm_gemmn_grid_desc);
-}
-
-// A: out
-// B: wei
-// C: in
-// Number of GEMMs = 1
-// GemmM = N * Ho * Wo
-// GemmN = C
-// GemmK = K
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          index_t GemmK1Value>
-__host__ __device__ constexpr auto
-transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk_1x1(
-    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
-    const TensorDescriptor<Wei...>& /* wei_k_y_x_c_grid_desc */,
-    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
-    const ConvStrides& conv_strides,
-    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 Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
-    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
-
-    const auto ConvStrideH = conv_strides[I0];
-    const auto ConvStrideW = conv_strides[I1];
-
-    const auto K1 = GemmK1;
-    const auto K0 = K / K1;
-
-    // A: output tensor
-    const auto out_gemmk0_gemmm_gemmk1_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_unmerge_transform(make_tuple(K0, K1))),
-                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                    make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
-
-    // B: weight tensor
-    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
-        make_tuple(make_unmerge_transform(make_tuple(K0, K1)), make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}),
-        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-    // C: input tensor
-    const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
-        in_n_hi_wi_c_grid_desc,
-        make_tuple(make_pass_through_transform(N),
-                   make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
-                   make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, 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_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
-        in_n_y_ho_x_wo_c_grid_desc,
-        make_tuple(make_freeze_transform(I0),
-                   make_freeze_transform(I0),
-                   make_merge_transform(make_tuple(N, Ho, Wo)),
-                   make_pass_through_transform(C)),
-        make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
-        make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
-
-    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
-                      wei_gemmk0_gemmn_gemmk1_grid_desc,
-                      in_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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

include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
-#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_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>
-__host__ __device__ constexpr auto
-transform_backward_weight_convolution_into_gemm_v4r4r2_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>)
-{
-    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 GemmK  = N * Ho * Wo;
-    const auto GemmK0 = GemmK / GemmK1;
-
-    // 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>{}));
-
-    // 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_gemmk_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_gemmk0_gemmn_gemmk1_grid_desc =
-        transform_tensor_descriptor(in_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>{}));
-
-    // output tensor
-    const auto out_gemmk_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_gemmk0_gemmm_gemmk1_grid_desc =
-        transform_tensor_descriptor(out_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>{}));
-
-    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
-                      in_gemmk0_gemmn_gemmk1_grid_desc,
-                      wei_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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

include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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

include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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

include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
 #define CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP

include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NCHW_KCYX_NKHW_HPP
-#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NCHW_KCYX_NKHW_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// GemmM = K
-// GemmN = N * Ho * Wo
-// GemmK = C * Y * X
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads>
-__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_pad(
-    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
-    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
-    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
-    const ConvStrides& conv_strides,
-    const ConvDilations& conv_dilations,
-    const InLeftPads& in_left_pads,
-    const InRightPads& in_right_pads)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
-    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
-    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
-
-    const auto Hi = in_n_c_hi_wi_global_desc.GetLength(I2);
-    const auto Wi = in_n_c_hi_wi_global_desc.GetLength(I3);
-
-    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
-    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
-
-    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
-    const auto X = wei_k_c_y_x_global_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];
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_global_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<1>{}, Sequence<0>{}));
-
-    // input tensor
-    const auto in_n_c_hip_wip_global_desc = transform_tensor_descriptor(
-        in_n_c_hi_wi_global_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_global_desc = transform_tensor_descriptor(
-        in_n_c_hip_wip_global_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_gemmk_gemmn_global_desc =
-        transform_tensor_descriptor(in_n_c_y_ho_x_wo_global_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<0>{}, Sequence<1>{}));
-
-    // output tensor
-    const auto out_gemmm_gemmn_global_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<0>{}, Sequence<1>{}));
-
-    return make_tuple(
-        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
-}
-
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads>
-__host__ __device__ constexpr auto
-transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_no_pad(
-    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
-    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
-    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
-    const ConvStrides& conv_strides,
-    const ConvDilations& conv_dilations,
-    const InLeftPads& in_left_pads,
-    const InRightPads& in_right_pads)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
-    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
-    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
-
-    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
-    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
-
-    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
-    const auto X = wei_k_c_y_x_global_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];
-
-    assert(InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 && InRightPadW == 0);
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_global_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<1>{}, Sequence<0>{}));
-
-    // input tensor
-    const auto in_n_c_y_ho_x_wo_global_desc = transform_tensor_descriptor(
-        in_n_c_hi_wi_global_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_gemmk_gemmn_global_desc =
-        transform_tensor_descriptor(in_n_c_y_ho_x_wo_global_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<0>{}, Sequence<1>{}));
-
-    // output tensor
-    const auto out_gemmm_gemmn_global_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<0>{}, Sequence<1>{}));
-
-    return make_tuple(
-        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
-}
-
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads>
-__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nchw_kcyx_nkhw_1x1(
-    const TensorDescriptor<Wei...>& wei_k_c_y_x_global_desc,
-    const TensorDescriptor<In...>& in_n_c_hi_wi_global_desc,
-    const TensorDescriptor<Out...>& out_n_k_ho_wo_global_desc,
-    const ConvStrides& conv_strides,
-    const ConvDilations& conv_dilations,
-    const InLeftPads& in_left_pads,
-    const InRightPads& in_right_pads)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    const auto N = in_n_c_hi_wi_global_desc.GetLength(I0);
-    const auto C = in_n_c_hi_wi_global_desc.GetLength(I1);
-    const auto K = out_n_k_ho_wo_global_desc.GetLength(I1);
-
-    const auto Ho = out_n_k_ho_wo_global_desc.GetLength(I2);
-    const auto Wo = out_n_k_ho_wo_global_desc.GetLength(I3);
-
-    const auto Y = wei_k_c_y_x_global_desc.GetLength(I2);
-    const auto X = wei_k_c_y_x_global_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];
-
-    assert(Y == 1 && X == 1 && ConvStrideH == 1 && ConvStrideW == 1 && ConvDilationH == 1 &&
-           ConvDilationW == 1 && InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 &&
-           InRightPadW == 0);
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_global_desc = transform_tensor_descriptor(
-        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
-        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}),
-        make_tuple(Sequence<1>{}, Sequence<0>{}));
-
-    // input tensor
-    const auto in_gemmk_gemmn_global_desc = transform_tensor_descriptor(
-        in_n_c_hi_wi_global_desc,
-        make_tuple(make_pass_through_transform(C), make_merge_transform(make_tuple(N, Ho, Wo))),
-        make_tuple(Sequence<1>{}, Sequence<0, 2, 3>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-    // output tensor
-    const auto out_gemmm_gemmn_global_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<0>{}, Sequence<1>{}));
-
-    return make_tuple(
-        wei_gemmk_gemmm_global_desc, in_gemmk_gemmn_global_desc, out_gemmm_gemmn_global_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NHWC_KYXC_NHWK_HPP
-#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4_NHWC_KYXC_NHWK_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// GemmM = K
-// GemmN = N * Ho * Wo
-// GemmK = C * Y * X
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads>
-__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk_pad(
-    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
-    const TensorDescriptor<In...>& in_n_hi_wi_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)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    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];
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_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>{}));
-
-    // 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_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<0>{}, Sequence<1>{}));
-
-    // output tensor
-    const auto out_gemmm_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<1>{}, Sequence<0>{}));
-
-    return make_tuple(
-        wei_gemmk_gemmm_grid_desc, in_gemmk_gemmn_grid_desc, out_gemmm_gemmn_grid_desc);
-}
-
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads>
-__host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4_nhwc_kyxc_nhwk_1x1(
-    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
-    const TensorDescriptor<In...>& in_n_hi_wi_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)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    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 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];
-
-    assert(Y == 1 && X == 1 && ConvStrideH == 1 && ConvStrideW == 1 && ConvDilationH == 1 &&
-           ConvDilationW == 1 && InLeftPadH == 0 && InLeftPadW == 0 && InRightPadH == 0 &&
-           InRightPadW == 0);
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
-        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
-        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}),
-        make_tuple(Sequence<1>{}, Sequence<0>{}));
-
-    // input tensor
-    const auto in_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
-        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C)),
-        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(C)),
-        make_tuple(Sequence<0>{}, Sequence<1>{}),
-        make_tuple(Sequence<1>{}, Sequence<0>{}));
-
-    // output tensor
-    const auto out_gemmm_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<1>{}, Sequence<0>{}));
-
-    return make_tuple(
-        wei_gemmk_gemmm_grid_desc, in_gemmk_gemmn_grid_desc, out_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nchw_kcyx_nkhw.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
-#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NCHW_KCYX_NKHW_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// GemmM = K
-// GemmN = N * Ho * Wo
-// GemmK = C * Y * X
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads,
-          index_t GemmK1Value>
-__host__ __device__ constexpr auto
-transform_forward_convolution_into_gemm_v4r4r2_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>)
-{
-    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  = N * Ho * Wo;
-    const auto GemmK  = C * Y * X;
-    const auto GemmK0 = GemmK / GemmK1;
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_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<1>{}, Sequence<0>{}));
-
-    const auto wei_gemmk0_gemmm_gemmk1_grid_desc =
-        transform_tensor_descriptor(wei_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>{}));
-
-    // 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_gemmk_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<0>{}, Sequence<1>{}));
-
-    const auto in_gemmk0_gemmn_gemmk1_grid_desc =
-        transform_tensor_descriptor(in_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>{}));
-
-    // output tensor
-    const auto out_gemmm_gemmn_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<0>{}, Sequence<1>{}));
-
-    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
-                      in_gemmk0_gemmn_gemmk1_grid_desc,
-                      out_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
-#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R2_NHWC_KYXC_NHWK_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// GemmM = K
-// GemmN = N * Ho * Wo
-// GemmK = C * Y * X
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads,
-          index_t GemmK1Value>
-__host__ __device__ constexpr auto
-transform_forward_convolution_into_gemm_v4r4r2_nhwc_kyxc_nhwk_pad(
-    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
-    const TensorDescriptor<In...>& in_n_hi_wi_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  = K;
-    const auto GemmN  = N * Ho * Wo;
-    const auto GemmK  = C * Y * X;
-    const auto GemmK0 = GemmK / GemmK1;
-
-    // weight tensor
-    const auto wei_gemmk_gemmm_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>{}));
-
-    const auto wei_gemmk0_gemmm_gemmk1_grid_desc =
-        transform_tensor_descriptor(wei_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>{}));
-
-    // 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_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<0>{}, Sequence<1>{}));
-
-    const auto in_gemmk0_gemmn_gemmk1_grid_desc =
-        transform_tensor_descriptor(in_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>{}));
-
-    // output tensor
-    const auto out_gemmm_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<1>{}, Sequence<0>{}));
-
-    return make_tuple(wei_gemmk0_gemmm_gemmk1_grid_desc,
-                      in_gemmk0_gemmn_gemmk1_grid_desc,
-                      out_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
-#define CK_TRANSFORM_FORWARD_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_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
-    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  = N * Ho * Wo;
-    const auto GemmN  = K;
-    const auto GemmK  = Y * X * C;
-    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<0>{}, Sequence<1>{}));
-
-    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: weight tensor
-    const auto wei_gemmk_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>{}));
-
-    const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
-        transform_tensor_descriptor(wei_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: output tensor
-    const auto out_gemmm_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>{}));
-
-    return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
-                      wei_gemmk0_gemmn_gemmk1_grid_desc,
-                      out_gemmm_gemmn_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/problem_transform/transform_forward_convolution_into_gemm_v6r1_nchw_kcyx_nkhw.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
-#define CK_TRANSFORM_FORWARD_CONVOLUTION_INTO_CONTRACTION_V6R1_NCHW_KCYX_NKHW_HPP
-
-#include "common_header.hpp"
-#include "tensor_descriptor.hpp"
-#include "tensor_descriptor_helper.hpp"
-
-namespace ck {
-
-// GemmM0 = 1
-// GemmM1 = K
-// GemmN0 = N0
-// GemmN1 = (N / N0) * Ho * Wo
-// GemmK0 = (C / C0) * Y * X
-// GemmK1 = C0
-template <typename... Wei,
-          typename... In,
-          typename... Out,
-          typename ConvStrides,
-          typename ConvDilations,
-          typename InLeftPads,
-          typename InRightPads,
-          typename N0Type,
-          typename C0Type>
-__host__ __device__ constexpr auto
-transform_forward_convolution_into_contraction_v6r1_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,
-    const N0Type& N0,
-    const C0Type& C0)
-{
-    constexpr auto I0 = Number<0>{};
-    constexpr auto I1 = Number<1>{};
-    constexpr auto I2 = Number<2>{};
-    constexpr auto I3 = Number<3>{};
-
-    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 N1 = N / N0;
-    const auto C1 = C / C0;
-
-    // weight tensor
-    const auto wei_gk0_gm0_gm1_gk1_grid_desc =
-        transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
-                                    make_tuple(make_unmerge_transform(make_tuple(I1, K)),
-                                               make_unmerge_transform(make_tuple(C0, C1 * Y * X))),
-                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                    make_tuple(Sequence<1, 2>{}, Sequence<3, 0>{}));
-
-    // 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_n0_n1_c0_c1_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
-        in_n_c_hip_wip_grid_desc,
-        make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
-                   make_unmerge_transform(make_tuple(C0, C1)),
-                   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, 1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}, Sequence<6, 7>{}));
-
-    const auto in_gk0_gn0_gn1_gk1_grid_desc = transform_tensor_descriptor(
-        in_n0_n1_c0_c1_y_ho_x_wo_grid_desc,
-        make_tuple(make_merge_transform(make_tuple(C1, Y, X)),
-                   make_pass_through_transform(N0),
-                   make_merge_transform(make_tuple(N1, Ho, Wo)),
-                   make_pass_through_transform(C0)),
-        make_tuple(Sequence<3, 4, 6>{}, Sequence<0>{}, Sequence<1, 5, 7>{}, Sequence<2>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
-
-    // output tensor
-    const auto out_n_k_howo_grid_desc =
-        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo));
-
-    const auto out_n0_n1_1_k_howo_grid_desc =
-        transform_tensor_descriptor(out_n_k_howo_grid_desc,
-                                    make_tuple(make_unmerge_transform(make_tuple(N0, N1)),
-                                               make_unmerge_transform(make_tuple(I1, K)),
-                                               make_pass_through_transform(Ho * Wo)),
-                                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-                                    make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}, Sequence<4>{}));
-
-    const auto out_gm0_gm1_gn0_gn1_grid_desc = transform_tensor_descriptor(
-        out_n0_n1_1_k_howo_grid_desc,
-        make_tuple(make_pass_through_transform(I1),
-                   make_pass_through_transform(K),
-                   make_pass_through_transform(N0),
-                   make_merge_transform_v2_magic_division(make_tuple(N1, Ho * Wo))),
-        make_tuple(Sequence<2>{}, Sequence<3>{}, Sequence<0>{}, Sequence<1, 4>{}),
-        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
-
-    return make_tuple(
-        wei_gk0_gm0_gm1_gk1_grid_desc, in_gk0_gn0_gn1_gk1_grid_desc, out_gm0_gm1_gn0_gn1_grid_desc);
-}
-
-} // namespace ck
-#endif

include/ck/stream_config.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/tensor/static_tensor.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #ifndef CK_STATIC_TENSOR_HPP
 #define CK_STATIC_TENSOR_HPP

include/ck/tensor_description/cluster_descriptor.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once