Merge remote-tracking branch 'origin/develop' into aosewski/drop_cshuffle

example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //  ##################| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|     DsData|     EData|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ##################|        |        |         |        |      Type|      Type|        Type|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ##################|        |        |         |        |          |          |            |           |          |   Operation|   Operation|    Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ##################|        |        |         |        |          |          |            |           |          |            |            |             |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmMultipleD_Dl< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }

example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //##############################|      A|      B|       Ds|      E| AData| BData| AccData| CShuffle|     DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|   Layout| Layout|  Type|  Type|    Type| DataType|       Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |         |       |      |      |        |         |           |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |         |       |      |      |        |         |           |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, DsLayout,    Row,   F16,   F16,     F32,      F16, DsDataType,   F16, PassThrough, PassThrough, CDEElementOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }

example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc

+#pragma once
+
+bool run_gemm_add_multiply(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        std::cout << "wrong! this device_op instance does not support this problem" << std::endl;
+        return true;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = 2_uz * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
+                            sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+    }
+
+    return true;
+}
+
+bool run_gemm_add_multiply_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) ||
+           run_gemm_add_multiply(problem_size, config);
+}

example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt

+if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+   add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)
+endif()

example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using C0DEElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using Acc0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using F16              = ck::half_t;
+using F32              = float;
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using D0DataType       = F16;
+using Acc0BiasDataType = ck::Tuple<D0DataType>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        C0DEElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    int G0      = 3;
+    int G1      = 2;
+    int M       = 1024;
+    int N       = 1024;
+    int K       = 64;
+    int O       = 64;
+    float alpha = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{
+        M * G1 * K, K, G1 * K, 1}; // A layout [G0, M, G1, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{
+        N * G1 * K, K, G1 * K, 1}; // B0 layout [G0, N, G1, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{
+        N * G1 * O, O, 1, G1 * O}; // B1 layout [G0, N, G1, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{
+        M * G1 * O, O, G1 * O, 1}; // C layout [G0, M, G1, O]
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-1, 1});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    DeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    DeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    DeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    DeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    auto argument = device_op.MakeArgument(
+        static_cast<const ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<const B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<const B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+        {},                                                    // p_acc1_biases
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+        {},                       // acc1_biases_gs_ms_os_lengths
+        {},                       // acc1_biases_gs_ms_os_strides
+        a_element_op,
+        b0_element_op,
+        c0de_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t BatchCount = G0 * G1;
+    std::size_t flop       = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype =
+        (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N + sizeof(B1DataType) * N * O +
+         sizeof(CDataType) * M * O + sizeof(D0DataType) * M * N) *
+        BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // masking
+        const auto mask = DeviceOpInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

example/48_pool3d_fwd/CMakeLists.txt

+add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
+

example/48_pool3d_fwd/pool3d_fwd_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool3d_test(bool do_verification,
+                 bool time_kernel,
+                 ck::index_t N,
+                 ck::index_t C,
+                 ck::index_t Z,
+                 ck::index_t Y,
+                 ck::index_t X,
+                 ck::index_t Di,
+                 ck::index_t Hi,
+                 ck::index_t Wi,
+                 ck::index_t window_stride_d,
+                 ck::index_t window_stride_h,
+                 ck::index_t window_stride_w,
+                 ck::index_t in_left_pad_d,
+                 ck::index_t in_left_pad_h,
+                 ck::index_t in_left_pad_w,
+                 ck::index_t in_right_pad_d,
+                 ck::index_t in_right_pad_h,
+                 ck::index_t in_right_pad_w)
+{
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<
+            InDataType,      // InDataType
+            OutDataType,     // OutDataType
+            IndexDataType,   // IndexDataType
+            ComputeDataType, // ComputeDataType
+            ReduceOpId,
+            OutputIndex,
+            64, // BlockSize
+            64, // ReduceMThreadClusterSize
+            1,  // ReduceKThreadClusterSize
+            4,  // ReduceMThreadSliceSize
+            1,  // ReduceKThreadSliceSize
+            4>; // InSrcOutDstVectorSize
+
+    const ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Z) / window_stride_d + 1;
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
+
+    const std::vector<ck::index_t> window_spatial_lengths{Z, Y, X};
+    const std::vector<ck::index_t> window_strides{
+        window_stride_d, window_stride_h, window_stride_w};
+    const std::vector<ck::index_t> input_left_pads{in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    const std::vector<ck::index_t> input_right_pads{in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    // tensor layout
+    auto f_host_tensor_descriptor = [](std::size_t N_,
+                                       std::size_t C_,
+                                       std::size_t D,
+                                       std::size_t H,
+                                       std::size_t W,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+        {
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
+        }
+        else if constexpr(ck::is_same<decltype(layout),
+                                      ck::tensor_layout::convolution::NDHWC>::value)
+        {
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        }
+    };
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi, InLayout{}));
+    Tensor<OutDataType> out_n_c_do_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_do_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_di_hi_wi: " << in_n_c_di_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_do_ho_wo: " << out_n_c_do_ho_wo_host.mDesc << std::endl;
+
+    in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_di_hi_wi.mData.data());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        {N, C, Di, Hi, Wi},
+        {Z, Y, X},
+        {N, C, Do, Ho, Wo},
+        {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C},
+        {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
+        {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
+        window_strides,
+        input_left_pads,
+        input_right_pads,
+        {2, 3, 4});
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    std::cout << "Perf: " << ave_time << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<5,
+                                                            3,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ReduceOpId,
+                                                            PropagateNan,
+                                                            OutputIndex>;
+
+        auto ref_pooling          = ReferencePoolingFwdInstance{};
+        auto ref_pooling_invoker  = ref_pooling.MakeInvoker();
+        auto ref_pooling_argument = ref_pooling.MakeArgument(in_n_c_di_hi_wi,
+                                                             out_n_c_do_ho_wo_host,
+                                                             out_indices_n_c_do_ho_wo_host,
+                                                             window_spatial_lengths,
+                                                             window_strides,
+                                                             input_left_pads,
+                                                             input_right_pads);
+
+        ref_pooling_invoker.Run(ref_pooling_argument);
+
+        out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_do_ho_wo_device, out_n_c_do_ho_wo_host);
+
+        if constexpr(OutputIndex)
+        {
+            out_indices_device_buf.FromDevice(out_indices_n_c_do_ho_wo_device.mData.data());
+
+            pass = pass && ck::utils::check_err(out_indices_n_c_do_ho_wo_device,
+                                                out_indices_n_c_do_ho_wo_host);
+        };
+    }
+
+    return (pass);
+};

example/48_pool3d_fwd/pool3d_fwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "pool3d_fwd_common.hpp"
+
+using InDataType      = ck::half_t;
+using OutDataType     = ck::half_t;
+using ComputeDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // Pool shape
+    ck::index_t N               = 2;
+    ck::index_t C               = 32;
+    ck::index_t Z               = 2;
+    ck::index_t Y               = 2;
+    ck::index_t X               = 2;
+    ck::index_t Di              = 30;
+    ck::index_t Hi              = 30;
+    ck::index_t Wi              = 30;
+    ck::index_t window_stride_d = 2;
+    ck::index_t window_stride_h = 2;
+    ck::index_t window_stride_w = 2;
+    ck::index_t in_left_pad_d   = 1;
+    ck::index_t in_left_pad_h   = 1;
+    ck::index_t in_left_pad_w   = 1;
+    ck::index_t in_right_pad_d  = 1;
+    ck::index_t in_right_pad_h  = 1;
+    ck::index_t in_right_pad_w  = 1;
+
+    bool pass = pool3d_test<InDataType,
+                            OutDataType,
+                            ComputeDataType,
+                            IndexDataType,
+                            InLayout,
+                            OutLayout,
+                            ReduceOpId,
+                            PropagateNan,
+                            OutputIndex>(do_verification,
+                                         time_kernel,
+                                         N,
+                                         C,
+                                         Z,
+                                         Y,
+                                         X,
+                                         Di,
+                                         Hi,
+                                         Wi,
+                                         window_stride_d,
+                                         window_stride_h,
+                                         window_stride_w,
+                                         in_left_pad_d,
+                                         in_left_pad_h,
+                                         in_left_pad_w,
+                                         in_right_pad_d,
+                                         in_right_pad_h,
+                                         in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}

include/ck/ck.hpp

@@ -27,32 +27,29 @@
 #define CK_WAVELET_MIN_BLOCK_PER_CU 2
 #endif
 
-// check GPU target
-#ifdef __HIP_DEVICE_COMPILE__
-#if !(defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
-      defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__))
-#error Not supported target
-#endif
+// for wavelet GEMM kernel
+#define CK_WAVELET_MAX_THREAD_PER_BLOCK 512
+#define CK_WAVELET_MIN_BLOCK_PER_CU 2
 #endif
 
 // buffer resource
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
 #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
-    defined(__gfx90a__) // for GPU code
+    defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
-#elif defined(__gfx1100__) // for GPU code
-#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
 // FMA instruction
 #ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
 #elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
 #define CK_USE_AMD_V_MAC_F32
-#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx1030__) // for GPU code
+#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
+    defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16
 #define CK_USE_AMD_V_DOT4_I32_I8
@@ -61,18 +58,22 @@
 // MFMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_MFMA
-#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_MFMA
 #endif
 
-#if defined(__gfx90a__)
+#if(defined(__gfx90a__) || defined(__gfx940__))
 #define CK_USE_AMD_MFMA_BF16_1K_OP
 #endif
 
+#if defined(__gfx940__)
+#define CK_USE_AMD_MFMA_GFX940
+#endif
+
 // WMMA instruction
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_WMMA
-#elif defined(__gfx1100__) // for GPU code
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
 #define CK_USE_AMD_WMMA
 #endif
 
@@ -88,13 +89,13 @@
 // buffer atomic add: floating point
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
-#elif defined(__gfx908__) || defined(__gfx90a__) // for GPU code
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
 #else // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
 #endif
 
-#if defined(__gfx90a__) // for GPU code
+#if(defined(__gfx90a__) || defined(__gfx940__)) // for GPU code
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
 #else
 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
@@ -168,12 +169,21 @@
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
-// workaround: a BF16 attention kernel for gfx908 is likely affected by a compiler issue
-#ifdef __gfx908__
-#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 1
-#else // __gfx90a__, ...
-#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 0
-#endif // __gfx908__
+// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
+#define CK_WORKAROUND_SWDEV_383542 1
+
+// workaround: compiler issue on gfx908
+#define CK_WORKAROUND_SWDEV_388832 1
+// flag to enable (1) or disable (0) the debugging output in some kernels
+#define DEBUG_LOG 0
+
+// denorm test fix, required to work around dissue
+#ifndef CK_WORKAROUND_DENORM_FIX
+#define CK_WORKAROUND_DENORM_FIX 0
+#elif
+// enable only on MI200
+#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
+#endif // CK_WORKAROUND_DENORM_FIX
 
 namespace ck {
 

include/ck/host_utility/kernel_launch.hpp

@@ -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/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_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