Merge branch 'develop' into lwpck-756

.github/CODEOWNERS

+# Documentation files
+docs/* @saadrahim @LisaDelaney
+*.md  @saadrahim @LisaDelaney
+*.rst  @saadrahim @LisaDelaney
+# Header directory
+library/include/*  @saadrahim @LisaDelaney

CMakeLists.txt

@@ -460,7 +460,7 @@ rocm_install(FILES
 )
 
 # Install CK version and configuration files
-install(FILES
+rocm_install(FILES
     ${PROJECT_BINARY_DIR}/include/ck/version.h
     ${PROJECT_BINARY_DIR}/include/ck/config.h
     DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck/

client_example/20_image_to_column/CMakeLists.txt

+add_executable(client_image_to_column image_to_column.cpp)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)

client_example/20_image_to_column/image_to_column.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout = ck::tensor_layout::convolution::GNHWC;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 28; // input H
+static constexpr ck::index_t Wi            = 28; // input W
+static constexpr ck::index_t Ho            = 28; // output H
+static constexpr ck::index_t Wo            = 28; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+
+    std::array<ck::index_t, 2> in_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
+    std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
+
+    // We have NHWGC in memory space (G is dummy)
+    // However, CK's API only accept length and stride with order of GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 2> out_strides{Y * X * C, 1};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * Y * X * C);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(OutDataType) * N * Ho * Wo * Y * X * C;
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}

cmake/DoxygenDoc.cmake

@@ -309,6 +309,8 @@ XML_OUTPUT
 XML_PROGRAMLISTING
 )
 
+set(WARN_AS_ERROR YES)
+
 set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file")
 
 function(add_doxygen_doc)

docs/Contributors_Guide.rst

@@ -2,7 +2,101 @@
 Contributor's Guide
 ===================
 
-Pull-request guidelines
-=======================
+This chapter explains how to get started contributing to the Composable Kernel project and what are
+the contributing rules.
 
-[TODO]
+Getting started
+===============
+
+#. **Documentation:** Before contributing to the library, familiarize yourself with the
+   `Composable Kernel User Guide <https://rocm.docs.amd.com/projects/composable_kernel/en/latest/>`_.
+   It provides insight into the core concepts, environment configuration, and steps to obtain or
+   build the library. You can also find some of this information in the
+   `README file <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/README.md>`_
+   on the project's GitHub page.
+#. **Additional reading:** We also recommend reading a `blog post
+   <https://community.amd.com/t5/instinct-accelerators/amd-composable-kernel-library-efficient-fused-kernels-for-ai/ba-p/553224>`_
+   from the AMD Community portal. It offers a deeper understanding of the library's objectives and
+   showcases its performance capabilities.
+#. **General information:** For broader information about AMD products, consider exploring the
+   `AMD Developer Central portal <https://www.amd.com/en/developer.html>`_.
+
+How do I contribute
+===================
+
+We deeply value contributions from our users. You can make an impact by reporting issues or
+proposing code enhancements through pull requests.
+
+Reporting issues
+----------------
+
+We use `Github issues <https://github.com/ROCmSoftwarePlatform/composable_kernel/issues>`_
+to track public bugs and enhancement requests.
+
+If you encounter an issue with the library, please check if the problem has already been
+reported by searching existing issues on GitHub. If your issue seems unique, please submit a new
+issue. All reported issues must include:
+
+* A comprehensive description of the problem, including:
+
+  * What did you observe?
+  * Why do you think it is a bug (if it seems like one)?
+  * What did you expect to happen? What would indicate the resolution of the problem?
+  * Are there any known workarounds?
+
+* Your configuration details, including:
+
+  * Which GPU are you using?
+  * Which OS version are you on?
+  * Which ROCm version are you using?
+  * Are you using a Docker image? If so, which one?
+
+* Steps to reproduce the issue, including:
+
+  * What actions trigger the issue? What are the reproduction steps?
+
+    * If you build the library from scratch, what CMake command did you use?
+
+  * How frequently does this issue happen? Does it reproduce every time? Or is it a sporadic issue?
+
+Before sumbitting any issue, ensure you have addressed all relevant questions from the checklist.
+
+Creating Pull Requests
+----------------------
+
+You can submit `Pull Requests (PR) on GitHub
+<https://github.com/ROCmSoftwarePlatform/composable_kernel/pulls>`_.
+
+All contributors are required to develop their changes on a separate branch and then create a
+pull requrest to merge their changes into the `develop` branch, which is the default
+development branch in the Composable Kernel project. All external contributors must use their own
+forks of the project to develop their changes.
+
+When submitting a Pull Request you should:
+
+* Describe the change providing information about the motivation for the change and a general
+  description of all code modifications.
+
+* Verify and test the change:
+
+  * Run any relevant existing tests.
+  * Write new tests if added functionality is not covered by current tests.
+
+* Ensure your changes align with the coding style defined in the ``.clang-format`` file located in
+  the project's root directory. We leverage `pre-commit` to run `clang-format` automatically. We
+  highly recommend contributors utilize this method to maintain consistent code formatting.
+  Instructions on setting up `pre-commit` can be found in the project's
+  `README file <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/README.md>`_
+
+* Link your PR to any related issues:
+
+  * If there is an issue that is resolved by your change, please provide a link to the issue in
+    the description of your pull request.
+
+* For larger contributions, structure your change into a sequence of smaller, focused commits, each
+  addressing a particular aspect or fix.
+
+Following the above guidelines ensures a seamless review process and faster assistance from our
+end.
+
+Thank you for your commitment to enhancing the Composable Kernel project! We look forward to collaborating with you.

example/01_gemm/CMakeLists.txt

@@ -6,8 +6,7 @@ if(DL_KERNELS)
   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
     add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
     add_dependencies(example_gemm_dl example_gemm_dl_fp16)
-    add_example_executable(example_gemm_dl_dpp8_fp16 gemm_dl_dpp8_fp16.cpp)
-    add_dependencies(example_gemm_dl example_gemm_dl_dpp8_fp16)
+    add_example_executable(example_gemm_dpp_fp16 gemm_dpp_fp16.cpp)
   endif()
   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
     add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)

example/01_gemm/gemm_dl_dpp8_fp16.cpp → example/01_gemm/gemm_dpp_fp16.cpp

@@ -3,31 +3,33 @@
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp"
 
 using ADataType   = ck::half_t;
 using BDataType   = ck::half_t;
-using CDataType   = ck::half_t;
 using AccDataType = float;
+using CDataType   = ck::half_t;
+
+using F16 = ck::half_t;
 
-using ALayout = Col;
-using BLayout = Row;
+using ALayout = Row;
+using BLayout = Col;
 using CLayout = Row;
 
 using AElementOp = PassThrough;
 using BElementOp = PassThrough;
 using CElementOp = PassThrough;
 
-static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
 
 // clang-format off
-using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDlDpp8
-// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           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|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  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|                |                   |
-// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
-         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      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<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 DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDpp
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MDpp|    NDpp|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|    |    |  Dpp|  Dpp| PerWave| PerWave|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |     |     |        |        | Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+// ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |    |    |     |     |        |        |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   128,    64,    64,    64,   8,   2,   32,    8,       2,       2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               5,               1>;
+// // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;

example/02_gemm_bilinear/CMakeLists.txt

@@ -5,6 +5,9 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
     add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
+    add_example_executable(example_gemm_bilinear_wmma_int8 gemm_bilinear_wmma_int8.cpp)
+endif()
+if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
    set(target 1)
  endif()
 endforeach()

example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(int alpha, int beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<std::int8_t, std::int32_t, std::int8_t>(
+        std::int8_t& e, const std::int32_t& c, const std::int8_t& d) const
+    {
+        e = ck::type_convert<std::int8_t>(alpha_ * c + beta_ * ck::type_convert<std::int32_t>(d));
+    };
+
+    int alpha_;
+    int beta_;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using DDataType        = I8;
+using EDataType        = I8;
+
+using ALayout = Row;
+using BLayout = Row;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<ALayout,
+                                                                    BLayout,
+                                                                    ck::Tuple<DLayout>,
+                                                                    ELayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    ck::Tuple<DDataType>,
+                                                                    EDataType,
+                                                                    AccDataType,
+                                                                    CShuffleDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    CDEElementOp,
+                                                                    GemmSpec,
+                                                                    32,
+                                                                    16,
+                                                                    16,
+                                                                    4,
+                                                                    16,
+                                                                    16,
+                                                                    16,
+                                                                    1,
+                                                                    1,
+                                                                    S<2, 16, 1>,
+                                                                    S<1, 0, 2>,
+                                                                    S<1, 0, 2>,
+                                                                    2,
+                                                                    16,
+                                                                    16,
+                                                                    1,
+                                                                    S<4, 1, 8>,
+                                                                    S<0, 2, 1>,
+                                                                    S<0, 2, 1>,
+                                                                    1,
+                                                                    16,
+                                                                    2,
+                                                                    1,
+                                                                    1,
+                                                                    1,
+                                                                    S<1, 16, 1, 2>,
+                                                                    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    int alpha = 1;
+    int beta  = 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 == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                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<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    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 << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(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});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_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());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * 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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        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), d_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) ? 0 : 1;
+    }
+
+    return 0;
+}

example/52_image_to_column/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_image_to_column)
+   add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
+   add_dependencies(example_image_to_column example_image_to_column_f32)
+   set(target 1)
+ endif()
+endforeach()

example/52_image_to_column/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/convolution_parameter.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/reference_tensor_operation/cpu/reference_image_to_column.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static inline constexpr ck::index_t NDimSpatial = 2;
+
+using FP32 = float;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParams                                                            \
+    ck::utils::conv::ConvParam                                                       \
+    {                                                                                \
+        NDimSpatial, 1, 32, 1, 1, {4, 4}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, { 0, 0 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params                       = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}

example/52_image_to_column/image_to_column_f32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = FP32;
+using OutDataType = FP32;
+
+using InLayout = ck::tensor_layout::convolution::GNHWC;
+
+// clang-format off
+using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumnImpl
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, InLayout, InDataType, OutDataType,   256,   128,   128, S<16, 16>,     1>;
+// clang-format on
+
+bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
+{
+
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck::index_t NDoHoWo =
+        N * ck::accumulate_n<ck::index_t>(
+                conv_params.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        C * ck::accumulate_n<ck::index_t>(
+                conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+    const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> output_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), output_m_k_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> in(in_desc);
+    Tensor<OutDataType> out_device(out_desc);
+    Tensor<OutDataType> out_host(out_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out_device.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+
+    // reset input to zero
+    out_device_buf.SetZero();
+
+    static_assert(std::is_default_constructible_v<DeviceImgToColInstance>);
+
+    // do conv
+    auto img2col  = DeviceImgToColInstance{};
+    auto invoker  = img2col.MakeInvoker();
+    auto argument = img2col.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                         out_device_buf.GetDeviceBuffer(),
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         input_g_n_c_wis_strides,
+                                         output_m_k_strides,
+                                         conv_filter_strides,
+                                         conv_filter_dilations,
+                                         input_left_pads,
+                                         input_right_pads);
+
+    if(!img2col.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_img2col with the specified compilation parameters does "
+                     "not support this img2col problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time        = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    std::size_t num_btype = NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_image_to_column = ck::tensor_operation::host::
+            ReferenceImageToColumn<NDimSpatial, InLayout, InDataType, OutDataType>();
+
+        auto ref_invoker = ref_image_to_column.MakeInvoker();
+
+        auto ref_argument = ref_image_to_column.MakeArgument(in,
+                                                             out_host,
+                                                             conv_params.filter_spatial_lengths_,
+                                                             conv_params.conv_filter_strides_,
+                                                             conv_params.conv_filter_dilations_,
+                                                             conv_params.input_left_pads_,
+                                                             conv_params.input_right_pads_);
+
+        if(!ref_image_to_column.IsSupportedArgument(&ref_argument))
+        {
+            std::cerr << "wrong! ref_img2col with the specified compilation parameters does "
+                         "not support this img2col problem"
+                      << std::endl;
+            return false;
+        }
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device.mData, out_host.mData);
+    }
+
+    return true;
+}
+
+int RunImageToColumnExample(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    return !RunImageToColumn(config, conv_params);
+}
+
+int main(int argc, char* argv[]) { return RunImageToColumnExample(argc, argv); }

include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/utility/amd_gemm_dpp.hpp"
-#include "ck/utility/common_header.hpp"
-#include "ck/tensor_description/tensor_adaptor.hpp"
-#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp"
-#include "ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp"
-
-namespace ck {
-
-/**
- * DPP8 version of blockwise GEMM algorithm. It uses DPP8 instruction modifier to limit
- * the data loaded from LDS to registers.
- *
- * The algorithm groups threads into groups of size `dpp8::lane_group_size` and splits the matrix C
- * between them in such a way that threads from the same group need the same chunk of either
- * matrix A (or B, respectively). Without the usage of DPP8, each thread would need to load the
- * whole chunk from LDS to its own register space.
- * Usage of DPP8 modifiers allow each thread to load less data, exactly `1 / dpp8::lane_group_size`
- * of the chunk, and then share that data with other threads from the same lane group.
- *
- * Assumptions coming from the usage of DPP8:
- *   1. `BM10BN10ThreadClusterBM10Xs[1] == dpp8::lane_group_size` or
- *      `BM10BN10ThreadClusterBN10Xs[1] == dpp8::lane_group_size` -
- *        - it makes consecutive `dpp8::lane_group_size` threads use the same chunk of either
- *          matrix A or B;
- *        - based on these values we determine which matrix to share.
- *   2. `BM1PerThreadBM11 % dpp8::lane_group_size == 0` (if sharing A) or
- *      `BN1PerThreadBN11 % dpp8::lane_group_size == 0` (if sharing B) -
- *        - we have to make sure that the data to split is divisible by the number of
- *          threads in the group.
- *
- * General algorithm:
- * C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
- * A and B are visible to the whole block, C is distributed among each thread
- * Assume:
- *   1. A:
- *     1. ABlockDesc_BK0_BM_BK1 is known at compile-time
- *     2. ABlockBuffer is DynamicBuffer
- *   2. B:
- *     1. BBlockDesc_BK0_BN_BK1 is known at compile-time
- *     2. BBlockBuffer is DynamicBuffer
- *   3. C:
- *     1. CThreadDesc_BM0_BM11_BN0_BN11 is known at compile-time
- *     2. CThreadBuffer is StaticBuffer
- *   4. BM10BN10ThreadClusterBM10Xs::Size() = BM10BN10ThreadClusterBN10Xs::Size() == 2
- */
-template <index_t BlockSize,
-          typename FloatA,
-          typename FloatB,
-          typename FloatC,
-          typename ABlockDesc_BK0_BM_BK1,
-          typename BBlockDesc_BK0_BN_BK1,
-          index_t BM1PerThreadBM11,
-          index_t BN1PerThreadBN11,
-          index_t BK0PerThread,
-          typename BM10BN10ThreadClusterBM10Xs, // Sequence<BM10BN10ThreadClusterBM100,
-                                                //          BM10BN10ThreadClusterBM101, ...>
-          typename BM10BN10ThreadClusterBN10Xs, // Sequence<BM10BN10ThreadClusterBN100,
-                                                //          BM10BN10ThreadClusterBN101, ...>
-          index_t AThreadCopyScalarPerVector_BM11,
-          index_t BThreadCopyScalarPerVector_BN11,
-          typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
-                                 BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
-                             bool>::type = false>
-struct BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0
-{
-    using AIndex = MultiIndex<4>;
-    using BIndex = MultiIndex<4>;
-    using CIndex = MultiIndex<4>;
-
-    static constexpr auto I0 = Number<0>{};
-    static constexpr auto I1 = Number<1>{};
-    static constexpr auto I2 = Number<2>{};
-    static constexpr auto I3 = Number<3>{};
-
-    static constexpr index_t BK0 = ABlockDesc_BK0_BM_BK1{}.GetLength(I0);
-    static constexpr index_t BK1 = ABlockDesc_BK0_BM_BK1{}.GetLength(I2);
-    static constexpr index_t BM  = ABlockDesc_BK0_BM_BK1{}.GetLength(I1);
-    static constexpr index_t BN  = BBlockDesc_BK0_BN_BK1{}.GetLength(I1);
-
-    static constexpr index_t BM100 = BM10BN10ThreadClusterBM10Xs{}[I0];
-    static constexpr index_t BN100 = BM10BN10ThreadClusterBN10Xs{}[I0];
-
-    static constexpr index_t BM101 = BM10BN10ThreadClusterBM10Xs{}[I1];
-    static constexpr index_t BN101 = BM10BN10ThreadClusterBN10Xs{}[I1];
-
-    static constexpr index_t BM11 = BM1PerThreadBM11;
-    static constexpr index_t BN11 = BN1PerThreadBN11;
-
-    static constexpr index_t BM1 = BM100 * BM101 * BM11;
-    static constexpr index_t BN1 = BN100 * BN101 * BN11;
-
-    static constexpr index_t BM0 = BM / BM1;
-    static constexpr index_t BN0 = BN / BN1;
-
-    // We assume that either `BM101` or `BN101` is equal to `dpp8::lane_group_size`. It makes all
-    // threads in a lane group need the same chunk of B or A matrices and we can share them using
-    // DPP.
-    static_assert(BM101 == dpp8::lane_group_size || BN101 == dpp8::lane_group_size);
-    static constexpr bool ShareB = BM101 == dpp8::lane_group_size ? true : false;
-    static constexpr bool ShareA = !ShareB;
-
-    // If DPP shares A (B, respectively), lane group gets `BM1PerThreadBM11` (`BN1PerThreadBN11`,
-    // respectively) elements, so we split them between threads in lane group so each thread loads
-    // less data from LDS.
-    static constexpr index_t BM1PerThread =
-        ShareA ? BM1PerThreadBM11 / dpp8::lane_group_size : BM1PerThreadBM11;
-    static constexpr index_t BN1PerThread =
-        ShareB ? BN1PerThreadBN11 / dpp8::lane_group_size : BN1PerThreadBN11;
-
-    __host__ __device__ static constexpr auto
-    MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
-    {
-        const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
-            a_block_desc_bk0_bm_bk1,
-            make_tuple(make_pass_through_transform(Number<BK0>{}),
-                       make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
-                       make_pass_through_transform(Number<BK1>{})),
-            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
-
-        return a_block_bk0_bm0_bm1_bk1;
-    }
-
-    __host__ __device__ static constexpr auto
-    MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
-    {
-        const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
-            b_block_desc_bk0_bn_bk1,
-            make_tuple(make_pass_through_transform(Number<BK0>{}),
-                       make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
-                       make_pass_through_transform(Number<BK1>{})),
-            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
-
-        return b_block_desc_bk0_bn0_bn1_bk1;
-    }
-
-    __host__ __device__ static constexpr auto
-    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM_BN()
-    {
-        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
-        // lower: [BM, BN]
-        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n =
-            make_single_stage_tensor_adaptor(
-                make_tuple(make_unmerge_transform(make_tuple(
-                               Number<BM0>{}, Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
-                           make_unmerge_transform(make_tuple(
-                               Number<BN0>{}, Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
-
-        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n;
-    }
-
-    __host__ __device__ static constexpr auto
-    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1()
-    {
-        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
-        // lower: [BM0, BM1, BN0, BN1]
-        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1 =
-            make_single_stage_tensor_adaptor(
-                make_tuple(make_pass_through_transform(Number<BM0>{}),
-                           make_unmerge_transform(
-                               make_tuple(Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
-                           make_pass_through_transform(Number<BN0>{}),
-                           make_unmerge_transform(
-                               make_tuple(Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
-
-        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1;
-    }
-
-    __host__ __device__ static constexpr auto GetCThreadTensorLengths_BM0_BM1_BN0_BN1()
-    {
-        return Sequence<BM0, BM11, BN0, BN11>{};
-    }
-
-    static constexpr auto a_block_desc_bk0_bm0_bm1_bk1_ =
-        MakeABlockDescriptor_BK0_BM0_BM1_BK1(ABlockDesc_BK0_BM_BK1{});
-
-    static constexpr auto b_block_desc_bk0_bn0_bn1_bk1_ =
-        MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
-
-    public:
-    __device__ BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0()
-        : c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
-              get_thread_local_1d_id())},
-          a_thread_copy_{CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()},
-          b_thread_copy_{CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()}
-    {
-        static_assert(ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
-                          BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
-                      "wrong! Desc should be known at compile-time");
-
-        static_assert(BM % BM1 == 0 && BN % BN1 == 0, "wrong!");
-
-        static_assert(ABlockDesc_BK0_BM_BK1{}.GetLength(I0) ==
-                          BBlockDesc_BK0_BN_BK1{}.GetLength(I0),
-                      "wrong! K dimension not consistent");
-
-        static_assert(BM10BN10ThreadClusterBM10Xs::Size() == 2 &&
-                          BM10BN10ThreadClusterBN10Xs::Size() == 2,
-                      "wrong!");
-    }
-
-    __device__ static CIndex CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(index_t thread_id)
-    {
-        // lower: [BM0, BM1, BN0, BN1]
-        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
-        constexpr auto adaptor0 =
-            MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1();
-
-        // lower: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
-        // upper: [Tid, BM0, BM11, BN0, BN11]
-        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
-            make_tuple(make_merge_transform(make_tuple(BM100, BN100, BM101, BN101)),
-                       make_pass_through_transform(BM0),
-                       make_pass_through_transform(BM11),
-                       make_pass_through_transform(BN0),
-                       make_pass_through_transform(BN11)),
-            make_tuple(
-                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
-            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
-
-        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
-
-        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
-    }
-
-    __device__ AIndex CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()
-    {
-        const auto offsetBM0 = c_thread_origin_data_idx_[I0];
-        // If sharing matrix A, we need a separate BM1 offset for each thread in lane group.
-        const auto offsetBM1 = ShareA ? c_thread_origin_data_idx_[I1] +
-                                            dpp8::get_thread_idx_in_lane_group() * BM1PerThread
-                                      : c_thread_origin_data_idx_[I1];
-        return make_tuple(0, offsetBM0, offsetBM1, 0);
-    }
-
-    __device__ BIndex CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()
-    {
-        const auto offsetBN0 = c_thread_origin_data_idx_[I2];
-        // If sharing matrix B, we need a separate BN1 offset for each thread in lane group.
-        const auto offsetBN1 = ShareB ? c_thread_origin_data_idx_[I3] +
-                                            dpp8::get_thread_idx_in_lane_group() * BN1PerThread
-                                      : c_thread_origin_data_idx_[I3];
-        return make_tuple(0, offsetBN0, offsetBN1, 0);
-    }
-
-    template <typename CThreadDesc_BM0_BM11_BN0_BN11,
-              typename ABlockBuffer,
-              typename BBlockBuffer,
-              typename CThreadBuffer>
-    __device__ void Run(const CThreadDesc_BM0_BM11_BN0_BN11&,
-                        const ABlockBuffer& a_block_buf,
-                        const BBlockBuffer& b_block_buf,
-                        CThreadBuffer& c_thread_buf) const
-    {
-        static_assert(CThreadDesc_BM0_BM11_BN0_BN11::IsKnownAtCompileTime(),
-                      "wrong! Desc should be known at compile-time");
-
-        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
-            a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
-            b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
-
-        constexpr auto threadwise_contraction =
-            ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
-                FloatA,
-                FloatB,
-                FloatC,
-                decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
-                decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
-                CThreadDesc_BM0_BM11_BN0_BN11,
-                Sequence<BK0PerThread, BK1>,
-                Sequence<1, BM1PerThreadBM11>,
-                Sequence<1, BN1PerThreadBN11>,
-                ShareA>{};
-
-        static_for<0, BN0, 1>{}([&](auto bn0) {
-            static_for<0, BM0, 1>{}([&](auto bm0) {
-                a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
-                                   make_tuple(I0, bm0, I0, I0),
-                                   a_block_buf,
-                                   a_thread_desc_bk0_bm0_bm1_bk1_,
-                                   make_tuple(I0, I0, I0, I0),
-                                   a_thread_buf);
-
-                b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
-                                   make_tuple(I0, bn0, I0, I0),
-                                   b_block_buf,
-                                   b_thread_desc_bk0_bn0_bn1_bk1_,
-                                   make_tuple(I0, I0, I0, I0),
-                                   b_thread_buf);
-
-                threadwise_contraction.Run(a_thread_buf,
-                                           make_tuple(I0, I0, I0, I0),
-                                           b_thread_buf,
-                                           make_tuple(I0, I0, I0, I0),
-                                           c_thread_buf,
-                                           make_tuple(bm0, I0, bn0, I0));
-
-                static_for<BK0PerThread, BK0, BK0PerThread>{}([&](auto bk0) {
-                    a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
-                                       make_tuple(bk0, bm0, I0, I0),
-                                       a_block_buf,
-                                       a_thread_desc_bk0_bm0_bm1_bk1_,
-                                       make_tuple(I0, I0, I0, I0),
-                                       a_thread_buf);
-
-                    b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
-                                       make_tuple(bk0, bn0, I0, I0),
-                                       b_block_buf,
-                                       b_thread_desc_bk0_bn0_bn1_bk1_,
-                                       make_tuple(I0, I0, I0, I0),
-                                       b_thread_buf);
-
-                    threadwise_contraction.Run(a_thread_buf,
-                                               make_tuple(I0, I0, I0, I0),
-                                               b_thread_buf,
-                                               make_tuple(I0, I0, I0, I0),
-                                               c_thread_buf,
-                                               make_tuple(bm0, I0, bn0, I0));
-                });
-            });
-        });
-    }
-
-    private:
-    // A[BK0, BM0, BM1, BK1]
-    static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ = make_naive_tensor_descriptor_packed(
-        make_tuple(Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThread>{}, Number<BK1>{}));
-
-    // B[BK0, BN0, BN1, BK1]
-    static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ = make_naive_tensor_descriptor_packed(
-        make_tuple(Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThread>{}, Number<BK1>{}));
-
-    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
-        FloatA,
-        FloatA,
-        decltype(a_block_desc_bk0_bm0_bm1_bk1_),
-        decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
-        Sequence<BK0PerThread, 1, BM1PerThread, BK1>, // SliceLengths
-        Sequence<0, 1, 2, 3>,                         // DimAccessOrder
-        Sequence<1, 1, BM1PerThread, BK1>,            // SrcVectorTensorLengths
-        Sequence<0, 1, 2, 3>>;                        // SrcVectorTensorContiguousDimOrder
-
-    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
-        FloatB,
-        FloatB,
-        decltype(b_block_desc_bk0_bn0_bn1_bk1_),
-        decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
-        Sequence<BK0PerThread, 1, BN1PerThread, BK1>, // SliceLengths
-        Sequence<0, 1, 2, 3>,                         // DimAccessOrder
-        Sequence<1, 1, BN1PerThread, BK1>,            // SrcVectorTensorLengths
-        Sequence<0, 1, 2, 3>>;                        // SrcVectorTensorContiguousDimOrder
-
-    CIndex c_thread_origin_data_idx_;
-
-    AThreadCopy a_thread_copy_;
-    BThreadCopy b_thread_copy_;
-};
-
-} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/dpp_gemm.hpp"
+
+namespace ck {
+
+/**
+ * Blockwise GEMM that uses DPP instruction modifier to limit the amount of data loaded for each
+ * thread by sharing the data between threads in a lanegroup.
+ *
+ * In every iteration, each wave calculates a C tile of size `MPerDpp` * `NPerDpp`, there are
+ * `MRepeat` iterations for `M` dimension and `NRepeat` for `N` one.
+ * In total, the algorithm runs using
+ * `MPerBlock / (MRepeat * MPerDpp) * NPerBlock / (NRepeat * NPerDpp)` waves.
+ */
+template <index_t BlockSize,
+          typename ABDataType,
+          typename AccDataType,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerDpp,
+          index_t NPerDpp,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t KPerBlock =
+        BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto dpp_gemm = DppGemm<ABDataType, MPerDpp, NPerDpp, KPack>{};
+
+    static constexpr index_t KPerThread = KPerBlock / dpp_gemm.K0PerDpp;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerDpp);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerDpp);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              AccDataType,
+                              MRepeat * NRepeat,
+                              dpp_gemm.GetRegSizePerDpp(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex_M0_M1_M2_K()
+    {
+        const auto wave_idx    = GetWaveIdx();
+        const auto waveId_m    = wave_idx[I0];
+        const auto dpp_a_idx   = dpp_gemm.CalculateAThreadOriginDataIndex_K_M();
+        const auto dpp_a_idx_k = dpp_a_idx[I0];
+        const auto dpp_a_idx_m = dpp_a_idx[I1];
+        return make_tuple(0, waveId_m, dpp_a_idx_m, KPerThread * dpp_a_idx_k);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex_N0_N1_N2_K()
+    {
+        const auto wave_idx    = GetWaveIdx();
+        const auto waveId_n    = wave_idx[I1];
+        const auto dpp_b_idx   = dpp_gemm.CalculateBThreadOriginDataIndex_K_N();
+        const auto dpp_b_idx_k = dpp_b_idx[I0];
+        const auto dpp_b_idx_n = dpp_b_idx[I1];
+        return make_tuple(0, waveId_n, dpp_b_idx_n, KPerThread * dpp_b_idx_k);
+    }
+
+    template <index_t m0, index_t n0>
+    __device__ static auto CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>)
+    {
+        const auto wave_idx = GetWaveIdx();
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx      = dpp_gemm.GetBeginOfThreadBlk();
+        const auto blk_m_offset = blk_idx[I0];
+        const auto blk_n_offset = blk_idx[I1];
+
+        constexpr auto mrepeat_mwave_MPerDpp_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerDpp))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_NPerDpp_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerDpp))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_MPerDpp_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_m_offset))[I0];
+        const index_t c_thread_n = nrepeat_nwave_NPerDpp_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_n_offset))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0NK1BlockDesc::IsKnownAtCompileTime(),
+                      "Wrong! Block descriptors should be known at the time of compilation.");
+
+#if defined(__HIP_DEVICE_COMPILE__)
+        // Host wave size can be different than the device one and this assert could fail for host,
+        // but it does matter only for device.
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+#endif
+
+        static_assert(MPerBlock % (MPerDpp * MRepeat) == 0,
+                      "Invalid parameters. MPerBlock must be divisible by MPerDpp * MRepeat.");
+        static_assert(NPerBlock % (NPerDpp * NRepeat) == 0,
+                      "Invalid parameters. NPerBlock must be divisible by NPerDpp * NRepeat.");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2()
+    {
+        constexpr auto c_m_n_tblk_lens = dpp_gemm.GetCMNThreadBlkLengths();
+        constexpr auto M               = c_m_n_tblk_lens[I0];
+        constexpr auto N               = c_m_n_tblk_lens[I1];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_N2()
+    {
+        constexpr auto c_m_n_tblk_lens = dpp_gemm.GetCMNThreadBlkLengths();
+        constexpr auto M               = c_m_n_tblk_lens[I0];
+        constexpr auto N               = c_m_n_tblk_lens[I1];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerDpp>{},
+                                                           Number<NPerDpp>{}));
+
+        return c_block_desc_m0_n0_m1_n1_m2_n2;
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerDpp>{},
+                                                           Number<NPerDpp>{}));
+        return c_block_desc_g_m0_n0_m1_n1_m2_n2;
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerDpp), MWaves, MPerDpp)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerDpp), NWaves, NPerDpp))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return c_grid_desc_m0_n0_m1_n1_m2_n2;
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerDpp), MWaves, MPerDpp)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerDpp), NWaves, NPerDpp))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return c_grid_desc_g_m0_n0_m1_n1_m2_n2;
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerDpp>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
+    {
+        return transform_tensor_descriptor(
+            BK0NK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerDpp>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+    static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ABDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ABDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf);
+
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<ABDataType, KPack> a_thread_vec;
+                    vector_type<ABDataType, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<ABDataType>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<ABDataType>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                    });
+
+                    using dpp_input_type =
+                        typename vector_type<ABDataType, dpp_gemm.K1PerDpp>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    dpp_gemm.template Run(a_thread_vec.template AsType<dpp_input_type>(),
+                                          b_thread_vec.template AsType<dpp_input_type>(),
+                                          c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // C[M, N, NumRegDpp]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, dpp_gemm.GetRegSizePerDpp()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ABDataType,
+                                                         ABDataType,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<ABDataType,
+                                                         ABDataType,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex_M0_M1_M2_K()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex_N0_N1_N2_K()};
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp

@@ -221,49 +221,102 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
         auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
             b_thread_desc_.GetElementSpaceSize());
 
-        static_for<0, KPerBlock / WmmaK, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
-            static_for<0, MRepeat, 1>{}([&](auto m0) {
-                // read A
-                a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
-                                   make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
-                                   a_block_buf,
-                                   a_thread_desc_,
-                                   make_tuple(I0, m0, I0, I0, I0),
-                                   a_thread_buf);
-
-                static_for<0, NRepeat, 1>{}([&](auto n0) {
-                    // read B
-                    b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
-                                       make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
-                                       b_block_buf,
-                                       b_thread_desc_,
-                                       make_tuple(I0, n0, I0, I0, I0),
-                                       b_thread_buf);
-                    vector_type<FloatA, WmmaK> a_thread_vec;
-                    vector_type<FloatB, WmmaK> b_thread_vec;
-
-                    static_for<0, WmmaK, 1>{}([&](auto i) {
-                        a_thread_vec.template AsType<FloatA>()(i) =
-                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
-                                make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
-                        b_thread_vec.template AsType<FloatB>()(i) =
-                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
-                                make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
+        // basic intrinsic to determine loopover direction
+        if constexpr(MRepeat < NRepeat)
+        {
+            static_for<0, KPerBlock / WmmaK, 1>{}(
+                [&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        // read A
+                        a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
+                                           make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(I0, m0, I0, I0, I0),
+                                           a_thread_buf);
+
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            // read B
+                            b_thread_copy_.Run(
+                                b_block_desc_k0_n0_n1_n2_k1,
+                                make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
+                                b_block_buf,
+                                b_thread_desc_,
+                                make_tuple(I0, n0, I0, I0, I0),
+                                b_thread_buf);
+                            vector_type<FloatA, WmmaK> a_thread_vec;
+                            vector_type<FloatB, WmmaK> b_thread_vec;
+
+                            static_for<0, WmmaK, 1>{}([&](auto i) {
+                                a_thread_vec.template AsType<FloatA>()(i) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
+                                b_thread_vec.template AsType<FloatB>()(i) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
+                            });
+
+                            using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                            using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            wmma_gemm.template Run(
+                                a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                                b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
                     });
-
-                    using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
-                    using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
-
-                    constexpr index_t c_offset =
-                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
-
-                    wmma_gemm.template Run(
-                        a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
-                        b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
-                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
                 });
-            });
-        });
+        }
+        else
+        {
+            static_for<0, KPerBlock / WmmaK, 1>{}(
+                [&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        // read B
+                        b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
+                                           make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(I0, n0, I0, I0, I0),
+                                           b_thread_buf);
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            // read A
+                            a_thread_copy_.Run(
+                                a_block_desc_k0_m0_m1_m2_k1,
+                                make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
+                                a_block_buf,
+                                a_thread_desc_,
+                                make_tuple(I0, m0, I0, I0, I0),
+                                a_thread_buf);
+                            vector_type<FloatA, WmmaK> a_thread_vec;
+                            vector_type<FloatB, WmmaK> b_thread_vec;
+
+                            static_for<0, WmmaK, 1>{}([&](auto i) {
+                                a_thread_vec.template AsType<FloatA>()(i) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
+                                b_thread_vec.template AsType<FloatB>()(i) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
+                            });
+
+                            using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                            using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                            wmma_gemm.template Run(
+                                a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                                b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                                c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        });
+                    });
+                });
+        }
     }
 
     protected:

include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp

@@ -4,27 +4,13 @@
 #pragma once
 
 #include "ck/utility/common_header.hpp"
+#include "ck/utility/loop_scheduler.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
 #include "ck/tensor_description/tensor_adaptor.hpp"
 
 namespace ck {
 
-enum struct LoopScheduler
-{
-    Default,
-    Interwave,
-};
-
-constexpr LoopScheduler make_default_loop_scheduler()
-{
-#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
-    return LoopScheduler::Interwave;
-#else
-    return LoopScheduler::Default;
-#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
-}
-
 template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
 __host__ __device__ static constexpr auto
 MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)

include/ck/tensor_operation/gpu/device/device_image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/**
+ * \brief Image to column.
+ *
+ * This Device operator converts image ([G, N, Di, Hi, Wi, C]) to the gemm
+ * problem([N * Do * Ho * Wo, Z *  Y * X * C]). G must be equal to 1.
+ *
+ * \tparam NDimSpatial Number of spatial dimensions.
+ * \tparam InputLayout Input Layout.
+ * \tparam InputDataType Input Data Type.
+ * \tparam OutputDataType Output Data Type.
+ */
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType>
+struct DeviceImageToColumn : public BaseOperator
+{
+
+    /**
+     * \brief Make argument pointer for image to column.
+     *
+     * \param p_in A pointer to the device memory of the input image.
+     * \param p_out A pointer to the device memory of the output.
+     * \param N Convolution batch size.
+     * \param C Convolution number of channels.
+     * \param input_spatial_lengths Input spatial lengths.
+     * \param filter_spatial_lengths Filter spatial lengths.
+     * \param output_spatial_lengths Output spatial lengths.
+     * \param input_g_n_c_wis_strides Input strides in order [G, N, C, D, H, W].
+     * \param output_m_k_strides Output strides.
+     * \param conv_filter_strides Convolution filter strides.
+     * \param conv_filter_dilations Convolution filter dilations.
+     * \param input_left_pads Convolution left pads.
+     * \param input_right_pads Convolution right pads.
+     * \return Pointer to the argument.
+     */
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_out,
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-
-enum struct GemmDlAlgorithm
-{
-    Default, // Uses DOT vector instructions
-    Dpp8,    // Uses DOT vector instructions with DPP8 SEL modifier to reduce data loads from LDS
-};
-
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp

@@ -185,7 +185,7 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
         AElementwiseOperation,
         BElementwiseOperation,
         CElementwiseOperation,
-        GemmSpecialization::MNPadding,
+        GemmSpecialization::MNKPadding,
         MPerBlock,
         NPerBlock,
         K0PerBlock,
@@ -315,11 +315,6 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
             return false;
         }
 
-        if(problem.K % K1 != 0)
-        {
-            return false;
-        }
-
         return GridwiseGemm::CheckValidity(problem);
     }
 
@@ -416,7 +411,12 @@ struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
             << BlockSize << ", "
             << MPerBlock << ", "
             << NPerBlock << ", "
-            << K0PerBlock
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
             << ">"
             << " NumGemmKPrefetchStage: "
             << NumGemmKPrefetchStage << ", "