Merge remote-tracking branch 'upstream/develop' into merge_upstream_1129

also fix regression

example/14_gemm_quantization/CMakeLists.txt

-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 # dlops
-if(DL_KERNELS)
-  add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
-endif()
-
+add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
 # xdlops
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
@@ -14,4 +10,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-endif()
\ No newline at end of file

example/15_grouped_gemm/CMakeLists.txt

 add_custom_target(example_grouped_gemm_xdl)
+add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
+
+add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp16)
+
+add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_dl_fp16)
+
+add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_splitk_fp16)
+
+add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16)
+
+add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_bias_fp16)
+
+add_example_executable(example_grouped_gemm_xdl_bf16 grouped_gemm_xdl_bf16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bf16)
+
+add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
+
+add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
 
-if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-  add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
-  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
-endif()
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
-  add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
-  add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
-  add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
-  add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
-  add_dependencies(example_grouped_gemm_xdl 
-                   example_grouped_gemm_xdl_fp16
-                   example_grouped_gemm_multiple_d_dl_fp16
-                   example_grouped_gemm_xdl_splitk_fp16
-                   example_grouped_gemm_xdl_fixed_nk_fp16
-                   example_grouped_gemm_xdl_fixed_nk_bias_fp16)
-endif()
-if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-  add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
-  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bfp16)
-endif()
-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-  add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
-  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
-endif()
 if(USE_BITINT_EXTENSION_INT4)
-  add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
-  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
+    add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
+    add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
 endif()

example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp

@@ -296,13 +296,11 @@ int main(int argc, char* argv[])
 
     problem_size.group_count = 16;
 
-    problem_size.Ms = {
-        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
-
     for(int i = 0; i < problem_size.group_count; i++)
     {
-        problem_size.Ns.push_back(768);
-        problem_size.Ks.push_back(4608);
+        problem_size.Ms.push_back(256 + 256 * i);
+        problem_size.Ns.push_back(128 + 128 * i);
+        problem_size.Ks.push_back(128 + 64 * i);
 
         problem_size.stride_As.push_back(problem_size.Ks[i]);
         problem_size.stride_Bs.push_back(problem_size.Ks[i]);

example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F8;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Fixed_NK
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>;
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    int k_batch          = 1;
+    bool time_kernel     = false;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<void*> p_Cs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 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});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
+                                      a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
+        c_tensors_device[i]->SetZero();
+
+        p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              1,
+                              problem_size.stride_Bs[i],
+                              1,
+                              {}});
+
+        grouped_gemm_kernel_args_.push_back({a_tensors_device[i]->GetDeviceBuffer(),
+                                             b_tensors_device[i]->GetDeviceBuffer(),
+                                             {},
+                                             c_tensors_device[i]->GetDeviceBuffer(),
+                                             problem_size.Ms[i],
+                                             problem_size.Ns[i],
+                                             problem_size.Ks[i],
+                                             problem_size.stride_As[i],
+                                             problem_size.stride_Bs[i],
+                                             {},
+                                             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<const void*> p_As                = {};
+    std::vector<const void*> p_Bs                = {};
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        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, " << gemm.GetTypeString() << std::endl;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(256 + 256 * i);
+        problem_size.Ns.push_back(128 + 128 * i);
+        problem_size.Ks.push_back(128 + 64 * i);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    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=n0, 1=yes)\n");
+        printf("arg4: k_batch (> 0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp

@@ -66,13 +66,11 @@ int main(int argc, char* argv[])
 
     problem_size.group_count = 16;
 
-    problem_size.Ms = {
-        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
-
     for(int i = 0; i < problem_size.group_count; i++)
     {
-        problem_size.Ns.push_back(768);
-        problem_size.Ks.push_back(4608);
+        problem_size.Ms.push_back(256 + 256 * i);
+        problem_size.Ns.push_back(128 + 128 * i);
+        problem_size.Ks.push_back(128 + 64 * i);
 
         problem_size.stride_As.push_back(problem_size.Ks[i]);
         problem_size.stride_Bs.push_back(problem_size.Ks[i]);

example/16_gemm_multi_d_multi_reduces/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_gemm_reduce_xdl)
-   add_custom_target(example_gemm_reduce_xdl_max)
-   add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
-   add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
-   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
-    add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
-    add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
-    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
-    add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
-    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
-   endif()
-   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-    add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
-    add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
-    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
-    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
-   endif()
-   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-    add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
-    add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
-    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
-    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
-   endif()
-   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
-    add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
-    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
-    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
-   endif()
-   
-   add_dependencies(example_gemm_reduce_xdl
-                 example_gemm_reduce_xdl_mean_meansquare
-                 example_gemm_reduce_xdl_max
-                 example_gemm_add_add_mean_meansquare_xdl)
-
-   if(USE_BITINT_EXTENSION_INT4)
-      add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
-      add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
-   endif()
-   set(target 1)
- endif()
+    if(gpu IN_LIST gpu_list AND target EQUAL 0)
+        add_custom_target(example_gemm_reduce_xdl)
+        add_custom_target(example_gemm_reduce_xdl_max)
+        add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
+        add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
+
+        add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
+
+        add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
+        add_example_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
+
+        add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
+
+        add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
+
+        add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
+
+        add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
+
+        add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
+
+        add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
+
+        add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
+        add_example_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
+
+        add_example_dependencies(example_gemm_reduce_xdl
+            example_gemm_reduce_xdl_mean_meansquare
+            example_gemm_reduce_xdl_max
+            example_gemm_add_add_mean_meansquare_xdl)
+
+        if(USE_BITINT_EXTENSION_INT4)
+            add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
+            add_example_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
+        endif()
+        set(target 1)
+    endif()
 endforeach()

example/17_convnd_bwd_data/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 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_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
-   target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
+   if(result EQUAL 0)
+      target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
+   endif()
    set(target 1)
  endif()
 endforeach()
-  if(DL_KERNELS)
-    add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
-    target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
-  endif()
+
+add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
+if(result EQUAL 0)
+  target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
 endif()

example/18_batched_gemm_reduce/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -7,4 +6,3 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-endif()

example/20_grouped_conv_bwd_weight/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list AND target EQUAL 0)
-   add_custom_target(example_grouped_conv_bwd_weight)
-   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
-    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
-   endif()
-   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
-    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
-   endif()
-   set(target 1)
- endif()
+    if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
+        add_custom_target(example_grouped_conv_bwd_weight)
+        add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
+        add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
+
+        add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
+        add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
+
+        add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8 grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp)
+        add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8)
+        set(target 1)
+    endif()
+
+    if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
+        add_custom_target(example_grouped_conv_bwd_weight)
+        add_example_executable(example_grouped_conv_bwd_weight_wmma_fp16 grouped_conv_bwd_weight_wmma_fp16.cpp)
+        add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_wmma_fp16)
+        set(target 1)
+    endif()
 endforeach()
 
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  if(DL_KERNELS)
-    add_custom_target(example_grouped_conv_bwd_weight_dl)
-    add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
-    add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
-  endif()
-endif()
\ No newline at end of file
+add_custom_target(example_grouped_conv_bwd_weight_dl)
+
+add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
+add_example_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)

example/20_grouped_conv_bwd_weight/common.hpp

@@ -23,6 +23,12 @@
 using BF16 = ck::bhalf_t;
 using F16  = ck::half_t;
 using F32  = float;
+#ifdef CK_ENABLE_FP8
+using F8 = ck::f8_t;
+#endif
+#ifdef CK_ENABLE_BF8
+using BF8 = ck::bf8_t;
+#endif
 
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
@@ -40,25 +46,21 @@ struct CommonLayoutSetting
     using OutputLayout = OutputLay;
 };
 
-template <ck::index_t NDimSpatial>
-struct CommonLayoutSettingSelector;
-
 namespace ctl = ck::tensor_layout::convolution;
-
-template <>
-struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
-{
-};
-
-template <>
-struct CommonLayoutSettingSelector<2> final
-    : CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
-{
-};
-
-template <>
-struct CommonLayoutSettingSelector<3> final
-    : CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector
+    : CommonLayoutSetting<ck::tuple_element_t<NDimSpatial - 1,
+                                              ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                                        ck::tensor_layout::convolution::GNHWC,
+                                                        ck::tensor_layout::convolution::GNDHWC>>,
+                          ck::tuple_element_t<NDimSpatial - 1,
+                                              ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                                        ck::tensor_layout::convolution::GKYXC,
+                                                        ck::tensor_layout::convolution::GKZYXC>>,
+                          ck::tuple_element_t<NDimSpatial - 1,
+                                              ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                                        ck::tensor_layout::convolution::GNHWK,
+                                                        ck::tensor_layout::convolution::GNDHWK>>>
 {
 };
 
@@ -78,10 +80,10 @@ struct ExecutionConfig final
     bool time_kernel     = false;
 };
 
-#define DefaultConvParam                                                      \
-    ck::utils::conv::ConvParam                                                \
-    {                                                                         \
-        2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
+#define DefaultConvParam                                                                         \
+    ck::utils::conv::ConvParam                                                                   \
+    {                                                                                            \
+        3, 4, 1, 128, 256, {3, 3, 3}, {14, 14, 14}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, { 1, 1, 1 } \
     }
 
 inline void print_help_msg()

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp

@@ -3,7 +3,7 @@
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp"
 
 using InDataType  = F16;
 using WeiDataType = F16;
@@ -15,45 +15,84 @@ using WeiElementOp = PassThrough;
 using OutElementOp = PassThrough;
 
 template <ck::index_t NDimSpatial>
-using DeviceConvBwdWeightInstance =
-    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl<
-        NDimSpatial,          // NDimSpatial
-        InDataType,           // InDataType
-        WeiDataType,          // WeiDataType
-        OutDataType,          // OutDataType
-        AccDataType,          // AccDataType
-        InElementOp,          // InElementwiseOperation
-        WeiElementOp,         // WeiElementwiseOperation
-        OutElementOp,         // OutElementwiseOperation
-        ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
-        256,                  // BlockSize
-        128,                  // MPerBlock
-        128,                  // NPerBlock
-        16,                   // K0PerBlock
-        2,                    // K1
-        4,                    // M1PerThread
-        4,                    // N1PerThread
-        1,                    // KPerThread
-        S<8, 2>,              // M1N1ThreadClusterM1Xs
-        S<8, 2>,              // M1N1ThreadClusterN1Xs
-        S<1, 8, 1, 1, 2>,     // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
-        S<1, 2, 1, 128, 1>,   // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
-        S<0, 2, 3, 1, 4>,     // ABlockTransferThreadClusterArrangeOrder
-        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcAccessOrder
-        S<1, 1, 1, 1, 1>,     // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
-        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcVectorTensorContiguousDimOrder
-        S<1, 1, 1, 1, 1>,     // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
-        S<1, 1, 1, 8, 2>,     // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
-        S<1, 16, 1, 16, 1>,   // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
-        S<0, 1, 4, 2, 3>,     // BBlockTransferThreadClusterArrangeOrder
-        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcAccessOrder
-        S<1, 1, 1, 8, 1>,     // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
-        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcVectorTensorContiguousDimOrder
-        S<1, 1, 1, 1, 2>,     // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
-        S<0, 1, 2, 3, 4, 5>,  // CThreadTransferSrcDstAccessOrder
-        5,                    // CThreadTransferSrcDstVectorDim
-        4>;                   // CThreadTransferDstScalarPerVector
+using DeviceConvBwdWeightInstance = ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Dl<
+    NDimSpatial, // NDimSpatial
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                  ck::tensor_layout::convolution::GNHWC,
+                                  ck::tensor_layout::convolution::GNDHWC>>, // InLayout
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                  ck::tensor_layout::convolution::GKYXC,
+                                  ck::tensor_layout::convolution::GKZYXC>>, // WeiLayout
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                  ck::tensor_layout::convolution::GNHWK,
+                                  ck::tensor_layout::convolution::GNDHWK>>, // OutLayout
+    InDataType,                                                             // InDataType
+    WeiDataType,                                                            // WeiDataType
+    OutDataType,                                                            // OutDataType
+    AccDataType,                                                            // AccDataType
+    InElementOp,          // InElementwiseOperation
+    WeiElementOp,         // WeiElementwiseOperation
+    OutElementOp,         // OutElementwiseOperation
+    ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
+    256,                  // BlockSize
+    128,                  // MPerBlock
+    128,                  // NPerBlock
+    16,                   // K0PerBlock
+    2,                    // K1
+    4,                    // M1PerThread
+    4,                    // N1PerThread
+    1,                    // KPerThread
+    S<8, 2>,              // M1N1ThreadClusterM1Xs
+    S<8, 2>,              // M1N1ThreadClusterN1Xs
+    S<1, 8, 1, 1, 2>,     // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+    S<1, 2, 1, 128, 1>,   // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+    S<0, 2, 3, 1, 4>,     // ABlockTransferThreadClusterArrangeOrder
+    S<0, 2, 3, 1, 4>,     // ABlockTransferSrcAccessOrder
+    S<1, 1, 1, 1, 1>,     // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+    S<0, 2, 3, 1, 4>,     // ABlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 1, 1>,     // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+    S<1, 1, 1, 8, 2>,     // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+    S<1, 16, 1, 16, 1>,   // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+    S<0, 1, 4, 2, 3>,     // BBlockTransferThreadClusterArrangeOrder
+    S<0, 1, 4, 2, 3>,     // BBlockTransferSrcAccessOrder
+    S<1, 1, 1, 8, 1>,     // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+    S<0, 1, 4, 2, 3>,     // BBlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 1, 2>,     // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+    S<0, 1, 2, 3, 4, 5>,  // CThreadTransferSrcDstAccessOrder
+    5,                    // CThreadTransferSrcDstVectorDim
+    4>;                   // CThreadTransferDstScalarPerVector
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
 
 #include "run_grouped_conv_bwd_weight_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp"
+
+using InDataType  = F16;
+using WeiDataType = F16;
+using OutDataType = F16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Wmma_CShuffle<
+        NDimSpatial,
+        ck::tensor_layout::convolution::GNDHWC,
+        ck::tensor_layout::convolution::GKZYXC,
+        ck::tensor_layout::convolution::GNDHWK,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        4,                    // K0PerBlock
+        8,                    // K1
+        16,                   // MPerWMMA
+        16,                   // NPerWMMA
+        4,                    // MRepeat
+        2,                    // NRepeat
+        S<4, 64, 1>,          // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<0, 2, 1>,           // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1>,           // ABlockTransferSrcAccessOrder
+        1,                    // ABlockTransferSrcVectorDim
+        1,                    // ABlockTransferSrcScalarPerVector
+        8,                    // ABlockTransferDstScalarPerVector_AK1
+        true,                 // ABlockLdsExtraM
+        S<4, 64, 1>,          // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<0, 2, 1>,           // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1>,           // BBlockTransferSrcAccessOrder
+        1,                    // BBlockTransferSrcVectorDim
+        1,                    // BBlockTransferSrcScalarPerVector
+        8,                    // BBlockTransferDstScalarPerVector_BK1
+        true,                 // BBlockLdsExtraN
+        4,
+        2,
+        S<1, 32, 1, 8>,
+        1>;
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp

@@ -67,6 +67,34 @@ using DeviceConvBwdWeightInstance =
         S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
 
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
 #include "run_grouped_conv_bwd_weight_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp

@@ -66,6 +66,34 @@ using DeviceConvBwdWeightInstance =
         S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
 
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
 #include "run_grouped_conv_bwd_weight_example.inc"
 
-int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+
+using InDataType   = F16;
+using WeiDataType  = F16;
+using OutDataType  = F16;
+using AccDataType  = F32;
+using ComputeTypeA = BF8;
+using ComputeTypeB = F8;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        4,                    // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<1, 4, 16, 4>,       // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,        // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        1,                    // ABlockTransferSrcScalarPerVector
+        1,                    // ABlockTransferDstScalarPerVector_K1
+        true,                 // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,       // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,        // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        1,                    // BBlockTransferSrcScalarPerVector
+        1,                    // BBlockTransferDstScalarPerVector_K1
+        true,                 // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        2,                    // CBlockTransferScalarPerVector_NWaveNPerXdl
+        ComputeTypeA,         // ComputeTypeA
+        ComputeTypeB>;        // ComputeTypeB
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp,
+                                                                                     ComputeTypeA,
+                                                                                     ComputeTypeB>;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
-template <ck::index_t NDimSpatial>
-using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
-                                                                                     InDataType,
-                                                                                     WeiDataType,
-                                                                                     OutDataType,
-                                                                                     InElementOp,
-                                                                                     WeiElementOp,
-                                                                                     OutElementOp>;
-
 template <ck::index_t NDimSpatial>
 bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
                                  const ck::utils::conv::ConvParam& conv_param)
 {
-    ck::index_t split_k;
-    // Set split_k = 2 for xdl op, split_k = 1 for dl
-    // Dl op doesn't support split_k > 1
-    // TODO: Add Dl op split_k > 1 support
-    if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
-         ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
-         ck::get_device_name() == "gfx1102"))
-    {
-        split_k = 2;
-    }
-    else
-    {
-        split_k = 1;
-    }
+    // Dl and WMMA ops don't support split_k > 1
+    constexpr ck::index_t split_k = 1;
 
     const auto in_g_n_c_wis_desc =
         ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
@@ -58,8 +37,8 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
         out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
         break;
     default:
-        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
-        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 0.2});
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.1, 0.1});
     }
 
     DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
@@ -125,18 +104,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
         return true;
     }
 
-    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
-
-    std::size_t flop      = conv_param.GetFlops();
-    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
-
-    float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
-
-    float gb_per_sec = num_btype / 1.E6 / avg_time;
-
-    std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
-              << std::endl
-              << "DeviceOp: " << conv.GetTypeString() << std::endl;
+    invoker.Run(argument, StreamConfig{nullptr, false});
 
     if(config.do_verification)
     {
@@ -160,25 +128,18 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
         return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
     }
 
-    return true;
-}
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
 
-bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
-{
-    ExecutionConfig config;
-    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
 
-    if(!parse_cmd_args(argc, argv, config, conv_param))
-    {
-        return false;
-    }
+    float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
 
-    switch(conv_param.num_dim_spatial_)
-    {
-    case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
-    case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
-    case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
-    }
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+    std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl
+              << "DeviceOp: " << conv.GetTypeString() << std::endl;
 
-    return false;
+    return true;
 }

example/21_gemm_layernorm/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
 list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -10,4 +9,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
    set(target 1)
  endif()
 endforeach()
-endif()
+

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp

@@ -114,12 +114,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
                                                                               BetaDataType,
                                                                               HDataType,
                                                                               AccDataType,
+                                                                              AccDataType,
                                                                               HElementOp,
                                                                               2,
                                                                               1>;
 
     Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
     Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+    Tensor<AccDataType> save_mean({M});
+    Tensor<AccDataType> save_inv_std({M});
 
     auto ref_gemm         = ReferenceGemm{};
     auto ref_gemm_invoker = ref_gemm.MakeInvoker();
@@ -145,7 +148,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
     auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
 
     auto ref_layernorm_argument = ref_layernorm.MakeArgument(
-        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
+        e_m_n, gamma_n, beta_n, h_m_n, save_mean, save_inv_std, h_element_op, {M, N}, {1}, epsilon);
     ref_layernorm_invoker.Run(ref_layernorm_argument);
 }
 

example/22_cgemm/CMakeLists.txt

 add_custom_target(example_cgemm_xdl)
 
-if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-  add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
-  add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
-endif()
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
-  add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
-endif()
-if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
+
+add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+
 add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
-add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
-endif()
-if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-  add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
-  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
-endif()
+add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+
+add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
+
 if(USE_BITINT_EXTENSION_INT4)
-  add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
-  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
+    add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
+    add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
 endif()