WMMA support for GEMM reduce (#2823)

Added gemm + reduce instance library for RDNA4. This includes:

- New device implementation running GEMM and reduction kernel
- instances for wmma (xdl parity)
- examples for wmma (xdl parity)
- tests for existing xdl and wmma

example/35_splitK_gemm/CMakeLists.txt

@@ -27,3 +27,16 @@ add_example_executable(example_gemm_xdl_splitk_reduce_multi_d_bf16 gemm_xdl_spli
 add_example_executable(example_gemm_xdl_splitk_reduce_bf16A_i8B gemm_xdl_splitk_reduce_bf16A_i8B.cpp)
 
 add_example_executable(example_gemm_xdl_splitk_reduce_bfp16 gemm_xdl_splitk_reduce_bf16.cpp)
+
+add_custom_target(example_splitK_gemm_wmma)
+add_example_executable(example_gemm_wmma_splitk_reduce_bf16 gemm_wmma_splitk_reduce_bf16.cpp)
+add_example_dependencies(example_splitK_gemm_wmma example_gemm_wmma_splitk_reduce_bf16)
+
+add_example_executable(example_gemm_wmma_splitk_reduce_bf16A_i8B gemm_wmma_splitk_reduce_bf16A_i8B.cpp)
+add_example_dependencies(example_splitK_gemm_wmma example_gemm_wmma_splitk_reduce_bf16A_i8B)
+
+add_example_executable(example_gemm_wmma_splitk_reduce_multi_d_bf16 gemm_wmma_splitk_reduce_multi_d_bf16.cpp)
+add_example_dependencies(example_splitK_gemm_wmma example_gemm_wmma_splitk_reduce_multi_d_bf16)
+
+add_example_executable(example_gemm_wmma_splitk_reduce_multi_d_fp16 gemm_wmma_splitk_reduce_multi_d_fp16.cpp)
+add_example_dependencies(example_splitK_gemm_wmma example_gemm_wmma_splitk_reduce_multi_d_fp16)

example/35_splitK_gemm/common.hpp

@@ -99,3 +99,85 @@ bool parse_cmd_args(int argc,
 
     return true;
 }
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}

example/35_splitK_gemm/gemm_wmma_splitk_reduce_bf16.cpp

@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = ck::bhalf_t;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceWmmaGemmInstance = 
+    ck::tensor_operation::device::DeviceGemm_Wmma_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault,
+        256,
+        128, 128, 32,
+        8, 8,
+        16, 16,
+        4, 2,
+        S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>,
+        1, 1, 8, true,
+        S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>,
+        1, 1, 8, true,
+        1, 1, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v1, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_wmma_splitk_reduce_example.inc"
+
+int main(int argc, char* argv[]) { return !run_wmma_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/gemm_wmma_splitk_reduce_bf16A_i8B.cpp

@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = int8_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceWmmaGemmInstance = 
+    ck::tensor_operation::device::DeviceGemm_Wmma_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault,
+        256,   
+        128,  128,  32,
+        8,    8,
+        16,   16,
+        4,    2,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v1, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_wmma_splitk_reduce_example.inc"
+
+int main(int argc, char* argv[]) { return !run_wmma_gemm_splitk_example(argc, argv); }

example/35_splitK_gemm/gemm_wmma_splitk_reduce_multi_d_bf16.cpp

@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::bhalf_t;
+using DsDataType       = ck::Tuple<D0DataType>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Wmma_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault,
+        256,
+        128,  128,  32,
+        8,    8,
+        16,   16,
+        4,    2,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v1, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_wmma_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_multi_d_example(argc, argv); }

example/35_splitK_gemm/gemm_wmma_splitk_reduce_multi_d_fp16.cpp

@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3r1.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_t;
+using ReduceDataType   = float;
+using D0DataType       = ck::half_t;
+using DsDataType       = ck::Tuple<D0DataType>;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using CLayout  = Row;
+using D0Layout = CLayout;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Wmma_CShuffleV3R1<
+        ALayout,   BLayout,   DsLayout,  CLayout,
+        ADataType,   BDataType, DsDataType,  CDataType, AccDataType,  CShuffleDataType,
+        AElementOp, BElementOp, CDEElementOp, GemmDefault,
+        256,
+        128,  256,  64,
+        8,    8,
+        16,   16,
+        4,    4,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        S<4, 64, 1>,     S<0, 2, 1>,    S<0, 2, 1>,
+        1,    1,    8,   true,
+        1,    1,    S<1, 32, 1, 8>,  8,
+        ck::BlockGemmPipelineScheduler::Intrawave,
+        ck::BlockGemmPipelineVersion::v1, ReduceDataType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_wmma_splitk_reduce_multi_d_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_multi_d_example(argc, argv); }

example/35_splitK_gemm/run_gemm_splitk_reduce_multi_d_example.inc

@@ -3,88 +3,6 @@
 
 #pragma once
 
-template <typename DataType>
-inline __host__ __device__ constexpr double get_rtol()
-{
-    if constexpr(std::is_same_v<DataType, float>)
-    {
-        return 1e-3;
-    }
-    else if constexpr(std::is_same_v<DataType, double>)
-    {
-        return 1e-6;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::half_t>)
-    {
-        return 1e-3;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
-    {
-        return 5e-2;
-    }
-    else if constexpr(std::is_same_v<DataType, int32_t>)
-    {
-        return 1e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, int8_t>)
-    {
-        return 1e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
-    {
-        return 1e-1; // 240 and 224 are acceptable
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
-    {
-        return 1.5e-1; // 57344 and 49152 are acceptable
-    }
-    else
-    {
-        return 1e-3;
-    }
-}
-
-template <typename DataType>
-inline __host__ __device__ constexpr double get_atol()
-{
-    if constexpr(std::is_same_v<DataType, float>)
-    {
-        return 1e-3;
-    }
-    else if constexpr(std::is_same_v<DataType, double>)
-    {
-        return 1e-6;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::half_t>)
-    {
-        return 1e-3;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
-    {
-        return 5e-2;
-    }
-    else if constexpr(std::is_same_v<DataType, int32_t>)
-    {
-        return 1e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, int8_t>)
-    {
-        return 1e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
-    {
-        return 16.1; // 240 and 224 are acceptable
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
-    {
-        return 8192.1; // 57344 and 49152 are acceptable
-    }
-    else
-    {
-        return 1e-3;
-    }
-}
-
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {

example/35_splitK_gemm/run_gemm_wmma_splitk_reduce_example.inc

@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <typename ProblemType>
+bool run_wmma_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto M       = problem_size.M;
+    auto N       = problem_size.N;
+    auto K       = problem_size.K;
+    auto StrideA = problem_size.StrideA;
+    auto StrideB = problem_size.StrideB;
+    auto StrideC = problem_size.StrideC;
+    auto KBatch  = problem_size.KBatch;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
+
+    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{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+    }
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "init method: " << config.init_method << std::endl;
+    std::cout << "KBatch: " << KBatch << std::endl;
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device GEMM
+    auto device_op = DeviceWmmaGemmInstance{};
+    auto invoker   = device_op.MakeInvoker();
+
+    auto argument =
+        device_op.MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      std::array<const void*, 0>{}, // empty D tensors
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      std::array<ck::index_t, 0>{}, // empty D strides
+                                      StrideC,
+                                      KBatch,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    // Allocate workspace for split-K reduction if needed
+    size_t workspace_size = device_op.GetWorkSpaceSize(argument.get());
+    DeviceMem workspace_buf(workspace_size);
+    std::cout << "Workspace size: " << workspace_size << " bytes" << std::endl;
+    if(workspace_size > 0)
+    {
+        argument->p_workspace_ = workspace_buf.GetDeviceBuffer();
+        std::cout << "Allocated workspace of size: " << workspace_size << " bytes" << std::endl;
+    }
+
+    if(!device_op.IsSupportedArgument(argument.get()))
+    {
+        std::cout << "The runtime argument is not supported!" << std::endl;
+        std::cout << "Debug info:" << std::endl;
+        std::cout << "  M=" << M << ", N=" << N << ", K=" << K << ", KBatch=" << KBatch
+                  << std::endl;
+        std::cout << "  StrideA=" << StrideA << ", StrideB=" << StrideB << ", StrideC=" << StrideC
+                  << std::endl;
+        return false;
+    }
+
+    bool pass      = true;
+    float ave_time = 0;
+
+    if(config.do_verification)
+    {
+        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_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        ave_time = invoker.Run(argument.get(), StreamConfig{nullptr, false});
+
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData,
+                                    c_m_n_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    get_rtol<CDataType>(),
+                                    get_atol<CDataType>());
+    }
+
+    if(config.time_kernel)
+    {
+        ave_time = invoker.Run(argument.get(), StreamConfig{nullptr, config.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(CDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E12 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E9 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << device_op.GetTypeString() << std::endl;
+    }
+
+    return pass;
+}
+
+bool run_wmma_gemm_splitk_example(int argc, char* argv[])
+{
+    ProblemSizeSplitK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_wmma_gemm(problem_size, config);
+}

example/35_splitK_gemm/run_gemm_wmma_splitk_reduce_multi_d_example.inc

@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <typename ProblemSize>
+bool run_wmma_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto M        = problem_size.M;
+    auto N        = problem_size.N;
+    auto K        = problem_size.K;
+    auto StrideA  = problem_size.StrideA;
+    auto StrideB  = problem_size.StrideB;
+    auto StrideC  = problem_size.StrideC;
+    auto StrideD0 = problem_size.StrideC;
+    auto KBatch   = problem_size.KBatch;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    StrideA  = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB  = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC  = f_get_default_stride(M, N, StrideC, CLayout{});
+    StrideD0 = f_get_default_stride(M, N, StrideD0, D0Layout{});
+
+    Tensor<ADataType> a_m_k(
+        f_host_tensor_descriptor(problem_size.M, problem_size.K, problem_size.StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(
+        f_host_tensor_descriptor(problem_size.K, problem_size.N, problem_size.StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(
+        f_host_tensor_descriptor(problem_size.M, problem_size.N, problem_size.StrideC, D0Layout{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    Tensor<CDataType> c_m_n_host_result(
+        f_host_tensor_descriptor(problem_size.M, problem_size.N, problem_size.StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(
+        f_host_tensor_descriptor(problem_size.M, problem_size.N, problem_size.StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "init method: " << config.init_method << std::endl;
+    std::cout << "KBatch: " << KBatch << std::endl;
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto gemm                                              = DeviceGemmV2Instance{};
+    auto invoker                                           = gemm.MakeInvoker();
+    constexpr auto kNum_DTensors                           = DsDataType::Size();
+    const std::array<const void*, kNum_DTensors> p_ds      = {d0_m_n_device_buf.GetDeviceBuffer()};
+    const std::array<ck::index_t, kNum_DTensors> d_strides = {problem_size.StrideC};
+
+    auto argument =
+        gemm.MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                 static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                 p_ds,
+                                 static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                 problem_size.M,
+                                 problem_size.N,
+                                 problem_size.K,
+                                 problem_size.StrideA,
+                                 problem_size.StrideB,
+                                 d_strides,
+                                 problem_size.StrideC,
+                                 problem_size.KBatch,
+                                 a_element_op,
+                                 b_element_op,
+                                 c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument.get()))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+        return false;
+    }
+
+    auto workspace_size = gemm.GetWorkSpaceSize(argument.get());
+    DeviceMem workspace_device_buf(workspace_size);
+
+    std::cout << "Workspace size: " << workspace_size << " bytes" << std::endl;
+    std::cout << "Allocated workspace of size: " << workspace_size << " bytes" << std::endl;
+
+    if(workspace_size > 0)
+    {
+        argument->p_workspace_ = workspace_device_buf.GetDeviceBuffer();
+    }
+
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstanceMultiD = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                      BDataType,
+                                                                                      CDataType,
+                                                                                      AccDataType,
+                                                                                      AElementOp,
+                                                                                      BElementOp,
+                                                                                      PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstanceMultiD{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        c_m_n_host_result.ForEach(
+            [&](auto& self, auto idx) { c_element_op(self(idx), self(idx), d0_m_n(idx)); });
+    }
+
+    std::cout << "init method: " << config.init_method << std::endl;
+    std::cout << "KBatch: " << problem_size.KBatch << std::endl;
+
+    float ave_time = invoker.Run(argument.get(), StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = std::size_t(2) * problem_size.M * problem_size.N * problem_size.K;
+    std::size_t num_btype = sizeof(ADataType) * problem_size.M * problem_size.K +
+                            sizeof(BDataType) * problem_size.K * problem_size.N +
+                            sizeof(CDataType) * problem_size.M * problem_size.N +
+                            sizeof(D0DataType) * problem_size.M * problem_size.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, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        double rtol = get_rtol<CDataType>();
+        double atol = get_atol<CDataType>();
+
+        return ck::utils::check_err(
+            c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results!", rtol, atol);
+    }
+
+    return true;
+}
+
+int run_gemm_splitk_multi_d_example(int argc, char* argv[])
+{
+    ProblemSizeSplitK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_wmma_gemm(problem_size, config);
+}