Universal gemm splitk using reduce (with multi-d) (#1341)

* init for reduce_threadwise multi_d

* add reduce_threadwise_multi_d

* add reduce_multi_d

* clean

* start add an other splitk device op

* add reduce template parameter to SplitKBatchOffset

* add reduce c matrix

* clean up code

* change example data type to bf16

* add bf16Ai8B example

* remove reduce template parameter

* add splitk atomic status to v4

* example add multi d parameters

* device op add multi-d parameters

* add multi-d to reduce

* fix kbach=1 bug

* change B layout to col in  bf16Ai8B example

* remove float adding struct

* change  multi-d interface

* change file and class name

* remove multi-d of bf16Ai8B example

* change IsReduce function to IsReduceAdd

* change example layout to RRR from RCR

* according layout to set ds stride

* reset parameter layout

* add gemm universal reduce instance

* add reduce factory

* add profile_gemm_universal_reduce

* add reduce to profiler

* fix reduce instance

* fix profiler reduce compiling bug

* format

* format library instance code

* add mem instance for reduce library

* fix call instance names

* add workspace for reduce in ckProfiler

* format

* add mnpading to reduce library instance

* add fp16 instance to reduce of profiler

* change copyright time

* restore profiler cmake file

* add reduce text to instances

* add DsLayout and DsDataType to instances template parameter

* fixed gemm_reduce_multi_d

* add an example without multi_d

* Update common.hpp

* Update gtest.cmake

* Update gemm_xdl_splitk_reduce_bf16.cpp

* clean

* Update gtest.cmake

* format

* fixe api

* format

* default parameter change to RRR

* add vector_len for multi_d

* format

* Update gtest.cmake

* fix bf16A iBB elementwiseop

* add ReduceDataType

* move ReduceDataType to end position

* format

* remove googletest git method  address

* fix copyright time

* update init data

---------
Co-authored-by: root <jizhan@amd.com>
Co-authored-by: letaoqin <letaoqin@amd.com>
Co-authored-by: Jing Zhang <jizhan@meta.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>

library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
+                                               Row,
+                                               DsLayout,
+                                               Row,
+                                               F16,
+                                               F16,
+                                               DsDataType,
+                                               F16,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave,
+                                                                            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
+                                               Row,
+                                               DsLayout,
+                                               Row,
+                                               F16,
+                                               F16,
+                                               DsDataType,
+                                               F16,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave,
+                                                                            GemmMNKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
+                                               Row,
+                                               DsLayout,
+                                               Row,
+                                               F16,
+                                               F16,
+                                               DsDataType,
+                                               F16,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances<Interwave,
+                                                                            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
+                                               Row,
+                                               DsLayout,
+                                               Row,
+                                               F16,
+                                               F16,
+                                               DsDataType,
+                                               F16,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances<Interwave,
+                                                                            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_reduce/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
+                                               Row,
+                                               DsLayout,
+                                               Row,
+                                               F16,
+                                               F16,
+                                               DsDataType,
+                                               F16,
+                                               PassThrough,
+                                               PassThrough,
+                                               PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_instances<Interwave,
+                                                                            GemmMNKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_gemm_universal_reduce_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3r1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_universal_reduce.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"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout>
+bool profile_gemm_universal_reduce_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        int M,
+                                        int N,
+                                        int K,
+                                        int StrideA,
+                                        int StrideB,
+                                        int StrideC,
+                                        int KBatch,
+                                        int n_warmup,
+                                        int n_iter,
+                                        uint64_t rotating = 0)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<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{}));
+
+    int total_gemm_needed = a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
+    int rotating_count    = std::max(
+        1,
+        std::min(n_iter,
+                 static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
+
+    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_device_result.mDesc << std::endl;
+    std::cout << "rotating count: " << rotating_count << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    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());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmV2R1<ALayout,
+                                                                  BLayout,
+                                                                  DsLayout,
+                                                                  CLayout,
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  DsDataType,
+                                                                  CDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        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, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38};
+
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
+
+        for(std::size_t i = 0; i < kbatch_list.size(); i++)
+        {
+            auto kbatch_curr = kbatch_list[i];
+
+            auto argument_ptr =
+                op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                            {},
+                                            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                            M,
+                                            N,
+                                            K,
+                                            StrideA,
+                                            StrideB,
+                                            {},
+                                            StrideC,
+                                            kbatch_curr,
+                                            a_element_op,
+                                            b_element_op,
+                                            c_element_op);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+
+                DeviceMem gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+                op_ptr->SetWorkSpacePointer(
+                    argument_ptr.get(), gemm_workspace_dev.GetDeviceBuffer(), StreamConfig{});
+
+                // re-init C to zero before profiling next kernel
+                c_device_buf.SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(),
+                                 StreamConfig{nullptr, false, 0, n_warmup, n_iter});
+
+                if(do_verification)
+                {
+                    c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+
+                std::string op_name = op_ptr->GetTypeString();
+
+                float ave_time = invoker_ptr->Run(argument_ptr.get(),
+                                                  StreamConfig{nullptr,
+                                                               time_kernel,
+                                                               0,
+                                                               n_warmup,
+                                                               n_iter,
+                                                               rotating_count > 1,
+                                                               rotating_count});
+
+                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.E9 / ave_time;
+
+                float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                          << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
+                          << kbatch_curr << std::endl;
+
+#if defined CK_ENABLE_FP8
+                // set softer tolerances for fp8
+                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                             is_same_v<CDataType, f8_t>)
+                {
+                    std::string msg = "Error: Incorrect results!";
+                    double rtol     = 1e-1;
+                    double atol     = 1e-1;
+                    pass            = pass & ck::utils::check_err(
+                                      c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
+                }
+                else
+                {
+#endif
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+#if defined CK_ENABLE_FP8
+                }
+#endif
+
+                if(tflops > best_tflops)
+                {
+                    best_op_name    = op_name;
+                    best_tflops     = tflops;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                    best_kbatch     = kbatch_curr;
+                }
+            }
+            else
+            {
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
+              << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/CMakeLists.txt

@@ -52,6 +52,7 @@ if(GPU_TARGETS MATCHES "gfx9")
   list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
@@ -122,6 +123,7 @@ if(GPU_TARGETS MATCHES "gfx9")
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)

profiler/src/profile_gemm_universal_reduce.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_universal_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+    F8_F16_F16,     // 4
+    BF16_I8_BF16,   // 5
+    F16_F16_F16_F8, // 6
+};
+
+#define OP_NAME "gemm_universal_reduce"
+#define OP_DESC "Universal GEMM"
+
+int profile_gemm_universal_reduce(int argc, char* argv[])
+{
+    if(argc != 15 && argc != 18)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@i8; 6: f16, "
+               "comp f8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        printf("optional:\n");
+        printf("arg15: number of warm-up cycles (default 1)\n");
+        printf("arg16: number of iterations (default 10)\n");
+        printf("arg17: memory for rotating buffer (default 0, size in MB)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+    const int KBatch  = std::stoi(argv[14]);
+
+    int n_warmup      = 1;
+    int n_iter        = 10;
+    uint64_t rotating = 0;
+    if(argc == 18)
+    {
+        n_warmup = std::stoi(argv[15]);
+        n_iter   = std::stoi(argv[16]);
+        rotating = std::stoull(argv[17]) * 1024 * 1024;
+    }
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using I8   = int8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+
+    using DsDataType = ck::Tuple<>;
+    using DsLayout   = ck::Tuple<>;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_universal_reduce_impl<ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     AccDataType,
+                                                                     CDataType,
+                                                                     ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideC < 0) ? DefaultStrideC : StrideC,
+            KBatch,
+            n_warmup,
+            n_iter,
+            rotating);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::BF16_I8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(BF16{}, I8{}, F32{}, BF16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_universal_reduce);