Merge branch 'develop' into grouped_gemm_multi_abd_fixed_nk_example

profiler/include/profiler/profile_layernorm_impl.hpp → profiler/include/profiler/profile_layernorm_fwd_impl.hpp

@@ -6,7 +6,7 @@
 #include <iomanip>
 
 #include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
@@ -94,7 +94,7 @@ bool profile_layernorm_impl(int do_verification,
     constexpr int NumReduceDim = Rank - 1;
 
     // add device normalization instances
-    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
                                                                           GammaDataType,
                                                                           BetaDataType,
                                                                           YDataType,

profiler/include/profiler/profile_transpose_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/transpose_3d.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"
+
+namespace ck {
+namespace profiler {
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nchwd, const HostTensorA& A_ncdhw, Functor functor)
+{
+    for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
+                for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
+                    for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
+                    {
+                        auto a_val = A_ncdhw(n, c, d, h, w);
+                        functor(B_nchwd(n, c, h, w, d), a_val);
+                    }
+}
+
+template <typename ADataType, typename BDataType, index_t NumDim>
+bool profile_transpose_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> lengths)
+{
+    bool pass = true;
+
+    index_t N = lengths[0];
+    index_t C = lengths[1];
+    index_t D = lengths[2];
+    index_t H = lengths[3];
+    index_t W = lengths[4];
+
+    std::vector<ck::index_t> ncdhw = {N, C, D, H, W};
+    std::vector<ck::index_t> ndhwc = {N, D, H, W, C};
+    Tensor<ADataType> a(ncdhw);
+    Tensor<BDataType> b(ndhwc);
+    Tensor<BDataType> host_b(ndhwc);
+
+    // a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
+    std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
+    std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
+
+    std::cout << "A: " << a.mDesc << std::endl;
+    std::cout << "B: " << b.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
+    default: a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+    }
+
+    using ElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // const auto element_op = ElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+    using DeviceOp                   = ck::tensor_operation::device::
+        DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, ElementOp, NumDim>;
+
+    // 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;
+
+    if(do_verification)
+    {
+        host_elementwise4D(host_b, a, ElementOp{});
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            ab_lengths, {a_strides}, {b_strides}, input, output, ElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            // re-init C to zero before profiling next kernel
+            b_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+            if(do_verification)
+            {
+                b_device_buf.FromDevice(b.mData.data());
+
+                pass &= ck::utils::check_err(
+                    b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
+                }
+            }
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop =
+                std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
+
+            std::size_t num_btype =
+                sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
+                sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
+
+            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 << std::endl;
+
+            // pass = pass & ck::utils::check_err(b_device_result, b_host_result);
+            pass &= ck::utils::check_err(
+                b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << " N = " << N << " C = " << C << " D = " << D << " H = " << H << " W = " << W
+              << " : " << 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

@@ -16,8 +16,8 @@ set(PROFILER_SOURCES
     profile_grouped_conv_fwd.cpp
     profile_grouped_conv_bwd_weight.cpp
     profile_reduce.cpp
-    profile_groupnorm.cpp
-    profile_layernorm.cpp
+    profile_groupnorm_fwd.cpp
+    profile_layernorm_fwd.cpp
     profile_max_pool3d_fwd.cpp
     profile_avg_pool3d_bwd.cpp
     profile_max_pool3d_bwd.cpp
@@ -77,7 +77,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_w
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
-target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)

profiler/src/profile_groupnorm.cpp → profiler/src/profile_groupnorm_fwd.cpp

@@ -6,7 +6,7 @@
 #include <unordered_map>
 
 #include "profiler/data_type_enum.hpp"
-#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler/profile_groupnorm_fwd_impl.hpp"
 #include "profiler_operation_registry.hpp"
 
 using ck::index_t;

profiler/src/profile_layernorm.cpp → profiler/src/profile_layernorm_fwd.cpp

@@ -6,7 +6,7 @@
 #include <unordered_map>
 
 #include "profiler/data_type_enum.hpp"
-#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler/profile_layernorm_fwd_impl.hpp"
 #include "profiler_operation_registry.hpp"
 
 using ck::index_t;
@@ -78,6 +78,9 @@ int profile_layernorm(int argc, char* argv[])
 
     using F16 = ck::half_t;
     using F32 = float;
+
+    if(length.size() == 2)
+    {
         constexpr int rank = 2;
 
         if(data_type == ck::DataTypeEnum::Half)
@@ -94,6 +97,30 @@ int profile_layernorm(int argc, char* argv[])
         {
             throw std::runtime_error("not implemented yet");
         }
+    }
+    else if(length.size() == 4)
+    {
+        constexpr int rank = 4;
+
+        if(data_type == ck::DataTypeEnum::Half)
+        {
+            ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
+                do_verification, init_method, do_log, time_kernel, length);
+        }
+        else if(data_type == ck::DataTypeEnum::Float)
+        {
+            ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, F32, false, rank>(
+                do_verification, init_method, do_log, time_kernel, length);
+        }
+        else
+        {
+            throw std::runtime_error("not implemented yet");
+        }
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
 
     return 0;
 }

profiler/src/profile_transpose.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 "profiler/profile_transpose_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct MatrixLayout
+{
+    NCDHW, // 0
+    NCHWD, // 1
+};
+
+enum struct DataType
+{
+    F32_F32_F32_F32_F32, // 0
+    F16_F16_F16_F16_F16, // 1
+};
+
+#define OP_NAME "transpose"
+#define OP_DESC "Transpose"
+
+int profile_transpose(int argc, char* argv[])
+{
+    if(argc != 15)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        // printf("arg3: matrix layout (NCDHW -> NDCHW);\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: N, C, D, H, W\n");
+        exit(1);
+    }
+
+    const auto data_type = static_cast<DataType>(std::stoi(argv[2]));
+    // const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification   = std::stoi(argv[3]);
+    const int init_method        = std::stoi(argv[4]);
+    const bool do_log            = std::stoi(argv[5]);
+    const bool time_kernel       = std::stoi(argv[6]);
+    std::vector<index_t> lengths = std::stoi(argv[7]);
+
+    /**const int N = std::stoi(argv[7]);
+    const int C = std::stoi(argv[8]);
+    const int D = std::stoi(argv[9]);
+    const int H = std::stoi(argv[10]);
+    const int W = std::stoi(argv[11]);**/
+
+    using F32 = float;
+    using F16 = ck::half_t;
+
+    auto profile = [&](auto a_type, auto b_type) {
+        using ADataType = decltype(a_type);
+        using BDataType = decltype(b_type);
+
+        bool pass = ck::profiler::profile_transpose_impl<ADataType, BDataType>(
+            do_verification, init_method, do_log, time_kernel, lengths);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32_F32_F32)
+    {
+        return profile(F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16_F16_F16)
+    {
+        return profile(F16{}, F16{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_transpose);

test/CMakeLists.txt

@@ -139,7 +139,7 @@ add_subdirectory(grouped_convnd_fwd)
 add_subdirectory(grouped_convnd_bwd_weight)
 add_subdirectory(block_to_ctile_map)
 add_subdirectory(softmax)
-add_subdirectory(normalization)
+add_subdirectory(normalization_fwd)
 add_subdirectory(data_type)
 add_subdirectory(elementwise_normalization)
 add_subdirectory(batchnorm)
@@ -148,6 +148,7 @@ add_subdirectory(pool)
 add_subdirectory(batched_gemm_multi_d)
 add_subdirectory(grouped_convnd_bwd_data)
 add_subdirectory(conv_tensor_rearrange)
+add_subdirectory(transpose)
 if(GPU_TARGETS MATCHES "gfx11")
     add_subdirectory(wmma_op)
 endif()

test/normalization/CMakeLists.txt

-add_custom_target(test_normalization)
-add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
-  add_dependencies(test_normalization test_layernorm2d_fp32)
-endif()
-add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
-  add_dependencies(test_normalization test_groupnorm_fp32)
-endif()
-add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
-  add_dependencies(test_normalization test_layernorm2d_fp16)
-endif()
-add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
-if(result EQUAL 0)
-  target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
-  add_dependencies(test_normalization test_groupnorm_fp16)
-endif()

test/normalization_fwd/CMakeLists.txt

+add_custom_target(test_normalization_fwd)
+add_gtest_executable(test_layernorm2d_fwd_fp32 test_layernorm2d_fwd_fp32.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_layernorm2d_fwd_fp32 PRIVATE utility device_normalization_fwd_instance)
+  add_dependencies(test_normalization_fwd test_layernorm2d_fwd_fp32)
+endif()
+
+add_gtest_executable(test_groupnorm_fwd_fp32 test_groupnorm_fwd_fp32.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_groupnorm_fwd_fp32 PRIVATE utility device_normalization_fwd_instance)
+  add_dependencies(test_normalization_fwd test_groupnorm_fwd_fp32)
+endif()
+
+add_gtest_executable(test_layernorm2d_fwd_fp16 test_layernorm2d_fwd_fp16.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_layernorm2d_fwd_fp16 PRIVATE utility device_normalization_fwd_instance)
+  add_dependencies(test_normalization_fwd test_layernorm2d_fwd_fp16)
+endif()
+
+add_gtest_executable(test_layernorm4d_fwd_fp16 test_layernorm4d_fwd_fp16.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_layernorm4d_fwd_fp16 PRIVATE utility device_normalization_fwd_instance)
+  add_dependencies(test_normalization_fwd test_layernorm4d_fwd_fp16)
+endif()
+
+add_gtest_executable(test_groupnorm_fwd_fp16 test_groupnorm_fwd_fp16.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_groupnorm_fwd_fp16 PRIVATE utility device_normalization_fwd_instance)
+  add_dependencies(test_normalization_fwd test_groupnorm_fwd_fp16)
+endif()

test/normalization/test_groupnorm_fp16.cpp → test/normalization_fwd/test_groupnorm_fwd_fp16.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "gtest/gtest.h"
-#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler/profile_groupnorm_fwd_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

test/normalization/test_groupnorm_fp32.cpp → test/normalization_fwd/test_groupnorm_fwd_fp32.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "gtest/gtest.h"
-#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler/profile_groupnorm_fwd_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

test/normalization/test_layernorm2d_fp16.cpp → test/normalization_fwd/test_layernorm2d_fwd_fp16.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "gtest/gtest.h"
-#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler/profile_layernorm_fwd_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

test/normalization/test_layernorm2d_fp32.cpp → test/normalization_fwd/test_layernorm2d_fwd_fp32.cpp

@@ -2,7 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "gtest/gtest.h"
-#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler/profile_layernorm_fwd_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

test/normalization_fwd/test_layernorm4d_fwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_fwd_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm4d : public ::testing::Test
+{
+    protected:
+    using XDataType              = std::tuple_element_t<0, Tuple>;
+    using GammaDataType          = std::tuple_element_t<1, Tuple>;
+    using BetaDataType           = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType        = std::tuple_element_t<3, Tuple>;
+    using YDataType              = std::tuple_element_t<4, Tuple>;
+    using SaveMeanInvStdDataType = std::tuple_element_t<5, Tuple>;
+
+    void Run()
+    {
+        // [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {1, 1, 1, 1}, {7, 7, 7, 7}, {256, 16, 16, 8}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_layernorm_impl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                ComputeDataType,
+                                                                YDataType,
+                                                                SaveMeanInvStdDataType,
+                                                                true,
+                                                                4>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    std::tuple<F16, F16, F16, F32, F16, F32>>;
+
+TYPED_TEST_SUITE(TestLayernorm4d, KernelTypes);
+TYPED_TEST(TestLayernorm4d, Test_FP16) { this->Run(); }

test/transpose/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_gtest_executable(test_transpose test_transpose.cpp)
+   target_link_libraries(test_transpose PRIVATE utility device_transpose_instance)
+   set(target 1)
+ endif()
+endforeach()

test/transpose/test_transpose.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "gtest/gtest.h"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "test_transpose_util.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <typename Tuple>
+class TestTranspose : public ::testing::Test
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<      F16,       F16>,
+    std::tuple<      F32,       F32>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestTranspose, KernelTypes);
+
+//#include "test_transpose_ut_cases.inc"

test/transpose/test_transpose_ut_cases.inc

+#pragma once
+
+TYPED_TEST(TestTranspose, Test1)
+{
+    // for 16, 8, 16, 32, 8
+    std::vector<int> Ms{1, 2, 3, 4, 5, 6};
+    std::vector<index_t> lengths{16, 8, 16, 32, 8};
+    /**constexpr int N = 16;
+    constexpr int C = 8;
+    constexpr int D = 16;
+    constexpr int H = 32;
+    constexpr int W = 8;**/
+
+    this->Run();
+}
+
+
+TYPED_TEST(TestTranpose, Test2)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    std::vector<index_t> lengths{16, 8, 16, 32, 16};
+    /**constexpr int N = 16;
+    constexpr int C = 8;
+    constexpr int D = 16;
+    constexpr int H = 32;
+    constexpr int W = 8;**/
+
+    this->Run();
+}
+

test/transpose/test_transpose_util.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "include/ck/utility/data_type.hpp"
+#include "profiler/profile_transpose_impl.hpp"
+
+namespace ck {
+namespace test {
+
+template <typename Tuple>
+class TestTranspose : public testing::Test
+{
+    using F32 = float;
+
+    protected:
+    using ADataType = std::tuple_element_t<0, Tuple>;
+    using BDataType = std::tuple_element_t<1, Tuple>;
+
+    public:
+    static constexpr bool verify_              = true;
+    static constexpr int init_method_          = 1; // decimal value initialization
+    static constexpr bool log_                 = false;
+    static constexpr bool bench_               = false; // measure kernel performance
+    std::vector<std::vector<index_t>> lengths_ = {{16, 32, 16, 32, 16}, {16, 8, 16, 32, 8}};
+
+    void Run()
+    {
+        for(auto length : this->lengths_)
+        {
+            this->RunSingle(length);
+        }
+    }
+
+    void RunSingle()
+    {
+        bool pass = ck::profiler::profile_transpose_impl<ADataType, BDataType, 5>(
+            verify_, init_method_, log_, bench_, lengths_);
+        EXPECT_TRUE(pass);
+    }
+};
+
+} // namespace test
+} // namespace ck