Merge branch 'amd-develop' into amd-master

profiler/src/profile_groupnorm_fwd.cpp

@@ -98,7 +98,7 @@ int profile_groupnorm(int argc, char* argv[])
     }
     else if(data_type == ck::DataTypeEnum::Half)
     {
-        ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F32, false>(
+        ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F16, false>(
             do_verification, init_method, do_log, time_kernel, length);
     }
     else

profiler/src/profile_layernorm_bwd_data.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_layernorm_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct layernormBwdDataArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_layernorm_bwd_data()
+{
+    // eg: ckProfiler layernorm_bwd_data 0 0 2 0 1 --length 1502 4096
+    std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 1024 1024) \n"
+              << std::endl;
+}
+
+int profile_layernorm_bwd_data(int argc, char* argv[])
+{
+    if(argc <= 2)
+    {
+        print_help_layernorm_bwd_data();
+        return 0;
+    }
+
+    layernormBwdDataArgParser arg_parser;
+
+    // short unnamed options
+    const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+    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]);
+
+    // parse the long options
+    arg_parser(argc, argv);
+    const std::vector<index_t> length = arg_parser.long_opts["length"];
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    if(length.size() == 2)
+    {
+        constexpr int rank = 2;
+
+        if(data_type == ck::DataTypeEnum::Half)
+        {
+            ck::profiler::profile_layernorm_bwd_data_impl<F16, F16, F16, F16, F32, F16, rank>(
+                do_verification, init_method, do_log, time_kernel, length);
+        }
+        else if(data_type == ck::DataTypeEnum::Float)
+        {
+            ck::profiler::profile_layernorm_bwd_data_impl<F32, F32, F32, F32, F32, F32, 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;
+}
+
+REGISTER_PROFILER_OPERATION("layernorm_bwd_data",
+                            "Layer Normalization",
+                            profile_layernorm_bwd_data);

profiler/src/profile_layernorm_fwd.cpp

@@ -104,7 +104,7 @@ int profile_layernorm(int argc, char* argv[])
 
         if(data_type == ck::DataTypeEnum::Half)
         {
-            ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
+            ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F16, false, rank>(
                 do_verification, init_method, do_log, time_kernel, length);
         }
         else if(data_type == ck::DataTypeEnum::Float)
@@ -125,4 +125,4 @@ int profile_layernorm(int argc, char* argv[])
     return 0;
 }
 
-REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);
+REGISTER_PROFILER_OPERATION("layernorm_fwd", "Layer Normalization", profile_layernorm);

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

@@ -140,6 +140,7 @@ add_subdirectory(grouped_convnd_bwd_weight)
 add_subdirectory(block_to_ctile_map)
 add_subdirectory(softmax)
 add_subdirectory(normalization_fwd)
+add_subdirectory(normalization_bwd_data)
 add_subdirectory(data_type)
 add_subdirectory(elementwise_normalization)
 add_subdirectory(batchnorm)
@@ -149,6 +150,8 @@ add_subdirectory(batched_gemm_multi_d)
 add_subdirectory(grouped_convnd_bwd_data)
 add_subdirectory(conv_tensor_rearrange)
 add_subdirectory(transpose)
+add_subdirectory(permute_scale)
+add_subdirectory(wrapper)
 if(GPU_TARGETS MATCHES "gfx11")
     add_subdirectory(wmma_op)
 endif()

test/normalization_bwd_data/CMakeLists.txt

+add_custom_target(test_normalization_bwd_data)
+add_gtest_executable(test_layernorm2d_bwd_data_fp32 test_layernorm2d_bwd_data_fp32.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_layernorm2d_bwd_data_fp32 PRIVATE utility device_normalization_bwd_data_instance)
+  add_dependencies(test_normalization_bwd_data test_layernorm2d_bwd_data_fp32)
+endif()
+
+add_gtest_executable(test_groupnorm_bwd_data_fp32 test_groupnorm_bwd_data_fp32.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_groupnorm_bwd_data_fp32 PRIVATE utility device_normalization_bwd_data_instance)
+  add_dependencies(test_normalization_bwd_data test_groupnorm_bwd_data_fp32)
+endif()
+

test/normalization_bwd_data/test_groupnorm_bwd_data_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_groupnorm_bwd_data_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestgroupnormBwdData : public ::testing::Test
+{
+    protected:
+    using DYDataType         = std::tuple_element_t<0, Tuple>;
+    using XDataType          = std::tuple_element_t<1, Tuple>;
+    using GammaDataType      = std::tuple_element_t<2, Tuple>;
+    using MeanInvStdDataType = std::tuple_element_t<3, Tuple>;
+    using ComputeDataType    = std::tuple_element_t<4, Tuple>;
+    using DXDataType         = std::tuple_element_t<5, Tuple>;
+
+    void Run()
+    {
+        // Bwd data: [N, H, W, G, C], reduce H, W, C
+        std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
+                                                         {1, 2, 3, 4, 5},
+                                                         {256, 9, 9, 9, 9},
+                                                         {1, 64, 64, 32, 10},
+                                                         {1, 32, 32, 32, 20},
+                                                         {1, 16, 16, 32, 40}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_groupnorm_bwd_data_impl<DYDataType,
+                                                                         XDataType,
+                                                                         GammaDataType,
+                                                                         MeanInvStdDataType,
+                                                                         ComputeDataType,
+                                                                         DXDataType>(
+                true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // DYDataType XDataType, GammaDataType, MeanInvStdDataType, ComputeDataType, DXDataType>
+    std::tuple<F32, F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestgroupnormBwdData, KernelTypes);
+TYPED_TEST(TestgroupnormBwdData, Test_FP32) { this->Run(); }

test/normalization_bwd_data/test_layernorm2d_bwd_data_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_bwd_data_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm2dBwdData : public ::testing::Test
+{
+    protected:
+    using DYDataType         = std::tuple_element_t<0, Tuple>;
+    using XDataType          = std::tuple_element_t<1, Tuple>;
+    using GammaDataType      = std::tuple_element_t<2, Tuple>;
+    using MeanInvStdDataType = std::tuple_element_t<3, Tuple>;
+    using ComputeDataType    = std::tuple_element_t<4, Tuple>;
+    using DXDataType         = std::tuple_element_t<5, Tuple>;
+
+    void Run()
+    {
+        // Bwd data: [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::profiler::profile_layernorm_bwd_data_impl<DYDataType,
+                                                              XDataType,
+                                                              GammaDataType,
+                                                              MeanInvStdDataType,
+                                                              ComputeDataType,
+                                                              DXDataType,
+                                                              2>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // DYDataType XDataType, GammaDataType, MeanInvStdDataType, ComputeDataType, DXDataType>
+    std::tuple<F32, F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestLayernorm2dBwdData, KernelTypes);
+TYPED_TEST(TestLayernorm2dBwdData, Test_FP32) { this->Run(); }

test/normalization_fwd/test_groupnorm_fwd_fp16.cpp

@@ -47,8 +47,8 @@ class TestGroupnorm : public ::testing::Test
 };
 
 using KernelTypes = ::testing::Types<
-    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
-    std::tuple<F16, F16, F16, F32, F16, F32>>;
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType>
+    std::tuple<F16, F16, F16, F32, F16, F16>>;
 
 TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);
 TYPED_TEST(TestGroupnorm, Test_FP16) { this->Run(); }

test/normalization_fwd/test_groupnorm_fwd_fp32.cpp

@@ -45,7 +45,7 @@ class TestGroupnorm : public ::testing::Test
 };
 
 using KernelTypes = ::testing::Types<
-    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType>
     std::tuple<F32, F32, F32, F32, F32, F32>>;
 
 TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);

test/normalization_fwd/test_layernorm2d_fwd_fp16.cpp

@@ -41,8 +41,8 @@ class TestLayernorm2d : public ::testing::Test
 };
 
 using KernelTypes = ::testing::Types<
-    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
-    std::tuple<F16, F16, F16, F32, F16, F32>>;
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType>
+    std::tuple<F16, F16, F16, F32, F16, F16>>;
 
 TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
 TYPED_TEST(TestLayernorm2d, Test_FP16) { this->Run(); }

test/normalization_fwd/test_layernorm4d_fwd_fp16.cpp

@@ -41,8 +41,8 @@ class TestLayernorm4d : public ::testing::Test
 };
 
 using KernelTypes = ::testing::Types<
-    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
-    std::tuple<F16, F16, F16, F32, F16, F32>>;
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType>
+    std::tuple<F16, F16, F16, F32, F16, F16>>;
 
 TYPED_TEST_SUITE(TestLayernorm4d, KernelTypes);
 TYPED_TEST(TestLayernorm4d, Test_FP16) { this->Run(); }

test/permute_scale/CMakeLists.txt

+add_custom_target(test_permute)
+add_gtest_executable(test_permute_scale test_permute_scale.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_permute_scale PRIVATE utility device_permute_scale_instance)
+  add_dependencies(test_permute test_permute_scale)
+endif()

test/permute_scale/test_permute_scale.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_permute_scale_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestPermute : public ::testing::Test
+{
+    protected:
+    using ADataType = std::tuple_element_t<0, Tuple>;
+    using BDataType = std::tuple_element_t<1, Tuple>;
+
+    void Run()
+    {
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 2, 1, 8}, {1, 1, 1, 1}, {16, 8, 32, 64}, {32, 64, 128, 128}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::test_permute_scale_impl<ADataType, BDataType, 4>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F16>, std::tuple<F32, F32>>;
+
+TYPED_TEST_SUITE(TestPermute, KernelTypes);
+TYPED_TEST(TestPermute, Test_FP16) { this->Run(); }
+TYPED_TEST(TestPermute, Test_FP32) { this->Run(); }

test/permute_scale/test_permute_scale_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale.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 {
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    std::size_t N = A_nchw.mDesc.GetLengths()[0];
+    std::size_t C = A_nchw.mDesc.GetLengths()[1];
+    std::size_t H = A_nchw.mDesc.GetLengths()[2];
+    std::size_t W = A_nchw.mDesc.GetLengths()[3];
+    for(std::size_t w = 0; w < W; ++w)
+        for(std::size_t h = 0; h < H; ++h)
+            for(std::size_t c = 0; c < C; ++c)
+                for(std::size_t n = 0; n < N; ++n)
+                {
+                    using tmp_type   = ck::remove_reference_t<decltype(B_nhwc(0, 0))>;
+                    tmp_type tmp_val = 0;
+                    auto a_val       = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
+                    functor_b(tmp_val, a_val);
+                    functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
+                              scale * tmp_val);
+                }
+}
+
+template <typename ADataType, typename BDataType, index_t NumDim>
+bool test_permute_scale_impl(int do_verification,
+                             int init_method,
+                             bool do_log,
+                             bool time_kernel,
+                             std::vector<index_t> lengths)
+{
+    bool pass = true;
+
+    using ElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryOp   = ck::tensor_operation::element_wise::UnarySquare;
+    using Scale     = ck::tensor_operation::element_wise::Scale;
+    float scale     = 2.f;
+
+    index_t N = lengths[0];
+    index_t C = lengths[1];
+    index_t H = lengths[2];
+    index_t W = lengths[3];
+
+    std::vector<ck::index_t> nchw = {N, C, H, W};
+    std::vector<ck::index_t> nhwc = {N, H, W, C};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+    Tensor<BDataType> host_b(nhwc);
+
+    std::array<ck::index_t, 4> ab_lengths;
+
+    std::array<ck::index_t, 4> a_strides = {1,
+                                            static_cast<int>(nchw[0]),
+                                            static_cast<int>(nchw[0] * nchw[1]),
+                                            static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
+
+    std::array<ck::index_t, 4> b_strides = {1,
+                                            static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
+                                            static_cast<int>(nhwc[0]),
+                                            static_cast<int>(nhwc[0] * nhwc[1])};
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    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}
+        std::mt19937 gen(11939);
+        std::uniform_int_distribution<int> dis(0, 1);
+        auto i = 0;
+        for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
+            for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
+                    for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
+                    {
+                        a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
+                                (h * nchw[3]) + w] = i;
+                        i                          = dis(gen);
+                    }
+    }
+
+    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,
+                                                                     UnaryOp,
+                                                                     Scale,
+                                                                     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;
+
+    std::string best_instance_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    if(do_verification)
+    {
+        host_elementwise4D(host_b, a, ElementOp{}, UnaryOp{}, scale);
+    }
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(ab_lengths,
+                                                        {a_strides},
+                                                        {b_strides},
+                                                        input,
+                                                        output,
+                                                        ElementOp{},
+                                                        UnaryOp{},
+                                                        Scale{scale});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            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) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+            std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                                    sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+            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;
+
+            if(tflops > best_tflops)
+            {
+                best_instance_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;
+        }
+    }
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", lengths, ",") << ", ";
+        std::cout << "best perf = " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    return true;
+}
+
+} // namespace ck

test/wrapper/CMakeLists.txt

+add_gtest_executable(test_layout test_layout.cpp)
+target_link_libraries(test_layout PRIVATE utility)
+add_gtest_executable(test_tensor test_tensor.cpp)
+target_link_libraries(test_tensor PRIVATE utility)

test/wrapper/test_layout.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/wrapper/layout.hpp"
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+
+class TestWrapperLayout : public ::testing::Test
+{
+    protected:
+    static constexpr auto I0 = ck::Number<0>{};
+    static constexpr auto I1 = ck::Number<1>{};
+
+    template <typename Desc,
+              typename Desc1d,
+              typename LayoutRuntime,
+              typename LayoutCompiletime,
+              typename Idxs>
+    void Run(Desc& desc,
+             Desc1d& desc_1d,
+             LayoutRuntime& layout_runtime,
+             LayoutCompiletime& layout_compiletime,
+             const std::vector<Idxs>& idxs)
+    {
+        // 1d check
+        EXPECT_EQ(desc_1d.GetLength(I0), ck::wrapper::size(layout_runtime));
+        // Check layout compiletime and runtime result consistency
+        EXPECT_EQ(ck::wrapper::size(layout_runtime), ck::wrapper::size(layout_compiletime));
+
+        for(ck::index_t i = 0; i < desc_1d.GetLength(I0); i++)
+        {
+            const ck::index_t layout_runtime_offset_1d     = layout_runtime(ck::make_tuple(i));
+            const ck::index_t layout_compiletime_offset_1d = layout_compiletime(ck::make_tuple(i));
+            const ck::index_t desc_offset_1d = desc_1d.CalculateOffset(ck::make_tuple(i));
+            EXPECT_EQ(layout_runtime_offset_1d, desc_offset_1d);
+            EXPECT_EQ(layout_compiletime_offset_1d, layout_runtime_offset_1d);
+        }
+        // size(layout)-d check, don't check if access is hierarchical
+        if constexpr(!IsNestedTuple(Idxs{}))
+        {
+            ck::static_for<0, Idxs::Size(), 1>{}([&](auto d) {
+                EXPECT_EQ(desc.GetLength(ck::Number<d>{}), ck::wrapper::size<d>(layout_runtime));
+                EXPECT_EQ(ck::wrapper::size<d>(layout_runtime),
+                          ck::wrapper::size<d>(layout_compiletime));
+            });
+        }
+        for(const auto idx : idxs)
+        {
+            const ck::index_t layout_runtime_offset     = layout_runtime(idx);
+            const ck::index_t layout_compiletime_offset = layout_compiletime(idx);
+            const ck::index_t desc_offset =
+                desc.CalculateOffset(UnrollNestedTuple(idx)); // Unroll if nested
+            EXPECT_EQ(layout_runtime_offset, desc_offset);
+            EXPECT_EQ(layout_runtime_offset, layout_compiletime_offset);
+        }
+    }
+};
+
+TEST_F(TestWrapperLayout, 2d)
+{
+    // dims:(4, 3) strides:(1, 4)
+    constexpr ck::index_t d1 = 4;
+    constexpr ck::index_t d0 = 3;
+    constexpr ck::index_t s1 = 1;
+    constexpr ck::index_t s0 = 4;
+    const auto desc =
+        ck::make_naive_tensor_descriptor(ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}),
+                                         ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
+    // Reverse due to column major
+    const auto desc_1d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1))),
+        ck::make_tuple(ck::Sequence<1, 0>{}),
+        ck::make_tuple(ck::Sequence<0>{}));
+    const auto layout_runtime = ck::wrapper::make_layout(ck::make_tuple(d1, d0));
+    const auto layout_compiletime =
+        ck::wrapper::make_layout(ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}));
+    std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs;
+
+    for(ck::index_t h = 0; h < d1; h++)
+    {
+        for(ck::index_t w = 0; w < d0; w++)
+        {
+            idxs.emplace_back(h, w);
+        }
+    }
+
+    this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs);
+}
+
+TEST_F(TestWrapperLayout, 3d_nested)
+{
+    // dims:((2, 3), 4, 3) strides:((2, 4), 12, 48)
+    constexpr ck::index_t d3 = 2;
+    constexpr ck::index_t d2 = 3;
+    constexpr ck::index_t d1 = 4;
+    constexpr ck::index_t d0 = 3;
+    constexpr ck::index_t s3 = 2;
+    constexpr ck::index_t s2 = 4;
+    constexpr ck::index_t s1 = 12;
+    constexpr ck::index_t s0 = 48;
+    const auto desc          = ck::make_naive_tensor_descriptor(
+        ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}, ck::Number<d1>{}, ck::Number<d0>{}),
+        ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}, ck::Number<s1>{}, ck::Number<s0>{}));
+    // Reverse due to column major
+    const auto desc_1d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3))),
+        ck::make_tuple(ck::Sequence<3, 2, 1, 0>{}),
+        ck::make_tuple(ck::Sequence<0>{}));
+    const auto desc_3d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3)),
+                       ck::make_pass_through_transform(d1),
+                       ck::make_pass_through_transform(d2)),
+        ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}, ck::Sequence<3>{}),
+        ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}, ck::Sequence<2>{}));
+    const auto layout_runtime =
+        ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(d3, d2), d1, d0),
+                                 ck::make_tuple(ck::make_tuple(s3, s2), s1, s0));
+    const auto layout_compiletime = ck::wrapper::make_layout(
+        ck::make_tuple(
+            ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}), ck::Number<d1>{}, ck::Number<d0>{}),
+        ck::make_tuple(ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}),
+                       ck::Number<s1>{},
+                       ck::Number<s0>{}));
+    std::vector<ck::Tuple<ck::index_t, ck::index_t, ck::index_t>> idxs_3d;
+
+    for(ck::index_t d = 0; d < d2 * d3; d++)
+    {
+        for(ck::index_t h = 0; h < d1; h++)
+        {
+            for(ck::index_t w = 0; w < d0; w++)
+            {
+                idxs_3d.emplace_back(d, h, w);
+            }
+        }
+    }
+    this->Run(desc_3d, desc_1d, layout_runtime, layout_compiletime, idxs_3d);
+
+    // Check also 4d iteration
+    std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t, ck::index_t>> idxs_4d;
+
+    for(ck::index_t e = 0; e < d3; e++)
+    {
+        for(ck::index_t d = 0; d < d2; d++)
+        {
+            for(ck::index_t h = 0; h < d1; h++)
+            {
+                for(ck::index_t w = 0; w < d0; w++)
+                {
+                    idxs_4d.emplace_back(ck::make_tuple(e, d), h, w);
+                }
+            }
+        }
+    }
+    this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_4d);
+}
+
+TEST_F(TestWrapperLayout, 2d_nested)
+{
+    // dims:((2, 3), (4, 3)) strides:((2, 4), (48, 12))
+    constexpr ck::index_t d3 = 2;
+    constexpr ck::index_t d2 = 3;
+    constexpr ck::index_t d1 = 4;
+    constexpr ck::index_t d0 = 3;
+    constexpr ck::index_t s3 = 2;
+    constexpr ck::index_t s2 = 4;
+    constexpr ck::index_t s1 = 48;
+    constexpr ck::index_t s0 = 12;
+    const auto desc          = ck::make_naive_tensor_descriptor(
+        ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}, ck::Number<d1>{}, ck::Number<d0>{}),
+        ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}, ck::Number<s1>{}, ck::Number<s0>{}));
+    // Reverse due to column major
+    const auto desc_1d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3))),
+        ck::make_tuple(ck::Sequence<3, 2, 1, 0>{}),
+        ck::make_tuple(ck::Sequence<0>{}));
+    const auto desc_2d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3)),
+                       ck::make_merge_transform(ck::make_tuple(d0, d1))),
+        ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<3, 2>{}),
+        ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
+    const auto layout_runtime =
+        ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(d3, d2), ck::make_tuple(d1, d0)),
+                                 ck::make_tuple(ck::make_tuple(s3, s2), ck::make_tuple(s1, s0)));
+    const auto layout_compiletime = ck::wrapper::make_layout(
+        ck::make_tuple(ck::make_tuple(ck::Number<d3>{}, ck::Number<d2>{}),
+                       ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})),
+        ck::make_tuple(ck::make_tuple(ck::Number<s3>{}, ck::Number<s2>{}),
+                       ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{})));
+    std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs_2d;
+
+    for(ck::index_t h = 0; h < d2 * d3; h++)
+    {
+        for(ck::index_t w = 0; w < d0 * d1; w++)
+        {
+            idxs_2d.emplace_back(h, w);
+        }
+    }
+    this->Run(desc_2d, desc_1d, layout_runtime, layout_compiletime, idxs_2d);
+    // Check also 4d iteration
+    std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::Tuple<ck::index_t, ck::index_t>>>
+        idxs_4d;
+
+    for(ck::index_t e = 0; e < d3; e++)
+    {
+        for(ck::index_t d = 0; d < d2; d++)
+        {
+            for(ck::index_t h = 0; h < d1; h++)
+            {
+                for(ck::index_t w = 0; w < d0; w++)
+                {
+                    idxs_4d.emplace_back(ck::make_tuple(e, d), ck::make_tuple(h, w));
+                }
+            }
+        }
+    }
+    this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_4d);
+}
+
+TEST_F(TestWrapperLayout, 3d_double_nested)
+{
+    // dims:(((2, 2), 3), (4, 3)) strides:(((2, 4), 8), (96, 24))
+    constexpr ck::index_t d4 = 2;
+    constexpr ck::index_t d3 = 2;
+    constexpr ck::index_t d2 = 3;
+    constexpr ck::index_t d1 = 4;
+    constexpr ck::index_t d0 = 3;
+    constexpr ck::index_t s4 = 2;
+    constexpr ck::index_t s3 = 4;
+    constexpr ck::index_t s2 = 8;
+    constexpr ck::index_t s1 = 96;
+    constexpr ck::index_t s0 = 24;
+    const auto desc          = ck::make_naive_tensor_descriptor(ck::make_tuple(ck::Number<d4>{},
+                                                                      ck::Number<d3>{},
+                                                                      ck::Number<d2>{},
+                                                                      ck::Number<d1>{},
+                                                                      ck::Number<d0>{}),
+                                                       ck::make_tuple(ck::Number<s4>{},
+                                                                      ck::Number<s3>{},
+                                                                      ck::Number<s2>{},
+                                                                      ck::Number<s1>{},
+                                                                      ck::Number<s0>{}));
+    // Reverse due to column major
+    const auto desc_1d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d0, d1, d2, d3, d4))),
+        ck::make_tuple(ck::Sequence<4, 3, 2, 1, 0>{}),
+        ck::make_tuple(ck::Sequence<0>{}));
+    const auto desc_3d = transform_tensor_descriptor(
+        desc,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d3, d4)),
+                       ck::make_pass_through_transform(d2),
+                       ck::make_merge_transform(ck::make_tuple(d0, d1))),
+        ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}, ck::Sequence<4, 3>{}),
+        ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}, ck::Sequence<2>{}));
+    const auto desc_2d = transform_tensor_descriptor(
+        desc_3d,
+        ck::make_tuple(ck::make_merge_transform(ck::make_tuple(d2, d3 * d4)),
+                       ck::make_pass_through_transform(d1 * d0)),
+        ck::make_tuple(ck::Sequence<1, 0>{}, ck::Sequence<2>{}),
+        ck::make_tuple(ck::Sequence<0>{}, ck::Sequence<1>{}));
+    const auto layout_runtime = ck::wrapper::make_layout(
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)),
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, s3), s2), ck::make_tuple(s1, s0)));
+    const auto layout_compiletime = ck::wrapper::make_layout(
+        ck::make_tuple(
+            ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
+            ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})),
+        ck::make_tuple(
+            ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<s3>{}), ck::Number<s2>{}),
+            ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{})));
+    std::vector<ck::Tuple<ck::index_t, ck::index_t>> idxs_2d;
+
+    for(ck::index_t h = 0; h < d2 * d3 * d4; h++)
+    {
+        for(ck::index_t w = 0; w < d0 * d1; w++)
+        {
+            idxs_2d.emplace_back(h, w);
+        }
+    }
+    this->Run(desc_2d, desc_1d, layout_runtime, layout_compiletime, idxs_2d);
+    // Check also 3d iteration
+    std::vector<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t>> idxs_3d;
+
+    for(ck::index_t d = 0; d < d3 * d4; d++)
+    {
+        for(ck::index_t h = 0; h < d2; h++)
+        {
+            for(ck::index_t w = 0; w < d1 * d0; w++)
+            {
+                idxs_3d.emplace_back(ck::make_tuple(d, h), w);
+            }
+        }
+    }
+    this->Run(desc_3d, desc_1d, layout_runtime, layout_compiletime, idxs_3d);
+    // Check also 5d iteration
+    std::vector<ck::Tuple<ck::Tuple<ck::Tuple<ck::index_t, ck::index_t>, ck::index_t>,
+                          ck::Tuple<ck::index_t, ck::index_t>>>
+        idxs_5d;
+
+    for(ck::index_t f = 0; f < d4; f++)
+    {
+        for(ck::index_t e = 0; e < d3; e++)
+        {
+            for(ck::index_t d = 0; d < d2; d++)
+            {
+                for(ck::index_t h = 0; h < d1; h++)
+                {
+                    for(ck::index_t w = 0; w < d0; w++)
+                    {
+                        idxs_5d.emplace_back(ck::make_tuple(ck::make_tuple(f, e), d),
+                                             ck::make_tuple(h, w));
+                    }
+                }
+            }
+        }
+    }
+    this->Run(desc, desc_1d, layout_runtime, layout_compiletime, idxs_5d);
+}
+
+TEST(TestLayoutHelpers, SizeAndGet)
+{
+    // dims:(((2, 2), 3), (4, 3))
+    constexpr ck::index_t d4  = 2;
+    constexpr ck::index_t d3  = 2;
+    constexpr ck::index_t d2  = 3;
+    constexpr ck::index_t d1  = 4;
+    constexpr ck::index_t d0  = 3;
+    const auto layout_runtime = ck::wrapper::make_layout(
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)));
+    const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
+        ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
+        ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
+
+    // Size of layout
+    EXPECT_EQ(ck::wrapper::size(layout_runtime), d4 * d3 * d2 * d1 * d0);
+    EXPECT_EQ(ck::wrapper::size(layout_compiletime), d4 * d3 * d2 * d1 * d0);
+
+    // Size of dims
+    EXPECT_EQ(ck::wrapper::size<0>(layout_runtime), d4 * d3 * d2);
+    EXPECT_EQ(ck::wrapper::size<0>(layout_compiletime), d4 * d3 * d2);
+    EXPECT_EQ(ck::wrapper::size<1>(layout_runtime), d1 * d0);
+    EXPECT_EQ(ck::wrapper::size<1>(layout_compiletime), d1 * d0);
+
+    // Access through new layout (using get with layout object)
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(layout_runtime)), d4 * d3);
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(layout_compiletime)), d4 * d3);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_runtime)), d2);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_compiletime)), d2);
+
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_runtime))), d4);
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_compiletime))),
+              d4);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_runtime))), d3);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(ck::wrapper::get<0>(layout_compiletime))),
+              d3);
+
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_runtime)), d2);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<0>(layout_compiletime)), d2);
+
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<1>(layout_runtime)), d1);
+    EXPECT_EQ(ck::wrapper::size<0>(ck::wrapper::get<1>(layout_compiletime)), d1);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<1>(layout_runtime)), d0);
+    EXPECT_EQ(ck::wrapper::size<1>(ck::wrapper::get<1>(layout_compiletime)), d0);
+}
+
+TEST(TestLayoutHelpers, DepthAndRank)
+{
+    // dims:(((2, 2), 3), (4, 3))
+    constexpr ck::index_t d4  = 2;
+    constexpr ck::index_t d3  = 2;
+    constexpr ck::index_t d2  = 3;
+    constexpr ck::index_t d1  = 4;
+    constexpr ck::index_t d0  = 3;
+    const auto layout_runtime = ck::wrapper::make_layout(
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0)));
+    const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
+        ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
+        ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
+
+    EXPECT_EQ(ck::wrapper::depth(layout_runtime), 3);
+    EXPECT_EQ(ck::wrapper::depth(layout_compiletime), 3);
+    EXPECT_EQ(ck::wrapper::depth(ck::make_tuple(ck::make_tuple(d4, d3), d2)), 2);
+    // Check for integer
+    EXPECT_EQ(ck::wrapper::depth(d0), 0);
+
+    EXPECT_EQ(ck::wrapper::rank(layout_runtime), 2);
+    EXPECT_EQ(ck::wrapper::rank(layout_compiletime), 2);
+    EXPECT_EQ(ck::wrapper::rank(ck::make_tuple(ck::make_tuple(d4, d3), d2)), 2);
+    // Check for integer
+    EXPECT_EQ(ck::wrapper::rank(d0), 1);
+}
+
+TEST(TestLayoutHelpers, ShapeAndStrides)
+{
+    // dims:(((2, 2), 3), (4, 3))
+    constexpr ck::index_t d4     = 2;
+    constexpr ck::index_t d3     = 2;
+    constexpr ck::index_t d2     = 3;
+    constexpr ck::index_t d1     = 4;
+    constexpr ck::index_t d0     = 3;
+    constexpr ck::index_t s4     = 2;
+    constexpr ck::index_t s3     = 4;
+    constexpr ck::index_t s2     = 8;
+    constexpr ck::index_t s1     = 96;
+    constexpr ck::index_t s0     = 24;
+    const auto shape_compiletime = ck::make_tuple(
+        ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
+        ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{}));
+    const auto strides_compiletime = ck::make_tuple(
+        ck::make_tuple(ck::make_tuple(ck::Number<s4>{}, ck::Number<s3>{}), ck::Number<s2>{}),
+        ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
+    const auto shape_runtime =
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0));
+    const auto strides_runtime =
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(s4, s3), s2), ck::make_tuple(s1, s0));
+    const auto layout_runtime = ck::wrapper::make_layout(shape_runtime, strides_runtime);
+    const auto layout_compiletime =
+        ck::wrapper::make_layout(shape_compiletime, strides_compiletime);
+
+    constexpr bool check_compiletime_shape =
+        std::is_same_v<decltype(shape_compiletime),
+                       std::remove_reference_t<decltype(shape(layout_compiletime))>>;
+    constexpr bool check_compiletime_strides =
+        std::is_same_v<decltype(strides_compiletime),
+                       std::remove_reference_t<decltype(stride(layout_compiletime))>>;
+    constexpr bool check_runtime_shape =
+        std::is_same_v<decltype(shape_runtime),
+                       std::remove_reference_t<decltype(shape(layout_runtime))>>;
+    constexpr bool check_runtime_strides =
+        std::is_same_v<decltype(strides_runtime),
+                       std::remove_reference_t<decltype(stride(layout_runtime))>>;
+    EXPECT_TRUE(check_compiletime_shape);
+    EXPECT_TRUE(check_compiletime_strides);
+    EXPECT_TRUE(check_runtime_shape);
+    EXPECT_TRUE(check_runtime_strides);
+}
+
+TEST(TestLayoutHelpers, Hierarchical)
+{
+    // dims:(((2, 2), 3), (4, 3))
+    constexpr ck::index_t d4 = 2;
+    constexpr ck::index_t d3 = 2;
+    constexpr ck::index_t d2 = 3;
+    constexpr ck::index_t d1 = 4;
+    constexpr ck::index_t d0 = 3;
+    const auto runtime_shape =
+        ck::make_tuple(ck::make_tuple(ck::make_tuple(d4, d3), d2), ck::make_tuple(d1, d0));
+    const auto layout_runtime     = ck::wrapper::make_layout(runtime_shape);
+    const auto layout_compiletime = ck::wrapper::make_layout(ck::make_tuple(
+        ck::make_tuple(ck::make_tuple(ck::Number<d4>{}, ck::Number<d3>{}), ck::Number<d2>{}),
+        ck::make_tuple(ck::Number<d1>{}, ck::Number<d0>{})));
+
+    EXPECT_EQ((ck::wrapper::rank<0, 0>(runtime_shape)), 2);
+    EXPECT_EQ((ck::wrapper::rank<0, 0>(layout_runtime)), 2);
+    EXPECT_EQ((ck::wrapper::rank<0, 0>(layout_compiletime)), 2);
+
+    EXPECT_EQ((ck::wrapper::depth<0, 0>(runtime_shape)), 1);
+    EXPECT_EQ((ck::wrapper::depth<0, 0>(layout_runtime)), 1);
+    EXPECT_EQ((ck::wrapper::depth<0, 0>(layout_compiletime)), 1);
+
+    EXPECT_EQ((ck::wrapper::size<0, 0>(runtime_shape)), d4 * d3);
+    EXPECT_EQ((ck::wrapper::size<0, 0>(layout_runtime)), d4 * d3);
+    EXPECT_EQ((ck::wrapper::size<0, 0>(layout_compiletime)), d4 * d3);
+
+    EXPECT_EQ((ck::wrapper::get<0, 0, 0>(runtime_shape)), d4);
+}

test/wrapper/test_tensor.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/library/utility/device_memory.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/wrapper/layout.hpp"
+#include "ck/wrapper/tensor.hpp"
+
+// Compare data in tensor with offset from layout.
+// Data and offset should match if physical memory has been initialized with
+// sequentially increasing values from 0.
+template <typename TensorType>
+__host__ __device__ bool TestTensorCheck3d(TensorType& tensor)
+{
+    const auto& layout = ck::wrapper::layout(tensor);
+    for(ck::index_t d = 0; d < ck::wrapper::size<0>(ck::wrapper::get<0>(layout)); d++)
+    {
+        for(ck::index_t h = 0; h < ck::wrapper::size<1>(ck::wrapper::get<0>(layout)); h++)
+        {
+            for(ck::index_t w = 0; w < ck::wrapper::size<1>(layout); w++)
+            {
+                const auto idx = ck::make_tuple(ck::make_tuple(d, h), w);
+                if(tensor(idx) != layout(idx))
+                {
+                    return false;
+                }
+            }
+        }
+    }
+    return true;
+}
+
+template <typename TensorType>
+__host__ __device__ bool TestTensorCheck1d(TensorType& tensor, ck::index_t start_offset = 0)
+{
+    const auto& layout = ck::wrapper::layout(tensor);
+    for(ck::index_t w = 0; w < ck::wrapper::size<0>(layout); w++)
+    {
+        if(tensor(w) - start_offset != layout(ck::make_tuple(w)))
+        {
+            return false;
+        }
+    }
+    return true;
+}
+
+template <ck::index_t nelems, typename TensorType>
+__host__ __device__ bool StaticTestTensorCheck1d(TensorType& tensor)
+{
+    const auto& layout = ck::wrapper::layout(tensor);
+    bool success       = true;
+    ck::static_for<0, nelems, 1>{}([&](auto w) {
+        if(tensor(ck::Number<w.value>{}) != layout(ck::make_tuple(w.value)))
+        {
+            success = false;
+        }
+    });
+    return success;
+}
+
+template <typename TensorType>
+__host__ __device__ void InitTensor(TensorType& tensor)
+{
+    for(ck::index_t i = 0; i < ck::wrapper::size(ck::wrapper::layout(tensor)); i++)
+    {
+        tensor(i) = i;
+    }
+}
+
+template <ck::index_t nelems, typename TensorType>
+__host__ __device__ void StaticInitTensor(TensorType& tensor)
+{
+
+    ck::static_for<0, nelems, 1>{}([&](auto i) { tensor(ck::Number<i.value>{}) = i.value; });
+}
+
+// Tests
+TEST(TestTensor, ReadWriteHostMemory)
+{
+    constexpr ck::index_t nelems = 8;
+
+    std::array<ck::index_t, nelems> data;
+    const auto layout = ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(2, 2), 2));
+    auto tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(&data[0], layout);
+    InitTensor(tensor);
+
+    EXPECT_TRUE(TestTensorCheck1d(tensor));
+    EXPECT_TRUE(TestTensorCheck3d(tensor));
+}
+
+__global__ void TestTensorReadWriteDevice(void* data, void* success)
+{
+    constexpr ck::index_t nelems            = 8;
+    constexpr ck::index_t scalar_per_vector = 1;
+    __shared__ ck::index_t p_shared[nelems];
+
+    ck::index_t* casted_data_ptr = static_cast<ck::index_t*>(data);
+    bool* casted_success_ptr     = static_cast<bool*>(success);
+
+    const auto layout = ck::wrapper::make_layout(ck::make_tuple(ck::make_tuple(2, 2), 2));
+    constexpr auto register_layout = ck::wrapper::make_layout(ck::make_tuple(ck::Number<8>{}));
+
+    auto tensor_global =
+        ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(casted_data_ptr, layout);
+    auto tensor_lds  = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(p_shared, layout);
+    auto tensor_vgpr = ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr,
+                                                         nelems,
+                                                         scalar_per_vector,
+                                                         ck::index_t>(register_layout);
+    auto tensor_sgpr = ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Sgpr,
+                                                         nelems,
+                                                         scalar_per_vector,
+                                                         ck::index_t>(register_layout);
+
+    InitTensor(tensor_global);
+    InitTensor(tensor_lds);
+    StaticInitTensor<nelems>(tensor_vgpr);
+    StaticInitTensor<nelems>(tensor_sgpr);
+
+    *casted_success_ptr &= TestTensorCheck1d(tensor_global);
+    *casted_success_ptr &= TestTensorCheck3d(tensor_global);
+
+    *casted_success_ptr &= TestTensorCheck1d(tensor_lds);
+    *casted_success_ptr &= TestTensorCheck3d(tensor_lds);
+
+    *casted_success_ptr &= StaticTestTensorCheck1d<nelems>(tensor_vgpr);
+
+    *casted_success_ptr &= StaticTestTensorCheck1d<nelems>(tensor_sgpr);
+}
+
+TEST(TestTensor, ReadWriteGlobalLdsRegistersMemory)
+{
+    constexpr ck::index_t nelems = 8;
+    std::array<ck::index_t, nelems> host_data;
+
+    DeviceMem data_buf(nelems * sizeof(ck::index_t));
+    data_buf.ToDevice(&host_data[0]);
+    DeviceMem success_buf(sizeof(bool));
+
+    launch_and_time_kernel(StreamConfig{},
+                           TestTensorReadWriteDevice,
+                           dim3(1),
+                           dim3(1),
+                           nelems * sizeof(ck::index_t),
+                           data_buf.GetDeviceBuffer(),
+                           success_buf.GetDeviceBuffer());
+
+    bool success;
+    success_buf.FromDevice(&success);
+    EXPECT_TRUE(success);
+}
+
+TEST(TestTensor, Slicing)
+{
+    constexpr ck::index_t nelems = 8;
+
+    std::array<ck::index_t, nelems> data;
+    const auto shape   = ck::make_tuple(ck::make_tuple(2, 2), 2);
+    const auto strides = ck::make_tuple(ck::make_tuple(1, 2), 4);
+    const auto layout  = ck::wrapper::make_layout(shape, strides);
+    auto tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(&data[0], layout);
+    InitTensor(tensor);
+
+    auto tensor2x2x2 =
+        tensor(ck::make_tuple(ck::wrapper::slice(2), ck::wrapper::slice(2)), ck::wrapper::slice(2));
+    EXPECT_EQ(ck::wrapper::rank(tensor2x2x2), 2);
+    EXPECT_EQ(ck::wrapper::depth(tensor2x2x2), 2);
+    EXPECT_EQ(ck::wrapper::size(tensor2x2x2), 8);
+    EXPECT_TRUE(TestTensorCheck1d(tensor2x2x2));
+
+    auto tensor2x2 = tensor(ck::make_tuple(1, ck::wrapper::slice(2)), ck::wrapper::slice(2));
+    EXPECT_EQ(ck::wrapper::rank(tensor2x2), 2);
+    EXPECT_EQ(ck::wrapper::depth(tensor2x2), 2);
+    EXPECT_EQ(ck::wrapper::size(tensor2x2), 4);
+    EXPECT_TRUE(TestTensorCheck1d(tensor2x2, layout(ck::make_tuple(ck::make_tuple(1, 0), 0))));
+
+    auto tensor1x1 = tensor(ck::make_tuple(1, ck::wrapper::slice(1, 2)), ck::wrapper::slice(1, 2));
+    EXPECT_EQ(rank(tensor1x1), 2);
+    EXPECT_EQ(depth(tensor1x1), 2);
+    EXPECT_EQ(size(tensor1x1), 1);
+    EXPECT_TRUE(TestTensorCheck1d(tensor1x1, layout(ck::make_tuple(ck::make_tuple(1, 1), 1))));
+
+    auto tensor2 = tensor(ck::make_tuple(1, 1), ck::wrapper::slice(0, 2));
+    EXPECT_EQ(ck::wrapper::rank(tensor2), 1);
+    EXPECT_EQ(ck::wrapper::depth(tensor2), 1);
+    EXPECT_EQ(ck::wrapper::size(tensor2), 2);
+    EXPECT_TRUE(TestTensorCheck1d(tensor2, layout(ck::make_tuple(ck::make_tuple(1, 1), 0))));
+
+    // negative indexing
+    auto tensor1x2 = tensor(ck::make_tuple(1, ck::wrapper::slice(0, -2)), ck::wrapper::slice());
+    EXPECT_EQ(rank(tensor1x2), 2);
+    EXPECT_EQ(depth(tensor1x2), 2);
+    EXPECT_EQ(size(tensor1x2), 2);
+    EXPECT_TRUE(TestTensorCheck1d(tensor1x2, layout(ck::make_tuple(ck::make_tuple(1, 0), 0))));
+}