Merge pull request #33 from ROCm/lwpck-1292

Merge from the public repo.

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/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"
+#include "profiler/profile_transpose_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;
+using ck::index_t;
 
 template <typename Tuple>
 class TestTranspose : 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, 16, 16, 32, 5}, {8, 16, 16, 32, 8} /**{32, 16, 16, 32, 8},**/};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_transpose_impl<ADataType, BDataType, 5>(
+                true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
 };
 
-// clang-format off
-using KernelTypes = ::testing::Types<
-    std::tuple<      F16,       F16>,
-    std::tuple<      F32,       F32>
-    >;
-// clang-format on
+using KernelTypes = ::testing::Types<std::tuple<F16, F16>, std::tuple<F32, F32>>;
 
 TYPED_TEST_SUITE(TestTranspose, KernelTypes);
-
-//#include "test_transpose_ut_cases.inc"
+TYPED_TEST(TestTranspose, Test_FP16) { this->Run(); }
+TYPED_TEST(TestTranspose, Test_FP32) { this->Run(); }

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

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)
+add_gtest_executable(test_copy test_copy.cpp)
+target_link_libraries(test_copy PRIVATE utility)
+add_gtest_executable(test_partition test_partition.cpp)
+target_link_libraries(test_partition PRIVATE utility)
+if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR
+   GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR
+   GPU_TARGETS MATCHES "gfx942")
+    add_gtest_executable(test_gemm test_gemm.cpp)
+    target_link_libraries(test_gemm PRIVATE utility)
+endif()

test/wrapper/test_copy.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/wrapper/layout.hpp"
+#include "ck/wrapper/tensor.hpp"
+#include "ck/wrapper/operations/copy.hpp"
+
+// Test copy from Global to Global through LDS and VGPR
+template <typename InputTensor,
+          typename OutputTensor,
+          typename BlockShape,
+          typename ThreadLayoutShape,
+          bool UseOptimizedCopy>
+__global__ void TestCopyDevice(const InputTensor input_tensor,
+                               OutputTensor output_tensor,
+                               const BlockShape tile_shape,
+                               const ThreadLayoutShape thread_layout)
+{
+    __shared__ ck::index_t p_shared[ck::wrapper::size(tile_shape)];
+    const auto tensor_lds = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
+        p_shared, ck::wrapper::make_layout(tile_shape));
+
+    const auto block_idx = static_cast<ck::index_t>(blockIdx.x);
+
+    // Get local tiles for global memory
+    const auto input_local_tile = ck::wrapper::make_local_tile(input_tensor, tile_shape, block_idx);
+    const auto output_local_tile =
+        ck::wrapper::make_local_tile(output_tensor, tile_shape, block_idx);
+
+    // Get partition per thread
+    const auto input_local_partition =
+        ck::wrapper::make_local_partition(input_local_tile, thread_layout, threadIdx.x);
+    auto lds_local_partition =
+        ck::wrapper::make_local_partition(tensor_lds, thread_layout, threadIdx.x);
+    auto output_local_partition =
+        ck::wrapper::make_local_partition(output_local_tile, thread_layout, threadIdx.x);
+
+    // Allocate VGPR
+    auto tensor_vgpr =
+        ck::wrapper::make_register_tensor<ck::wrapper::MemoryTypeEnum::Vgpr, ck::index_t>(
+            layout(lds_local_partition));
+
+    // Perform copy
+    if constexpr(UseOptimizedCopy)
+    {
+        using DimAccessOrder                    = ck::Tuple<ck::Number<1>, ck::Number<0>>;
+        constexpr ck::index_t vector_dim        = 0;
+        constexpr ck::index_t scalar_per_vector = 2;
+        ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(input_local_partition,
+                                                                         lds_local_partition);
+        // TODO: Enable optimized copy for static buffers
+        ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(lds_local_partition,
+                                                                         tensor_vgpr);
+        ck::wrapper::copy<DimAccessOrder, vector_dim, scalar_per_vector>(tensor_vgpr,
+                                                                         output_local_partition);
+    }
+    else
+    {
+        ck::wrapper::copy(input_local_partition, lds_local_partition);
+        ck::wrapper::copy(lds_local_partition, tensor_vgpr);
+        ck::wrapper::copy(tensor_vgpr, output_local_partition);
+    }
+}
+
+template <bool UseOptimizedCopy>
+void PerformCopyGlobalToGlobalViaLDS()
+{
+    const auto shape =
+        ck::make_tuple(ck::make_tuple(ck::Number<2>{}, ck::Number<2>{}), ck::Number<256>{});
+    const auto strides =
+        ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<2>{}), ck::Number<4>{});
+    const auto layout = ck::wrapper::make_layout(shape, strides);
+
+    // 0, 1, 2, ..., size(shape) - 1
+    std::vector<ck::index_t> input_data(ck::wrapper::size(shape));
+    std::iota(input_data.begin(), input_data.end(), 0);
+
+    // Global memory buffers
+    DeviceMem in_buf(ck::wrapper::size(layout) * sizeof(ck::index_t));
+    DeviceMem out_buf(ck::wrapper::size(layout) * sizeof(ck::index_t));
+
+    in_buf.ToDevice(input_data.data());
+    out_buf.SetZero();
+
+    // Create tensors for global memory
+    const auto input_tensor_global = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
+        static_cast<const ck::index_t*>(in_buf.GetDeviceBuffer()), layout);
+    auto output_tensor_global = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
+        static_cast<ck::index_t*>(out_buf.GetDeviceBuffer()), layout);
+
+    const auto thread_layout = ck::make_tuple(ck::Number<1>{}, ck::Number<32>{});
+    const auto tile_shape    = ck::make_tuple(ck::Number<4>{}, ck::Number<64>{});
+
+    const ck::index_t grid_size = ck::math::integer_divide_ceil(
+        ck::wrapper::size(input_tensor_global), ck::wrapper::size(tile_shape));
+
+    const auto kernel = TestCopyDevice<decltype(input_tensor_global),
+                                       decltype(output_tensor_global),
+                                       decltype(tile_shape),
+                                       decltype(thread_layout),
+                                       UseOptimizedCopy>;
+    launch_and_time_kernel(StreamConfig{},
+                           kernel,
+                           dim3(grid_size),
+                           dim3(ck::wrapper::size(thread_layout)),
+                           0,
+                           input_tensor_global,
+                           output_tensor_global,
+                           tile_shape,
+                           thread_layout);
+
+    // Verify results
+    std::vector<ck::index_t> output_data(ck::wrapper::size(shape));
+    out_buf.FromDevice(output_data.data());
+    EXPECT_TRUE(ck::utils::check_err(output_data, input_data));
+}
+
+TEST(TestCopyGlobalToGlobalViaLDS, GenericCopy) { PerformCopyGlobalToGlobalViaLDS<false>(); }
+TEST(TestCopyGlobalToGlobalViaLDS, OptimizedCopy) { PerformCopyGlobalToGlobalViaLDS<true>(); }

test/wrapper/test_gemm.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/library/utility/host_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/wrapper/layout.hpp"
+#include "ck/wrapper/tensor.hpp"
+#include "ck/wrapper/operations/copy.hpp"
+#include "ck/wrapper/operations/gemm.hpp"
+
+template <typename DataType>
+void CheckResult(const std::vector<DataType>& a_data,
+                 const std::vector<DataType>& b_data,
+                 std::vector<DataType>& c_m_n_device_result,
+                 const ck::index_t M,
+                 const ck::index_t N,
+                 const ck::index_t K)
+{
+    using PassThrough           = ck::tensor_operation::element_wise::PassThrough;
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<DataType, DataType, DataType, float, PassThrough, PassThrough, PassThrough>;
+
+    Tensor<DataType> a_m_k(HostTensorDescriptor({M, K}));
+    Tensor<DataType> b_k_n(HostTensorDescriptor({K, N}, {1, K}));
+    Tensor<DataType> c_m_n_host_result(HostTensorDescriptor({M, N}));
+
+    a_m_k.mData = a_data;
+    b_k_n.mData = b_data;
+
+    auto ref_op       = ReferenceGemmInstance{};
+    auto ref_invoker  = ref_op.MakeInvoker();
+    auto ref_argument = ref_op.MakeArgument(
+        a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
+
+    ref_invoker.Run(ref_argument);
+    EXPECT_TRUE(ck::utils::check_err(c_m_n_device_result, c_m_n_host_result.mData));
+}
+
+template <typename DataType,
+          typename GemmTraits,
+          ck::index_t scalar_per_vector,
+          typename BlockShape,
+          typename ThreadLayoutShape>
+__global__ void DeviceGemm(const void* p_a,
+                           const void* p_b,
+                           void* p_c,
+                           const ck::index_t M,
+                           const ck::index_t N,
+                           const ck::index_t K,
+                           const BlockShape tile_shape,
+                           const ThreadLayoutShape thread_layout)
+{
+    constexpr auto MPerBlock = ck::wrapper::size<0>(tile_shape);
+    constexpr auto NPerBlock = ck::wrapper::size<1>(tile_shape);
+    constexpr auto KPerBlock = ck::wrapper::size<2>(tile_shape);
+
+    const auto a_global_layout =
+        ck::wrapper::make_layout(ck::make_tuple(M, K), ck::make_tuple(K, 1));
+    const auto b_global_layout =
+        ck::wrapper::make_layout(ck::make_tuple(N, K), ck::make_tuple(K, 1));
+    const auto c_global_layout =
+        ck::wrapper::make_layout(ck::make_tuple(M, N), ck::make_tuple(N, 1));
+
+    constexpr auto a_tile_layout = ck::wrapper::make_layout(
+        ck::make_tuple(MPerBlock, KPerBlock), ck::make_tuple(KPerBlock, ck::Number<1>{}));
+    constexpr auto b_tile_layout = ck::wrapper::make_layout(
+        ck::make_tuple(NPerBlock, KPerBlock), ck::make_tuple(KPerBlock, ck::Number<1>{}));
+    constexpr auto c_tile_layout = ck::wrapper::make_layout(
+        ck::make_tuple(MPerBlock, NPerBlock), ck::make_tuple(NPerBlock, ck::Number<1>{}));
+
+    auto a_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
+        static_cast<const DataType*>(p_a), a_global_layout);
+    auto b_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
+        static_cast<const DataType*>(p_b), b_global_layout);
+    auto c_global_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Global>(
+        static_cast<DataType*>(p_c), c_global_layout);
+
+    auto a_padded_global_tensor = ck::wrapper::pad(a_global_tensor, shape(a_tile_layout));
+    auto b_padded_global_tensor = ck::wrapper::pad(b_global_tensor, shape(b_tile_layout));
+    auto c_padded_global_tensor = ck::wrapper::pad(c_global_tensor, shape(c_tile_layout));
+
+    __shared__ DataType lds_a[ck::wrapper::size(a_tile_layout)];
+    __shared__ DataType lds_b[ck::wrapper::size(b_tile_layout)];
+
+    auto a_lds_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
+        static_cast<DataType*>(lds_a), a_tile_layout);
+    auto b_lds_tensor = ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Lds>(
+        static_cast<DataType*>(lds_b), b_tile_layout);
+
+    const ck::index_t block_idx      = static_cast<ck::index_t>(blockIdx.x);
+    using DimAccessOrder             = ck::Tuple<ck::Number<0>, ck::Number<1>>;
+    constexpr ck::index_t vector_dim = 1;
+
+    auto c_global_local_tile = ck::wrapper::make_local_tile(
+        c_padded_global_tensor,
+        tile_shape,
+        block_idx,
+        make_tuple(ck::Number<1>{}, ck::Number<1>{}, ck::wrapper::slice(KPerBlock)));
+    auto c_global_local_partition =
+        ck::wrapper::make_blockwise_gemm_xdl_c_local_partition<DataType,
+                                                               decltype(a_tile_layout),
+                                                               decltype(b_tile_layout),
+                                                               ck::wrapper::size(thread_layout),
+                                                               GemmTraits>(c_global_local_tile);
+    auto c_vgpr_reg = ck::wrapper::make_blockwise_gemm_xdl_c_vgpr<DataType,
+                                                                  decltype(a_tile_layout),
+                                                                  decltype(b_tile_layout),
+                                                                  ck::wrapper::size(thread_layout),
+                                                                  GemmTraits>();
+    ck::wrapper::clear(c_vgpr_reg);
+
+    const ck::index_t num_loop = ck::math::integer_divide_ceil(K, KPerBlock);
+    ck::index_t i              = 0;
+    do
+    {
+        const auto k_slice = ck::wrapper::slice(i * KPerBlock, (i + 1) * KPerBlock);
+        auto a_padded_global_tensor_k_slice = a_padded_global_tensor(ck::wrapper::slice(), k_slice);
+        auto b_padded_global_tensor_k_slice = b_padded_global_tensor(ck::wrapper::slice(), k_slice);
+        auto a_global_local_tile            = ck::wrapper::make_local_tile(
+            a_padded_global_tensor_k_slice,
+            tile_shape,
+            block_idx,
+            make_tuple(ck::Number<1>{}, ck::wrapper::slice(N), ck::Number<1>{}));
+        auto b_global_local_tile = ck::wrapper::make_local_tile(
+            b_padded_global_tensor_k_slice,
+            tile_shape,
+            block_idx,
+            make_tuple(ck::wrapper::slice(M), ck::Number<1>{}, ck::Number<1>{}));
+
+        ck::wrapper::blockwise_copy<DimAccessOrder, vector_dim, scalar_per_vector>(
+            a_global_local_tile, a_lds_tensor, thread_layout);
+        ck::wrapper::blockwise_copy<DimAccessOrder, vector_dim, scalar_per_vector>(
+            b_global_local_tile, b_lds_tensor, thread_layout);
+        ck::block_sync_lds();
+        ck::wrapper::blockwise_gemm_xdl<DataType, ck::wrapper::size(thread_layout), GemmTraits>(
+            a_lds_tensor, b_lds_tensor, c_vgpr_reg);
+
+        ++i;
+    } while(i < num_loop);
+
+    ck::wrapper::copy(c_vgpr_reg, c_global_local_partition);
+}
+
+template <typename DataType,
+          typename GemmTraits,
+          ck::index_t scalar_per_vector,
+          typename BlockShape,
+          typename ThreadLayoutShape>
+void PerformGemm(const ck::index_t M,
+                 const ck::index_t N,
+                 const ck::index_t K,
+                 const BlockShape& tile_shape,
+                 const ThreadLayoutShape& thread_layout)
+{
+    // Global memory buffers
+    DeviceMem a_mem(M * K * sizeof(DataType));
+    DeviceMem b_mem(K * N * sizeof(DataType));
+    DeviceMem c_mem(M * N * sizeof(DataType));
+
+    std::vector<DataType> a_data(M * K);
+    std::vector<DataType> b_data(K * N);
+    ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(a_data);
+    ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(b_data);
+
+    a_mem.ToDevice(a_data.data());
+    b_mem.ToDevice(b_data.data());
+    c_mem.SetZero();
+
+    const ck::index_t grid_size =
+        ck::math::integer_divide_ceil(M, ck::wrapper::size<0>(tile_shape)) *
+        ck::math::integer_divide_ceil(N, ck::wrapper::size<1>(tile_shape));
+
+    const auto kernel =
+        DeviceGemm<DataType, GemmTraits, scalar_per_vector, BlockShape, ThreadLayoutShape>;
+    launch_and_time_kernel(StreamConfig{nullptr},
+                           kernel,
+                           dim3(grid_size),
+                           dim3(ck::wrapper::size(thread_layout)),
+                           0,
+                           a_mem.GetDeviceBuffer(),
+                           b_mem.GetDeviceBuffer(),
+                           c_mem.GetDeviceBuffer(),
+                           M,
+                           N,
+                           K,
+                           tile_shape,
+                           thread_layout);
+
+    std::vector<DataType> c_data(M * N);
+    c_mem.FromDevice(c_data.data());
+
+    CheckResult<DataType>(a_data, b_data, c_data, M, N, K);
+}
+
+TEST(TestGemm, Float)
+{
+    using DataType           = float;
+    const auto thread_layout = ck::make_tuple(ck::Number<16>{}, ck::Number<16>{});
+    const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<64>{});
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_4K1, 4>(
+        512, 512, 128, tile_shape, thread_layout);
+    // Irregular case
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_4K1, 1>(
+        129, 129, 67, tile_shape, thread_layout);
+}
+
+TEST(TestGemm, Int8)
+{
+    using DataType           = int8_t;
+    const auto thread_layout = ck::make_tuple(ck::Number<64>{}, ck::Number<4>{});
+    const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<64>{});
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_16K1, 16>(
+        512, 512, 128, tile_shape, thread_layout);
+    // Irregular case
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_16K1, 1>(
+        129, 129, 67, tile_shape, thread_layout);
+}
+
+TEST(TestGemm, Half)
+{
+    using DataType           = ck::half_t;
+    const auto thread_layout = ck::make_tuple(ck::Number<32>{}, ck::Number<8>{});
+    const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<64>{});
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_8K1, 8>(
+        512, 512, 128, tile_shape, thread_layout);
+    // Irregular case
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x2XdlPerWave_8K1, 1>(
+        129, 129, 67, tile_shape, thread_layout);
+}
+
+TEST(TestGemm, Float_2x4_4x2_XdlPerWave)
+{
+    using DataType                            = float;
+    const auto thread_layout_4x2_xdl_per_wave = ck::make_tuple(ck::Number<16>{}, ck::Number<8>{});
+    const auto thread_layout_2x4_xdl_per_wave = ck::make_tuple(ck::Number<8>{}, ck::Number<16>{});
+    const auto tile_shape = ck::make_tuple(ck::Number<128>{}, ck::Number<128>{}, ck::Number<64>{});
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_4x2XdlPerWave_4K1, 4>(
+        512, 512, 128, tile_shape, thread_layout_4x2_xdl_per_wave);
+    PerformGemm<DataType, ck::wrapper::BlockwisGemmXdlTraits_32x32Xdl_2x4XdlPerWave_4K1, 4>(
+        512, 512, 128, tile_shape, thread_layout_2x4_xdl_per_wave);
+}

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>{}),
+                                 ck::make_tuple(ck::Number<s1>{}, ck::Number<s0>{}));
+    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_runtime_shape =
+        std::is_same_v<decltype(shape_runtime),
+                       std::remove_reference_t<decltype(shape(layout_runtime))>>;
+    EXPECT_TRUE(check_compiletime_shape);
+    EXPECT_TRUE(check_runtime_shape);
+}
+
+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_partition.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/wrapper/layout.hpp"
+#include "ck/wrapper/tensor.hpp"
+
+TEST(TestPartition, LocalPartition)
+{
+    const auto shape =
+        ck::make_tuple(ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}), ck::Number<4>{});
+    const auto strides =
+        ck::make_tuple(ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}), ck::Number<64>{});
+    const auto layout = ck::wrapper::make_layout(shape, strides);
+
+    std::vector<ck::index_t> data(ck::wrapper::size(layout));
+    std::iota(data.begin(), data.end(), 0);
+
+    const auto tensor =
+        ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
+
+    const auto thread_steps  = ck::make_tuple(ck::Number<1>{}, ck::Number<8>{}, ck::Number<1>{});
+    const auto thread_layout = ck::make_tuple(ck::Number<4>{}, ck::Number<8>{}, ck::Number<1>{});
+    // 3d partition on 2d shape (calculate partition on 3d thread layout, and then skip first dim)
+    const auto thread_projection =
+        ck::make_tuple(ck::wrapper::slice(4), ck::Number<1>{}, ck::Number<1>{});
+    constexpr ck::index_t projection_thread_length = ck::Number<4>{};
+
+    for(ck::index_t thread_id = 0;
+        thread_id < ck::wrapper::size(thread_layout) / projection_thread_length;
+        thread_id++)
+    {
+        const auto packed_partition =
+            ck::wrapper::make_local_partition(tensor, thread_layout, thread_id, thread_projection);
+
+        const auto expected_partition_size =
+            ck::wrapper::size(tensor) /
+            (ck::wrapper::size(thread_layout) / projection_thread_length);
+        const auto expected_partition_first_val  = thread_id * ck::wrapper::size<1>(thread_steps);
+        const auto expected_partition_second_val = expected_partition_first_val + 1;
+        EXPECT_EQ(ck::wrapper::size(packed_partition), expected_partition_size);
+        EXPECT_EQ(packed_partition(0), expected_partition_first_val);
+        EXPECT_EQ(packed_partition(1), expected_partition_second_val);
+    }
+}
+
+TEST(TestPartition, LocalTile)
+{
+    const auto shape   = ck::make_tuple(ck::Number<16>{}, ck::Number<4>{}, ck::Number<4>{});
+    const auto strides = ck::make_tuple(ck::Number<1>{}, ck::Number<16>{}, ck::Number<64>{});
+    const auto layout  = ck::wrapper::make_layout(shape, strides);
+
+    std::vector<ck::index_t> data(ck::wrapper::size(layout));
+    std::iota(data.begin(), data.end(), 0);
+
+    const auto tensor =
+        ck::wrapper::make_tensor<ck::wrapper::MemoryTypeEnum::Generic>(data.data(), layout);
+    // 4d tile partitioning on 3d shape (calculate tile on 4d tile layout, and then skip last dim)
+    const auto block_shape =
+        ck::make_tuple(ck::Number<2>{}, ck::Number<4>{}, ck::Number<2>{}, ck::Number<2>{});
+    const auto block_projection =
+        ck::make_tuple(ck::Number<1>{}, ck::Number<1>{}, ck::Number<1>{}, ck::wrapper::slice(2));
+    constexpr ck::index_t projection_block_dim = ck::Number<2>{};
+    const auto num_blocks =
+        ck::make_tuple(ck::wrapper::size<0>(shape) / ck::wrapper::size<0>(block_shape),
+                       ck::wrapper::size<1>(shape) / ck::wrapper::size<1>(block_shape),
+                       ck::wrapper::size<2>(shape) / ck::wrapper::size<2>(block_shape));
+    std::vector<ck::index_t> block_idxs(ck::wrapper::size(num_blocks));
+    std::iota(block_idxs.begin(), block_idxs.end(), 0);
+
+    for(auto block_idx : block_idxs)
+    {
+        const auto packed_tile =
+            ck::wrapper::make_local_tile(tensor, block_shape, block_idx, block_projection);
+
+        const auto expected_tile_size = ck::wrapper::size(block_shape) / projection_block_dim;
+        auto expected_tile_first_val  = (block_idx % ck::wrapper::size<2>(num_blocks)) *
+                                       ck::wrapper::size<2>(block_shape) *
+                                       ck::wrapper::size<2>(strides);
+        block_idx /= ck::wrapper::size<2>(num_blocks);
+        expected_tile_first_val += (block_idx % ck::wrapper::size<1>(num_blocks)) *
+                                   ck::wrapper::size<1>(block_shape) *
+                                   ck::wrapper::size<1>(strides);
+        block_idx /= ck::wrapper::size<1>(num_blocks);
+        expected_tile_first_val += (block_idx % ck::wrapper::size<0>(num_blocks)) *
+                                   ck::wrapper::size<0>(block_shape) *
+                                   ck::wrapper::size<0>(strides);
+
+        const auto expected_tile_second_val = expected_tile_first_val + 1;
+        EXPECT_EQ(ck::wrapper::size(packed_tile), expected_tile_size);
+        EXPECT_EQ(packed_tile(0), expected_tile_first_val);
+        EXPECT_EQ(packed_tile(1), expected_tile_second_val);
+    }
+}

test/wrapper/test_tensor.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, 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;
+    __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 vgpr_layout =
+        ck::wrapper::make_layout(make_tuple(ck::Number<nelems>{}), make_tuple(ck::Number<1>{}));
+
+    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, ck::index_t>(
+            vgpr_layout);
+
+    InitTensor(tensor_global);
+    InitTensor(tensor_lds);
+    StaticInitTensor<nelems>(tensor_vgpr);
+
+    *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);
+}
+
+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),
+                           0,
+                           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(tensor2x2x2(0), layout(ck::make_tuple(ck::make_tuple(0, 0), 0)));
+    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(tensor2x2(0), layout(ck::make_tuple(ck::make_tuple(1, 0), 0)));
+    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));
+
+    auto tensor1x1 = tensor(ck::make_tuple(1, ck::wrapper::slice(1, 2)), ck::wrapper::slice(1, 2));
+    EXPECT_EQ(tensor1x1(0), layout(ck::make_tuple(ck::make_tuple(1, 1), 1)));
+    EXPECT_EQ(rank(tensor1x1), 2);
+    EXPECT_EQ(depth(tensor1x1), 2);
+    EXPECT_EQ(size(tensor1x1), 1);
+    EXPECT_TRUE(TestTensorCheck1d(tensor1x1));
+
+    auto tensor2 = tensor(ck::make_tuple(1, 1), ck::wrapper::slice(0, 2));
+    EXPECT_EQ(tensor2(0), layout(ck::make_tuple(ck::make_tuple(1, 1), 0)));
+    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));
+
+    auto tensor2_v2 = tensor(2, ck::wrapper::slice(0, 2));
+    EXPECT_EQ(tensor2_v2(0), layout(ck::make_tuple(2, 0)));
+    EXPECT_EQ(ck::wrapper::rank(tensor2_v2), 1);
+    EXPECT_EQ(ck::wrapper::depth(tensor2_v2), 1);
+    EXPECT_EQ(ck::wrapper::size(tensor2_v2), 2);
+    EXPECT_TRUE(TestTensorCheck1d(tensor2_v2));
+
+    // negative indexing
+    auto tensor1x2 = tensor(ck::make_tuple(1, ck::wrapper::slice(0, -2)), ck::wrapper::slice());
+    EXPECT_EQ(tensor1x2(0), layout(ck::make_tuple(ck::make_tuple(1, 0), 0)));
+    EXPECT_EQ(rank(tensor1x2), 2);
+    EXPECT_EQ(depth(tensor1x2), 2);
+    EXPECT_EQ(size(tensor1x2), 2);
+    EXPECT_TRUE(TestTensorCheck1d(tensor1x2));
+}