Tile program init bulk PR (#4)

Tile Program init bulk PR

---------
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com>

cmake/EnableCompilerWarnings.cmake

@@ -93,6 +93,7 @@ else()
                 -Wno-unused-command-line-argument
                 -Wno-weak-vtables
                 -Wno-covered-switch-default
+                -Wno-unused-lambda-capture
                 -Wno-unsafe-buffer-usage
             )
         else()

example/01_gemm/gemm_xdl_fp16.cpp

@@ -39,7 +39,9 @@ using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffl
 // ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
 // ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
 // ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, 
+         ck::make_default_loop_scheduler(),
+         ck::PipelineVersion::v2>;
 // clang-format on
 
 using DeviceGemmInstance = DeviceGemmInstance1;

example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp

@@ -21,6 +21,7 @@ Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/literals.hpp"

example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc

@@ -139,9 +139,9 @@ bool run_batched_gemm_gemm_example(int argc, char* argv[])
     {
     case 0: break;
     case 1:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-3.f, 3.f}(a_g_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<B0DataType>{-3.f, 3.f}(b0_g_k_n);
+        ck::utils::FillUniformDistributionIntegerValue<B1DataType>{-3.f, 3.f}(b1_g_n_o);
         break;
     case 2:
         a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});

example/91_tile_program/CMakeLists.txt

+add_example_executable(example_hello_world hello_world.cpp)
+add_example_executable(example_im2col im2col.cpp)
+add_example_executable(example_gemm gemm.cpp)
+add_example_executable(example_gemm_gemm gemm_gemm.cpp)
+add_example_executable(example_reduce reduce.cpp)

example/91_tile_program/gemm.cpp

+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_gemm.hpp"
+#include "gemm.hpp"
+
+// elementwise lambda
+struct AElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+struct BElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+struct CElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    using ADataType   = ck::half_t;
+    using BDataType   = ck::half_t;
+    using AccDataType = float;
+    using CDataType   = ck::half_t;
+
+    ck::index_t M = 3328;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    if(argc == 4)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+    }
+
+    std::array<ck::index_t, 2> a_lengths{M, K};
+    std::array<ck::index_t, 2> a_strides{K, 1};
+
+    std::array<ck::index_t, 2> b_lengths{N, K};
+    std::array<ck::index_t, 2> b_strides{K, 1};
+
+    std::array<ck::index_t, 2> c_lengths{M, N};
+    std::array<ck::index_t, 2> c_strides{N, 1};
+
+    // host verify
+    Tensor<ADataType> a_host(a_lengths, a_strides);
+    Tensor<BDataType> b_host(b_lengths, b_strides);
+    Tensor<CDataType> c_host_ref(c_lengths, c_strides);
+    Tensor<CDataType> c_host_dev(c_lengths, c_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_host);
+    ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_host);
+
+    // reference gemm
+    reference_gemm<ADataType, ADataType, CDataType, float>(a_host, b_host, c_host_ref);
+
+    DeviceMem a_buf(sizeof(ADataType) * a_host.GetElementSpaceSize());
+    DeviceMem b_buf(sizeof(BDataType) * b_host.GetElementSpaceSize());
+    DeviceMem c_buf(sizeof(CDataType) * c_host_dev.GetElementSpaceSize());
+
+    a_buf.ToDevice(a_host.mData.data());
+    b_buf.ToDevice(b_host.mData.data());
+
+    constexpr ck::index_t kGemmMPerBlock = 256;
+    constexpr ck::index_t kGemmNPerBlock = 128;
+    constexpr ck::index_t kGemmKPerBlock = 32;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M / kGemmMPerBlock) * (N / kGemmNPerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    const auto gemm_kernel = Gemm<ADataType,
+                                  BDataType,
+                                  AccDataType,
+                                  CDataType,
+                                  ck::tensor_layout::gemm::RowMajor,
+                                  ck::tensor_layout::gemm::ColumnMajor,
+                                  ck::tensor_layout::gemm::RowMajor,
+                                  AElementFunction,
+                                  BElementFunction,
+                                  CElementFunction,
+                                  kBlockSize,
+                                  kGemmMPerBlock,
+                                  kGemmNPerBlock,
+                                  kGemmKPerBlock>{};
+
+    float ave_time = launch(ProgramServer{},
+                            gemm_kernel,
+                            kGridSize,
+                            kBlockSize,
+                            static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
+                            static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
+                            static_cast<CDataType*>(c_buf.GetDeviceBuffer()),
+                            M,
+                            N,
+                            K,
+                            K,
+                            K,
+                            N,
+                            AElementFunction{},
+                            BElementFunction{},
+                            CElementFunction{});
+
+    c_buf.FromDevice(c_host_dev.mData.data());
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(c_host_dev, c_host_ref);
+}

example/91_tile_program/gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "tile_program.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/grid/grid_gemm.hpp"
+#include "ck/tile_program/grid/grid_gemm_policy.hpp"
+#include "ck/tile_program/grid/grid_gemm_problem.hpp"
+
+// C = A * B
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementFunction,
+          typename BElementFunction,
+          typename CElementFunction,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kNPerBlock,
+          ck::index_t kKPerBlock>
+struct Gemm
+{
+    static_assert(std::is_same_v<ALayout, ck::tensor_layout::gemm::RowMajor> &&
+                  std::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor> &&
+                  std::is_same_v<CLayout, ck::tensor_layout::gemm::RowMajor>);
+
+    using Problem = ck::tile_program::grid::GridGemmProblem<ADataType,
+                                                            BDataType,
+                                                            AccDataType,
+                                                            CDataType,
+                                                            AElementFunction,
+                                                            BElementFunction,
+                                                            CElementFunction>;
+
+    using Policy = ck::tile_program::grid::GridGemmPolicy<
+        kBlockSize,
+        kMPerBlock,
+        kNPerBlock,
+        kKPerBlock,
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2,
+        ck::Tuple<ck::tile_program::grid::DefaultBlock2TileMap,
+                  ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>>;
+
+    using GridGemm = ck::tile_program::grid::GridGemm<Problem, Policy>;
+
+    __host__ __device__ void operator()(ProgramServer& ps,
+                                        const ADataType* p_a,
+                                        const BDataType* p_b,
+                                        CDataType* p_c,
+                                        ck::index_t M,
+                                        ck::index_t N,
+                                        ck::index_t K,
+                                        ck::index_t Lda,
+                                        ck::index_t Ldb,
+                                        ck::index_t Ldc,
+                                        const AElementFunction& a_element_func,
+                                        const BElementFunction& b_element_func,
+                                        const CElementFunction& c_element_func) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        // FIXME: assume RCR layout
+        const auto a_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a, make_tuple(M, K), make_tuple(Lda, 1), Number<32>{}, Number<1>{});
+
+        const auto b_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b, make_tuple(N, K), make_tuple(Ldb, 1), Number<32>{}, Number<1>{});
+
+        auto c_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_c, make_tuple(M, N), make_tuple(Ldc, 1), Number<32>{}, Number<1>{});
+
+        GridGemm{}(ps,
+                   a_dram_grid,
+                   b_dram_grid,
+                   c_dram_grid,
+                   a_element_func,
+                   b_element_func,
+                   c_element_func);
+    }
+};

example/91_tile_program/gemm_gemm.cpp

+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_gemm.hpp"
+#include "gemm_gemm.hpp"
+
+int main(int argc, char* argv[])
+{
+    using A0DataType   = ck::half_t;
+    using B0DataType   = ck::half_t;
+    using Acc0DataType = float;
+    using C0DataType   = ck::half_t;
+    using B1DataType   = ck::half_t;
+    using Acc1DataType = float;
+    using C1DataType   = ck::half_t;
+
+    ck::index_t M0 = 13312;
+    ck::index_t N0 = 4096;
+    ck::index_t K0 = 128;
+    ck::index_t N1 = 128;
+
+    if(argc == 5)
+    {
+        M0 = std::stoi(argv[1]);
+        N0 = std::stoi(argv[2]);
+        K0 = std::stoi(argv[3]);
+        N1 = std::stoi(argv[4]);
+    }
+
+    std::array<ck::index_t, 2> a0_lengths{M0, K0};
+    std::array<ck::index_t, 2> a0_strides{K0, 1};
+
+    std::array<ck::index_t, 2> b0_lengths{N0, K0};
+    std::array<ck::index_t, 2> b0_strides{K0, 1};
+
+    std::array<ck::index_t, 2> c0_lengths{M0, N0};
+    std::array<ck::index_t, 2> c0_strides{N0, 1};
+
+    std::array<ck::index_t, 2> b1_lengths{N1, N0};
+    std::array<ck::index_t, 2> b1_strides{N0, 1};
+
+    std::array<ck::index_t, 2> c1_lengths{M0, N1};
+    std::array<ck::index_t, 2> c1_strides{N1, 1};
+
+    // host verify
+    Tensor<A0DataType> a0_host(a0_lengths, a0_strides);
+    Tensor<B0DataType> b0_host(b0_lengths, b0_strides);
+    Tensor<B1DataType> b1_host(b1_lengths, b1_strides);
+    Tensor<C0DataType> c0_host_ref(c0_lengths, c0_strides);
+    Tensor<C1DataType> c1_host_ref(c1_lengths, c1_strides);
+    Tensor<C1DataType> c1_host_dev(c1_lengths, c1_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<A0DataType>{-3.f, 3.f}(a0_host);
+    ck::utils::FillUniformDistributionIntegerValue<B0DataType>{-3.f, 3.f}(b0_host);
+    ck::utils::FillUniformDistributionIntegerValue<B1DataType>{-3.f, 3.f}(b1_host);
+
+    // reference gemm
+    reference_gemm<A0DataType, B0DataType, C0DataType, float>(a0_host, b0_host, c0_host_ref);
+    reference_gemm<C0DataType, B1DataType, C1DataType, float>(c0_host_ref, b1_host, c1_host_ref);
+
+    DeviceMem a0_buf(sizeof(A0DataType) * a0_host.GetElementSpaceSize());
+    DeviceMem b0_buf(sizeof(B0DataType) * b0_host.GetElementSpaceSize());
+    DeviceMem b1_buf(sizeof(B1DataType) * b1_host.GetElementSpaceSize());
+    DeviceMem c1_buf(sizeof(C1DataType) * c1_host_ref.GetElementSpaceSize());
+
+    a0_buf.ToDevice(a0_host.mData.data());
+    b0_buf.ToDevice(b0_host.mData.data());
+    b1_buf.ToDevice(b1_host.mData.data());
+
+    constexpr ck::index_t kM0PerBlock = 128;
+    constexpr ck::index_t kN0PerBlock = 128;
+    constexpr ck::index_t kK0PerBlock = 32;
+    constexpr ck::index_t kN1PerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M0 / kM0PerBlock) * (N1 / kN1PerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    float ave_time = launch(ProgramServer{},
+                            GemmGemm<A0DataType,
+                                     B0DataType,
+                                     Acc0DataType,
+                                     C0DataType,
+                                     B1DataType,
+                                     Acc1DataType,
+                                     C1DataType,
+                                     kBlockSize,
+                                     kM0PerBlock,
+                                     kN0PerBlock,
+                                     kK0PerBlock,
+                                     kN1PerBlock>{},
+                            kGridSize,
+                            kBlockSize,
+                            static_cast<A0DataType*>(a0_buf.GetDeviceBuffer()),
+                            static_cast<B0DataType*>(b0_buf.GetDeviceBuffer()),
+                            static_cast<B1DataType*>(b1_buf.GetDeviceBuffer()),
+                            static_cast<C1DataType*>(c1_buf.GetDeviceBuffer()),
+                            M0,
+                            N0,
+                            K0,
+                            N1,
+                            K0,  // Lda0
+                            K0,  // Ldb0
+                            N0,  // Ldb1
+                            N1); // Ldc1
+
+    c1_buf.FromDevice(c1_host_dev.mData.data());
+
+    std::size_t flop      = std::size_t(2) * M0 * N0 * K0 + std::size_t(2) * M0 * N1 * N0;
+    std::size_t num_btype = sizeof(A0DataType) * M0 * K0 + sizeof(B0DataType) * N0 * K0 +
+                            sizeof(B1DataType) * N1 * N0 + sizeof(C1DataType) * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(c1_host_dev, c1_host_ref);
+}

example/91_tile_program/gemm_gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "tile_program.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+
+// C1 = A0 * B0 * B1
+template <typename A0DataType,
+          typename B0DataType,
+          typename Acc0DataType,
+          typename C0DataType,
+          typename B1DataType,
+          typename Acc1DataType,
+          typename C1DataType,
+          ck::index_t kBlockSize,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock>
+struct GemmGemm
+{
+    // block gemm0 pipeline
+    using BlockGemm0Pipeline = ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2<
+        ck::tile_program::block::BlockGemmPipelineProblem<
+            A0DataType,
+            B0DataType,
+            Acc0DataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN0PerBlock, kK0PerBlock>>,
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>;
+
+    // block gemm1
+    using BlockGemm1 = ck::tile_program::block::BlockGemmARegBSmemCRegV1<
+        ck::tile_program::block::BlockGemmARegBSmemCRegV1Problem<
+            C0DataType,
+            B1DataType,
+            Acc1DataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN1PerBlock, kN0PerBlock>>,
+        ck::tile_program::block::BlockGemmARegBSmemCRegV1DefaultPolicy>;
+
+#if 0
+    // 2d
+    __host__ __device__ static constexpr auto MakeB1LdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kN0PerBlock;
+
+        constexpr auto b_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), Number<32>{});
+
+        return b_lds_block_desc;
+    }
+#else
+    // fake XOR
+    __host__ __device__ static constexpr auto MakeB1LdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kN0PerBlock;
+
+        constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(kNPerBlock / 2, 2, kKPerBlock), Number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(BDataType);
+
+        constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            b_lds_block_desc_d1_d2_d3,
+            make_tuple(make_xor_transform(make_tuple(kNPerBlock / 2, kKPerBlock), kK1),
+                       make_pass_through_transform(2)),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+            b_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(kNPerBlock / 2, 2)),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc_n_k;
+    }
+#endif
+
+    __host__ __device__ static constexpr auto MakeB1DramTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kN0PerBlock;
+
+        constexpr index_t K1 = 16 / sizeof(BDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        using namespace ck;
+
+        return math::max(BlockGemm0Pipeline::GetStaticLdsSize(),
+                         static_cast<index_t>(MakeB1LdsBlockDescriptor().GetElementSpaceSize() *
+                                              sizeof(B1DataType)));
+    }
+
+    __host__ __device__ void operator()(ProgramServer& ps,
+                                        const A0DataType* p_a0,
+                                        const B0DataType* p_b0,
+                                        const B1DataType* p_b1,
+                                        C1DataType* p_c1,
+                                        ck::index_t M0,
+                                        ck::index_t N0,
+                                        ck::index_t K0,
+                                        ck::index_t N1,
+                                        ck::index_t Lda0,
+                                        ck::index_t Ldb0,
+                                        ck::index_t Ldb1,
+                                        ck::index_t Ldc1)
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        // FIXME: assume layout A0[M0, K0], B0[N0, K0], B1[N1, N0], C1[M0, N1]
+        const auto a0_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a0, make_tuple(M0, K0), make_tuple(Lda0, 1), Number<32>{}, Number<1>{});
+
+        const auto b0_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b0, make_tuple(N0, K0), make_tuple(Ldb0, 1), Number<32>{}, Number<1>{});
+
+        const auto b1_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b1, make_tuple(N1, N0), make_tuple(Ldb1, 1), Number<32>{}, Number<1>{});
+
+        // divide problem
+        const auto id_block = ps.get_block_id();
+
+        const auto num_tile_m0 = M0 / kM0PerBlock;
+        const auto num_tile_n1 = N1 / kN1PerBlock;
+
+        const auto block2tile = ps(make_cluster_descriptor(make_tuple(num_tile_m0, num_tile_n1)));
+
+        const auto id_tile = block2tile.CalculateBottomIndex(make_tuple(id_block));
+
+        const auto iM0 = ps.read_first_lane(id_tile.At<0>() * kM0PerBlock);
+        const auto iN1 = ps.read_first_lane(id_tile.At<1>() * kN1PerBlock);
+
+        __shared__ char p_smem_char[GetStaticLdsSize()];
+
+        // A0 DRAM block window
+        auto a0_dram_block_window = make_tile_window(
+            a0_dram_grid, make_tuple(Number<kM0PerBlock>{}, Number<kK0PerBlock>{}), {iM0, 0});
+
+        // B0 DRAM block window
+        auto b0_dram_block_window = make_tile_window(
+            b0_dram_grid, make_tuple(Number<kN0PerBlock>{}, Number<kK0PerBlock>{}), {0, 0});
+
+        // Block GEMM0 pipeline
+        constexpr auto block_gemm0_pipeline = BlockGemm0Pipeline{};
+
+        // B1 DRAM window
+        auto b1_dram_block_window =
+            make_tile_window(b1_dram_grid,
+                             make_tuple(Number<kN1PerBlock>{}, Number<kN0PerBlock>{}),
+                             {iN1, 0},
+                             MakeB1DramTileDistribution());
+
+        // B1 LDS tensor view: occupies the same LDS allocation as block_gemm0_pipeline
+        auto b1_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<B1DataType*>(p_smem_char), MakeB1LdsBlockDescriptor());
+
+        auto b1_lds_block_window = make_tile_window(
+            b1_lds_block, make_tuple(Number<kN1PerBlock>{}, Number<kN0PerBlock>{}), {0, 0});
+
+        // Bock GEMM1
+        constexpr auto block_gemm1 = BlockGemm1{};
+
+        // Acc1 tile
+        auto acc1_block_tile = decltype(block_gemm1(
+            tile_elementwise_in(
+                type_convert<C0DataType, Acc0DataType>,
+                block_gemm0_pipeline(a0_dram_block_window, b0_dram_block_window, 0, nullptr)),
+            b1_dram_block_window)){};
+
+        // init Acc1
+        tile_elementwise_inout([](auto& acc1) { acc1 = 0; }, acc1_block_tile);
+
+#if 0
+        index_t iN0 = 0;
+
+        do
+        {
+            // Block GEMM0 pipeline: acc0 = a0 * b0
+            const auto acc0_block_tile = block_gemm0_pipeline(
+                a0_dram_block_window, b0_dram_block_window, K0 / kK0PerBlock, p_smem_char);
+
+            // type cast acc0 into c0
+            const auto c0_block_tile =
+                tile_elementwise_in(type_convert<C0DataType, Acc0DataType>, acc0_block_tile);
+
+            // Block GEMM1: acc1 += c0 * b1
+            {
+                // load b1
+                const auto b1_block_tile = load_tile(b1_dram_block_window);
+
+                // wait for block gemm0 pipeline to finish
+                ps.block_sync_lds();
+
+                store_tile(b1_lds_block_window, b1_block_tile);
+
+                // wait for store_tile to finish
+                ps.block_sync_lds();
+
+                // acc1 += c0 * b1
+                block_gemm1(acc1_block_tile, c0_block_tile, b1_lds_block_window);
+
+                // wait for block gemm1 to finish
+                ps.block_sync_lds();
+            }
+
+            // move tile windows
+            move_tile_window(b0_dram_block_window, {kN0PerBlock, 0});
+            move_tile_window(b1_dram_block_window, {0, kN0PerBlock});
+
+            iN0 += kN0PerBlock;
+
+        } while(iN0 < N0);
+#else
+        index_t iN0 = 0;
+
+        do
+        {
+            // load b1
+            const auto b1_block_tile = load_tile(b1_dram_block_window);
+
+            // Block GEMM0 pipeline: acc0 = a0 * b0
+            const auto acc0_block_tile = block_gemm0_pipeline(
+                a0_dram_block_window, b0_dram_block_window, K0 / kK0PerBlock, p_smem_char);
+
+            // type cast acc0 into c0
+            const auto c0_block_tile =
+                tile_elementwise_in(type_convert<C0DataType, Acc0DataType>, acc0_block_tile);
+
+            // Block GEMM1: acc1 += c0 * b1
+            {
+                // wait for block gemm0 pipeline to finish
+                ps.block_sync_lds();
+
+                store_tile(b1_lds_block_window, b1_block_tile);
+
+                // wait for store_tile to finish
+                ps.block_sync_lds();
+
+                // acc1 += c0 * b1
+                block_gemm1(acc1_block_tile, c0_block_tile, b1_lds_block_window);
+
+                // wait for block gemm1 to finish
+                ps.block_sync_lds();
+            }
+
+            // move tile windows
+            move_tile_window(b0_dram_block_window, {kN0PerBlock, 0});
+            move_tile_window(b1_dram_block_window, {0, kN0PerBlock});
+
+            iN0 += kN0PerBlock;
+
+        } while(iN0 < N0);
+#endif
+
+        // type cast acc1 into c1
+        const auto c1_block_tile =
+            tile_elementwise_in(type_convert<C1DataType, Acc1DataType>, acc1_block_tile);
+
+        // store c1
+        auto c1_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_c1, make_tuple(M0, N1), make_tuple(Ldc1, 1), Number<32>{}, Number<1>{});
+
+        auto c1_dram_window =
+            make_tile_window(c1_dram_grid,
+                             make_tuple(Number<kM0PerBlock>{}, Number<kN1PerBlock>{}),
+                             {iM0, iN1},
+                             c1_block_tile.GetTileDistribution());
+
+        store_tile(c1_dram_window, c1_block_tile);
+    }
+};

example/91_tile_program/hello_world.cpp

+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+#include "tile_program.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+
+// ProgramServer contains a "meta data buffer"
+// host evaluate the expression inside ps(), and push the result into meta data buffer
+// ProgramServer send meta data buffer to GPU as kernel arguement
+// device read (not evaluate) the value of the expression inside ps() from meta data buffer
+struct HelloWorld
+{
+    __host__ __device__ void operator()(ProgramServer& ps, int x, int y, int* res)
+    {
+#if 1
+        auto r0 = ps(x + y);
+        auto r1 = ps(x - y);
+
+        res[0] = r0;
+        res[1] = r1;
+#elif 1
+        (void)x;
+        (void)y;
+
+        auto r0 = ps.get_thread_id();
+        auto r1 = ps.warp_shuffle_up(r0, 1);
+        auto r2 = ps.warp_shuffle_down(r0, 1);
+
+        printf("tid %d, r0 %d, r1 %d, r2 %d\n", ps.get_thread_id(), r0, r1, r2);
+
+        res[0] = r0;
+        res[1] = r2;
+#endif
+    }
+};
+
+int main()
+{
+    int x = 100;
+    int y = 101;
+
+    DeviceMem res_dev_buf(2 * sizeof(int));
+
+    launch(ProgramServer{},
+           HelloWorld{},
+           1,
+           64,
+           x,
+           y,
+           static_cast<int*>(res_dev_buf.GetDeviceBuffer()));
+
+    int res_host[2];
+
+    res_dev_buf.FromDevice(&res_host);
+
+    printf("x+y=: %d\n", res_host[0]);
+    printf("x-y=: %d\n", res_host[1]);
+
+    return 0;
+}

example/91_tile_program/im2col.cpp

+#include <string_view>
+#include <tuple>
+#include <array>
+#include <utility>
+#include <type_traits>
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+
+#include "tile_program.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <typename T>
+void reference_im2col(Tensor<T>& in_mtx_host_ref,
+                      const Tensor<T>& in_host,
+                      int /*N*/,
+                      int /*K*/,
+                      int C,
+                      int /*Y*/,
+                      int X,
+                      int Hi,
+                      int Wi,
+                      int Ho,
+                      int Wo,
+                      int ConvStrideH,
+                      int ConvStrideW,
+                      int ConvDilationH,
+                      int ConvDilationW,
+                      int InLeftPadH,
+                      int InLeftPadW,
+                      int /*InRightPadH*/,
+                      int /*InRightPadW*/)
+{
+    int GemmM = in_mtx_host_ref.GetLengths()[0];
+    int GemmK = in_mtx_host_ref.GetLengths()[1];
+
+    for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
+    {
+        int mtmp = gemm_m;
+        int n    = mtmp / (Ho * Wo);
+        mtmp -= n * Ho * Wo;
+        int ho = mtmp / Wo;
+        int wo = mtmp - ho * Wo;
+
+        for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
+        {
+            int ktmp = gemm_k;
+            int y    = ktmp / (X * C);
+            ktmp -= y * X * C;
+            int x = ktmp / C;
+            int c = ktmp - x * C;
+
+            int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
+            int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
+
+            bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
+
+            in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
+        }
+    }
+}
+
+template <ck::index_t NDimSpatial,
+          typename T,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kKPerBlock>
+struct Im2Col
+{
+    __host__ __device__ static constexpr auto MakeBlockCopyTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+        constexpr index_t K1 = 16 / sizeof(T);
+        constexpr index_t K0 = kKPerBlock / K1;
+
+        constexpr index_t M2 = get_warp_size() / K0;
+        constexpr index_t M1 = NumWarp;
+        constexpr index_t M0 = kMPerBlock / (M1 * M2);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    template <typename Server>
+    __host__ __device__ void
+    operator()(Server& ps,
+               const std::array<ck::index_t, NDimSpatial + 2>& a_n_wis_c_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* a_n_wis_c_strides */,
+               const std::array<ck::index_t, NDimSpatial + 2>& b_k_xs_c_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* b_k_xs_c_strides */,
+               const std::array<ck::index_t, NDimSpatial + 2>& c_n_wos_k_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* c_n_wos_k_strides */,
+               const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+               const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+               const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+               const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+               //
+               const std::array<ck::index_t, 2> a_gemmm_gemmk_lengths,
+               const std::array<ck::index_t, 2> a_gemmm_gemmk_strides,
+               //
+               const T* p_a_img,
+               T* p_a_mtx)
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        const index_t N = a_n_wis_c_lengths[0];
+        const index_t C = a_n_wis_c_lengths[3];
+
+        const index_t Hi = a_n_wis_c_lengths[1];
+        const index_t Wi = a_n_wis_c_lengths[2];
+
+        const index_t Ho = c_n_wos_k_lengths[1];
+        const index_t Wo = c_n_wos_k_lengths[2];
+
+        const index_t Y = b_k_xs_c_lengths[1];
+        const index_t X = b_k_xs_c_lengths[2];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const auto a_n_hi_wi_c = make_naive_tensor_view_packed<AddressSpaceEnum::Global>(
+            p_a_img, make_tuple(N, Hi, Wi, C), Number<32>{});
+
+        const auto a_n_hip_wip_c = transform_tensor_view(
+            a_n_hi_wi_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto a_n_y_ho_x_wo_c = transform_tensor_view(
+            a_n_hip_wip_c,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto src_gemmm_gemmk =
+            transform_tensor_view(a_n_y_ho_x_wo_c,
+                                  make_tuple(ps(make_merge_transform(make_tuple(N, Ho, Wo))),
+                                             ps(make_merge_transform(make_tuple(Y, X, C)))),
+                                  make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                  make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        auto dst_gemmm_gemmk = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a_mtx,
+            make_tuple(a_gemmm_gemmk_lengths[0], a_gemmm_gemmk_lengths[1]),
+            make_tuple(a_gemmm_gemmk_strides[0], a_gemmm_gemmk_strides[1]),
+            Number<32>{},
+            Number<1>{});
+
+        const auto numGemmM = a_gemmm_gemmk_lengths[0];
+        const auto numGemmK = a_gemmm_gemmk_lengths[1];
+
+        const auto id_block = ps.get_block_id();
+
+        const auto num_tile_m = ps.read_first_lane(numGemmM / kMPerBlock);
+
+        const auto block2tile = ps(make_cluster_descriptor(make_tuple(num_tile_m)));
+
+        const auto i_gemmm_gemmk = block2tile.CalculateBottomIndex(make_multi_index(id_block));
+
+        const auto iGemmM = ps.read_first_lane(i_gemmm_gemmk[0]) * kMPerBlock;
+
+        // src window
+        auto src_block_window =
+            make_tile_window(src_gemmm_gemmk,
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             {iGemmM, 0},
+                             MakeBlockCopyTileDistribution());
+
+        // dst window
+        auto dst_block_window = make_tile_window(
+            dst_gemmm_gemmk, make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}), {iGemmM, 0});
+
+        index_t iGemmK = 0;
+
+        do
+        {
+            const auto block_tile = load_tile(src_block_window);
+
+            store_tile(dst_block_window, block_tile);
+
+            move_tile_window(src_block_window, {0, kKPerBlock});
+            move_tile_window(dst_block_window, {0, kKPerBlock});
+
+            iGemmK += kKPerBlock;
+
+        } while(iGemmK < numGemmK);
+    }
+};
+
+int main()
+{
+    using DataType = ck::half_t;
+
+    constexpr ck::index_t NumDimSpatial = 2;
+
+    ck::index_t N  = 128;
+    ck::index_t K  = 1;
+    ck::index_t C  = 256;
+    ck::index_t Y  = 3;
+    ck::index_t X  = 3;
+    ck::index_t Hi = 28;
+    ck::index_t Wi = 28;
+    ck::index_t Ho = 14;
+    ck::index_t Wo = 14;
+
+    std::array<ck::index_t, NumDimSpatial + 2> in_lengths{N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 2> in_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 2> wei_lengths{K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 2> wei_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 2> out_lengths{N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 2> out_strides{0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{2, 2};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    // matrix
+    std::array<ck::index_t, 2> in_mtx_lengths{N * Ho * Wo, Y * X * C};
+    std::array<ck::index_t, 2> in_mtx_strides{0, 1};
+
+    std::partial_sum(rbegin(in_mtx_lengths),
+                     std::prev(rend(in_mtx_lengths)),
+                     std::next(rbegin(in_mtx_strides)),
+                     std::multiplies<>{});
+
+    // host verify
+    Tensor<DataType> in_host(in_lengths, in_strides);
+    Tensor<DataType> in_mtx_host_ref(in_mtx_lengths, in_mtx_strides);
+    Tensor<DataType> in_mtx_host_dev(in_mtx_lengths, in_mtx_strides);
+
+    std::cout << " image tensor element size: " << in_host.GetElementSize() << std::endl;
+    std::cout << "matrix tensor element size: " << in_mtx_host_ref.GetElementSize() << std::endl;
+
+    std::cout << " image tensor element space size: " << in_host.GetElementSpaceSize() << std::endl;
+    std::cout << "matrix tensor element sapce size: " << in_mtx_host_ref.GetElementSpaceSize()
+              << std::endl;
+
+    ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(in_host);
+
+    reference_im2col(in_mtx_host_ref,
+                     in_host,
+                     N,
+                     K,
+                     C,
+                     Y,
+                     X,
+                     Hi,
+                     Wi,
+                     Ho,
+                     Wo,
+                     filter_strides[0],
+                     filter_strides[1],
+                     filter_dilations[0],
+                     filter_dilations[1],
+                     input_left_pads[0],
+                     input_left_pads[1],
+                     input_right_pads[0],
+                     input_right_pads[1]);
+
+    DeviceMem in_buf(sizeof(DataType) * in_host.GetElementSpaceSize());
+    DeviceMem in_mtx_buf(sizeof(DataType) * in_mtx_host_ref.GetElementSpaceSize());
+
+    in_buf.ToDevice(in_host.mData.data());
+
+    constexpr ck::index_t kBlockSize = 256;
+
+    constexpr ck::index_t kGemmMPerBlock = 256;
+    constexpr ck::index_t kGemmKPerBlock = 128;
+
+    ck::index_t kGridSize = (N * Ho * Wo) / kGemmMPerBlock;
+
+    float ave_time = launch(ProgramServer{},
+                            Im2Col<2, DataType, kBlockSize, kGemmMPerBlock, kGemmKPerBlock>{},
+                            kGridSize,
+                            kBlockSize,
+                            in_lengths,
+                            in_strides,
+                            wei_lengths,
+                            wei_strides,
+                            out_lengths,
+                            out_strides,
+                            filter_strides,
+                            filter_dilations,
+                            input_left_pads,
+                            input_right_pads,
+                            //
+                            in_mtx_lengths,
+                            in_mtx_strides,
+                            //
+                            static_cast<DataType*>(in_buf.GetDeviceBuffer()),
+                            static_cast<DataType*>(in_mtx_buf.GetDeviceBuffer()));
+
+    std::size_t num_btype = sizeof(DataType) * in_host.GetElementSize() +
+                            sizeof(DataType) * in_mtx_host_ref.GetElementSize();
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    in_mtx_buf.FromDevice(in_mtx_host_dev.mData.data());
+
+    return !ck::utils::check_err(in_mtx_host_dev, in_mtx_host_ref);
+}

example/91_tile_program/reduce.cpp

+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reduce.hpp"
+
+template <typename ADataType, typename AccDataType, typename BDataType>
+void reference_reduce(const Tensor<ADataType>& a_m_n, Tensor<BDataType>& b_m)
+{
+    auto f = [&](auto m) {
+        const int N = a_m_n.mDesc.GetLengths()[1];
+
+        AccDataType v_acc = 0;
+
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_m_n(m, n);
+
+            v_acc += v_a;
+        }
+
+        b_m(m) = ck::type_convert<BDataType>(v_acc);
+    };
+
+    make_ParallelTensorFunctor(f, b_m.mDesc.GetLengths()[0])(std::thread::hardware_concurrency());
+}
+
+int main(int argc, char* argv[])
+{
+    using ADataType   = float;
+    using AccDataType = float;
+    using BDataType   = float;
+
+    ck::index_t M = 3328;
+    ck::index_t N = 4096;
+
+    if(argc == 3)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+    }
+
+    std::array<ck::index_t, 2> a_lengths{M, N};
+    std::array<ck::index_t, 2> a_strides{N, 1};
+
+    std::array<ck::index_t, 2> b_lengths{M};
+    std::array<ck::index_t, 2> b_strides{1};
+
+    // host verify
+    Tensor<ADataType> a_host(a_lengths, a_strides);
+    Tensor<BDataType> b_host_ref(b_lengths, b_strides);
+    Tensor<BDataType> b_host_dev(b_lengths, b_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_host);
+
+    // reference
+    reference_reduce<ADataType, AccDataType, BDataType>(a_host, b_host_ref);
+
+    DeviceMem a_buf(sizeof(ADataType) * a_host.GetElementSpaceSize());
+    DeviceMem b_buf(sizeof(BDataType) * b_host_ref.GetElementSpaceSize());
+
+    a_buf.ToDevice(a_host.mData.data());
+
+    constexpr ck::index_t kMPerBlock = 128;
+    constexpr ck::index_t kNPerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M / kMPerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    const auto kernel =
+        Reduce<ADataType, AccDataType, BDataType, kBlockSize, kMPerBlock, kNPerBlock>{};
+
+    float ave_time = launch(ProgramServer{},
+                            kernel,
+                            kGridSize,
+                            kBlockSize,
+                            static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
+                            static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
+                            M,
+                            N);
+
+    b_buf.FromDevice(b_host_dev.mData.data());
+
+    std::size_t num_btype = sizeof(ADataType) * M * N + sizeof(BDataType) * M;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    return !ck::utils::check_err(b_host_dev, b_host_ref);
+}

example/91_tile_program/reduce.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "tile_program.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+template <typename ADataType,
+          typename AccDataType,
+          typename BDataType,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kNPerBlock>
+struct Reduce
+{
+#if 0
+    __host__ __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 2x2 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<2, 2, 4, 2, 4>, Sequence<2, 2, 32>>,
+                                           Tuple<Sequence<1, 2>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<1, 2, 1, 1>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#elif 0
+    __host__ __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 2x2 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<2, 2, 32>, Sequence<2, 2, 4, 2, 4>>,
+                                           Tuple<Sequence<2, 1>, Sequence<2, 1>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<2, 1, 2, 2>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#elif 1
+    __host__ __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 4x1 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<1, 4, 4, 2, 4>, Sequence<4, 1, 32>>,
+                                           Tuple<Sequence<1, 2>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<1, 2, 1, 1>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#endif
+
+    __host__ __device__ void operator()(
+        ProgramServer& ps, const ADataType* p_a, BDataType* p_b, ck::index_t M, ck::index_t N) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        const auto a_m_n = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        const auto iM = ps.get_block_id() * kMPerBlock;
+
+        // A window
+        auto a_block_window =
+            make_tile_window(a_m_n,
+                             make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}),
+                             {iM, 0},
+                             MakeABlockTileDistribution());
+
+        const auto f_reduce = [](const auto& v0, const auto& v1) { return v0 + v1; };
+
+        const ADataType reduce_init_value = 0;
+
+        constexpr auto reduce_dims = Sequence<1>{};
+
+        // Acc tile
+        // FIXME: support cross warp reduction
+        auto acc_block_tensor = decltype(block_tile_reduce<AccDataType>(
+            load_tile(a_block_window), reduce_dims, f_reduce, reduce_init_value)){};
+
+        // init Acc tile
+        tile_elementwise_inout(
+            [&](auto& acc) { acc = type_convert<AccDataType>(reduce_init_value); },
+            acc_block_tensor);
+
+        // loop
+        index_t iN = 0;
+
+        do
+        {
+            const auto a_block_tensor = load_tile(a_block_window);
+
+            // FIXME: support cross warp reduction
+            block_tile_reduce(acc_block_tensor, a_block_tensor, reduce_dims, f_reduce);
+
+            move_tile_window(a_block_window, {0, kNPerBlock});
+
+            iN += kNPerBlock;
+
+        } while(iN < N);
+
+        // FIXME: support cross warp reduction
+        block_tile_reduce_sync(acc_block_tensor, f_reduce);
+
+        // convert acc_block_tensor to b_block_tensor
+        const auto b_block_tensor = tile_elementwise_in(
+            [](const auto& acc) { return type_convert<BDataType>(acc); }, acc_block_tensor);
+
+        // B
+        const auto b_m = make_naive_tensor_view_packed<AddressSpaceEnum::Global>(
+            p_b, make_tuple(M), Number<32>{});
+
+        // B window
+        auto b_block_window = make_tile_window(b_m, make_tuple(Number<kMPerBlock>{}), {iM});
+
+        // store B tile
+        store_tile(b_block_window, b_block_tensor);
+    }
+};

example/91_tile_program/reference_gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <typename ADataType, typename BDataType, typename CDataType, typename AccDataType>
+void reference_gemm(const Tensor<ADataType>& a_m_k,
+                    const Tensor<BDataType>& b_n_k,
+                    Tensor<CDataType>& c_m_n)
+{
+    auto f_mk_kn_mn = [&](auto m, auto n) {
+        const int K = a_m_k.mDesc.GetLengths()[1];
+
+        AccDataType v_acc = 0;
+
+        for(int k = 0; k < K; ++k)
+        {
+            ADataType v_a = a_m_k(m, k);
+            BDataType v_b = b_n_k(n, k);
+
+            v_acc += ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+        }
+
+        c_m_n(m, n) = ck::type_convert<CDataType>(v_acc);
+    };
+
+    make_ParallelTensorFunctor(f_mk_kn_mn,
+                               c_m_n.mDesc.GetLengths()[0],
+                               c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
+}

example/91_tile_program/tile_program.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+// Meta data for GPU
+// TODO: do we need to take care of data alignment in code or it's done by compiler?
+template <ck::index_t kSize>
+struct MetaData
+{
+    char p_data_[kSize];
+
+    ck::index_t size_ = 0;
+    ck::index_t pos_  = 0;
+
+    __host__ __device__ void reset()
+    {
+        size_ = 0;
+        pos_  = 0;
+    }
+
+    __device__ void reset_pos() { pos_ = 0; }
+
+    // push meta data on host
+    // TODO: correct forwarding?
+    template <typename T>
+    __host__ auto push(T&& a)
+    {
+        using Type = ck::remove_cvref_t<T>;
+
+        static_assert(std::is_trivially_copy_constructible_v<Type> &&
+                      std::is_trivially_destructible_v<Type>);
+
+        assert(size_ + sizeof(Type) <= kSize);
+
+        // use placement new to create object copy
+        new(p_data_ + size_) Type(std::forward<T>(a));
+
+        size_ += sizeof(Type);
+
+        return ck::forwarder{}(a);
+    }
+
+    // pull meta data on device
+    // TODO: correct forwarding?
+    template <typename T>
+    __device__ auto pull()
+    {
+        using Type = ck::remove_cvref_t<T>;
+
+        static_assert(std::is_trivially_copy_constructible_v<Type> &&
+                      std::is_trivially_destructible_v<Type>);
+
+        Type a(*reinterpret_cast<Type*>(p_data_ + pos_));
+
+        pos_ += sizeof(Type);
+
+        return a;
+    }
+};
+
+// namespace tp (for tile programming)
+struct ProgramServer
+{
+    // meta data on device
+    MetaData<1024> meta_data_;
+
+    __host__ void cpu_init() { meta_data_.reset(); }
+
+    __device__ void gpu_init() { meta_data_.reset_pos(); }
+
+    // push meta data on host
+    template <typename T>
+    __host__ auto operator()(T&& a)
+    {
+        return ck::forwarder{}(meta_data_.push(a));
+    }
+
+    // push meta data on host
+    template <typename T>
+    __device__ auto operator()(T&&)
+    {
+        return ck::forwarder{}(meta_data_.pull<T>());
+    }
+
+    //
+    __host__ static ck::index_t get_block_id() { return -1; }
+
+    __host__ static ck::index_t get_thread_id() { return -1; }
+
+    __host__ static ck::index_t get_grid_size() { return -1; }
+
+    __host__ static void block_sync_lds() {}
+
+    // TODO: correct forwarding?
+    template <typename T>
+    __host__ static constexpr auto read_first_lane(T&& a)
+    {
+        return ck::forwarder{}(a);
+    }
+
+    template <typename T>
+    __host__ T warp_shuffle_up(T, uint32_t)
+    {
+        return 0;
+    }
+
+    template <typename T>
+    __host__ T warp_shuffle_down(T, uint32_t)
+    {
+        return 0;
+    }
+
+    //
+    __device__ static ck::index_t get_block_id() { return ck::get_block_id(); }
+
+    __device__ static ck::index_t get_thread_id() { return ck::get_thread_id(); }
+
+    __device__ static ck::index_t get_grid_size() { return ck::get_grid_size(); }
+
+    __device__ static void block_sync_lds() { ck::block_sync_lds(); }
+
+    template <typename T>
+    __device__ static constexpr auto read_first_lane(T&& a)
+    {
+        return __builtin_amdgcn_readfirstlane(a);
+    }
+
+    template <typename T>
+    __device__ T warp_shuffle_up(const T& var, uint32_t delta)
+    {
+        return ck::warp_shuffle_up(var, delta);
+    }
+
+    template <typename T>
+    __device__ T warp_shuffle_down(const T& var, uint32_t delta)
+    {
+        return ck::warp_shuffle_down(var, delta);
+    }
+};
+
+template <typename Server, typename Program, typename... Xs>
+#if CK_USE_LAUNCH_BOUNDS
+__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+    __global__ void gpu_program_wrapper(Server server, Program f, Xs... xs)
+{
+    server.gpu_init();
+    f(server, xs...);
+}
+
+template <typename Server, typename Program, typename... Xs>
+float launch(Server server, Program f, dim3 grid_dim, dim3 block_dim, Xs... xs)
+{
+    server.cpu_init();
+
+    f(server, xs...);
+
+    printf("meta data size %d\n", server.meta_data_.size_);
+
+    printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
+           __func__,
+           grid_dim.x,
+           grid_dim.y,
+           grid_dim.z,
+           block_dim.x,
+           block_dim.y,
+           block_dim.z);
+#if 0
+    gpu_program_wrapper<Server, Program><<<grid_dim, block_dim, 0, nullptr>>>(server, f, xs...);
+#else
+    return launch_and_time_kernel(StreamConfig{nullptr, true, 0},
+                                  gpu_program_wrapper<Server, Program, Xs...>,
+                                  grid_dim,
+                                  block_dim,
+                                  0,
+                                  server,
+                                  f,
+                                  xs...);
+#endif
+}

include/ck/ck.hpp

@@ -134,7 +134,7 @@
 #define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
 
 // experimental feature: multi index implemented as array
-#define CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 0
+#define CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 1
 
 // experimental feature: static tensor descriptor
 #define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0
@@ -150,9 +150,6 @@
 #define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
 #define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
 
-// experimental feature: in-regsiter sub-dword transpose
-#define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1
-
 // experimental feature: merge transformation use magic number division
 #define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
 
@@ -160,9 +157,6 @@
 // pointer of scalar
 #define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 0
 
-// experimental feature: use __builtin_memcpy instead of union to do bit_cast
-#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
-
 // experimental feature: optimize for inter-wave scheduling policy
 #define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
 #define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
@@ -215,6 +209,11 @@
 
 namespace ck {
 
+enum struct ArchEnum
+{
+    Gfx90a,
+};
+
 enum struct InMemoryDataOperationEnum
 {
     Set,

include/ck/tensor/static_tensor.hpp

@@ -78,7 +78,9 @@ struct StaticTensor
 
     StaticBuffer<AddressSpace, T, element_space_size_, true> data_;
     static constexpr T zero_scalar_value_ = T{0};
+    // for read access of invalid element
     const T invalid_element_scalar_value_;
+    // for write access of invalid element
     T ignored_element_scalar_;
 };
 
@@ -101,12 +103,12 @@ struct StaticTensorTupleOfVectorBuffer
     using V = vector_type<S, ScalarPerVector>;
 
     __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer()
-        : invalid_element_scalar_value_{0}
+        : invalid_element_scalar_value_{0}, ignored_element_scalar_{0}
     {
     }
 
     __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer(S invalid_element_value)
-        : invalid_element_scalar_value_{invalid_element_value}
+        : invalid_element_scalar_value_{invalid_element_value}, ignored_element_scalar_{0}
     {
     }
 
@@ -244,7 +246,9 @@ struct StaticTensorTupleOfVectorBuffer
 
     StaticBufferTupleOfVector<AddressSpace, S, num_of_vector_, ScalarPerVector, true> data_;
     static constexpr S zero_scalar_value_ = S{0};
+    // for read access of invalid element
     const S invalid_element_scalar_value_ = S{0};
+    // for write access of invalid element
     S ignored_element_scalar_;
 };
 

include/ck/tensor/tensor_view.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_coordinate.hpp"
+
+namespace ck {
+
+template <typename BufferView_, typename TensorDesc_>
+struct TensorView
+{
+    using BufferView  = remove_reference_t<BufferView_>;
+    using DataType    = typename BufferView::type;
+    using TensorDesc  = remove_cvref_t<TensorDesc_>;
+    using TensorIndex = Array<index_t, TensorDesc::GetNumOfTopDimension()>;
+    using TensorCoord = decltype(make_tensor_coordinate(TensorDesc{}, TensorIndex{}));
+
+    __host__ __device__ constexpr TensorView() = default;
+
+    __host__ __device__ constexpr TensorView(const BufferView& buffer_view, const TensorDesc& desc)
+        : buf_{buffer_view}, desc_{desc}
+    {
+    }
+
+    __host__ __device__ constexpr auto& GetTensorDescriptor() const { return desc_; }
+
+    __host__ __device__ static constexpr index_t GetNumOfDimension()
+    {
+        return TensorDesc::GetNumOfTopDimension();
+    }
+
+    __host__ __device__ constexpr const auto& GetBufferView() const { return buf_; }
+
+    __host__ __device__ constexpr auto& GetBufferView() { return buf_; }
+
+    __host__ __device__ constexpr DataType GetElement(const TensorCoord& coord) const
+    {
+        return buf_.template Get<DataType>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord));
+    }
+
+    __host__ __device__ constexpr void SetElement(const TensorCoord& coord, const DataType& x)
+    {
+        buf_.template Set<DataType>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord),
+            x);
+    }
+
+    // X is vector of DataType.
+    // "coord" is coordinate of DataType, not X. "coord" should be aligned to X
+    template <typename X,
+              typename enable_if<is_same_v<typename scalar_type<remove_cvref_t<X>>::type,
+                                           typename scalar_type<remove_cvref_t<DataType>>::type>,
+                                 bool>::type = false>
+    __host__ __device__ constexpr remove_cvref_t<X>
+    GetVectorizedElements(const TensorCoord& coord) const
+    {
+        return buf_.template Get<X>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord));
+    }
+
+    // X is vector of DataType.
+    // "coord" is coordinate of DataType, not X. "coord" should be aligned to X
+    template <typename X,
+              typename enable_if<is_same_v<typename scalar_type<remove_cvref_t<X>>::type,
+                                           typename scalar_type<remove_cvref_t<DataType>>::type>,
+                                 bool>::type = false>
+    __host__ __device__ constexpr void SetVectorizedElements(const TensorCoord& coord, const X& x)
+    {
+        buf_.template Set<X>(coord.GetOffset(),
+                             coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord),
+                             x);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("TensorView{");
+
+        // buf_
+        printf("buf_: ");
+        print(buf_);
+        printf(", ");
+
+        // desc_
+        printf("desc_: ");
+        print(desc_);
+
+        printf("}");
+    }
+
+    // member
+    BufferView buf_;
+    TensorDesc desc_;
+};
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Ts>
+__host__ __device__ constexpr auto make_tensor_view(DataType* p,
+                                                    const TensorDescriptor<Ts...>& desc)
+{
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Lengths,
+          typename... Strides,
+          index_t GuaranteedLastDimensionVectorLength                              = -1,
+          index_t GuaranteedLastDimensionVectorStride                              = -1,
+          typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
+__host__ __device__ constexpr auto
+make_naive_tensor_view(DataType* p,
+                       const Tuple<Lengths...>& lengths,
+                       const Tuple<Strides...>& strides,
+                       Number<GuaranteedLastDimensionVectorLength> = Number<-1>{},
+                       Number<GuaranteedLastDimensionVectorStride> = Number<-1>{})
+{
+    auto desc = make_naive_tensor_descriptor(lengths,
+                                             strides,
+                                             Number<GuaranteedLastDimensionVectorLength>{},
+                                             Number<GuaranteedLastDimensionVectorStride>{});
+
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Lengths,
+          index_t GuaranteedLastDimensionVectorLength = -1>
+__host__ __device__ constexpr auto
+make_naive_tensor_view_packed(DataType* p,
+                              const Tuple<Lengths...>& lengths,
+                              Number<GuaranteedLastDimensionVectorLength> = Number<-1>{})
+{
+    auto desc =
+        make_naive_tensor_descriptor_packed(lengths, Number<GuaranteedLastDimensionVectorLength>{});
+
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <typename OldTensorView,
+          typename NewTransforms,
+          typename NewLowerDimensionOldVisibleIdss,
+          typename NewUpperDimensionNewVisibleIdss>
+__host__ __device__ constexpr auto transform_tensor_view(const OldTensorView& old_tensor_view,
+                                                         const NewTransforms& new_transforms,
+                                                         NewLowerDimensionOldVisibleIdss,
+                                                         NewUpperDimensionNewVisibleIdss)
+{
+    auto new_desc = transform_tensor_descriptor(old_tensor_view.desc_,
+                                                new_transforms,
+                                                NewLowerDimensionOldVisibleIdss{},
+                                                NewUpperDimensionNewVisibleIdss{});
+
+    return TensorView<typename OldTensorView::BufferView, remove_cvref_t<decltype(new_desc)>>{
+        old_tensor_view.buf_, new_desc};
+}
+
+} // namespace ck

include/ck/tensor/thread_private_tensor.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/buffer_view.hpp"
+
+namespace ck {
+
+template <typename DataType, typename TensorDescriptor>
+struct ThreadPrivateTensor
+{
+    using DataType_         = DataType;
+    using T                 = DataType_;
+    using TensorDescriptor_ = remove_cvref_t<TensorDescriptor>;
+
+    __host__ __device__ constexpr ThreadPrivateTensor()
+        // FIXME: remove static_cast
+        : desc_{}, buf_view_{p_data_, static_cast<index_t>(desc_.GetElementSpaceSize())}
+    {
+    }
+
+    __host__ __device__ constexpr ThreadPrivateTensor(const TensorDescriptor_& desc)
+        // FIXME: remove static_cast
+        : desc_{desc}, buf_view_{p_data_, static_cast<index_t>(desc_.GetElementSpaceSize())}
+    {
+    }
+
+    // member
+    TensorDescriptor_ desc_;
+
+    //
+    static constexpr index_t kMaxBufferSize_ = 32;
+    DataType_ p_data_[kMaxBufferSize_];
+
+    // FIXME: remove assumption that type of BufferSize should be index_t
+    BufferView<AddressSpaceEnum::Vgpr, DataType_, index_t, true> buf_view_;
+
+    // function
+    // FIXME: doesn't do is_valid check
+    template <typename Idx>
+    __device__ constexpr const T& operator[](const Idx& idx) const
+    {
+#if 1 // debug
+      // FIXME: remove to_multi_index
+        const auto coord = make_tensor_coordinate(desc_, to_multi_index(idx));
+
+        const index_t offset = coord.GetOffset();
+
+        return buf_view_[offset];
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return buf_view_[offset];
+#endif
+    }
+
+    // FIXME: doesn't do is_valid check
+    template <typename Idx>
+    __device__ constexpr T& operator()(const Idx& idx)
+    {
+#if 1 // debug
+      // FIXME: remove to_multi_index
+        const auto coord = make_tensor_coordinate(desc_, to_multi_index(idx));
+
+        const index_t offset = coord.GetOffset();
+
+        return buf_view_(offset);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return buf_view_(offset);
+#endif
+    }
+
+    // idx is the index of T, not X. idx should be aligned to X
+    template <typename X, typename Idx>
+    __device__ constexpr X Get(const Idx& idx) const
+    {
+#if 1 // debug
+        const auto coord = make_tensor_coordinate(desc_, idx);
+
+        const index_t offset = coord.GetOffset();
+
+        const bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        return buf_view_.template Get<X>(offset, is_valid);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(TensorDescriptor_{}, coord);
+
+        return buf_view_.template Get<X>(offset, is_valid);
+#endif
+    }
+
+    // idx is the index of T, not X. idx should be aligned to X
+    template <typename X, typename Idx>
+    __device__ void Set(const Idx& idx, const X& x)
+    {
+
+#if 1 // debug
+        const auto coord = make_tensor_coordinate(desc_, idx);
+
+        const index_t offset = coord.GetOffset();
+
+        const bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        buf_view_.template Set<X>(offset, is_valid, x);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(TensorDescriptor_{}, coord);
+
+        buf_view_.template Set<X>(offset, is_valid, x);
+#endif
+    }
+};
+
+} // namespace ck