Merge branch 'develop' into dl_conv_multiple_d

CMakeLists.txt

@@ -262,6 +262,16 @@ rocm_package_setup_component(tests
         PACKAGE_NAME tests # Prevent -static suffix on package name
 )
 
+rocm_package_setup_component(examples
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME examples
+)
+
+rocm_package_setup_component(profiler
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME ckProfiler
+)
+
 add_subdirectory(library)
 add_subdirectory(example)
 add_subdirectory(test)

client_example/02_gemm_add_add_fastgelu/CMakeLists.txt

+add_custom_target(client_gemm_fastgelu_examples)
+
 add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
 target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
+target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
+target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
+                 client_gemm_fastgelu)

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            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(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // 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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            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(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/04_contraction/contraction_bilinear.cpp

@@ -12,6 +12,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
 
 using F32 = float;
 
@@ -192,20 +193,14 @@ int main(int argc, char* argv[])
         {
             float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
 
-            ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
-                                            e_ms_ns_lengths.begin() + NumDimM,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
 
-            ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
-                                            e_ms_ns_lengths.begin() + NumDimM + NumDimN,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
 
-            ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
-                                            a_ms_ks_lengths.begin() + NumDimM + NumDimK,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
 
             std::size_t flop      = std::size_t(2) * M * N * K;
             std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +

client_example/04_contraction/contraction_scale.cpp

@@ -12,6 +12,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
 
 using F32 = float;
 
@@ -178,20 +179,14 @@ int main(int argc, char* argv[])
         {
             float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
 
-            ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
-                                            e_ms_ns_lengths.begin() + NumDimM,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
 
-            ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
-                                            e_ms_ns_lengths.begin() + NumDimM + NumDimN,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
 
-            ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
-                                            a_ms_ks_lengths.begin() + NumDimM + NumDimK,
-                                            ck::index_t{1},
-                                            std::multiplies<ck::index_t>{});
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
 
             std::size_t flop = std::size_t(2) * M * N * K;
             std::size_t num_btype =

client_example/07_conv2d_fwd/CMakeLists.txt

-add_executable(client_conv2d_fwd conv2d_fwd.cpp)
-target_link_libraries(client_conv2d_fwd PRIVATE composable_kernel::device_operations)

client_example/07_grouped_conv2d_fwd/CMakeLists.txt

+add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
+target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations)

client_example/07_grouped_conv2d_fwd/grouped_conv2d_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 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<>{});
+
+    // transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
+    std::rotate(
+        rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
+    std::rotate(
+        rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
+    std::rotate(
+        rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
+    std::rotate(
+        rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
+    std::rotate(
+        rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
+    std::rotate(
+        rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    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};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 PassThrough>;
+
+    // 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_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}

client_example/09_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp → client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp

@@ -98,16 +98,18 @@ int main()
     SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
     SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
 
-    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdData<NumDimSpatial,
-                                                                            InLayout,
-                                                                            WeiLayout,
-                                                                            OutLayout,
-                                                                            InDataType,
-                                                                            WeiDataType,
-                                                                            OutDataType,
-                                                                            PassThrough,
-                                                                            PassThrough,
-                                                                            PassThrough>;
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     PassThrough>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();
@@ -126,15 +128,18 @@ int main()
     for(int i = 0; i < op_ptrs.size(); ++i)
     {
         auto& op_ptr        = op_ptrs[i];
-        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
                                                         wei.GetDeviceBuffer(),
-                                                        out.GetDeviceBuffer(),
-                                                        in_lengths,
-                                                        in_strides,
-                                                        wei_lengths,
-                                                        wei_strides,
+                                                        {},
+                                                        in.GetDeviceBuffer(),
                                                         out_lengths,
                                                         out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
                                                         filter_strides,
                                                         filter_dilations,
                                                         input_left_pads,
@@ -189,15 +194,18 @@ int main()
         auto& op_ptr = op_ptrs[best_op_id];
         std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
                   << std::endl;
-        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+        auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
                                                         wei.GetDeviceBuffer(),
-                                                        out.GetDeviceBuffer(),
-                                                        in_lengths,
-                                                        in_strides,
-                                                        wei_lengths,
-                                                        wei_strides,
+                                                        {},
+                                                        in.GetDeviceBuffer(),
                                                         out_lengths,
                                                         out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
                                                         filter_strides,
                                                         filter_dilations,
                                                         input_left_pads,

client_example/11_grouped_conv_bwd_weight/CMakeLists.txt

+add_executable(client_grouped_conv2d_bwd_weight grouped_conv2d_bwd_weight.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_weight PRIVATE composable_kernel::device_operations)

client_example/07_conv2d_fwd/conv2d_fwd.cpp → client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight.cpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
+#include <cstdlib>
 #include <iomanip>
 #include <iostream>
-#include <vector>
+#include <iterator>
+#include <numeric>
 
 #include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/gpu/convolution_forward.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
@@ -15,21 +17,22 @@ using InDataType  = ck::half_t;
 using WeiDataType = ck::half_t;
 using OutDataType = ck::half_t;
 
-using InLayout    = ck::tensor_layout::convolution::NHWC;
-using WeiLayout   = ck::tensor_layout::convolution::KYXC;
-using OutLayout   = ck::tensor_layout::convolution::NHWK;
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
 static constexpr ck::index_t NumDimSpatial = 2;
-static constexpr ck::index_t N             = 16;
-static constexpr ck::index_t K             = 32;
-static constexpr ck::index_t C             = 3;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
 static constexpr ck::index_t Y             = 3;
 static constexpr ck::index_t X             = 3;
-static constexpr ck::index_t Hi            = 224;
-static constexpr ck::index_t Wi            = 224;
-static constexpr ck::index_t Ho            = 113;
-static constexpr ck::index_t Wo            = 113;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
 
 struct SimpleDeviceMem
 {
@@ -47,30 +50,33 @@ struct SimpleDeviceMem
     void* p_mem_;
 };
 
-int main(int argc, char* argv[])
+int main()
 {
-    std::vector<ck::index_t> in_spatial_lengths{Hi, Wi};
-    std::vector<ck::index_t> filter_spatial_lengths{Y, X};
-    std::vector<ck::index_t> out_spatial_lengths{Ho, Wo};
-    std::vector<ck::index_t> filter_strides{2, 2};
-    std::vector<ck::index_t> filter_dilations{1, 1};
-    std::vector<ck::index_t> input_left_pads{2, 2};
-    std::vector<ck::index_t> input_right_pads{2, 2};
-
-    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
-    SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
-    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
-
-    using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NumDimSpatial,
-                                                                 InLayout,
-                                                                 WeiLayout,
-                                                                 OutLayout,
-                                                                 InDataType,
-                                                                 WeiDataType,
-                                                                 OutDataType,
-                                                                 PassThrough,
-                                                                 PassThrough,
-                                                                 PassThrough>;
+    std::array<ck::index_t, NumDimSpatial> input_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, NumDimSpatial> filter_spatial_lengths{Y, X};
+    std::array<ck::index_t, NumDimSpatial> output_spatial_lengths{Ho, Wo};
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> conv_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};
+
+    ck::index_t split_k = 2;
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              PassThrough,
+                                                                              PassThrough,
+                                                                              PassThrough>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();
@@ -92,19 +98,21 @@ int main(int argc, char* argv[])
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         wei.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         K,
                                                         C,
-                                                        in_spatial_lengths,
+                                                        input_spatial_lengths,
                                                         filter_spatial_lengths,
-                                                        out_spatial_lengths,
-                                                        filter_strides,
-                                                        filter_dilations,
+                                                        output_spatial_lengths,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
                                                         input_left_pads,
                                                         input_right_pads,
                                                         PassThrough{},
                                                         PassThrough{},
-                                                        PassThrough{});
+                                                        PassThrough{},
+                                                        split_k);
         auto invoker_ptr    = op_ptr->MakeInvokerPointer();
         std::string op_name = op_ptr->GetTypeString();
 
@@ -112,10 +120,10 @@ int main(int argc, char* argv[])
         {
             float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
 
-            std::size_t flop      = 2 * N * K * C * Ho * Wo * Y * X;
-            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * C +
-                                    sizeof(WeiDataType) * K * Y * X * C +
-                                    sizeof(OutDataType) * N * Ho * Wo * K;
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
 
             float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
             float gb_per_sec = num_bytes / 1.E6 / avg_time;
@@ -134,10 +142,16 @@ int main(int argc, char* argv[])
         }
         else
         {
-            std::cout << op_name << " does not support this problem" << std::endl;
+            std::cerr << op_name << " does not support this problem" << std::endl;
         }
     }
 
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
     std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
               << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
@@ -149,21 +163,22 @@ int main(int argc, char* argv[])
         auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
                                                         wei.GetDeviceBuffer(),
                                                         out.GetDeviceBuffer(),
+                                                        G,
                                                         N,
                                                         K,
                                                         C,
-                                                        in_spatial_lengths,
+                                                        input_spatial_lengths,
                                                         filter_spatial_lengths,
-                                                        out_spatial_lengths,
-                                                        filter_strides,
-                                                        filter_dilations,
+                                                        output_spatial_lengths,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
                                                         input_left_pads,
                                                         input_right_pads,
                                                         PassThrough{},
                                                         PassThrough{},
-                                                        PassThrough{});
-
-        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr  = op_ptr->MakeInvokerPointer();
 
         if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
@@ -172,6 +187,4 @@ int main(int argc, char* argv[])
 
         std::cout << "Done" << std::endl;
     }
-
-    return 0;
-}
+}
\ No newline at end of file

client_example/12_elementwise_normalization/CMakeLists.txt

+add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
+target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_operations)

client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto mn_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * mn_size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * mn_size);
+    SimpleDeviceMem gamma_dev_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_dev_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_dev_buf(sizeof(YDataType) * mn_size);
+
+    std::array<const void*, 2> ab_input               = {a_dev_buf.GetDeviceBuffer(),
+                                           b_dev_buf.GetDeviceBuffer()};
+    std::vector<ck::index_t> abStride                 = {Stride, 1};
+    std::array<std::vector<ck::index_t>, 2> abStrides = {abStride, abStride};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        XElementwiseOperation,
+        YElementwiseOperation,
+        Rank,
+        NumReduceDim>;
+
+    // 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_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(ADataType) * M * N + sizeof(BDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(BetaDataType) * N +
+                                   sizeof(YDataType) * M * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/13_batchnorm/CMakeLists.txt

+add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
+target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)

client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp

@@ -6,6 +6,8 @@
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp"
 
+#include "ck/library/utility/literals.hpp"
+
 using F16 = ck::half_t;
 using F32 = float;
 
@@ -135,15 +137,15 @@ int main(int argc, char* argv[])
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -240,7 +242,7 @@ int main(int argc, char* argv[])
             show_2d_matrix(std::cout << "c_host  :", c_m_n_host_result) << std::endl;
         }
 #endif
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
     }
 
     return 0;

example/01_gemm/run_gemm_example.inc

@@ -32,14 +32,12 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
     {
     case 0: break;
     case 1:
-        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(),
-                                                                             a_m_k.end());
-        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
-                                                                             b_k_n.end());
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
         break;
     default:
-        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end());
-        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end());
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
     }
 
     Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
@@ -133,11 +131,11 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
 
         c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
 
-        return ck::utils::check_err(c_m_n_device_result_converted.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
 #else
         c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
 #endif
     }
 

example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp

@@ -14,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 #include "ck/library/utility/check_err.hpp"
 
@@ -177,15 +178,15 @@ int main(int argc, char* argv[])
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -271,8 +272,7 @@ int main(int argc, char* argv[])
 
     if(do_verification)
     {
-        Tensor<CShuffleDataType> c_m_n(HostTensorDescriptor(
-            std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
+        Tensor<CShuffleDataType> c_m_n({M, N});
 
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
@@ -299,7 +299,7 @@ int main(int argc, char* argv[])
 
         e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
-        return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
     }
 
     return 0;