merge develop

client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp

@@ -204,6 +204,7 @@ int main(int argc, char* argv[])
               << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
 

client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#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_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 AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = 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 StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        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]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, 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 constexpr(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 d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    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, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // 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{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr =
+        op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                                               d1_m_n_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    M,
+                                    N,
+                                    K,
+                                    StrideA,
+                                    StrideB,
+                                    std::array<ck::index_t, 2>{StrideD0, StrideD1},
+                                    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp

@@ -197,6 +197,7 @@ int main(int argc, char* argv[])
               << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
 

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#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[7]);
+    }
+    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 constexpr(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{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp

@@ -190,6 +190,7 @@ int main(int argc, char* argv[])
               << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
 

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#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[6]);
+    }
+    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 constexpr(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{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    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});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}

client_example/03_gemm_layernorm/CMakeLists.txt

 add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
-target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)
 
 add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
-target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_gemm_operations composable_kernel::device_other_operations)

client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp

@@ -172,18 +172,19 @@ int main()
             BLayout,
             CLayout>();
 
-    const auto normalize_ptrs =
-        ck::tensor_operation::device::instance::get_device_normalize_from_mean_meansquare_instances<
-            CDataType,
-            ReduceDataType,
-            ReduceDataType,
-            GammaDataType,
-            BetaDataType,
-            LayerNormOutDataType>();
-
     std::cout << "found " << gemm_reduce_ptrs.size()
               << " gemm_reduceMean_reduceSquareMean instances" << std::endl;
 
+    using NormalizeDeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<CDataType, ReduceDataType, ReduceDataType, GammaDataType, BetaDataType>,
+        ck::Tuple<LayerNormOutDataType>,
+        ck::tensor_operation::element_wise::Normalize,
+        2>;
+
+    const auto normalize_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            NormalizeDeviceOp>::GetInstances();
+
     std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
 
     auto f_matrix_space_size =

client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp

@@ -200,6 +200,7 @@ int main(int argc, char* argv[])
               << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
 

client_example/04_contraction/CMakeLists.txt

 add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
-target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
 
 add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
-target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
 
 add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
-target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
 
 add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
-target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
 
 add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
-target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_other_operations composable_kernel::device_contraction_operations composable_kernel::device_gemm_operations)
 

client_example/05_layernorm/CMakeLists.txt

-add_executable(client_layernorm2d layernorm2d.cpp)
-target_link_libraries(client_layernorm2d PRIVATE composable_kernel::device_operations)
+add_executable(client_layernorm2d_bwd_data layernorm2d_bwd_data.cpp)
+target_link_libraries(client_layernorm2d_bwd_data PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm2d_bwd_gamma_beta layernorm2d_bwd_gamma_beta.cpp)
+target_link_libraries(client_layernorm2d_bwd_gamma_beta PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm2d_fwd layernorm2d_fwd.cpp)
+target_link_libraries(client_layernorm2d_fwd PRIVATE composable_kernel::device_other_operations)
+
+add_executable(client_layernorm4d_fwd layernorm4d_fwd.cpp)
+target_link_libraries(client_layernorm4d_fwd PRIVATE composable_kernel::device_other_operations)

client_example/05_layernorm/layernorm2d_bwd_data.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_normalization_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DXDataType         = float;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * M * N);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * M * N);
+    SimpleDeviceMem gamma_dev(sizeof(GammaDataType) * N);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem dx_dev(sizeof(DXDataType) * M * N);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
+                                                                              XDataType,
+                                                                              GammaDataType,
+                                                                              MeanInvStdDataType,
+                                                                              DXDataType,
+                                                                              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
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {0, 1}, // gammaStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N, 1}, // dxStrides
+                                                        {1},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        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_byte = sizeof(DYDataType) * M * N + sizeof(XDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(MeanInvStdDataType) * M * 2 +
+                                   sizeof(DXDataType) * 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
+    if(found)
+    {
+        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
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {0, 1}, // gammaStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N, 1}, // dxStrides
+                                                        {1},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        gamma_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dx_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        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());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/05_layernorm/layernorm2d_bwd_gamma_beta.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DGammaDataType     = float;
+using DBetaDataType      = float;
+
+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(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+
+    SimpleDeviceMem dy_dev(sizeof(DYDataType) * M * N);
+    SimpleDeviceMem x_dev(sizeof(XDataType) * M * N);
+    SimpleDeviceMem mean_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * M);
+    SimpleDeviceMem dgamma_dev(sizeof(DGammaDataType) * N);
+    SimpleDeviceMem dbeta_dev(sizeof(DBetaDataType) * N);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
+                                                                      XDataType,
+                                                                      MeanInvStdDataType,
+                                                                      DGammaDataType,
+                                                                      DBetaDataType,
+                                                                      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;
+
+    std::size_t num_bytes = sizeof(DYDataType) * M * N + sizeof(XDataType) * M * N +
+                            sizeof(MeanInvStdDataType) * M * 2 + sizeof(DGammaDataType) * N +
+                            sizeof(DBetaDataType) * N;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // inLengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N},    // outLengths
+                                                        {1},    // dgammaStrides
+                                                        {1},    // dbetaStrides
+                                                        {0},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        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});
+            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;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        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}, // inLengths
+                                                        {N, 1}, // dyStrides
+                                                        {N, 1}, // xStrides
+                                                        {1, 0}, // meanStrides
+                                                        {1, 0}, // invStdStrides
+                                                        {N},    // outLengths
+                                                        {1},    // dgammaStrides
+                                                        {1},    // dbetaStrides
+                                                        {0},    // reduceDims
+                                                        dy_dev.GetDeviceBuffer(),
+                                                        x_dev.GetDeviceBuffer(),
+                                                        mean_dev.GetDeviceBuffer(),
+                                                        inv_std_dev.GetDeviceBuffer(),
+                                                        dgamma_dev.GetDeviceBuffer(),
+                                                        dbeta_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        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());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/05_layernorm/layernorm2d.cpp → client_example/05_layernorm/layernorm2d_fwd.cpp

@@ -7,17 +7,19 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
 
-using XDataType       = ck::half_t;
-using GammaDataType   = ck::half_t;
-using BetaDataType    = ck::half_t;
-using YDataType       = ck::half_t;
-using ComputeDataType = float;
-using PassThrough     = ck::tensor_operation::element_wise::PassThrough;
+using XDataType              = ck::half_t;
+using GammaDataType          = ck::half_t;
+using BetaDataType           = ck::half_t;
+using YDataType              = ck::half_t;
+using SaveMeanInvStdDataType = ck::half_t;
+using PassThrough            = ck::tensor_operation::element_wise::PassThrough;
+
+#define SAVE_MEAN_INV_STD
 
 constexpr int Rank         = 2;
 constexpr int NumReduceDim = 1;
@@ -50,15 +52,19 @@ int main(int argc, char* argv[])
     SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
     SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
     SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
-
-    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
-                                                                       GammaDataType,
-                                                                       BetaDataType,
-                                                                       ComputeDataType,
-                                                                       YDataType,
-                                                                       PassThrough,
-                                                                       Rank,
-                                                                       NumReduceDim>;
+#ifdef SAVE_MEAN_INV_STD
+    SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * M);
+    SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * M);
+#endif
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
+                                                                          GammaDataType,
+                                                                          BetaDataType,
+                                                                          YDataType,
+                                                                          SaveMeanInvStdDataType,
+                                                                          PassThrough,
+                                                                          Rank,
+                                                                          NumReduceDim>;
 
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
@@ -84,14 +90,21 @@ int main(int argc, char* argv[])
                                                         {0, 1},      // gammaStrides
                                                         {0, 1},      // betaStrides
                                                         {Stride, 1}, // yStrides
+                                                        {1},         // save_mean Strides
+                                                        {1},         // save_inv_std Strides
                                                         {1},         // reduceDims
                                                         1e-4,
                                                         x_device_buf.GetDeviceBuffer(),
                                                         gamma_device_buf.GetDeviceBuffer(),
                                                         beta_device_buf.GetDeviceBuffer(),
                                                         y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                                        save_mean_device_buf.GetDeviceBuffer(),
+                                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
                                                         nullptr,
                                                         nullptr,
+#endif
                                                         PassThrough{});
 
         auto invoker_ptr = op_ptr->MakeInvokerPointer();
@@ -100,11 +113,19 @@ int main(int argc, char* argv[])
 
         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_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
                                    sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
 
+#ifdef SAVE_MEAN_INV_STD
+            num_byte += sizeof(SaveMeanInvStdDataType) * M * 2;
+#endif
+
             float gb_per_sec = num_byte / 1.E6 / ave_time;
 
             std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
@@ -129,6 +150,7 @@ int main(int argc, char* argv[])
               << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
         std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
@@ -136,23 +158,34 @@ int main(int argc, char* argv[])
 
         auto argument_ptr = op_ptr->MakeArgumentPointer({M, N},      // lengths
                                                         {Stride, 1}, // xStrides
-                                                        {1},         // gammaStrides
-                                                        {1},         // betaStrides
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
                                                         {Stride, 1}, // yStrides
+                                                        {1},         // save_mean Strides
+                                                        {1},         // save_inv_std Strides
                                                         {1},         // reduceDims
                                                         1e-4,
                                                         x_device_buf.GetDeviceBuffer(),
                                                         gamma_device_buf.GetDeviceBuffer(),
                                                         beta_device_buf.GetDeviceBuffer(),
                                                         y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                                        save_mean_device_buf.GetDeviceBuffer(),
+                                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
                                                         nullptr,
                                                         nullptr,
+#endif
                                                         PassThrough{});
 
         auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
         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());
+
             invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
         }
 

client_example/05_layernorm/layernorm4d_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, 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_normalization_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp"
+
+using XDataType              = ck::half_t;
+using GammaDataType          = ck::half_t;
+using BetaDataType           = ck::half_t;
+using YDataType              = ck::half_t;
+using SaveMeanInvStdDataType = ck::half_t;
+using PassThrough            = ck::tensor_operation::element_wise::PassThrough;
+
+#define SAVE_MEAN_INV_STD
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+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[])
+{
+    ck::index_t N = 256;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t C = 8;
+
+    std::vector<ck::index_t> strideXY             = {H * W * C, W * C, C, 1};
+    std::vector<ck::index_t> strideGammaBeta      = {0, W * C, C, 1};
+    std::vector<ck::index_t> strideSaveMeanInvStd = {1};
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * N * H * W * C);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * H * W * C);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * H * W * C);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * N * H * W * C);
+#ifdef SAVE_MEAN_INV_STD
+    SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N);
+    SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N);
+#endif
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
+                                                                          GammaDataType,
+                                                                          BetaDataType,
+                                                                          YDataType,
+                                                                          SaveMeanInvStdDataType,
+                                                                          PassThrough,
+                                                                          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({N, H, W, C},         // lengths
+                                        strideXY,             // xStrides
+                                        strideGammaBeta,      // gammaStrides
+                                        strideGammaBeta,      // betaStrides
+                                        strideXY,             // yStrides
+                                        strideSaveMeanInvStd, // save_mean Strides
+                                        strideSaveMeanInvStd, // save_inv_std Strides
+                                        {1, 2, 3},            // reduceDims
+                                        1e-4,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        PassThrough{});
+
+        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_byte =
+                sizeof(XDataType) * N * H * W * C + sizeof(GammaDataType) * H * W * C +
+                sizeof(BetaDataType) * H * W * C + sizeof(YDataType) * N * H * W * C;
+
+#ifdef SAVE_MEAN_INV_STD
+            num_byte += sizeof(SaveMeanInvStdDataType) * N * 2;
+#endif
+
+            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
+    if(found)
+    {
+        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({N, H, W, C},         // lengths
+                                        strideXY,             // xStrides
+                                        strideGammaBeta,      // gammaStrides
+                                        strideGammaBeta,      // betaStrides
+                                        strideXY,             // yStrides
+                                        strideSaveMeanInvStd, // save_mean Strides
+                                        strideSaveMeanInvStd, // save_inv_std Strides
+                                        {1, 2, 3},            // reduceDims
+                                        1e-4,
+                                        x_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer(),
+                                        y_device_buf.GetDeviceBuffer(),
+#ifdef SAVE_MEAN_INV_STD
+                                        save_mean_device_buf.GetDeviceBuffer(),
+                                        save_inv_std_device_buf.GetDeviceBuffer(),
+#else
+                                        nullptr,
+                                        nullptr,
+#endif
+                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        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());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/06_softmax/CMakeLists.txt

 add_executable(client_softmax4d softmax4d.cpp)
-target_link_libraries(client_softmax4d PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_softmax4d PRIVATE composable_kernel::device_other_operations composable_kernel::device_reduction_operations)

client_example/06_softmax/softmax4d.cpp

@@ -53,12 +53,35 @@ int main(int argc, char* argv[])
     SimpleDeviceMem in(sizeof(InDataType) * num_elements);
     SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
 
-    using DeviceOp = ck::tensor_operation::device::
-        DeviceSoftmax<InDataType, AccDataType, OutDataType, PassThrough, PassThrough, Rank>;
+    using DeviceOp = ck::tensor_operation::device::DeviceSoftmax<InDataType,
+                                                                 AccDataType,
+                                                                 OutDataType,
+                                                                 PassThrough,
+                                                                 PassThrough,
+                                                                 Rank,
+                                                                 NumReduceDim>;
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
         DeviceOp>::GetInstances();
 
+    auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr = generic_op_ptr->MakeArgumentPointer(in_lengths,
+                                                                    in_strides,
+                                                                    reduce_dims,
+                                                                    alpha,
+                                                                    beta,
+                                                                    in.GetDeviceBuffer(),
+                                                                    out.GetDeviceBuffer(),
+                                                                    PassThrough{},
+                                                                    PassThrough{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
     std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
     std::string best_op_name;
@@ -74,11 +97,6 @@ int main(int argc, char* argv[])
     {
         auto& op_ptr = op_ptrs[i];
 
-        if(op_ptr->GetRank() != Rank || op_ptr->GetNumReduceDim() != NumReduceDim)
-        {
-            continue;
-        }
-
         auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
                                                         in_strides,
                                                         reduce_dims,
@@ -122,6 +140,7 @@ int main(int argc, char* argv[])
               << best_op_name << std::endl;
 
     // run the best intance
+    if(found)
     {
         auto& op_ptr = op_ptrs[best_op_id];
         std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()

client_example/07_grouped_convnd_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)
+target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_conv_operations)
 
 add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
-target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_conv_operations)

client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp

@@ -100,18 +100,18 @@ int main()
     SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
     SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K);
 
-    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
-                                                                                 InLayout,
-                                                                                 WeiLayout,
-                                                                                 ck::Tuple<>,
-                                                                                 OutLayout,
-                                                                                 InDataType,
-                                                                                 WeiDataType,
-                                                                                 ck::Tuple<>,
-                                                                                 OutDataType,
-                                                                                 PassThrough,
-                                                                                 PassThrough,
-                                                                                 PassThrough>;
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<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<

client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp

@@ -71,18 +71,18 @@ int main()
     SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
     SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
 
-    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
-                                                                                 InLayout,
-                                                                                 WeiLayout,
-                                                                                 ck::Tuple<>,
-                                                                                 OutLayout,
-                                                                                 InDataType,
-                                                                                 WeiDataType,
-                                                                                 ck::Tuple<>,
-                                                                                 OutDataType,
-                                                                                 PassThrough,
-                                                                                 PassThrough,
-                                                                                 PassThrough>;
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<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<