Merge branch 'develop' into gemm_layernorm_welford

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;
+}

example/17_convnd_bwd_data/CMakeLists.txt

 add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
 target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
+
+add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
+target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)

example/17_convnd_bwd_data/convnd_bwd_data_common.hpp

@@ -61,9 +61,13 @@ int run_conv_bwd_data(bool do_verification,
         out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
         wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
         break;
-    default:
+    case 2:
         out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
         wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
     }
 
     DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
@@ -98,9 +102,8 @@ int run_conv_bwd_data(bool do_verification,
 
     if(!conv.IsSupportedArgument(argument))
     {
-        throw std::runtime_error(
-            "wrong! device_conv with the specified compilation parameters does "
-            "not support this Conv problem");
+        std::cout << "Not support,please check parameters or device";
+        return 0;
     }
 
     float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_bwd_data_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+template <ck::index_t NDimSpatial>
+// clang-format off
+using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Dl<
+//        ######|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|    Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+//        ######|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|        Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+//        ######|           |           |            |            |            |    Operation|     Operation|     Operation| Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+//        ######|           |           |            |            |            |             |              |              |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+                 NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        using InLayout  = ctc::GNWC;
+        using WeiLayout = ctc::GKXC;
+        using OutLayout = ctc::GNWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<1,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<1>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        using InLayout  = ctc::GNHWC;
+        using WeiLayout = ctc::GKYXC;
+        using OutLayout = ctc::GNHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<2,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<2>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        using InLayout  = ctc::GNDHWC;
+        using WeiLayout = ctc::GKZYXC;
+        using OutLayout = ctc::GNDHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<3,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<3>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+
+    return 0;
+}

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

 add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
 add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
 add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
 add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
 add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
 
 add_custom_target(example_gemm_scale_softmax_gemm)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
+add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
 add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
 add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: bf16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, bf16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: bf16 in, bf16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_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 ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp

@@ -139,261 +139,6 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
                                                                                 B1ElementOp,
                                                                                 CElementOp>;
 
-int main(int argc, char* argv[])
-{
-    bool do_verification = true;
-    int init_method      = 1;
-    bool time_kernel     = false;
+#include "run_batched_gemm_scale_softmax_gemm.inc"
 
-    // GEMM shape
-    ck::index_t M             = 1020;
-    ck::index_t N             = 1020;
-    ck::index_t K             = 64;
-    ck::index_t O             = 128;
-    ck::index_t BatchCount    = 4;
-    ck::index_t StrideA       = -1;
-    ck::index_t StrideB0      = -1;
-    ck::index_t StrideB1      = -1;
-    ck::index_t StrideC       = -1;
-    ck::index_t BatchStrideA  = -1;
-    ck::index_t BatchStrideB0 = -1;
-    ck::index_t BatchStrideB1 = -1;
-    ck::index_t BatchStrideC  = -1;
-    float alpha               = 1;
-
-    if(argc == 1)
-    {
-        // use default case
-    }
-    else if(argc == 4)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        time_kernel     = std::stoi(argv[3]);
-    }
-    else if(argc == 9)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        time_kernel     = std::stoi(argv[3]);
-
-        M = std::stoi(argv[4]);
-        N = std::stoi(argv[5]);
-        K = std::stoi(argv[6]);
-        O = std::stoi(argv[7]);
-
-        BatchCount = std::stoi(argv[8]);
-    }
-    else if(argc == 18)
-    {
-        do_verification = std::stoi(argv[1]);
-        init_method     = std::stoi(argv[2]);
-        time_kernel     = std::stoi(argv[3]);
-
-        M = std::stoi(argv[4]);
-        N = std::stoi(argv[5]);
-        K = std::stoi(argv[6]);
-        O = std::stoi(argv[7]);
-
-        BatchCount = std::stoi(argv[8]);
-
-        StrideA  = std::stoi(argv[9]);
-        StrideB0 = std::stoi(argv[10]);
-        StrideB1 = std::stoi(argv[11]);
-        StrideC  = std::stoi(argv[12]);
-
-        BatchStrideA  = std::stoi(argv[13]);
-        BatchStrideB0 = std::stoi(argv[14]);
-        BatchStrideB1 = std::stoi(argv[15]);
-        BatchStrideC  = std::stoi(argv[16]);
-
-        alpha = std::stof(argv[17]);
-    }
-    else
-    {
-        printf("arg1: verification (0=no, 1=yes)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: time kernel (0=no, 1=yes)\n");
-        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
-               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
-        printf("arg17: scale (alpha)\n");
-        exit(0);
-    }
-
-    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
-    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
-    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
-    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
-
-    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
-    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
-    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
-    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
-
-    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
-    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
-    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
-    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
-
-    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
-    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
-    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
-    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
-
-    auto f_host_tensor_descriptor = [](std::size_t batch_count,
-                                       std::size_t row,
-                                       std::size_t col,
-                                       std::size_t stride,
-                                       std::size_t batch_stride,
-                                       auto layout) {
-        using namespace ck::literals;
-
-        if(std::is_same<decltype(layout), Row>::value)
-        {
-            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
-        }
-        else
-        {
-            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
-        }
-    };
-
-    // C_m_o = A_m_k * B0_k_n * B1_n_o
-    Tensor<ADataType> a_g_m_k(
-        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
-    Tensor<B0DataType> b0_g_k_n(
-        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
-    Tensor<B1DataType> b1_g_n_o(
-        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
-    Tensor<CDataType> c_g_m_o_host_result(
-        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
-    Tensor<CDataType> c_g_m_o_device_result(
-        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
-
-    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
-    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
-    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
-    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
-        break;
-    case 2:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
-        break;
-    case 3:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-        break;
-    default:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
-    }
-
-    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
-    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
-    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
-    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
-                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
-
-    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
-    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
-    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
-
-    auto a_element_op    = AElementOp{};
-    auto b0_element_op   = B0ElementOp{};
-    auto acc0_element_op = Acc0ElementOp{alpha};
-    auto b1_element_op   = B1ElementOp{};
-    auto c_element_op    = CElementOp{};
-
-    // do GEMM
-    auto gemm    = DeviceGemmInstance{};
-    auto invoker = gemm.MakeInvoker();
-    auto argument =
-        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
-                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
-                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
-                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
-                          M,
-                          N,
-                          K,
-                          O,
-                          BatchCount,
-                          StrideA,
-                          StrideB0,
-                          StrideB1,
-                          StrideC,
-                          BatchStrideA,
-                          BatchStrideB0,
-                          BatchStrideB1,
-                          BatchStrideC,
-                          a_element_op,
-                          b0_element_op,
-                          acc0_element_op,
-                          b1_element_op,
-                          c_element_op);
-
-    if(!gemm.IsSupportedArgument(argument))
-    {
-        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
-
-        return 0;
-    }
-
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
-
-    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
-    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
-                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
-                            BatchCount;
-
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
-
-    float gb_per_sec = num_btype / 1.E6 / ave_time;
-
-    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
-              << gemm.GetTypeString() << std::endl;
-
-    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
-
-    if(do_verification)
-    {
-        // Output of Gemm0 is input A of Gemm1
-        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-
-        auto ref_gemm0          = ReferenceGemm0Instance{};
-        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
-        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
-            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
-
-        ref_gemm0_invoker.Run(ref_gemm0_argument);
-
-        auto ref_softmax          = ReferenceSoftmaxInstance{};
-        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
-        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
-
-        ref_softmax_invoker.Run(ref_softmax_argument);
-
-        auto ref_gemm1          = ReferenceGemm1Instance{};
-        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
-        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
-            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
-
-        ref_gemm1_invoker.Run(ref_gemm1_argument);
-
-        return ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result) ? 0 : 1;
-    }
-
-    return 0;
-}
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1020;
+    ck::index_t N             = 1020;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+    float alpha               = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+
+        alpha = std::stof(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        printf("arg17: scale (alpha)\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA,
+                          StrideB0,
+                          StrideB1,
+                          StrideC,
+                          BatchStrideA,
+                          BatchStrideB0,
+                          BatchStrideB1,
+                          BatchStrideC,
+                          a_element_op,
+                          b0_element_op,
+                          acc0_element_op,
+                          b1_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}

example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc

@@ -253,7 +253,23 @@ int run(int argc, char* argv[])
             self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
         });
 
-        return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        // when BF16 is taken, set absolute error and relative error to 0.01
+        if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
+           std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
+        {
+            rtol = 1e-2;
+            atol = 1e-2;
+        }
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
                    ? 0
                    : 1;
     }

include/ck/ck.hpp

@@ -154,6 +154,13 @@
 // tuning parameter
 #define CK_WORKAROUND_SWDEV_325164 0
 
+// workaround: a BF16 attention kernel for gfx908 is likely affected by a compiler issue
+#ifdef __gfx908__
+#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 1
+#else // __gfx90a__, ...
+#define CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE 0
+#endif // __gfx908__
+
 namespace ck {
 
 enum struct InMemoryDataOperationEnum

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -194,21 +194,36 @@ struct Relu
     }
 };
 
-// https://paperswithcode.com/method/gelu
-// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+// Y = FastGelu(X)
 struct FastGelu
 {
-    template <typename Y, typename X>
-    __host__ __device__ void operator()(Y& y, const X& x) const;
-
-    template <>
-    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    // Fast GeLU
+    // https://paperswithcode.com/method/gelu
+    // y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+    __host__ __device__ static constexpr float GetFastGeLU(float x)
     {
-        const float u   = float(2) * x * (float(0.035677) * x * x + float(0.797885));
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
         const float emu = exp(-u);
-        const float cdf = float(0.5) + float(0.5) * (float(2) / (float(1) + emu) - float(1));
+        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
+        return x * cdf;
+    }
+
+    template <typename T>
+    static inline constexpr bool is_valid_param_type_v =
+        std::is_same_v<T, float> || std::is_same_v<T, half_t> || std::is_same_v<T, bhalf_t> ||
+        std::is_same_v<T, int32_t> || std::is_same_v<T, int8_t>
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        || std::is_same_v<T, ck::int4_t>
+#endif
+        ;
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        static_assert(is_valid_param_type_v<Y> && is_valid_param_type_v<X>);
 
-        y = x * cdf;
+        const float tmp_y = GetFastGeLU(type_convert<float>(x));
+        y                 = type_convert<Y>(tmp_y);
     }
 };
 

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -874,6 +874,14 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
                 }
             } // end gemm1
 
+            // workaround compiler issue; see ck/ck.hpp
+            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
+                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
+                         Gemm1NPerBlock == 128)
+            {
+                __builtin_amdgcn_sched_barrier(0);
+            }
+
             constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
                 gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
             constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);

include/ck/utility/amd_xdlops.hpp

@@ -254,7 +254,7 @@ struct intrin_mfma_f32_16x16x8bf16<16, 16>
     template <class FloatC>
     __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
     {
-        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16(
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x8bf16(
             reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
     }
 };

library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp

@@ -86,8 +86,8 @@ struct ReferenceSoftmax : public device::BaseOperator
             };
 
             arg.in_.ForEach([&](auto& self, auto idx) {
-                reduce_max(to_sm_scalar_idx(idx)) = std::max(reduce_max(to_sm_scalar_idx(idx)),
-                                                             static_cast<AccDataType>(self(idx)));
+                reduce_max(to_sm_scalar_idx(idx)) = std::max(
+                    reduce_max(to_sm_scalar_idx(idx)), ck::type_convert<AccDataType>(self(idx)));
             });
 
             // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
@@ -96,7 +96,7 @@ struct ReferenceSoftmax : public device::BaseOperator
             Tensor<AccDataType> in_stable(arg.in_.mDesc);
             in_stable.ForEach([&](auto& self, auto idx) {
                 // numerator = exp(x - max(x))
-                self(idx) = std::exp(static_cast<AccDataType>(arg.in_(idx)) -
+                self(idx) = std::exp(ck::type_convert<AccDataType>(arg.in_(idx)) -
                                      reduce_max(to_sm_scalar_idx(idx)));
             });
 
@@ -111,8 +111,10 @@ struct ReferenceSoftmax : public device::BaseOperator
             // std::endl;
 
             arg.out_.ForEach([&](auto& self, auto idx) {
-                self(idx) = arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
-                            arg.beta_ * self(idx);
+                AccDataType temp_result =
+                    arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
+                    arg.beta_ * self(idx);
+                self(idx) = ck::type_convert<OutDataType>(temp_result);
             });
 
             // LogRangeAsType<float>(std::cout << "out: ", arg.out_.mData, ",") << std::endl;