Add gemm_bais_add_reduce to ckProfiler

profiler/CMakeLists.txt

@@ -29,6 +29,7 @@ set(PROFILER_SOURCE
     src/profile_gemm_bias_relu.cpp
     src/profile_gemm_bias_relu_add.cpp
     src/profile_gemm_reduce.cpp
+    src/profile_gemm_bias_add_reduce.cpp
     src/profile_batched_gemm.cpp
     src/profile_conv_fwd_bias_relu.cpp
     src/profile_conv_fwd_bias_relu_add.cpp
@@ -46,6 +47,7 @@ add_executable(ckProfiler ${PROFILER_SOURCE})
 target_link_libraries(ckProfiler PRIVATE host_tensor)
 target_link_libraries(ckProfiler PRIVATE conv_util)
 target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)

profiler/include/profile_gemm_bias_add_reduce_impl.hpp

+#pragma once
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_conv.hpp"
+#include "tensor_layout.hpp"
+#include "device_tensor.hpp"
+#include "element_wise_operation.hpp"
+#include "reduction_operator.hpp"
+#include "device_gemm_reduce.hpp"
+#include "reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_gemm_instance {
+
+using F32            = float;
+using F16            = ck::half_t;
+using DPtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
+using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
+using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
+using DInElementOps  = ck::Tuple<Identity, Square>;
+using DOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmBiasAddReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmBiasAddReducePtr<
+    DPtrsGlobal,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    DInElementOps,
+    DOutElementOps>;
+
+void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+} // namespace device_gemm_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename C1DataType,
+          typename DDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_bias_add_reduce_impl(int do_verification,
+                                       int init_method,
+                                       bool do_log,
+                                       bool time_kernel,
+                                       int M,
+                                       int N,
+                                       int K,
+                                       int StrideA,
+                                       int StrideB,
+                                       int StrideC,
+                                       int StrideC1)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d0_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d0_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
+    std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        c1_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
+        c1_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using PassThrough       = ck::tensor_operation::element_wise::PassThrough;
+    using AElementOp        = PassThrough;
+    using BElementOp        = PassThrough;
+    using CElementOp        = PassThrough;
+    using D0ReduceOp        = ck::reduce::Add<float>;
+    using D1ReduceOp        = ck::reduce::Add<float>;
+    using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
+    using UnaryIdenticElementOp =
+        ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
+    using UnarySquareElementOp =
+        ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
+    using DxsInElementOps  = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+    using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+    const auto d0_reduce_op = D0ReduceOp{};
+    const auto d1_reduce_op = D1ReduceOp{};
+
+    auto dxs_in_element_op  = DxsInElementOps{};
+    auto dxs_out_element_op = DxsOutElementOps{M, M};
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                DDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+            for(int n = 0; n < N; ++n)
+            {
+                float acc =
+                    static_cast<float>(c_m_n_host_result(m, n)) + static_cast<float>(bias_n(n));
+                c_element_op(acc, acc);
+                acc += static_cast<float>(c1_m_n(m, n));
+                c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
+            }
+
+        for(int m = 0; m < M; ++m)
+        {
+            float d0_acc = d0_reduce_op.GetIdentityValue();
+            float d1_acc = d1_reduce_op.GetIdentityValue();
+
+            for(int n = 0; n < N; ++n)
+            {
+                float c_val  = ck::type_convert<float>(c_m_n_host_result(m, n));
+                float d0_val = 0;
+                float d1_val = 0;
+
+                dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
+                dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
+                d0_reduce_op(d0_acc, d0_val);
+                d1_reduce_op(d1_acc, d1_val);
+            }
+
+            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
+            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
+            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
+            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
+    DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
+    DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
+    DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
+
+    auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    c1_device_buf.ToDevice(c1_m_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasAddReduceNoOpPtr>
+        gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::device_gemm_instance::
+                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::device_gemm_instance::
+                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::device_gemm_instance::
+                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::device_gemm_instance::
+                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
+            static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
+            dxs_global,
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            StrideC,
+            StrideC1,
+            a_element_op,
+            b_element_op,
+            c_element_op,
+            dxs_in_element_op,
+            dxs_out_element_op);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            d0_device_buf.SetZero();
+            d1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+
+            std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                   sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
+                                   sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
+                                   sizeof(DDataType) * M;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+                d0_device_buf.FromDevice(d0_m_device_result.mData.data());
+                d1_device_buf.FromDevice(d1_m_device_result.mData.data());
+
+                float c_error  = check_error(c_m_n_host_result, c_m_n_device_result);
+                float d0_error = check_error(d0_m_host_result, d0_m_device_result);
+                float d1_error = check_error(d1_m_host_result, d1_m_device_result);
+
+                pass = pass && (c_error < 1E-6);
+                pass = pass && (d0_error < 1E-6);
+                pass = pass && (d1_error < 1E-6);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/profile_gemm_bias_add_reduce.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include "profile_gemm_bias_add_reduce_impl.hpp"
+
+int profile_gemm_bias_add_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 14 || argc == 15))
+    {
+        printf("arg1: tensor operation (gemm: GEMM+bias+add+Reduce)\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideC  = std::stoi(argv[13]);
+    const int StrideC1 = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+       layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}

profiler/src/profiler.cpp

@@ -11,6 +11,7 @@ int profile_gemm_bias_2d(int, char*[]);
 int profile_gemm_bias_relu(int, char*[]);
 int profile_gemm_bias_relu_add(int, char*[]);
 int profile_gemm_reduce(int, char*[]);
+int profile_gemm_bias_add_reduce(int, char*[]);
 int profile_batched_gemm(int, char*[]);
 int profile_grouped_gemm(int, char*[]);
 int profile_conv_fwd(int, char*[]);
@@ -44,6 +45,10 @@ int main(int argc, char* argv[])
     {
         return profile_gemm_reduce(argc, argv);
     }
+    else if(strcmp(argv[1], "gemm_bias_add_reduce") == 0)
+    {
+        return profile_gemm_bias_add_reduce(argc, argv);
+    }
     else if(strcmp(argv[1], "batched_gemm") == 0)
     {
         return profile_batched_gemm(argc, argv);