Batched gemm and reduction (#156)

* adding batched_gemm_and_reduction

* batched_gemm_reduce works with bactch_count=1

* fix a bug in grid_size; batched_gemm_reduce works for batch_count > 1

* adding profiler for batched_gemm_fp16

* fixed a bug in declaration of d1 and d0; both example and profiler work

* clang-format

* cleanup

* batched_gemm_reduce: add test

* minor change

* fixed some typo in function names

profiler/include/profile_batched_gemm_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 "element_wise_reduce_operation.hpp"
+#include "device_gemm_reduce.hpp"
+#include "reference_batched_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_gemm_instance {
+
+using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::ReduceSum,
+    ck::tensor_operation::element_wise::ReduceSquareSum>;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+} // namespace device_gemm_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename DDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_batched_gemm_reduce_impl(int do_verification,
+                                      int init_method,
+                                      bool do_log,
+                                      int nrepeat,
+                                      int M,
+                                      int N,
+                                      int K,
+                                      int StrideA,
+                                      int StrideB,
+                                      int StrideC,
+                                      int BatchCount)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({row * stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({col * stride, 1, stride}));
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
+        {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
+        {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
+
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
+        {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
+        {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
+    std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
+    using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
+
+    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{};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(int batch = 0; batch < BatchCount; ++batch)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                float d0_acc = d0_reduce_op.GetReduceZeroValue();
+                float d1_acc = d1_reduce_op.GetReduceZeroValue();
+
+                for(int n = 0; n < N; ++n)
+                {
+                    d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n));
+                    d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n));
+                }
+
+                d0_g_m_host_result(batch, m) = d0_acc;
+                d1_g_m_host_result(batch, m) = d1_acc;
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
+    DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
+        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_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_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_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_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_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_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_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_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<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                          static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          d0_reduce_op,
+                                          d1_reduce_op,
+                                          BatchCount);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // warm up
+            invoker_ptr->Run(argument_ptr.get());
+
+            // timing
+            float total_time = 0;
+
+            for(int i = 0; i < nrepeat; ++i)
+            {
+                // init DO, D1 to 0
+                d0_device_buf.SetZero();
+                d1_device_buf.SetZero();
+
+                KernelTimer timer;
+
+                timer.Start();
+
+                invoker_ptr->Run(argument_ptr.get());
+
+                timer.End();
+
+                total_time += timer.GetElapsedTime();
+            }
+
+            float ave_time = total_time / nrepeat;
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
+                                    sizeof(BDataType) * BatchCount * K * N +
+                                    sizeof(CDataType) * BatchCount * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << 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_g_m_n_device_result.mData.data());
+                d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+                d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+
+                float c_error  = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
+                float d0_error = check_error(d0_g_m_host_result, d0_g_m_device_result);
+                float d1_error = check_error(d1_g_m_host_result, d1_g_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_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", d0_g_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", d1_g_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_batched_gemm_reduce.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+
+#include "profile_batched_gemm_reduce_impl.hpp"
+
+int profile_batched_gemm_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout_t
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType_t
+    {
+        F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 15 || argc == 16))
+    {
+        printf("arg1: tensor operation (batched_gemm: BatchedGEMM+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("arg8: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: run kernel # of times (>1)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
+        printf("arg15: split k into  mulitiple batch\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType_t>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout_t>(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 int nrepeat          = 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 BatchCount = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
+       layout == GemmMatrixLayout_t::MK_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<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,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout_t::MK_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<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,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout_t::KM_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<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,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType_t::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout_t::KM_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<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,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 1;
+}

profiler/src/profiler.cpp

@@ -17,6 +17,7 @@ int profile_conv_fwd_bias_relu_add(int, char*[]);
 int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
 int profile_convnd_bwd_data(int, char*[], int);
 int profile_reduce(int, char*[]);
+int profile_batched_gemm_reduce(int, char*[]);
 
 int main(int argc, char* argv[])
 {
@@ -44,6 +45,10 @@ int main(int argc, char* argv[])
     {
         return profile_batched_gemm(argc, argv);
     }
+    else if(strcmp(argv[1], "batched_gemm_reduce") == 0)
+    {
+        return profile_batched_gemm_reduce(argc, argv);
+    }
     else if(strcmp(argv[1], "grouped_gemm") == 0)
     {
         profile_grouped_gemm(argc, argv);

test/CMakeLists.txt

@@ -39,6 +39,7 @@ add_subdirectory(gemm)
 add_subdirectory(gemm_split_k)
 add_subdirectory(gemm_reduce)
 add_subdirectory(batched_gemm)
+add_subdirectory(batched_gemm_reduce)
 add_subdirectory(grouped_gemm)
 add_subdirectory(convnd_fwd)
 add_subdirectory(reduce)

test/batched_gemm_reduce/CMakeLists.txt

+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/profiler/include
+    ${PROJECT_SOURCE_DIR}/test/include
+    ${PROJECT_SOURCE_DIR}/external/include/half
+)
+
+add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
+target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE host_tensor)
+target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)

test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp

+#include <iostream>
+
+#include "profile_batched_gemm_reduce_impl.hpp"
+
+int main()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    int M = 512;
+    int N = 256;
+    int K = 128;
+
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, 1, M, N, K, K, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, 1, M, N, K, K, K, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, 1, M, N, K, M, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, 1, M, N, K, M, K, N, BatchCount);
+
+    if(pass)
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Fail" << std::endl;
+        return -1;
+    }
+}

test/gemm_reduce/gemm_reduce_fp16.cpp

-#include <algorithm>
-#include <cstdlib>
-#include <half.hpp>
 #include <iostream>
-#include <numeric>
-#include <tuple>
-#include <vector>
 
 #include "profile_gemm_reduce_impl.hpp"