Merge branch 'develop' into wavelet_model

profiler/include/profile_batched_gemm_softmax_gemm_impl.hpp → profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp

@@ -49,7 +49,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                             int BatchStrideB0 = -1,
                                             int BatchStrideB1 = -1,
                                             int BatchStrideC  = -1,
-                                            float alpha       = 1.f)
+                                            float alpha       = -1.f)
 
 {
 
@@ -187,6 +187,10 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
     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());
 
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
     auto a_element_op    = AElementOp{};
     auto b0_element_op   = B0ElementOp{};
     auto acc0_element_op = Acc0ElementOp{alpha};

profiler/include/profile_batched_gemm_softmax_gemm_permute_impl.hpp → profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp

@@ -45,7 +45,7 @@ bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
                                                     int O,
                                                     int G0,
                                                     int G1,
-                                                    float alpha = 1.f)
+                                                    float alpha = -1.f)
 
 {
 
@@ -154,6 +154,10 @@ bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
     b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
     b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
 
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
     auto a_element_op    = AElementOp{};
     auto b0_element_op   = B0ElementOp{};
     auto acc0_element_op = Acc0ElementOp{alpha};

profiler/include/profiler/profile_batchnorm_backward_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.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/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_backward_impl(bool do_verification,
+                                     int init_method,
+                                     bool do_dumpout,
+                                     bool time_kernel,
+                                     const std::vector<size_t> inOutLengths,
+                                     const std::vector<int> reduceDims,
+                                     bool haveSavedMeanInvVar,
+                                     double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<DyDataType> dy(inOutLengths);
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<MeanVarDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<MeanVarDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<DxDataType> dx_ref(inOutLengths);
+    Tensor<DxDataType> dx(inOutLengths);
+
+    Tensor<DscaleDbiasDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<DscaleDbiasDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<DyDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<DyDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<DyDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<DyDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(DyDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(MeanVarDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(MeanVarDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(DxDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(DscaleDbiasDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(DscaleDbiasDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // add device batchnorm-backward instances
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DxDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThroughOp,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              DxDataType,
+                                                              DyDataType,
+                                                              AccDataType,
+                                                              ScaleDataType,
+                                                              DscaleDbiasDataType,
+                                                              MeanVarDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumBatchNormReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x_dev.GetDeviceBuffer(),
+            dy_dev.GetDeviceBuffer(),
+            bnScale_dev.GetDeviceBuffer(),
+            haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_dev.GetDeviceBuffer(),
+            dscale_dev.GetDeviceBuffer(),
+            dbias_dev.GetDeviceBuffer());
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+        DeviceMem workspace_dev(workspace_sz);
+
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                     invariant_length * sizeof(DscaleDbiasDataType) * 2;
+
+        // inputting of savedMean, savedInvVariance
+        if(haveSavedMeanInvVar)
+            num_bytes += invariant_length * sizeof(MeanVarDataType) * 2;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass = true;
+
+            dx_dev.FromDevice(dx.mData.data());
+            dscale_dev.FromDevice(dscale.data());
+            dbias_dev.FromDevice(dbias.data());
+
+            // clang-format off
+            single_pass = single_pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:", 5e-4, 5e-4);
+            single_pass = single_pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 3e-3, 3e-3);
+            single_pass = single_pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 3e-3, 3e-3);
+            // clang-format on
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dy.bin", dy.mData.data(), dy.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx.bin", dx.mData.data(), dx.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx_ref.bin", dx_ref.mData.data(), dx_ref.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale.bin", dscale.mData.data(), dscale.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale_ref.bin", dscale_ref.mData.data(), dscale_ref.mDesc.GetElementSize());
+            // clang-format off
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#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"
+
+#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/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_fastgelu_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        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, 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)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        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()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            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)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        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;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profiler/profile_gemm_fastgelu_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#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"
+
+#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/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout>
+bool profile_gemm_fastgelu_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    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;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    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;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        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, 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)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        std::array<ck::index_t, 0>{},
+                                                        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()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            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)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        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;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck