Merge branch 'add_a8w8_preshuffle_ckprofiler' of...

Merge branch 'add_a8w8_preshuffle_ckprofiler' of https://github.com/ROCm/composable_kernel into update_cka8w8_uc

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p1_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p2_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances<
+            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16/device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_padding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitKBPreShuffle<Row,
+                                                                     Col,
+                                                                     Tuple<Row, Col>,
+                                                                     Row,
+                                                                     F8,
+                                                                     F8,
+                                                                     Tuple<F32, F32>,
+                                                                     F16,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_weight_preshuffle_xdl_f8_f8_f16_mk_mfma_mn_p3_instances<
+            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply_weight_preshuffle.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 InOutDataType>
+void preShuffleBuffer(const InOutDataType* src,
+                      InOutDataType* dst,
+                      int N,
+                      int K,
+                      int NRepeat,
+                      int KRepeat,
+                      int NWave,
+                      int KLane,
+                      int NLane,
+                      int KPack)
+{
+    int K0 = K / (KRepeat * KLane * KPack);
+    // K -> src: K0 KLane KRepeat KPack  -> dst: K0 KRpeat KLane KPack, move klane inner to make all
+    // lanes contiguous N -> N0 NRepeat NWave NLane  // todo : is NRepeat outer or inner? now it's 1
+    int tempn, tempk;
+    for(int n = 0; n < N; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0 = n / (NRepeat * NLane * NWave);
+            int k0 = k / (KRepeat * KLane * KPack);
+            tempn  = n % (NRepeat * NLane * NWave);
+            tempk  = k % (KRepeat * KLane * KPack);
+
+            int n1 = tempn / (NLane * NWave);
+            int k1 = tempk / (KRepeat * KPack); // Klane
+            tempn  = tempn % (NLane * NWave);
+            tempk  = tempk % (KRepeat * KPack);
+            int n2 = tempn / NLane;
+            int k2 = tempk / KPack; // KRepeat
+            int n3 = tempn % NLane;
+            int k3 = tempk % KPack; // Kpack
+
+            int outputIndex = n0 * KPack * NLane * KLane * NWave * KRepeat * K0 * NRepeat +
+                              n1 * KPack * NLane * KLane * NWave * KRepeat * K0 +
+                              k0 * KPack * NLane * KLane * NWave * KRepeat +
+                              k2 * KPack * NLane * KLane * NWave + n2 * KPack * NLane * KLane +
+                              k1 * KPack * NLane + n3 * KPack + k3;
+
+            dst[outputIndex] = src[n * K + k];
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_multiply_multiply_weight_preshuffle_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 StrideD1,
+                                                           int StrideE,
+                                                           int KBatch,
+                                                           int n_warmup,
+                                                           int n_iter,
+                                                           uint64_t rotating = 0)
+{
+    bool pass = true;
+
+    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<BDataType> b_preshuffled(
+        f_host_tensor_descriptor(K, N, StrideB, BLayout{})); // use layout only for size
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    int total_gemm_needed =
+        a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes() +
+        d0_m_n.GetElementSpaceSizeInBytes() + d1_m_n.GetElementSpaceSizeInBytes();
+    int rotating_count = std::max(
+        1,
+        std::min(n_iter,
+                 static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
+
+    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 << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+    std::cout << "rotating count: " << rotating_count << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        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});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough      = ck::tensor_operation::element_wise::PassThrough;
+    using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply;
+
+    using AElementOp = PassThrough;
+    using BElementOp = PassThrough;
+    using CElementOp = MultiplyMultiply;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDSplitKBPreShuffle<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        AElementOp,
+        BElementOp,
+        CElementOp>;
+
+    // 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 GEMM
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough,
+                                                                                ComputeDataType>;
+
+        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, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                c_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto preshuffle_params = op_ptr->GetPreShuffleParameters();
+
+        preShuffleBuffer(b_k_n.mData.data(),
+                         b_preshuffled.mData.data(),
+                         N,
+                         K,
+                         preshuffle_params[0],
+                         preshuffle_params[1],
+                         preshuffle_params[2],
+                         preshuffle_params[3],
+                         preshuffle_params[4],
+                         preshuffle_params[5]);
+
+        b_device_buf.ToDevice(b_preshuffled.mData.data());
+
+        std::vector<int> kbatch_list = {1, 2, 4, 8};
+
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
+
+        for(std::size_t i = 0; i < kbatch_list.size(); i++)
+        {
+            auto kbatch_curr = kbatch_list[i];
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(
+                static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                std::array<const void*, 2>{d0_device_buf.GetDeviceBuffer(),
+                                           d1_device_buf.GetDeviceBuffer()},
+                static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
+                M,
+                N,
+                K,
+                StrideA,
+                StrideB,
+                std::array<ck::index_t, 2>{StrideD0, StrideD1},
+                StrideE,
+                kbatch_curr,
+                a_element_op,
+                b_element_op,
+                c_element_op);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                c_device_buf.SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+                if(do_verification)
+                {
+                    c_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
+                    // set softer tolerances for fp8
+                    if constexpr((is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                                  is_same_v<EDataType, f8_t>) ||
+                                 (is_same_v<ADataType, int8_t> || is_same_v<BDataType, int8_t> ||
+                                  is_same_v<EDataType, int8_t>))
+                    {
+                        std::string msg = "Error: Incorrect results!";
+                        double rtol     = 1e-1;
+                        double atol     = 1e-1;
+                        pass            = pass & ck::utils::check_err(
+                                          e_m_n_device_result, e_m_n_host_result, msg, rtol, atol);
+                    }
+                    else
+                    {
+#endif
+                        pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
+                    }
+#endif
+
+                    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  : ", e_m_n_host_result.mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", e_m_n_device_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+
+                std::string op_name = op_ptr->GetTypeString();
+
+                float ave_time = invoker_ptr->Run(argument_ptr.get(),
+                                                  StreamConfig{nullptr,
+                                                               time_kernel,
+                                                               0,
+                                                               n_warmup,
+                                                               n_iter,
+                                                               rotating_count > 1,
+                                                               rotating_count});
+
+                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 << ", KBatch "
+                          << kbatch_curr << std::endl;
+
+                if(tflops > best_tflops && ave_time > 1e-10)
+                {
+                    best_op_name    = op_name;
+                    best_tflops     = tflops;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                    best_kbatch     = kbatch_curr;
+                }
+            }
+            else
+            {
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if constexpr(is_same<EDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<EDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<EDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<EDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideE = " << StrideE << " KBatch = " << best_kbatch
+              << " : " << 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/src/CMakeLists.txt

@@ -51,6 +51,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
   # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
     list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
+    list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_weight_preshuffle.cpp)
   #   list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
   # endif()
   # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
@@ -137,6 +138,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
   # if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
+    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_weight_preshuffle_instance)
   #   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
   # endif()
   # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)

profiler/src/profile_gemm_multiply_multiply_weight_preshuffle.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_multiply_multiply_weight_preshuffle_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_MFMA_MN, // 0
+};
+
+enum struct GemmDataType
+{
+    F8_F8_F16,  // 0
+    F8_F8_BF16, // 1
+};
+
+#define OP_NAME "gemm_multiply_multiply_weight_preshuffle"
+#define OP_DESC "GEMM_Multiply_Multiply_Weight_PreShuffle"
+
+int profile_gemm_multiply_multiply_weight_preshuffle(int argc, char* argv[])
+{
+    if(argc != 16 && argc != 20)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: f8->f16; 1: f8->bf16;\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[MFMA] = 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=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        printf("optional:\n");
+        printf("arg16: number of kbatch (default 1)\n");
+        printf("arg17: number of warm-up cycles (default 1)\n");
+        printf("arg18: number of iterations (default 10)\n");
+        printf("arg19: memory for rotating buffer (default 0, size in MB)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(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 StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideE  = std::stoi(argv[15]);
+
+    int n_warmup      = 1;
+    int n_iter        = 10;
+    uint64_t rotating = 0;
+    int KBatch        = 1;
+    if(argc == 20)
+    {
+        KBatch   = std::stoi(argv[16]);
+        n_warmup = std::stoi(argv[17]);
+        n_iter   = std::stoi(argv[18]);
+        rotating = std::stoull(argv[19]) * 1024 * 1024;
+    }
+
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F16  = ck::half_t;
+    using F8   = ck::f8_t;
+    // using I8   = int8_t;
+    // using I32  = int;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto comp_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto e_layout) {
+        using ADataType       = decltype(a_type);
+        using BDataType       = decltype(b_type);
+        using ComputeDataType = decltype(comp_type);
+        using D0DataType      = decltype(d0_type);
+        using D1DataType      = decltype(d1_type);
+        using AccDataType     = decltype(acc_type);
+        using EDataType       = decltype(c_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass =
+            ck::profiler::profile_gemm_multiply_multiply_weight_preshuffle_impl<ADataType,
+                                                                                BDataType,
+                                                                                ComputeDataType,
+                                                                                AccDataType,
+                                                                                D0DataType,
+                                                                                D1DataType,
+                                                                                EDataType,
+                                                                                ALayout,
+                                                                                BLayout,
+                                                                                D0Layout,
+                                                                                D1Layout,
+                                                                                ELayout>(
+                do_verification,
+                init_method,
+                do_log,
+                time_kernel,
+                M,
+                N,
+                K,
+                (StrideA < 0) ? DefaultStrideA : StrideA,
+                (StrideB < 0) ? DefaultStrideB : StrideB,
+                (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+                (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+                (StrideE < 0) ? DefaultStrideE : StrideE,
+                KBatch,
+                n_warmup,
+                n_iter,
+                rotating);
+
+        return pass ? 0 : 1;
+    };
+    if(data_type == GemmDataType::F8_F8_F16 && layout == GemmMatrixLayout::MK_MFMA_MN)
+    {
+        return profile(
+            F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Col{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_MFMA_MN)
+    {
+        return profile(
+            F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_multiply_multiply_weight_preshuffle);