tmp save

client_example/22_grouped_gemm/CMakeLists.txt

@@ -6,3 +6,6 @@ target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel
 
 add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
 target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)
+
+add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
+target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using I8   = int8_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = BF16;
+using BDataType  = I8;
+using DsDataType = ck::Tuple<>;
+using EDataType  = BF16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    const int group_count = 16;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * i);
+        Ns.push_back(128 + 128 * i);
+        Ks.push_back(128 + 64 * i);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 1);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -594,10 +594,16 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
         {
             bool has_main_k_block_loop = true;
 
+            index_t kbatch = arg.k_batch_;
+            if constexpr(std::is_same<ADataType, ck::bhalf_t>::value)
+            {
+                kbatch = 1;
+            }
+
             for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
             {
                 const auto KPad =
-                    GridwiseGemm::CalculateKPadded(arg.gemm_desc_kernel_arg_[i].K_, arg.k_batch_);
+                    GridwiseGemm::CalculateKPadded(arg.gemm_desc_kernel_arg_[i].K_, kbatch);
 
                 if(GridwiseGemm::CalculateHasMainKBlockLoop(KPad) != has_main_k_block_loop)
                 {
@@ -641,7 +647,7 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
                     arg.barrier_size_grp_,
                     arg.gemm_desc_kernel_arg_.size(),
                     arg.grid_size_grp_,
-                    arg.k_batch_,
+                    kbatch,
                     arg.a_element_op_,
                     arg.b_element_op_,
                     arg.c_element_op_);
@@ -650,32 +656,50 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
             constexpr auto AtomicAdd = InMemoryDataOperationEnum::AtomicAdd;
             constexpr auto Set       = InMemoryDataOperationEnum::Set;
 
-            if(arg.k_batch_ > 1)
+            if constexpr(std::is_same<ADataType, ck::bhalf_t>::value)
             {
                 if(has_main_k_block_loop)
                 {
-                    ave_time =
-                        launch_kernel(integral_constant<bool, true>{},
-                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                    ave_time = launch_kernel(integral_constant<bool, true>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
                 }
                 else
                 {
-                    ave_time =
-                        launch_kernel(integral_constant<bool, false>{},
-                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                    ave_time = launch_kernel(integral_constant<bool, false>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
                 }
             }
             else
             {
-                if(has_main_k_block_loop)
+                if(kbatch > 1)
                 {
-                    ave_time = launch_kernel(integral_constant<bool, true>{},
-                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                    if(has_main_k_block_loop)
+                    {
+                        ave_time = launch_kernel(
+                            integral_constant<bool, true>{},
+                            integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                    }
+                    else
+                    {
+                        ave_time = launch_kernel(
+                            integral_constant<bool, false>{},
+                            integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                    }
                 }
                 else
                 {
-                    ave_time = launch_kernel(integral_constant<bool, false>{},
-                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                    if(has_main_k_block_loop)
+                    {
+                        ave_time =
+                            launch_kernel(integral_constant<bool, true>{},
+                                          integral_constant<InMemoryDataOperationEnum, Set>{});
+                    }
+                    else
+                    {
+                        ave_time =
+                            launch_kernel(integral_constant<bool, false>{},
+                                          integral_constant<InMemoryDataOperationEnum, Set>{});
+                    }
                 }
             }
 

library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp

@@ -15,9 +15,9 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-using I8  = int8_t;
+using I8   = int8_t;
 using BF16 = ck::bhalf_t;
-using F32 = float;
+using F32  = float;
 
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -52,7 +52,7 @@ using device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_irregular_tile_inst
     // clang-format on
     >;
 
-void add_device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instances(
+void add_device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(
     std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
                                                          Row,
                                                          DsLayout,
@@ -66,7 +66,8 @@ void add_device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instances(
                                                          PassThrough>>>& instances)
 {
     add_device_operation_instances(
-        instances, device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances{});
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances{});
 }
 
 } // namespace instance

profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.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/utility/fill.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        const std::vector<int>& Ms,
+                                        const std::vector<int>& Ns,
+                                        const std::vector<int>& Ks,
+                                        const std::vector<int>& StrideAs,
+                                        const std::vector<int>& StrideBs,
+                                        const std::vector<int>& StrideCs,
+                                        int kbatch   = 1,
+                                        int n_warmup = 1,
+                                        int n_iter   = 10)
+{
+    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});
+            }
+        };
+
+    std::size_t group_count = Ms.size();
+
+    if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
+         group_count == StrideBs.size() && group_count == StrideCs.size()))
+    {
+        throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
+    }
+
+    std::vector<Tensor<ADataType>> a_m_k;
+    std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_host_results;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_m_k.push_back(
+            Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
+        b_k_n.push_back(
+            Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
+
+        c_m_n_device_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        c_m_n_host_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+#if DEBUG_LOG
+        std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
+                  << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
+                  << "]:" << c_m_n_device_results[i].mDesc << std::endl;
+#endif // DEBUG_LOG
+        std::size_t num_thread = 1;
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+            b_k_n[i].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;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
+
+    a_device_buf.reserve(group_count);
+    b_device_buf.reserve(group_count);
+    c_device_buf.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_c.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
+        b_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
+        c_device_buf.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
+
+        a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
+        b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+
+        p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
+        p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
+        p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_device_buf[i]->GetDeviceBuffer(),
+                                             b_device_buf[i]->GetDeviceBuffer(),
+                                             {},
+                                             c_device_buf[i]->GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideCs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            ck::Tuple<>,
+                                                                            CLayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            ck::Tuple<>,
+                                                                            CDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(op_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;
+    float best_kbatch     = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                    BDataType,
+                                                                                    CDataType,
+                                                                                    AccDataType,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CElementOp>;
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                      b_k_n[i],
+                                                      c_m_n_host_results[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+        }
+    }
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            gemm_ptr->MakeArgumentPointer(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_c,
+                                          gemm_descs,
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        DeviceMem grouped_gemm_kernel_args_dev(
+            gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        gemm_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                      grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        std::string gemm_name = gemm_ptr->GetTypeString();
+
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
+
+        if(kbatch > 0)
+        {
+            kbatch_list = {kbatch};
+        }
+
+        for(std::size_t j = 0; j < kbatch_list.size(); j++)
+        {
+
+            auto kbatch_curr = kbatch_list[j];
+
+            gemm_ptr->SetKBatch(argument_ptr.get(), kbatch_curr);
+
+            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    c_device_buf[i]->SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(),
+                                 StreamConfig{nullptr, false, 0, n_warmup, n_iter});
+
+                if(do_verification)
+                {
+                    bool instance_pass = true;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+
+                        c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
+
+                        if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i],
+                                                                      "Error: Incorrect results!",
+                                                                      0.06);
+                        }
+                        else
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i]);
+                        }
+
+                        if(do_log)
+                        {
+                            LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_host  : ", c_m_n_host_results[i].mData, ",")
+                                << std::endl;
+                        }
+                    }
+
+                    std::cout << "Instance: " << gemm_name << " verification "
+                              << (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
+
+                    pass = pass && instance_pass;
+                }
+
+                float ave_time = invoker_ptr->Run(
+                    argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
+
+                if(time_kernel)
+                {
+                    std::size_t flop = 0, num_btype = 0;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+                        flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+
+                        num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
+                                     sizeof(BDataType) * Ks[i] * Ns[i] +
+                                     sizeof(CDataType) * Ms[i] * Ns[i];
+                    }
+
+                    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, " << gemm_name << ", KBatch "
+                              << kbatch_curr << 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;
+                        best_kbatch     = kbatch_curr;
+                    }
+                }
+            }
+            else
+            {
+                std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+                  << best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
+                  << std::endl;
+    }
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/CMakeLists.txt

@@ -51,6 +51,7 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
   list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
   list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
   list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
   list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
 endif()
 
@@ -124,6 +125,7 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
 endif()
 

profiler/src/profile_grouped_gemm_fixed_nk.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_grouped_gemm_fixed_nk_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+};
+
+enum struct GemmDataType
+{
+    BF16_I8_BF16, // 0
+    F16_F16_F16,  // 1
+    F16_F8_F16,   // 2
+    F16_I8_F16,   // 3
+
+};
+
+#define OP_NAME "grouped_gemm_fixed_nk"
+#define OP_DESC "Grouped GEMM Fixed NK"
+
+namespace {
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+
+    std::istringstream in(input);
+
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+
+    return out;
+}
+
+int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
+{
+    if(argc < 14)
+    {
+        std::cout
+            << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+            << "arg2: data type (0: bf16@int8; 1: fp16; 2: fp16@fp8; 3: fp16@int8)\n"
+            << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
+            << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
+            << "arg4: verification (0: no; 1: yes)\n"
+            << "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+            << "arg6: print tensor value (0: no; 1: yes)\n"
+            << "arg7: time kernel (0=n0, 1=yes)\n"
+            << "arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n"
+            << "arg15: kbatch value (default 1)\n"
+            << "optional:\n"
+            << "arg16: number of warm-up cycles (default 1)\n"
+            << "arg17: number of iterations (default 10)\n"
+            << std::endl;
+
+        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 auto Ms = argToIntArray(argv[8]);
+    const auto Ns = argToIntArray(argv[9]);
+    const auto Ks = argToIntArray(argv[10]);
+
+    const auto StrideAs = argToIntArray(argv[11]);
+    const auto StrideBs = argToIntArray(argv[12]);
+    const auto StrideCs = argToIntArray(argv[13]);
+    const int kbatch    = argc == 15 ? std::stoi(argv[14]) : 1;
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using F8   = ck::f8_t;
+    using BF16 = ck::bhalf_t;
+    using I8   = int8_t;
+
+    int n_warmup = 1;
+    int n_iter   = 10;
+    if(argc == 17)
+    {
+        n_warmup = std::stoi(argv[16]);
+        n_iter   = std::stoi(argv[17]);
+    }
+
+#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
+    if(data_type == GemmDataType::BF16_I8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        if(kbatch != 1)
+        {
+            throw std::runtime_error("wrong! BF16&INT8 gemm_fixed_nk supports only kbatch = 1");
+        }
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<BF16,
+                                                         I8,
+                                                         BF16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+    else if(data_type == GemmDataType::BF16_I8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<BF16,
+                                                         I8,
+                                                         BF16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+#endif
+#if defined(CK_ENABLE_FP16)
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         F16,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         F16,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+#endif
+#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8)
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         F8,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         F8,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+#endif
+#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_INT8)
+    else if(data_type == GemmDataType::F16_I8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         I8,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            kbatch,
+            n_warmup,
+            n_iter);
+    }
+    else if(data_type == GemmDataType::F16_I8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fixed_nk_impl<F16,
+                                                         I8,
+                                                         F16,
+                                                         F32,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs,
+            1,
+            n_warmup,
+            n_iter);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
+    }
+    return 0;
+}
+
+} // anonymous namespace
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_fixed_nk);