Universal streamk with atomics (#1360)

* universal streamk with atomics with ckprofiler support. grid_size and streamk strategy are tunable. grid_size of -1 leads to #WGs = maximum occupancy X num_CUs. implementation supports many different streamk policies: 1-tile, 2-tile, 3-tile and 4-tile. streamk strategy of -1 leads to default streamk policy (4-tile). 

* Update README.md

* fixing clang-format issues

* removed conflicts in struct members between streamk and universal streamk

* corrected arg parsing for streamk and universal streamk

* added stream-k policies for 3 tile and 4 tile

* fixed argument type issue with parsing cmd args

* changes suggested in PR review are made- removing comments and correcting copyright

* file permissions updated

* added default value support for grid_size and streamk-policy selection set to -1

* print messages for arguments

* print messages for arguments

* print messages for arguments1

example/01_gemm/CMakeLists.txt

@@ -22,6 +22,8 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
 add_example_executable(example_gemm_xdl_fp16_v2 gemm_xdl_fp16_v2.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
 
+add_example_executable(example_gemm_xdl_fp16_streamk_v3 gemm_xdl_fp16_streamk_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_streamk_v3)
 add_example_executable(example_gemm_xdl_fp16_v3 gemm_xdl_fp16_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
 add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)

example/01_gemm/README.md

@@ -7,3 +7,21 @@
 #arg3: run kernel # of times (>1)
 ./bin/example_gemm_xdl 0 1 5
 ```
+
+# Instructions for ```example_gemm_xdl_fp16_streamk_v3```
+
+## Run ```example_gemm_xdl_fp16_streamk_v3```
+```bash
+arg1: verification (0=no, 1=yes)
+arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+arg3: time kernel (0=no, 1=yes)
+arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
+arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)
+arg11: Grid_size(-1 for max occupancy)
+bin/example_gemm_xdl_fp16_streamk_v3 1 2 1 3840 4096 4096 4096 4096 4096 1 -1
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+problem {M:3840, N:4096, K:4096, SA:4096, SB:4096, SC:4096, MP:4032, NP:4096, KRead:4096, KP:4096, AK0:512, BK0:2048, MBlock: 18, NBlock: 16, Stream-K Selection:1, Grid size:-1}
+Perf: 0.292022 ms, 441.23 TFlops, 330.348 GB/s, DeviceGemmXdlUniversal<MNPadding, RRR> BlkSize: 256, BlkTile: 224x256x64, WaveTile: 16x16, WaveMap: 7x8, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2
+```

example/01_gemm/common.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
 
@@ -45,6 +45,19 @@ struct ProblemSizeStreamK final
 
     ck::index_t NumSKBlocks = -1;
 };
+struct ProblemSizeStreamK_universal final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    ck::index_t Grid_size   = -1; // defaults to max occupancy
+    ck::index_t Streamk_sel = 1;  // defaults to 1-tile SK
+};
 
 struct ProblemSizeSplitK final
 {
@@ -123,6 +136,57 @@ bool parse_cmd_args<ProblemSize>(int argc,
     return true;
 }
 
+template <>
+bool parse_cmd_args<ProblemSizeStreamK_universal>(int argc,
+                                                  char* argv[],
+                                                  ProblemSizeStreamK_universal& problem_size,
+                                                  ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc >= 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+
+        if(argc >= 11)
+        {
+            problem_size.Streamk_sel = std::stoi(argv[10]);
+            problem_size.Grid_size   = std::stoi(argv[11]);
+        }
+    }
+    else
+    {
+        std::cerr
+            << "arg1: verification (0=no, 1=yes)" << std::endl
+            << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)" << std::endl
+            << "arg3: time kernel (0=no, 1=yes)" << std::endl
+            << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
+            << "arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
+            << "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
 template <>
 bool parse_cmd_args<ProblemSizeStreamK>(int argc,
                                         char* argv[],
@@ -165,7 +229,8 @@ bool parse_cmd_args<ProblemSizeStreamK>(int argc,
                   << std::endl
                   << "arg3: time kernel (0=no, 1=yes)" << std::endl
                   << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
-                  << "arg10: NumSKBlocks(optional)" << std::endl;
+                  << "arg10: stream-k select (0: all DP, 1: 1-tile SK, 2: 2-tile SK)"
+                  << "\narg11: Grid_size(-1 for max occupancy)" << std::endl;
         return false;
     }
 

example/01_gemm/gemm_xdl_fp16_streamk_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmV2_Streamk_Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_Streamk_V3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType,   BDataType,  CDataType,  AccDataType,  CShuffleDataType, 
+        PassThrough, PassThrough, PassThrough, GemmDefault, 
+        256,
+        224, 256, 
+        64, 8, 2,
+        16,   16,
+        7,    8,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        S<8, 32, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
+        1, 8, 2, 0,
+        1, 2, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example_streamk_v2.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_universal_streamk_example(argc, argv); }

example/01_gemm/run_gemm_example_streamk_v2.inc

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename ProblemType>
+bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    using namespace ck::literals;
+
+    auto M           = problem_size.M;
+    auto N           = problem_size.N;
+    auto K           = problem_size.K;
+    auto StrideA     = problem_size.StrideA;
+    auto StrideB     = problem_size.StrideB;
+    auto StrideC     = problem_size.StrideC;
+    auto Grid_size   = problem_size.Grid_size;
+    auto Streamk_sel = problem_size.Streamk_sel;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
+            if(stride == -1)
+            {
+                // give a chance if stride is -1, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return static_cast<std::size_t>(col);
+                }
+                else
+                {
+                    return static_cast<std::size_t>(row);
+                }
+            }
+            else
+                return static_cast<std::size_t>(stride);
+        };
+
+    auto f_get_default_streamk_policy = [](ck::index_t streamk_sel) {
+        if(streamk_sel == -1)
+        {
+            return static_cast<std::size_t>(4);
+        }
+        else
+            return static_cast<std::size_t>(streamk_sel);
+    };
+
+    StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
+
+    Streamk_sel = f_get_default_streamk_policy(Streamk_sel);
+
+    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{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{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});
+    }
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
+                               c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+    DeviceMem workspace;
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm      = DeviceGemmV2_Streamk_Instance{};
+    auto invoker   = gemm.MakeInvoker();
+    float ave_time = 0;
+
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#endif
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        StrideC,
+        Streamk_sel,
+        Grid_size,
+        a_element_op,
+        b_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return true;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 1});
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
+
+        c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
+
+        c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
+
+        return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
+#else
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        pass &= ck::utils::check_err(c_m_n_device_result,
+                                     c_m_n_host_result,
+                                     "Error: Incorrect results!",
+                                     get_rtol<CDataType>(),
+                                     get_atol<CDataType>());
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop = 2_uz * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+    return pass;
+}
+
+bool run_gemm_universal_streamk_example(int argc, char* argv[])
+{
+    ProblemSizeStreamK_universal problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}

include/ck/tensor_operation/gpu/device/device_gemm_streamk_v2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemm_Streamk_V2 : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t Streamk_sel,
+                        ck::index_t Grid_size,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp

@@ -1404,4 +1404,326 @@ struct BlockToCTileMap_GemmStreamK
     }
 };
 
+template <uint32_t MPerBlock_,
+          uint32_t NPerBlock_,
+          uint32_t KPerBlock_,
+          StreamKReductionStrategy ReductionStrategy_ = StreamKReductionStrategy::Atomic,
+          uint32_t TileSwizzleSubM_                   = 8,
+          index_t GroupNum                            = 8,
+          index_t M01_                                = 4>
+struct BlockToCTileMap_GemmStreamK_v2
+{
+    static constexpr uint32_t min_k_iters_per_sk_block          = 2;
+    static constexpr uint32_t MPerBlock                         = MPerBlock_;
+    static constexpr uint32_t NPerBlock                         = NPerBlock_;
+    static constexpr uint32_t KPerBlock                         = KPerBlock_;
+    static constexpr StreamKReductionStrategy ReductionStrategy = ReductionStrategy_;
+    static constexpr uint32_t tile_swizzle_sub_m                = TileSwizzleSubM_;
+
+    //--------------------------------------
+    // pass to device
+    mutable uint32_t sk_num_blocks;
+    uint32_t sk_num_big_blocks;
+    uint32_t dp_start_block_idx;
+    uint32_t reduction_start_block_idx;
+    uint32_t k_iters_per_big_block;
+    MDiv2 n_tiles;
+    MDiv k_iters_per_tile;
+    MDiv equiv_tiles_big;    // for reduction
+    MDiv equiv_tiles_little; // for reduction
+
+    // prefer construct on host
+    __host__ __device__ BlockToCTileMap_GemmStreamK_v2(
+        uint32_t m, uint32_t n, uint32_t k, uint32_t grid_size = 1, uint32_t streamk_sel = 1)
+    {
+        // total output tiles
+        uint32_t num_tiles =
+            math::integer_divide_ceil(m, MPerBlock) * math::integer_divide_ceil(n, NPerBlock);
+        k_iters_per_tile = MDiv(math::integer_divide_ceil(k, KPerBlock));
+
+        uint32_t dp_tiles, dp_num_blocks, sk_total_iters;
+
+        // default to regular DP GEMM if sk blocks == 0
+        if(streamk_sel == 0)
+        {
+            sk_num_blocks         = 0;
+            dp_tiles              = num_tiles;
+            sk_num_big_blocks     = 0;
+            k_iters_per_big_block = 0;
+
+            dp_num_blocks      = num_tiles; // all tile to be dp block
+            dp_start_block_idx = 0;
+            sk_total_iters     = 0; // clear this tiles
+        }
+        // 2-tile sk + DP GEMM
+        else
+        {
+
+            // check if there's enough work for DP+ stream-k
+            bool bigEnough = num_tiles > grid_size;
+            // select between stream-k strategies
+            uint32_t sk_tiles = 0;
+            if(streamk_sel == 1) // 1 tile stream-k
+            {
+                sk_tiles = bigEnough ? (num_tiles % grid_size) : num_tiles;
+            }
+            else if(streamk_sel == 2) // 2-tile stream-k
+            {
+                sk_tiles = bigEnough ? (grid_size + num_tiles % grid_size) : num_tiles;
+            }
+            else if(streamk_sel == 3) // 3-tile stream-k
+            {
+                sk_tiles = (num_tiles > (2 * grid_size)) ? (2 * grid_size + num_tiles % grid_size)
+                                                         : num_tiles;
+            }
+            else if(streamk_sel == 4) // 4-tile stream-k
+            {
+                sk_tiles = (num_tiles > (3 * grid_size)) ? (3 * grid_size + num_tiles % grid_size)
+                                                         : num_tiles;
+            }
+            sk_num_blocks = sk_tiles;
+            // remaining tiles are DP tiles
+            dp_tiles = bigEnough ? (num_tiles - sk_tiles) : 0;
+
+            sk_total_iters = k_iters_per_tile.get() * sk_tiles;
+
+            // k_iters_per_sk_block is the floor of avg each ck block loop over tiles.
+            // we need to decide how many iters for each sk block
+            // let m = k_iters_per_sk_block
+            // some of the sk block (little) will cover m iters, some (big) will cover m+1
+            // we have
+            // 1) l + b = sk_blocks
+            // 2) l * m + b * (m + 1) = sk_total_iters
+            //      => (l + b) * m + b = sk_total_iters
+            //      => sk_blocks * m + b = sk_total_iters
+            //      => b = sk_total_iters - m * sk_blocks
+            //      NOTE: big could be zero
+            uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
+            sk_num_big_blocks             = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
+            k_iters_per_big_block         = k_iters_per_sk_block + 1;
+
+            dp_num_blocks      = dp_tiles;
+            dp_start_block_idx = sk_num_blocks;
+        }
+
+        n_tiles = MDiv2(math::integer_divide_ceil(n, NPerBlock));
+        // using multiple blocks for parallel reduction
+        reduction_start_block_idx = dp_start_block_idx + dp_num_blocks;
+
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            uint32_t upper_big    = math::lcm(k_iters_per_big_block, k_iters_per_tile.get());
+            uint32_t upper_little = math::lcm(k_iters_per_big_block - 1, k_iters_per_tile.get());
+            equiv_tiles_big       = MDiv(upper_big / k_iters_per_tile.get());
+            equiv_tiles_little    = MDiv(upper_little / k_iters_per_tile.get());
+        }
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+        return M0 * N0;
+    }
+    __host__ __device__ uint32_t get_sk_total_iters() const
+    {
+        uint32_t sk_total_iters = sk_num_big_blocks * k_iters_per_big_block +
+                                  (sk_num_blocks - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+        return sk_total_iters;
+    }
+
+    __host__ __device__ uint32_t get_sk_tiles() const
+    {
+        // tiles for sk
+        uint32_t sk_total_iters = get_sk_total_iters();
+        return k_iters_per_tile.div(sk_total_iters);
+    }
+
+    __host__ __device__ index_t get_grid_dims() const
+    {
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            // return dim3(reduction_start_block_idx + get_sk_tiles(), 1, 1);
+            return reduction_start_block_idx + get_sk_tiles();
+        }
+        else
+            return reduction_start_block_idx;
+    }
+
+    __device__ uint32_t get_block_idx() const
+    {
+        // TODO: swizzle block index for better locality
+        return __builtin_amdgcn_readfirstlane(blockIdx.x);
+    }
+
+    __device__ void
+    get_block_itr(uint32_t block_idx, uint32_t& iter_start, uint32_t& iter_end) const
+    {
+        if(block_idx < sk_num_big_blocks)
+        {
+            iter_start = block_idx * k_iters_per_big_block;
+            iter_end   = iter_start + k_iters_per_big_block;
+        }
+        else if(block_idx < sk_num_blocks)
+        {
+            iter_start = (sk_num_big_blocks * k_iters_per_big_block) +
+                         (block_idx - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+            iter_end = iter_start + (k_iters_per_big_block - 1);
+        }
+        else if(block_idx >= dp_start_block_idx)
+        {
+            uint32_t sk_total_iters     = get_sk_total_iters();
+            uint32_t dp_iters_per_block = k_iters_per_tile.get();
+            iter_start = sk_total_iters + (block_idx - dp_start_block_idx) * dp_iters_per_block;
+            iter_end   = iter_start + dp_iters_per_block;
+        }
+    }
+
+    __device__ uint32_t get_current_iter_length(uint32_t iter_start,
+                                                uint32_t iter_end,
+                                                uint32_t total_iter_length) const
+    {
+        uint32_t iter_length_mod, iter_length_quo /*unused*/;
+        k_iters_per_tile.divmod(iter_end, iter_length_quo, iter_length_mod);
+        uint32_t current_iter_length = math::min(
+            iter_length_mod == 0 ? (iter_end - iter_start) : iter_length_mod, total_iter_length);
+        return current_iter_length;
+    }
+
+    __device__ uint32_t get_tile_idx(uint32_t iter) const { return k_iters_per_tile.div(iter); }
+
+    __device__ void
+    get_tile_idx_with_offset(uint32_t iter, uint32_t& tile_idx, uint32_t& iter_offset) const
+    {
+        k_iters_per_tile.divmod(iter, tile_idx, iter_offset);
+    }
+
+    __device__ auto tile_to_spatial(uint32_t tile_idx, uint32_t m, uint32_t n) const
+    {
+        uint32_t m_tile_idx, n_tile_idx;
+        uint32_t n_tiles_value = math::integer_divide_ceil(n, NPerBlock);
+        n_tiles.divmod(tile_idx, n_tiles_value, m_tile_idx, n_tile_idx);
+
+        // // swizzle tile
+        uint32_t m_tiles = math::integer_divide_ceil(m, MPerBlock);
+
+        uint32_t tile_swizzle_sub_m_rem = m_tiles % tile_swizzle_sub_m;
+
+        const auto sub_m_adapt = (m_tile_idx < (m_tiles - tile_swizzle_sub_m_rem))
+                                     ? tile_swizzle_sub_m
+                                     : tile_swizzle_sub_m_rem;
+
+        uint32_t m_tile_idx_sub0, m_tile_idx_sub1;
+        m_tile_idx_sub0 = m_tile_idx / tile_swizzle_sub_m;
+        m_tile_idx_sub1 = m_tile_idx % tile_swizzle_sub_m;
+
+        uint32_t tile_idx_local = n_tile_idx + m_tile_idx_sub1 * n_tiles_value;
+
+        uint32_t m_tile_idx_with_adapt, n_tile_idx_with_adapt;
+
+        n_tile_idx_with_adapt = tile_idx_local / sub_m_adapt;
+        m_tile_idx_with_adapt = tile_idx_local % sub_m_adapt;
+        return make_tuple(m_tile_idx_with_adapt + m_tile_idx_sub0 * tile_swizzle_sub_m,
+                          n_tile_idx_with_adapt);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_acc(uint32_t acc_element_bytes) const
+    {
+        static constexpr uint32_t alignment = 128;
+        uint32_t acc_buffer_bytes =
+            MPerBlock * NPerBlock * get_total_acc_buffers() * acc_element_bytes;
+        return (acc_buffer_bytes + alignment - 1) / alignment * alignment;
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_semaphore() const
+    {
+        return get_sk_tiles() * sizeof(uint32_t);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size(uint32_t acc_element_bytes) const
+    {
+        return get_workspace_size_for_acc(acc_element_bytes) + get_workspace_size_for_semaphore();
+    }
+
+    __host__ __device__ uint32_t get_tile_intersections(uint32_t tiles_,
+                                                        const MDiv& equiv_tiles_) const
+    {
+        uint32_t tile_idx_        = tiles_ == 0 ? 0 : (tiles_ - 1);
+        uint32_t max_equiv_tiles_ = equiv_tiles_.get() - 1;
+        uint32_t quo_, rem_;
+        equiv_tiles_.divmod(tile_idx_, quo_, rem_);
+        return quo_ * max_equiv_tiles_ + rem_;
+    }
+
+    __host__ __device__ uint32_t get_tiles_cover_sk_block(uint32_t num_sk_blocks_,
+                                                          uint32_t iters_per_sk_block_) const
+    {
+        return k_iters_per_tile.div(num_sk_blocks_ * iters_per_sk_block_ + k_iters_per_tile.get() -
+                                    1);
+    }
+
+    __host__ __device__ uint32_t get_total_acc_buffers() const
+    {
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        uint32_t tiles_cover_little_blocks =
+            get_tiles_cover_sk_block(sk_num_blocks - sk_num_big_blocks, k_iters_per_big_block - 1);
+
+        uint32_t total_intersec_big =
+            get_tile_intersections(tiles_cover_big_blocks, equiv_tiles_big);
+        uint32_t total_intersec_little =
+            get_tile_intersections(tiles_cover_little_blocks, equiv_tiles_little);
+
+        return sk_num_blocks + total_intersec_big + total_intersec_little;
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_tile(uint32_t tile_idx_) const
+    {
+        // TODO: from big to little
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        if(tile_idx_ < tiles_cover_big_blocks)
+        {
+            uint32_t touched_sk_blocks =
+                (tile_idx_ * k_iters_per_tile.get() + k_iters_per_big_block - 1) /
+                k_iters_per_big_block;
+            uint32_t current_intersec = get_tile_intersections(tile_idx_, equiv_tiles_big);
+            return touched_sk_blocks + current_intersec;
+        }
+        else
+        {
+            uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+            uint32_t tile_idx_little_reverse   = get_sk_tiles() - tile_idx_;
+            uint32_t touched_sk_blocks =
+                (tile_idx_little_reverse * k_iters_per_tile.get() + iters_per_little_sk_block - 1) /
+                iters_per_little_sk_block;
+            uint32_t current_intersec =
+                get_tile_intersections(tile_idx_little_reverse, equiv_tiles_little);
+            return get_total_acc_buffers() - (touched_sk_blocks + current_intersec);
+        }
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_block(uint32_t block_idx_) const
+    {
+        uint32_t iters_per_big_sk_block    = k_iters_per_big_block;
+        uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+        if(block_idx_ < sk_num_big_blocks)
+        {
+            uint32_t touched_tiles    = k_iters_per_tile.div(block_idx_ * iters_per_big_sk_block +
+                                                          k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, equiv_tiles_big);
+            return block_idx_ + current_intersec;
+        }
+        else
+        {
+            uint32_t block_idx_little_reverse = sk_num_blocks - block_idx_;
+            uint32_t touched_tiles            = k_iters_per_tile.div(
+                block_idx_little_reverse * iters_per_little_sk_block + k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, equiv_tiles_little);
+            return get_total_acc_buffers() - (block_idx_little_reverse + current_intersec);
+        }
+    }
+};
+
 } // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_streamk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+#endif
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemm_Streamk_V2<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemm_Streamk_V2<ALayout,
+                                           BLayout,
+                                           CLayout,
+                                           ADataType,
+                                           BDataType,
+                                           CDataType,
+                                           ck::tensor_operation::element_wise::PassThrough,
+                                           ck::tensor_operation::element_wise::PassThrough,
+                                           ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+                add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/CMakeLists.txt

+# ONLY XDL_KERNELS
+set(GEMM_UNIVERSAL_STREAMK_INSTANCES)
+
+list(APPEND GEMM_UNIVERSAL_STREAMK_INSTANCES 
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_default_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+        device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instance.cpp)
+
+add_instance_library(device_gemm_universal_streamk_instance ${GEMM_UNIVERSAL_STREAMK_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances<GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances<GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances<GemmMNKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances<GemmMNPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave,
+                                                                             GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave,
+                                                                             GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemm_Streamk_V2<Row,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances<Intrawave,
+                                                                             GemmMNKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck