Grouped Gemm + SplitK + simplified Kernel Args (#669)

* simplify karg in device/grid split-k op

* fix mk_kn_mn instances

* add more instances

* B2C with 3D grid for KSplit

* Remove unused code.

* Use default B2C (3D grid) in grid gemm v2r4r2.

* Device gemm splitk use B2C map.

* Device GroupedGemmXdlSplitKCShuffle

* Example for GroupedGemm Xdl SplitK

* Introduce Device GroupedGemmSplitK

* Fix updating kbatch size.

* Add instance mk-nk-mn

* Enable set kbatch in profiler.

* Add GGemmSplitK mk-kn-mn instances

* Add more instances & split into multiple files.

* minor fix

* tuning

* clean

* disabled failed instances

* use pipe v2

* Ignore arg on not supported arch.

* fix warning

---------
Co-authored-by: carlushuang <carlus.huang@amd.com>
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: zjing14 <zhangjing14@gmail.com>
Co-authored-by: Jing Zhang <jizhan@amd.com>
Co-authored-by: root <root@ctr-ubbsmc15.amd.com>

example/15_grouped_gemm/CMakeLists.txt

@@ -5,6 +5,7 @@ add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
 add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
 add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
 add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
+add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
 
 
 add_dependencies(example_grouped_gemm_xdl
@@ -12,7 +13,8 @@ add_dependencies(example_grouped_gemm_xdl
                  example_grouped_gemm_xdl_fp16
                  example_grouped_gemm_xdl_bfp16
                  example_grouped_gemm_xdl_int8
-                 example_grouped_gemm_multiple_d_dl_fp16)
+                 example_grouped_gemm_multiple_d_dl_fp16
+                 example_grouped_gemm_xdl_splitk_fp16)
 
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)

example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_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        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdlSplitKCShuffle
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    problem_size.Ms = {
+        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ns.push_back(768);
+        problem_size.Ks.push_back(4608);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    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
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/15_grouped_gemm/run_grouped_gemm_example.inc

@@ -147,6 +147,7 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
 #else
         a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
         b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+	c_tensors_device[i]->SetZero();
 #endif
 
         p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());

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

@@ -31,7 +31,7 @@ struct DeviceGroupedGemm : public BaseOperator
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
-    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsistent NumDTensor");
 
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(std::vector<const void*>& p_a,

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

+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemmSplitK : public DeviceGroupedGemm<ALayout,
+                                                          BLayout,
+                                                          DsLayout,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          DsDataType,
+                                                          EDataType,
+                                                          AElementwiseOperation,
+                                                          BElementwiseOperation,
+                                                          CElementwiseOperation>
+{
+    virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -114,7 +114,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
         CBlockTransferScalarPerVector_NWaveNPerXDL,
         CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
 
-    using Argument = typename GridwiseGemm::Argument;
+    using Argument              = typename GridwiseGemm::Argument;
+    using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
 
     // Invoker
     struct Invoker : public BaseInvoker
@@ -138,8 +139,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     "setting");
             }
 
+            const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(karg);
+            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0           = karg.K0;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
@@ -152,7 +154,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                         hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType)));
 
                 ave_time = launch_and_time_kernel(
-                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
             };
 
             if(has_main_k0_block_loop)
@@ -162,7 +164,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              true,
-                                                             InMemoryDataOperationEnum::Set>;
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -171,7 +174,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              true,
-                                                             InMemoryDataOperationEnum::AtomicAdd>;
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -183,7 +187,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              false,
-                                                             InMemoryDataOperationEnum::Set>;
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -192,7 +197,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              false,
-                                                             InMemoryDataOperationEnum::AtomicAdd>;
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/ck.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl_splitk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                       const index_t group_count)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+    __shared__ uint8_t p_shared[shared_size];
+
+    const index_t block_id = get_block_1d_id();
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].block_start_ &&
+             block_id < gemm_desc_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
+        gemm_desc_ptr[group_id].karg_,
+        static_cast<void*>(p_shared),
+        gemm_desc_ptr[group_id].block_2_ctile_map_);
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler(),
+          // Current implementation does not support multiple D fusions.
+          enable_if_t<AK1 == BK1 && is_same_v<DsLayout, ck::Tuple<>> &&
+                          is_same_v<DsDataType, ck::Tuple<>>,
+                      bool> = false>
+struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayout,
+                                                                           BLayout,
+                                                                           DsLayout,
+                                                                           ELayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           DsDataType,
+                                                                           EDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CDEElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static_assert(KPerBlock % AK1 == 0);
+    static constexpr index_t K0PerBlock = KPerBlock / AK1;
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        EDataType,
+        ALayout,
+        BLayout,
+        ELayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        GemmSpec,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        AK1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
+    using Block2ETileMapKSplit =
+        BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>;
+    // Block2CTileMap configuration parameter.
+    static constexpr index_t B2E_M01 = 8;
+    using GroupedGemmBlock2ETileMap  = OffsettedBlockToCTileMap<Block2ETileMapKSplit>;
+    using KernelArgument             = typename GridwiseGemm::Argument;
+
+    struct GemmTransKernelArg
+    {
+        KernelArgument karg_;
+        GroupedGemmBlock2ETileMap block_2_ctile_map_;
+        index_t block_start_, block_end_;
+
+        GemmTransKernelArg() = default;
+        GemmTransKernelArg(KernelArgument&& karg,
+                           GroupedGemmBlock2ETileMap&& b2c_map,
+                           index_t block_start,
+                           index_t block_end)
+            : karg_{karg},
+              block_2_ctile_map_{b2c_map},
+              block_start_{block_start},
+              block_end_{block_end}
+        {
+        }
+    };
+
+    static constexpr index_t DefaultKBatch = 1;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs)
+            : Argument(p_As, p_Bs, p_Es, gemm_descs, DefaultKBatch)
+        {
+            // TODO: use occupancy api to calculate appropriate batch size.
+        }
+
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 index_t kbatch)
+            : K_BATCH{kbatch}
+        {
+            grid_size_   = 0;
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_kernel_args_.reserve(group_count_);
+
+            skipped_group_count_ = 0;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); ++i)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                if(M == 0)
+                {
+                    skipped_group_count_++;
+                    continue;
+                }
+
+                const index_t stride_a = gemm_descs[i].stride_A_;
+                const index_t stride_b = gemm_descs[i].stride_B_;
+                const index_t stride_c = gemm_descs[i].stride_C_;
+
+                const index_t m_padded = GridwiseGemm::CalculateMPadded(M);
+                const index_t n_padded = GridwiseGemm::CalculateNPadded(N);
+                const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
+                const index_t k0       = GridwiseGemm::CalculateK0(K, K_BATCH);
+
+                const auto c_grid_desc_m_n =
+                    GridwiseGemm::MakeCGridDescriptor_M_N(M, N, m_padded, n_padded, stride_c);
+
+                const auto local_b2c_tile_map =
+                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+                const index_t block_start = grid_size_;
+                const index_t block_end   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                auto grouped_block_2_ctile_map =
+                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
+
+                auto karg = KernelArgument{type_convert<const ADataType*>(p_As[i]),
+                                           type_convert<const BDataType*>(p_Bs[i]),
+                                           type_convert<EDataType*>(p_Es[i]),
+                                           M,
+                                           N,
+                                           K,
+                                           stride_a,
+                                           stride_b,
+                                           stride_c,
+                                           m_padded,
+                                           n_padded,
+                                           k_padded,
+                                           k0,
+                                           K_BATCH};
+
+                gemm_kernel_args_.emplace_back(
+                    std::move(karg), std::move(grouped_block_2_ctile_map), block_start, block_end);
+            }
+        }
+
+        /**
+         * @brief      Recalculate group grid size for all gemms and update B2C maps.
+         *
+         * @param[in]  kbatch  The new splitK parameter value.
+         */
+        void UpdateKBatch(index_t kbatch)
+        {
+            K_BATCH    = kbatch;
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < gemm_kernel_args_.size(); ++i)
+            {
+
+                auto& karg = gemm_kernel_args_[i].karg_;
+
+                const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
+                const index_t k0       = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
+
+                const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(
+                    karg.M, karg.N, karg.MPadded, karg.NPadded, karg.StrideC);
+
+                const auto local_b2c_tile_map =
+                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+                const index_t block_start = grid_size_;
+                const index_t block_end   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                auto grouped_block_2_ctile_map =
+                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
+
+                karg.KPadded                            = k_padded;
+                karg.K0                                 = k0;
+                karg.k_batch                            = K_BATCH;
+                gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
+                gemm_kernel_args_[i].block_start_       = block_start;
+                gemm_kernel_args_[i].block_end_         = block_end;
+            }
+        }
+
+        //  private:
+        index_t K_BATCH;
+        index_t group_count_;
+        index_t skipped_group_count_;
+
+        std::vector<GemmTransKernelArg> gemm_kernel_args_;
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0;
+            bool all_have_kbatch_gt_one      = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
+            bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
+            {
+                const auto& karg = arg.gemm_kernel_args_[i].karg_;
+                if(stream_config.log_level_ > 0)
+                {
+                    karg.Print();
+                }
+
+                auto kbatch = karg.k_batch;
+
+                if(!GridwiseGemm::CheckValidity(karg))
+                {
+                    std::ostringstream err;
+                    err << "Group id: " << i << " has invalid GridwiseGemm settings!" << __FILE__
+                        << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+
+                K0 = karg.K0;
+                bool not_all_have_main_k0_block_loop_same =
+                    all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+                bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);
+
+                if(not_all_have_main_k0_block_loop_same)
+                {
+                    std::ostringstream err;
+                    err << "Not all gemms have same value for main_k0_block_loop! in " << __FILE__
+                        << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+
+                if(not_all_have_kbatch_value_same)
+                {
+                    std::ostringstream err;
+                    err << "Not all gemms have same kbatch value (=1 or >1)! "
+                        << "group [" << i << "], kbatch: " << kbatch
+                        << ", group [0], kbatch: " << arg.gemm_kernel_args_[0].karg_.k_batch
+                        << " in " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+            }
+
+            hip_check_error(hipMemcpy(arg.p_workspace_,
+                                      arg.gemm_kernel_args_.data(),
+                                      arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
+                                      hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                if(all_have_kbatch_gt_one)
+                {
+                    for(const auto& trans_arg : arg.gemm_kernel_args_)
+                    {
+                        const auto& karg = trans_arg.karg_;
+                        hip_check_error(
+                            hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(EDataType)));
+                    }
+                }
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(arg.grid_size_),
+                                           dim3(BlockSize),
+                                           0,
+                                           cast_pointer_to_constant_address_space(arg.p_workspace_),
+                                           arg.gemm_kernel_args_.size());
+            };
+
+            if(all_have_main_k0_block_loop)
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       true,
+                                                       InMemoryDataOperationEnum::AtomicAdd>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       true,
+                                                       InMemoryDataOperationEnum::Set>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       false,
+                                                       InMemoryDataOperationEnum::AtomicAdd>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       false,
+                                                       InMemoryDataOperationEnum::Set>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if((ck::type_convert<ck::index_t>(arg.gemm_kernel_args_.size()) +
+            arg.skipped_group_count_) != arg.group_count_)
+        {
+            return false;
+        }
+
+        bool supported = true;
+        for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
+        {
+            const auto& a        = arg.gemm_kernel_args_[i].karg_;
+            bool group_arg_valid = GridwiseGemm::CheckValidity(a);
+#if DEBUG_LOG
+            if(not group_arg_valid)
+            {
+                std::cout << "[" << __func__ << "] group id: " << i << " is not supported!\n";
+                a.Print();
+            }
+#endif // DEBUG_LOG
+            supported &= group_arg_valid;
+        }
+        return supported;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>&,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CDEElementwiseOperation)
+    {
+        return Argument{p_As, p_Bs, p_Es, gemm_descs};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>&,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CDEElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(p_As, p_Bs, p_Es, gemm_descs);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_XdlSplitK"
+            << "<"
+            << std::string(ALayout::name)[0] << ","
+            << std::string(BLayout::name)[0] << ","
+            << std::string(ELayout::name)[0] << ","
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->gemm_kernel_args_.size() *
+               sizeof(GemmTransKernelArg);
+    }
+
+    static void SetKBatchSize(Argument& arg, index_t kbatch) { arg.UpdateKBatch(kbatch); }
+
+    // polymorphic
+    void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const override
+    {
+        return SetKBatchSize(*dynamic_cast<Argument*>(p_arg), kbatch);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -587,4 +587,52 @@ struct OffsettedBlockToCTileMap
     index_t block_start_;
 };
 
+/**
+ * @brief      Simple tile mapping which creates 3D grid of block of threads.
+ *
+ * @paragraph  Description
+ *             This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
+ *             blocks. The first 2D are regular 2D tiles created by division of output GEMM
+ *             dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
+ *             which denotes the number of blocks we use to divide work on GEMM K dimension onto.
+ *
+ * @tparam     MPerBlock  Output block tile size in M dimension.
+ * @tparam     NPerBlock  Output block tile size in N dimension.
+ */
+template <index_t MPerBlock, index_t NPerBlock>
+struct BlockToCTileMap_3DGrid_KSplit
+{
+
+    __host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
+
+    __host__ __device__ constexpr auto
+    CalculateGridSize(index_t M, index_t N, index_t k_split) const
+    {
+        // Create 3D grid
+        const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+        return std::make_tuple(N0, M0, k_split);
+    }
+
+    template <typename TopIdx>
+    __device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
+    {
+        return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+    {
+        return true;
+    }
+};
+
 } // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp

@@ -15,16 +15,20 @@
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+
 namespace ck {
 
 template <typename GridwiseGemm,
           bool HasMainKBlockLoop,
-          InMemoryDataOperationEnum CGlobalMemoryDataOperation>
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename Block2CTileMap>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg)
+        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
+                                             const Block2CTileMap& b2c_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
@@ -32,9 +36,10 @@ __global__ void
     __shared__ uint8_t p_shared[shared_size];
 
     GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
-        karg, static_cast<void*>(p_shared));
+        karg, static_cast<void*>(p_shared), b2c_map);
 #else
     ignore = karg;
+    ignore = b2c_map;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
 
@@ -478,8 +483,21 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                        Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
     }
 
-    template <bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation>
-    __device__ static void Run(const Argument& karg, void* __restrict__ p_shared_block)
+    // return block_id to C matrix tile idx (m0, n0, k_split) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap()
+    {
+        return BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>();
+    }
+
+    using CGridDesc_M_N         = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
+    using DefaultBlock2CTileMap = remove_cvref_t<decltype(MakeDefaultBlock2CTileMap())>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              typename Block2CTileMap>
+    __device__ static void Run(const Argument& karg,
+                               void* __restrict__ p_shared_block,
+                               const Block2CTileMap& block_2_ctile_map)
     {
         const FloatAB* p_a_grid          = karg.p_a_grid;
         const FloatAB* p_b_grid          = karg.p_b_grid;
@@ -504,11 +522,21 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
-        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        // divide block work by [KBatch, M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
 
-        const index_t block_m_id = __builtin_amdgcn_readfirstlane(blockIdx.y);
-        const index_t block_n_id = __builtin_amdgcn_readfirstlane(blockIdx.x);
-        const index_t k_batch_id = __builtin_amdgcn_readfirstlane(blockIdx.z);
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
+        const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
 
         // HACK: this force m/n_block_data_idx_on_grid into SGPR
         const index_t m_block_data_idx_on_grid =
@@ -651,6 +679,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
+#if 1
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                 FloatAB,
@@ -662,6 +691,20 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                                                 MRepeat,
                                                                 NRepeat,
                                                                 K1>{};
+#else
+        auto blockwise_gemm = BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_k0_m_k1_block_desc),
+            decltype(b_k0_n_k1_block_desc),
+            MPerXDL,
+            NPerXDL,
+            MRepeat,
+            NRepeat,
+            K1>{};
+
+#endif
 
         auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
 
@@ -680,6 +723,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
 
+#if 0
         // preload data into LDS
         {
             a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
@@ -725,6 +769,31 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
 
             blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
         }
+#else
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVersion::v2, 1, LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
+            (K0PerBlock * K1));
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
+                                                               a_b_k0_m_k1_block_desc,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_b_k0_n_k1_grid_desc,
+                                                               b_b_k0_n_k1_block_desc,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+#endif
 
         // output: register to global memory
         {

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

@@ -68,6 +68,58 @@ void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
                                                   PassThrough,
                                                   PassThrough>>>& instances);
 
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
 template <typename ALayout,
           typename BLayout,
           typename ELayout,
@@ -109,11 +161,17 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                          is_same_v<ELayout, Row>)
             {
                 add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+                    op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
                               is_same_v<ELayout, Row>)
             {
                 add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+                    op_ptrs);
             }
             else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
                               is_same_v<ELayout, Row>)

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

@@ -3,4 +3,8 @@ add_instance_library(device_grouped_gemm_instance
    device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
    device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
    device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
 )

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // Currently AK1 must equal BK1 !
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // Currently AK1 must equal BK1 !
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+//      DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+//      DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+//      DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                 = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+//      DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+//      DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_grouped_gemm_impl.hpp

@@ -8,6 +8,7 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
@@ -39,7 +40,8 @@ bool profile_grouped_gemm_impl(int do_verification,
                                const std::vector<int>& Ks,
                                const std::vector<int>& StrideAs,
                                const std::vector<int>& StrideBs,
-                               const std::vector<int>& StrideCs)
+                               const std::vector<int>& StrideCs,
+                               int kbatch = 1)
 {
 
     bool pass = true;
@@ -96,8 +98,6 @@ bool profile_grouped_gemm_impl(int do_verification,
             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);
         }
-
-        c_m_n_device_results[i].GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
     }
 
     using AElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -132,13 +132,12 @@ bool profile_grouped_gemm_impl(int do_verification,
             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());
-        c_device_buf[i]->ToDevice(c_m_n_device_results[i].mData.data());
+        c_device_buf[i]->SetZero();
 
         gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
 
@@ -197,6 +196,28 @@ bool profile_grouped_gemm_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
+            if(kbatch > 1)
+            {
+                using DeviceOpSplitK =
+                    ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
+                                                                          BLayout,
+                                                                          ck::Tuple<>,
+                                                                          CLayout,
+                                                                          ADataType,
+                                                                          BDataType,
+                                                                          ck::Tuple<>,
+                                                                          CDataType,
+                                                                          AElementOp,
+                                                                          BElementOp,
+                                                                          CElementOp>;
+
+                if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) != nullptr)
+                {
+                    dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
+                        ->SetKBatchSize(argument_ptr.get(), kbatch);
+                }
+            }
+
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 

profiler/src/profile_grouped_gemm.cpp

@@ -52,20 +52,24 @@ std::vector<int> argToIntArray(char* input)
 
 int profile_grouped_gemm(int argc, char* argv[])
 {
-    if(!(argc == 14))
+    if(argc < 14)
     {
-        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
-        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
-        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
-        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
-        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
-        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
-        printf("arg4: verification (0: no; 1: yes)\n");
-        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
-        printf("arg6: print tensor value (0: no; 1: yes)\n");
-        printf("arg7: time kernel (0=n0, 1=yes)\n");
-        printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
-               "64,64 64,64 128,128)\n");
+        std::cout
+            << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+            << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: 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"
+            << "                     2: A[k, m] * B[k, n] = C[m, n];\n"
+            << "                     3: A[k, m] * 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 4)\n"
+            << std::endl;
+
         exit(1);
     }
 
@@ -83,6 +87,7 @@ int profile_grouped_gemm(int argc, char* argv[])
     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;
 
     if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
@@ -101,7 +106,8 @@ int profile_grouped_gemm(int argc, char* argv[])
                                                                                    Ks,
                                                                                    StrideAs,
                                                                                    StrideBs,
-                                                                                   StrideCs);
+                                                                                   StrideCs,
+                                                                                   kbatch);
     }
     else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
     {
@@ -120,7 +126,8 @@ int profile_grouped_gemm(int argc, char* argv[])
                                                                                    Ks,
                                                                                    StrideAs,
                                                                                    StrideBs,
-                                                                                   StrideCs);
+                                                                                   StrideCs,
+                                                                                   kbatch);
     }
     else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
     {
@@ -139,7 +146,8 @@ int profile_grouped_gemm(int argc, char* argv[])
                                                                                    Ks,
                                                                                    StrideAs,
                                                                                    StrideBs,
-                                                                                   StrideCs);
+                                                                                   StrideCs,
+                                                                                   kbatch);
     }
     else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
     {
@@ -158,7 +166,8 @@ int profile_grouped_gemm(int argc, char* argv[])
                                                                                    Ks,
                                                                                    StrideAs,
                                                                                    StrideBs,
-                                                                                   StrideCs);
+                                                                                   StrideCs,
+                                                                                   kbatch);
     }
     else
     {