init of grouped_gemm

composable_kernel/include/config.hpp

@@ -171,5 +171,12 @@ enum ActivTypeEnum_t
 using index_t      = int32_t;
 using long_index_t = int64_t;
 
+struct gemm_desc
+{
+    ck::index_t M, N, K;
+    ck::index_t StrideA, StrideB, StrideC;
+    ck::index_t OffsetA, OffsetB, OffsetC;
+};
+
 } // namespace ck
 #endif

composable_kernel/include/tensor_operation/gridwise_grouped_gemm_xdlops_v2r3.hpp

+#ifndef CK_GRIDWISE_GROUPED_GEMM_XDLOPS_V2R3_HPP
+#define CK_GRIDWISE_GROUPED_GEMM_XDLOPS_V2R3_HPP
+
+#include "common_header.hpp"
+#include "multi_index_transform_helper.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "blockwise_gemm_xdlops.hpp"
+#include "blockwise_tensor_slice_transfer_v4r1.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "gridwise_gemm_pipeline_v1.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const index_t group_id = 0;
+
+    constexpr auto I0 = Number<0>{};
+
+    if(group_id == 0)
+        GridwiseGemm::template Run<HasMainK0BlockLoop>(
+            p_a_grid,
+            p_b_grid,
+            p_c_grid,
+            p_shared,
+            a_grid_desc_k0_m_k1[Number<0>{}],
+            b_grid_desc_k0_n_k1[Number<0>{}],
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2[Number<0>{}],
+            a_element_op,
+            b_element_op,
+            c_element_op,
+            block_2_ctile_map[Number<0>{}],
+            block_id);
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          index_t NumPrefetch = 1>
+struct GridwiseGroupedGemm_k0mk1_k0nk1_mn_xdlops_v2r3
+{
+    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 constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned + b_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check NumPrefetch
+        if constexpr(NumPrefetch == 1)
+        {
+            // 1-stage prefetch always supported
+        }
+        else if constexpr(NumPrefetch == 2)
+        {
+            // 2-stage prefetch currently only support even number of K0 loop
+            // TODO: add support for odd number of K0 loop
+            if(!((K0 / K0PerBlock) % 2 == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // TODO move this function into GEMM-pipeline class
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = (K0 / (NumPrefetch * K0PerBlock)) > 1;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map,
+        const index_t block_id)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(block_id));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4r1<BlockSize,
+                                              AElementwiseOperation,
+                                              ck::tensor_operation::element_wise::PassThrough,
+                                              InMemoryDataOperationEnum_t::Set,
+                                              Sequence<K0PerBlock, MPerBlock, K1>,
+                                              ABlockTransferThreadClusterLengths_K0_M_K1,
+                                              ABlockTransferThreadClusterArrangeOrder,
+                                              FloatAB,
+                                              FloatAB,
+                                              decltype(a_grid_desc_k0_m_k1),
+                                              decltype(a_block_desc_k0_m_k1),
+                                              ABlockTransferSrcAccessOrder,
+                                              Sequence<1, 0, 2>,
+                                              ABlockTransferSrcVectorDim,
+                                              2,
+                                              ABlockTransferSrcScalarPerVector,
+                                              ABlockTransferDstScalarPerVector_K1,
+                                              1,
+                                              1,
+                                              AThreadTransferSrcResetCoordinateAfterRun,
+                                              true,
+                                              NumPrefetch>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            BlockwiseTensorSliceTransfer_v4r1<BlockSize,
+                                              BElementwiseOperation,
+                                              ck::tensor_operation::element_wise::PassThrough,
+                                              InMemoryDataOperationEnum_t::Set,
+                                              Sequence<K0PerBlock, NPerBlock, K1>,
+                                              BBlockTransferThreadClusterLengths_K0_N_K1,
+                                              BBlockTransferThreadClusterArrangeOrder,
+                                              FloatAB,
+                                              FloatAB,
+                                              decltype(b_grid_desc_k0_n_k1),
+                                              decltype(b_block_desc_k0_n_k1),
+                                              BBlockTransferSrcAccessOrder,
+                                              Sequence<1, 0, 2>,
+                                              BBlockTransferSrcVectorDim,
+                                              2,
+                                              BBlockTransferSrcScalarPerVector,
+                                              BBlockTransferDstScalarPerVector_K1,
+                                              1,
+                                              1,
+                                              BThreadTransferSrcResetCoordinateAfterRun,
+                                              true,
+                                              NumPrefetch>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1<remove_cvref_t<decltype(a_grid_desc_k0_m_k1)>,
+                                    remove_cvref_t<decltype(a_block_desc_k0_m_k1)>,
+                                    remove_cvref_t<decltype(a_blockwise_copy)>,
+                                    remove_cvref_t<decltype(a_grid_buf)>,
+                                    remove_cvref_t<decltype(a_block_buf)>,
+                                    remove_cvref_t<decltype(a_block_slice_copy_step)>,
+                                    remove_cvref_t<decltype(b_grid_desc_k0_n_k1)>,
+                                    remove_cvref_t<decltype(b_block_desc_k0_n_k1)>,
+                                    remove_cvref_t<decltype(b_blockwise_copy)>,
+                                    remove_cvref_t<decltype(b_grid_buf)>,
+                                    remove_cvref_t<decltype(b_block_buf)>,
+                                    remove_cvref_t<decltype(b_block_slice_copy_step)>,
+                                    remove_cvref_t<decltype(blockwise_gemm)>,
+                                    remove_cvref_t<decltype(c_thread_buf)>,
+                                    NumPrefetch,
+                                    HasMainK0BlockLoop>{};
+
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        gridwise_gemm_pipeline.Run(a_grid_desc_k0_m_k1,
+                                   a_block_desc_k0_m_k1,
+                                   a_blockwise_copy,
+                                   a_grid_buf,
+                                   a_block_buf,
+                                   a_block_slice_copy_step,
+                                   b_grid_desc_k0_n_k1,
+                                   b_block_desc_k0_n_k1,
+                                   b_blockwise_copy,
+                                   b_grid_buf,
+                                   b_block_buf,
+                                   b_block_slice_copy_step,
+                                   blockwise_gemm,
+                                   c_thread_buf,
+                                   K0BlockMainLoop);
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
+#endif

device_operation/include/device_gemm.hpp

@@ -59,6 +59,23 @@ struct DeviceGemm : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                              const void* p_b,
+                                                              void* p_c,
+                                                              std::vector<gemm_desc> gemm_shapes,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
 template <typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>

device_operation/include/device_grouped_gemm_xdl.hpp

+#ifndef DEVICE_GROUPED_GEMM_XDL_HPP
+#define DEVICE_GROUPED_GEMM_XDL_HPP
+
+#include <iostream>
+#include <sstream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_gemm.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_grouped_gemm_xdlops_v2r3.hpp"
+#include "gemm_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization_t GemmSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          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 ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumPrefetch = 1,
+          ck::index_t GroupCount  = 1>
+struct DeviceGroupedGemmXdl
+    : public DeviceGroupedGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpecialization == GemmSpecialization_t::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGroupedGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        NumPrefetch>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 std::vector<gemm_desc> gemm_shapes,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            const index_t i = 0;
+
+            const index_t M = gemm_shapes[Number<0>{}].M;
+            const index_t N = gemm_shapes[Number<0>{}].N;
+            const index_t K = gemm_shapes[Number<0>{}].K;
+
+            const index_t StrideA = gemm_shapes[Number<0>{}].StrideA;
+            const index_t StrideB = gemm_shapes[Number<0>{}].StrideB;
+            const index_t StrideC = gemm_shapes[Number<0>{}].StrideC;
+
+            a_grid_desc_k0_m_k1_(Number<0>{}) =
+                DeviceGroupedGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_(Number<0>{}) =
+                DeviceGroupedGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_(Number<0>{}) =
+                DeviceGroupedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_[Number<0>{}],
+                                           b_grid_desc_k0_n_k1_[Number<0>{}],
+                                           c_grid_desc_m_n_[Number<0>{}],
+                                           M01_,
+                                           N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_(Number<0>{}) =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(
+                        c_grid_desc_m_n_[Number<0>{}]);
+
+                block_2_ctile_map_(Number<0>{}) = GridwiseGemm::MakeDefaultBlock2CTileMap(
+                    c_grid_desc_m_n_[Number<0>{}], M01, N01);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        StaticallyIndexedArray<AGridDesc_K0_M_K1, GroupCount> a_grid_desc_k0_m_k1_;
+        StaticallyIndexedArray<BGridDesc_K0_N_K1, GroupCount> b_grid_desc_k0_n_k1_;
+        StaticallyIndexedArray<CGridDesc_M_N, GroupCount> c_grid_desc_m_n_;
+        StaticallyIndexedArray<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2, GroupCount>
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        StaticallyIndexedArray<typename GridwiseGemm::DefaultBlock2CTileMap, GroupCount>
+            block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGroupedGemmXdl::Argument;
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            std::cout << "Gemm " << i << ":" << std::endl;
+            std::cout << "arg.a_grid_desc_k0_m_k1_{"
+                      << arg.a_grid_desc_k0_m_k1_[Number<0>{}].GetLength(I0) << ", "
+                      << arg.a_grid_desc_k0_m_k1_[Number<0>{}].GetLength(I1) << ", "
+                      << arg.a_grid_desc_k0_m_k1_[Number<0>{}].GetLength(I2) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_k0_n_k1_{"
+                      << arg.b_grid_desc_k0_n_k1_[Number<0>{}].GetLength(I0) << ", "
+                      << arg.b_grid_desc_k0_n_k1_[Number<0>{}].GetLength(I1) << ", "
+                      << arg.b_grid_desc_k0_n_k1_[Number<0>{}].GetLength(I2) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_[Number<0>{}].GetLength(I0)
+                      << ", " << arg.c_grid_desc_m_n_[Number<0>{}].GetLength(I1) << "}"
+                      << std::endl;
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_[Number<0>{}],
+                                            arg.b_grid_desc_k0_n_k1_[Number<0>{}],
+                                            arg.c_grid_desc_m_n_[Number<0>{}],
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_[Number<0>{}]);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_[Number<0>{}].GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+#if 1
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<
+                        StaticallyIndexedArray<DeviceGroupedGemmXdl::AGridDesc_K0_M_K1,
+                                               GroupCount>>,
+                    remove_reference_t<
+                        StaticallyIndexedArray<DeviceGroupedGemmXdl::BGridDesc_K0_N_K1,
+                                               GroupCount>>,
+                    remove_reference_t<StaticallyIndexedArray<
+                        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+                        GroupCount>>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<
+                        StaticallyIndexedArray<typename GridwiseGemm::DefaultBlock2CTileMap,
+                                               GroupCount>>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<
+                        StaticallyIndexedArray<DeviceGroupedGemmXdl::AGridDesc_K0_M_K1,
+                                               GroupCount>>,
+                    remove_reference_t<
+                        StaticallyIndexedArray<DeviceGroupedGemmXdl::BGridDesc_K0_N_K1,
+                                               GroupCount>>,
+                    remove_reference_t<StaticallyIndexedArray<
+                        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+                        GroupCount>>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<
+                        StaticallyIndexedArray<typename GridwiseGemm::DefaultBlock2CTileMap,
+                                               GroupCount>>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(kernel,
+                                                  nrepeat,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+#endif
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        const index_t i = 0;
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_[Number<0>{}],
+                                           arg.b_grid_desc_k0_n_k1_[Number<0>{}],
+                                           arg.c_grid_desc_m_n_[Number<0>{}],
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             std::vector<gemm_desc> gemm_shapes,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a, p_b, p_c, gemm_shapes, 1, 1, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      std::vector<gemm_desc> gemm_shapes,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      index_t /* KBatch */ = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          gemm_shapes,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // 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 << "DeviceGroupedGemmXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

example/12_grouped_gemm_xdl/README.md

+# Instructions for ```gemm_xdl``` Example
+
+## Docker script
+```bash
+docker run                                                                   \
+-it                                                                          \
+--rm                                                                         \
+--privileged                                                                 \
+--group-add sudo                                                             \
+-w /root/workspace                                                           \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                                \
+rocm/tensorflow:rocm4.3.1-tf2.6-dev                                          \
+/bin/bash
+```
+
+## Build ```gemm_xdl```
+```bash
+mkdir build && cd build
+```
+
+```bash
+# Need to specify target ID, example below is gfx908
+cmake                                                                  \
+-D BUILD_DEV=OFF                                                       \
+-D CMAKE_BUILD_TYPE=Release                                            \
+-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 "   \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                              \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                         \
+..
+```
+
+```bash
+ make -j gemm_xdl
+```
+
+## Run ```gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./example/gemm_xdl 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
+```

example/12_grouped_gemm_xdl/grouped_gemm_xdl.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_gemm.hpp"
+#include "device_tensor.hpp"
+#include "device_grouped_gemm_xdl.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.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   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
+// static constexpr auto GemmMNPadding =
+// ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
+//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|      Num|
+//######|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|
+//######|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|         |
+//######|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |
+        <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    int nrepeat          = 5;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = 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: run kernel # of times (>1)\n");
+        exit(0);
+    }
+
+    int group_count = 1;
+
+    // GEMM shape
+    std::vector<ck::gemm_desc> gemm_shapes;
+
+    int A_size = 0, B_size = 0, C_size = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = 256;
+        int N = 512;
+        int K = 1024;
+
+        gemm_shapes.push_back({M, N, K, K, K, N, A_size, B_size, C_size});
+
+        A_size += M * K;
+        B_size += N * K;
+        C_size += M * N;
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<CDataType>> c_host_tensors;
+    std::vector<Tensor<CDataType>> c_device_tensors;
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
+        c_host_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+        c_device_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+    }
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    DeviceMem a_tensors_device_buf(sizeof(ADataType) * A_size);
+    DeviceMem b_tensors_device_buf(sizeof(BDataType) * B_size);
+    DeviceMem c_tensors_device_buf(sizeof(CDataType) * C_size);
+
+    std::vector<ADataType> a_tensors_data, b_tensors_data, c_tensors_data;
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        a_tensors_data.insert(
+            a_tensors_data.end(), a_tensors[i].mData.begin(), a_tensors[i].mData.end());
+        b_tensors_data.insert(
+            b_tensors_data.end(), b_tensors[i].mData.begin(), b_tensors[i].mData.end());
+    }
+
+    a_tensors_device_buf.ToDevice(a_tensors_data.data());
+    b_tensors_device_buf.ToDevice(b_tensors_data.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_tensors_device_buf.GetDeviceBuffer()),
+                          static_cast<BDataType*>(b_tensors_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_tensors_device_buf.GetDeviceBuffer()),
+                          gemm_shapes,
+                          a_element_op,
+                          b_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, nrepeat);
+
+    c_tensors_data.resize(C_size);
+
+    c_tensors_device_buf.FromDevice(c_tensors_data.data());
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        memcpy(c_device_tensors[i].mData.data(),
+               c_tensors_data.data() + gemm_shapes[i].OffsetC,
+               c_device_tensors[i].mData.size() * sizeof(CDataType));
+    }
+
+    if(do_verification)
+    {
+        for(int i = 0; i < gemm_shapes.size(); i++)
+        {
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            check_error(c_host_tensors[i], c_device_tensors[i]);
+        }
+    }
+#if 0
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, nrepeat);
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        check_error(c_m_n_host_result, c_m_n_device_result);
+    }
+#endif
+
+    return 0;
+}

example/CMakeLists.txt

@@ -24,8 +24,10 @@ set(GEMM_XDL_ALPHA_BETA_SOURCE 8_gemm_xdl_alpha_beta/gemm_xdl_alpha_beta.cpp)
 set(CONV2D_FWD_XDL_INT8_SOURCE 9_conv2d_fwd_xdl_int8/conv2d_fwd_xdl_int8.cpp)
 set(CONV3D_FWD_XDL_SOURCE 10_conv3d_fwd_xdl/conv3d_fwd_xdl.cpp)
 set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp)
+set(GROUPED_GEMM_XDL_SOURCE 12_grouped_gemm_xdl/grouped_gemm_xdl.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
+add_executable(grouped_gemm_xdl ${GROUPED_GEMM_XDL_SOURCE})
 add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
 add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
 add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
@@ -38,6 +40,7 @@ add_executable(conv3d_fwd_xdl ${CONV3D_FWD_XDL_SOURCE})
 add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
+target_link_libraries(grouped_gemm_xdl PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)