Initial int4 moe, compile pass, function not check.

example/01_gemm/gemm_xdl_fp8_pk_i4_bpreshuffle_v3.cpp

@@ -5,7 +5,7 @@
 
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_preshuffle.hpp"
 
-using F8 = ck::f8_t;
+using F8  = ck::f8_t;
 using I4  = ck::pk_i4_t;
 using F16 = ck::half_t;
 using F32 = float;

example/65_gemm_multiply_multiply/CMakeLists.txt

@@ -5,3 +5,4 @@ add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
 add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
 add_example_executable(example_moe_gemm1 moe_gemm1.cpp)
 add_example_executable(example_moe_gemm2 moe_gemm2.cpp)
+add_example_executable(example_moe_pk_i4_gemm1 moe_pk_i4_gemm1.cpp)

example/65_gemm_multiply_multiply/moe_pk_i4_gemm1.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_moe_gemm.hpp"
+// #include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.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_moe_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I4  = ck::pk_i4_t;
+using F16 = ck::half_t;
+// using BF16 = ck::bhalf_t;
+using F8  = ck::f8_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F8;
+using B0DataType       = I4;
+using EDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using ELayout  = Row;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+
+// for gate, a_scale, b_scale
+struct MulABScale
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float, float>
+                                                                           (EDataType& e,
+                                                                            const float& c,
+                                                                            const float& d0,
+                                                                            const float& d1) const
+    {
+        e = ck::type_convert<EDataType>(c * d1 * d0);
+    }
+};
+
+// for gate, a_scale, b_scale, fuse silu, 
+struct MulABScaleSilu
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float>
+                                                                           (EDataType& e,
+                                                                            const float& c,
+                                                                            const float& d0,
+                                                                            const float& d1) const
+    {
+        // act
+        float x0 = 0;
+        ck::tensor_operation::element_wise::Silu{}(x0, c * d1 * d0);
+        e = ck::type_convert<EDataType>(x0);
+    }
+};
+
+// using DsLayout = DsLayoutGate;
+// using DsDataType       = DsDataTypeGate;
+using CDEElementOp = MulABScale;
+// using CDEElementOp = MulABScaleSiluMulGate;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t MNPerXDL = 32;
+static constexpr ck::index_t CShuffleMXDLPerWave = MPerBlock / 32;
+static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
+static constexpr ck::index_t MXDLPerWave = MPerBlock / 32; //todo fix this constraint
+static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType);
+static constexpr ck::index_t BK1 = 32 / sizeof(B0DataType);
+static constexpr ck::index_t EVec = 16 / sizeof(EDataType);
+static constexpr ck::index_t D0Vec = 1;
+static constexpr ck::index_t D1Vec = 1;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm<
+            Row, Col, DsLayout, ELayout, 
+            A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
+            AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   
+            256,   MPerBlock,   128,    KPerBlock,
+            AK1,   BK1,
+            MNPerXDL,   MNPerXDL,
+            MXDLPerWave,    1,
+            S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
+            S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
+            CShuffleMXDLPerWave,    1,   S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
+            ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, true, A0DataType>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+// tokens = 1
+// topk = 1
+// experts = 8
+// per expert: 
+    // GEMM shape
+    ck::index_t N = 6144;
+    ck::index_t K = 8192;
+    ck::index_t experts = 8;
+    ck::index_t sorted_tile_num = 8;
+    ck::index_t sorted_tile_size = MPerBlock;
+    ck::index_t SORTED_SIZE = sorted_tile_num * sorted_tile_size;
+    ck::index_t tokens = 128;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        N = std::stoi(argv[4]);
+        K = std::stoi(argv[5]);
+    }
+    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=no, 1=yes)\n");
+        printf(
+            "arg4 to 5: N, K\n");
+        exit(0);
+    }
+
+    ck::index_t StrideA              = K;
+    ck::index_t StrideB              = K;
+    ck::index_t StrideE              = N;
+    ck::index_t batch_stride_B       = K * N;
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+    constexpr auto StrideDs = std::array<ck::index_t, NumDTensor>{0, 0};
+
+    ck::index_t KBatch = 1;
+
+    // const ck::index_t experts = 8;
+    Tensor<ck::index_t> expert_ids(HostTensorDescriptor({experts}, {1}));
+    Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({SORTED_SIZE}, {1}));
+    for (int i = 0; i < sorted_tile_num; i++) {
+        expert_ids.mData[i] = i;
+    }
+    int token_per_tile = tokens / sorted_tile_num;
+    int tokenid = 0;
+    // sorted_token_ids.mData[0] = 0;
+    for (int i = 0; i < SORTED_SIZE; i++) {
+        int tile_off = i % sorted_tile_size;
+        if(tile_off < token_per_tile)
+            sorted_token_ids.mData[i] = tokenid++;
+        else
+            sorted_token_ids.mData[i] = tokens;
+    }
+    expert_ids.savetxt("expert_ids.txt", "int");
+    sorted_token_ids.savetxt("sorted_token_ids.txt", "int");
+    Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
+    Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, N, K}, {N*K, K, 1}));
+    Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, N, K}, {N*K, K, 1}));
+    Tensor<D0DataType> d0_t_n(HostTensorDescriptor({tokens, N}, {StrideDs[0], 0}));
+    Tensor<D1DataType> d1_e_n(HostTensorDescriptor({experts, N}, {1, StrideDs[1]}));
+    Tensor<EDataType> e_m_n_host_result(HostTensorDescriptor({SORTED_SIZE, N}, {N, 1}));
+    Tensor<EDataType> e_m_n_device_result(HostTensorDescriptor({SORTED_SIZE, N}, {N, 1}));
+
+    std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
+    std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
+    std::cout << "d1_e_n: " << d1_e_n.mDesc << std::endl;
+    std::cout << "d0_t_n: " << d0_t_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_t_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{1, 3});
+        d1_e_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{1, 3});
+        break;
+    case 2:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        d0_t_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{});
+        d1_e_n.GenerateTensorValue(GeneratorTensor_1<D1DataType>{});
+        break;
+    default:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_t_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_e_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+    d0_t_n.savetxt("d0_t_n.txt", "int");
+    d1_e_n.savetxt("d1_e_n.txt", "int");
+    DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.mDesc.GetElementSpaceSize());
+    DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_t_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_e_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+    a0_t_k.savetxt("a.txt");
+    sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
+    expert_ids_dev.ToDevice(expert_ids.mData.data());
+    a0_device_buf.ToDevice(a0_t_k.mData.data());
+    d0_device_buf.ToDevice(d0_t_n.mData.data());
+    d1_device_buf.ToDevice(d1_e_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    // preShuffleBuffer(b0_e_n_k.mData.data(), b0_preshuffled.mData.data(), N * experts, K, NPerXdl);
+    printf("Start PreShuffle\n");
+    // weight pre-shuffle
+    int KPack = 32; // int4 -> 32, fp8 -> 16, fp16 -> 8
+    int NLane = device_op.GetPreShuffleParameters();
+    int KLane = 64 / NLane;
+
+    int K0 = K / (KLane * KPack);
+    // K -> K0 KLane KPack
+    // N -> N0 NLane
+    // N, K -> N0 K0 KLane NLane KPack
+    int tempk;
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                int n0 = n / NLane;
+                int n1 = n % NLane;
+
+                int k0 = k / (KLane * KPack);
+                tempk  = k % (KLane * KPack);
+                int k1 = tempk / KPack;
+                int k2 = tempk % KPack;
+
+                int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
+                                  k1 * KPack * NLane + n1 * KPack + k2;
+
+                b0_preshuffled(e * batch_stride_B + outputIndex) =
+                    b0_e_n_k(e * batch_stride_B + n * K + k);
+            }
+        }
+    }
+    printf("End PreShuffle, and Start vector permute\n");
+    // vector pk_i4x4 permute
+    for(int e = 0; e < experts; e++)
+    {
+        for(int i = 0; i < N; i++)
+        {
+            for(int j = 0; j < K; j++)
+            {
+                int input[8];
+
+                for(int k = 0; k < 4; k++)
+                {
+                    int i4x2         = b0_preshuffled(e, j + k * 2, i).data;
+                    input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
+                    input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
+                }
+
+                // permute 01234567->20643175
+                {
+                    int hi   = input[2];
+                    int lo   = input[0];
+                    int i4x2 = (hi << 4) | lo;
+
+                    b0_preshuffled(e, j + 0, i) = i4x2;
+                }
+
+                {
+                    int hi   = input[6];
+                    int lo   = input[4];
+                    int i4x2 = (hi << 4) | lo;
+
+                    b0_preshuffled(e, j + 2, i) = i4x2;
+                }
+
+                {
+                    int hi   = input[3];
+                    int lo   = input[1];
+                    int i4x2 = (hi << 4) | lo;
+
+                    b0_preshuffled(e, j + 4, i) = i4x2;
+                }
+
+                {
+                    int hi   = input[7];
+                    int lo   = input[5];
+                    int i4x2 = (hi << 4) | lo;
+
+                    b0_preshuffled(e, j + 6, i) = i4x2;
+                }
+            }
+        }
+    }
+
+    b0_device_buf.ToDevice(b0_preshuffled.mData.data());
+
+    printf("End Permute and Start GEMM\n");
+    auto invoker = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(sorted_token_ids_dev.GetDeviceBuffer(),
+                               expert_ids_dev.GetDeviceBuffer(),
+                               a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               tokens,
+                               SORTED_SIZE,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               StrideDs,
+                               StrideE,
+                               KBatch,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+    if (time_kernel) {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+        std::size_t flop = std::size_t(2) * SORTED_SIZE * N * K;
+        std::size_t num_btype =
+            sizeof(A0DataType) * SORTED_SIZE * K + sizeof(B0DataType) * K * N * experts + sizeof(EDataType) * SORTED_SIZE * 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"
+                << std::endl;
+    }
+
+    if(do_verification)
+    {
+        invoker.Run(argument, StreamConfig{nullptr, false, 0 ,0,1});
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<CShuffleDataType> c_m_n({SORTED_SIZE, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMoeGemm<A0DataType,
+                                                                                   B0DataType,
+                                                                                   CShuffleDataType,
+                                                                                   AccDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   PassThrough>;
+        auto ref_moe_gemm           = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_moe_gemm.MakeInvoker();
+
+        auto ref_argument = ref_moe_gemm.MakeArgument(
+           sorted_token_ids, expert_ids, sorted_tile_size, a0_t_k, b0_e_n_k, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+        for(int m = 0; m < SORTED_SIZE; ++m)
+        {
+            
+            const int t = sorted_token_ids(m);
+            const int e = expert_ids(m / sorted_tile_size);
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_t_n(t, n), d1_e_n(e, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+        e_m_n_device_result.savetxt("out.txt");
+        e_m_n_host_result.savetxt("ref.txt");
+        return ck::utils::check_err(
+                   e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_mod8.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#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/block/thread_group_tensor_slice_transfer_v7r3_scatter.hpp"
+
+#define DEBUG_LOG 0
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Even>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_moe_gemm_gather(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_sorted_token_ids,
+        karg.p_sorted_expert_ids,
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx9__))
+}
+
+// template <typename GridwiseGemm,
+//           bool HasMainKBlockLoop,
+//           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+//           index_t MinimumOccupancy = 1,
+//           TailNumber TailNum       = TailNumber::Even>
+// __global__ void
+// #if CK_USE_LAUNCH_BOUNDS
+//     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+// #endif
+//     // __attribute__((amdgpu_waves_per_eu(1, 1)))
+//     kernel_moe_gemm_gather_2lds(typename GridwiseGemm::Argument karg)
+// {
+// #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+//     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+//     __shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+//     auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
+
+//     GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+//         karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+//         karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+//         karg.p_ds_grid,
+//         karg.p_c_grid,
+//         p_shared,
+//         p_shared1,
+//         karg,
+//         karg.a_element_op,
+//         karg.b_element_op,
+//         karg.c_element_op);
+// #else
+//     ignore = karg;
+// #endif // end of if (defined(__gfx9__))
+// }
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ADataType,
+          typename LDSTypeB                           = BDataType>
+struct GridwiseMoeGemmGather
+{
+    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>{};
+
+    static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock =
+        CDEShuffleBlockTransferScalarPerVectors{}[I0];
+    // K1 should be Number<...>
+    static constexpr auto AK0Number       = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number       = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number       = Number<AK1Value>{};
+    static constexpr auto BK1Number       = Number<BK1Value>{};
+    static constexpr auto BlockSizeNumber = Number<BlockSize>{};
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>;
+    static constexpr index_t KPack =
+        math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
+    static constexpr index_t KLane =
+        mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
+    static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
+    static constexpr index_t NLane   = NPerXdl;
+    static constexpr index_t NWave   = NPerBlock / NPerXdl / NXdlPerWave;
+    static_assert(NWave * warpSize == BlockSize);
+    // static constexpr index_t NumTokens = 1;
+    static constexpr index_t SortedTileSize   = MPerBlock;
+
+
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t APackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr index_t BPackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(math::integer_divide_ceil(N, NPerBlock),
+                               math::integer_divide_ceil(M, MPerBlock),
+                               1);
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateBN0Shuffled(index_t N)
+    {
+        return math::integer_divide_ceil(N, NLane);
+    }
+    __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
+    {
+        return math::integer_divide_ceil(K, KLane * KPack);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
+    }
+
+    __host__ __device__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
+    {
+        constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+        auto K_t                = K_Batch * KReadVec;
+        return (K + K_t - 1) / K_t * KReadVec;
+    }
+
+    __host__ __device__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
+    {
+        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack>{};
+        return make_naive_tensor_descriptor(
+            make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber),
+            make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        static_assert(!(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> &&
+                        GemmSpec != GemmSpecialization::Default),
+                      "pk_i4_t does not support padding");
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWave, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ELayout>
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    template <typename DLayout>
+    __host__ __device__ static auto
+    MakeDGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I0));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I0, StrideC));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto MakeDsGridDescriptor_M_N(
+        index_t M, index_t MPad, index_t N, index_t NPad, std::array<index_t, NumDTensor> StrideDs)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                return MakeDGridDescriptor_M_N<DLayout>(M, MPad, N, NPad, StrideDs[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    template <typename DsGridDesc>
+    __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n[i], MBlock, NBlock);
+            },
+            Number<NumDTensor>{});
+    }
+
+    struct Problem
+    {
+        __host__ __device__ Problem(index_t NumTokens_,
+                                    index_t M_,
+                                    index_t N_,
+                                    index_t K_,
+                                    index_t StrideA_,
+                                    index_t StrideB_,
+                                    std::array<index_t, NumDTensor> StrideDs_,
+                                    index_t StrideC_,
+                                    index_t KBatch_)
+            : 
+              NumTokens{NumTokens_},
+              M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideDs{StrideDs_},
+              StrideC{StrideC_},
+              KBatch{KBatch_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KRead{CalculateKRead(K_, KBatch_)},
+              KPadded{CalculateKPadded(K_, KBatch_)},
+              AK0{CalculateAK0Padded(K_, KBatch_)},
+              BK0{CalculateBK0Padded(K_, KBatch_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)},
+              BN0Shuffled{CalculateBN0Shuffled(N_)},
+              BK0Shuffled{CalculateBK0Shuffled(K_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "NumTokens:" << NumTokens << ", "
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KRead:" << KRead << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t NumTokens;
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        std::array<index_t, NumDTensor> StrideDs;
+        index_t StrideC;
+        index_t KBatch;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KRead;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+        // FOR PRESHUFFLE ONLY
+        index_t BN0Shuffled;
+        index_t BK0Shuffled;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(
+                          const index_t* p_sorted_token_ids_,
+                          const index_t* p_sorted_expert_ids_,
+                          const ADataType* p_a_grid_,
+                          const BDataType* p_b_grid_,
+                          std::array<const void*, NumDTensor> p_ds_grid_,
+                          CDataType* p_c_grid_,
+                          index_t NumTokens_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          std::array<index_t, NumDTensor> StrideDs_,
+                          index_t StrideC_,
+                          index_t k_batch_,
+                          AElementwiseOperation a_element_op_,
+                          BElementwiseOperation b_element_op_,
+                          CElementwiseOperation c_element_op_)
+            : Problem{NumTokens_, M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
+            
+        p_sorted_token_ids{p_sorted_token_ids_},
+        p_sorted_expert_ids{p_sorted_expert_ids_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_ds_grid{},
+              p_c_grid{p_c_grid_},
+              a_element_op{a_element_op_},
+              b_element_op{b_element_op_},
+              c_element_op{c_element_op_}
+        {
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType_ = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid(i) = static_cast<const DDataType_*>(p_ds_grid_[i]);
+            });
+        }
+
+        const index_t * p_sorted_token_ids;
+        const index_t * p_sorted_expert_ids;
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        DsGridPointer p_ds_grid;
+        CDataType* p_c_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg, index_t k_id)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead / APackedSize;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead * karg.StrideA;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = k_id * karg.KRead * karg.StrideB;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                // KPack * NLane * KLane * K0 * N0
+                b_k_split_offset = k_id * karg.KRead * NLane / BPackedSize;
+            }
+
+            if(k_id < karg.KBatch - 1)
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+        }
+
+        index_t a_k_split_offset;
+        index_t b_k_split_offset;
+    };
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        if constexpr(ABlockLdsExtraM)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
+        }
+        // xor tensor transformation request more unnecessary vgpr usage, would cause register spill
+        // in some cases.
+        else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(LDSTypeA) /APackedSize < 1
+                                                  ? 1
+                                                  : 32 * 4 / KPerBlock / sizeof(LDSTypeA);
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_with_modulo_transform(make_tuple(
+                               Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
+                           make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_ak0_mldslayer_m_ak1,
+                make_tuple(make_pass_through_transform(AK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / MPerXdl;
+            constexpr auto K0PerThreadRead  = AK0Number / KThreadRead;
+
+            constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : 128 / (AK1Number * M0 * sizeof(LDSTypeA));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=n0
+            constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0
+                                              ? M0
+                                              : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * M1>{},
+                           Number<kfold * M0 / mpair>{},
+                           Number<mpair>{},
+                           AK1Number));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_with_modulo_transform(
+                        make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}));
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    using BlockwiseGemmPipe =
+        remove_cvref_t<decltype(BlockGemmBPreshufflePipeline_Selector<
+                                BlkGemmPipelineVer,
+                                BlkGemmPipeSched,
+                                BlockSize,
+                                ADataType,
+                                BDataType,
+                                ComputeTypeA,
+                                AccDataType,
+                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
+                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
+                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                                ABlockTransferSrcScalarPerVector,
+                                BBlockTransferSrcScalarPerVector,
+                                MPerBlock,
+                                NPerBlock,
+                                KPerBlock,
+                                MPerXdl,
+                                NPerXdl,
+                                MXdlPerWave,
+                                NXdlPerWave,
+                                KPack>())>;
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(LDSTypeA) / APackedSize,
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     !(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     (is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+
+            auto K_t = karg.KBatch * KPerBlock;
+            if(!(karg.K % K_t == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
+                          << karg.K << " " << __FILE__ << ":" << __LINE__
+                          << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+            auto K_t                = karg.KBatch * KReadVec;
+            auto KReadPadSplited    = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
+            if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+#if 1
+        const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
+
+        if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
+        {
+            return false;
+        }
+#endif
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ __device__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    // using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(
+                               const index_t* p_sorted_token_ids,
+                               const index_t* p_sorted_expert_ids,
+                               const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               DsGridPointer& p_ds_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               const Problem& problem,
+                               AElementwiseOperation a_element_op,
+                               BElementwiseOperation b_element_op,
+                               CElementwiseOperation c_element_op)
+    {
+        ignore                           = b_element_op;
+        const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+            problem.NumTokens, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+
+        const auto b_grid_desc_bpreshuffled =
+            MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+        // printf("tido %d size %d %d MNBLOCK %d %d %d %d\n", threadIdx.x, problem.StrideC, c_grid_desc_m_n.GetElementSpaceSize(),
+        // problem.MBlock, problem.NBlock, MPerBlock, NPerBlock);
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(blockIdx.x);
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(blockIdx.y);
+        const index_t expert_id = __builtin_amdgcn_readfirstlane(p_sorted_expert_ids[block_m_id]);
+        
+        // constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+        constexpr auto AMThreads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+        constexpr auto AK0Threads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+        constexpr auto AK1Threads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I2);
+        constexpr auto AKThreads = AK0Threads * AK1Threads;
+        constexpr auto AMRepeats = MPerBlock / AMThreads;
+        // static_assert(MLoadRepeats == 1, "only support 1 line per thread now!");
+        const index_t token_pos = block_m_id * MPerBlock + threadIdx.x / AKThreads * AMRepeats;
+        
+        const index_t t0 = (p_sorted_token_ids[block_m_id * MPerBlock] & 0xffffff);
+        if(t0 >= problem.NumTokens)
+            return;
+
+        StaticallyIndexedArray<index_t, AMRepeats> gather_offsets; //= p_sorted_token_ids[token_pos];
+        static_for<0, AMRepeats, 1>{}([&](auto m0) {
+            gather_offsets(m0) = (p_sorted_token_ids[token_pos + m0] & 0xffffff) * problem.K;
+            // printf("init off tid %d m %d off %d\n", threadIdx.x, m0(), gather_offsets(m0));
+        });
+        // const index_t m_block_data_idx_on_grid =
+        //     __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+        const index_t expert_stride = __builtin_amdgcn_readfirstlane(problem.N * problem.K);
+
+        // N0, K0, Blocksize*KPack
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid + expert_id * expert_stride, b_grid_desc_bpreshuffled.GetElementSpaceSize());
+        // if(threadIdx.x==0)
+        // printf("tid %d eid %d expert_stride %d bufsize %d\n",
+        // threadIdx.x, expert_id, expert_stride, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        // dummy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1_mod8<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                LDSTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                1,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{},
+                gather_offsets);
+
+        // Thread-wise copy
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        auto b_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            BDataType,
+            BDataType,
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(b_block_desc_bk0_n_bk1),
+            Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
+            Sequence<0, 1, 2, 3>,
+            3,
+            BBlockTransferSrcScalarPerVector,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true>(b_grid_desc_bpreshuffled,
+                  make_multi_index(n_block_data_idx_on_grid,
+                                   get_warp_local_1d_id(),
+                                   0,
+                                   KPack * (get_thread_local_1d_id() % warpSize)));
+
+        // LDS allocation for A and B: be careful of alignment
+        // Cast after lds
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_buf,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bpreshuffled,
+                                                                         b_blockwise_copy,
+                                                                         b_grid_buf,
+                                                                         b_block_buf,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_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_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_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_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            using EDataType = CDataType;
+
+            const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
+                problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
+
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+            const auto ds_grid_buf = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                    const DDataType *ptr_ = p_ds_grid[i];
+                    // hack logic here to support different kind of strides. todo fix it.
+                    // ascale t, 1; bscale E, N, 1, move ptr to E
+                    if (i.value == 1) 
+                    {
+                        ptr_ += expert_id * (problem.StrideDs[1]? problem.StrideDs[1] * problem.N : 1);
+                        // if ( threadIdx.x  % 16 ==0)
+                        //     printf("bid %d eid %d b eoff %d %f\n", blockIdx.y, expert_id, expert_id * (problem.StrideDs[1]? problem.StrideDs[1] * problem.N : 1), ptr_[0]);
+                    }
+                    return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                        ptr_, ds_grid_desc_m_n[i].GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_m_id, 0, block_n_id, 0);
+                        // return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                c_grid_desc_mblock_mperblock_nblock_nperblock;
+
+            using CDEBlockTransferCluster =
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
+            const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
+            constexpr auto EMThreads = CDEBlockTransferCluster{}.At(I0) * CDEBlockTransferCluster{}.At(I1);
+            constexpr auto EMRepeats = MPerBlock / EMThreads;
+            constexpr auto ENThreads = CDEBlockTransferCluster{}.At(I2) * CDEBlockTransferCluster{}.At(I3);
+            const index_t c_token_pos = block_m_id * MPerBlock + threadIdx.x / ENThreads * EMRepeats;
+            StaticallyIndexedArray<index_t, EMRepeats> scatter_offsets; //= p_sorted_token_ids[c_token_pos];
+            StaticallyIndexedArray<float, EMRepeats> scatter_weights; //= for topk
+            // too hack here, 2 specific for topk weights, fixme
+            const float *p_sorted_weights = p_ds_grid[I0];
+            static_for<0, EMRepeats, 1>{}([&](auto m0) {
+                scatter_offsets(m0) = 0;
+                scatter_weights(m0) = p_sorted_weights[(c_token_pos + m0) * problem.StrideDs[0]];
+                // if(threadIdx.x % 16 == 0)
+                // printf("init off bid %d tid %d m %d off %d\n", blockIdx.y, threadIdx.x, m0(), scatter_offsets(m0));
+            });
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferCluster,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
+                3,                    // index_t SrcVectorDim,
+                3,                    // index_t DstVectorDim,
+                CDEShuffleBlockTransferScalarPerVectors,
+                CShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>,      // ThreadTransferDstResetCoordinateAfterRunFlags
+                1,                    //ScatterDim
+                false,                //OutputScatter: false, only use scatter weights
+                1                     // ScatterWeightIdx: ascale
+                >                    
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)),
+                 c_element_op,
+                 scatter_offsets,
+                 scatter_weights};
+
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(c_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+
+    // template <bool HasMainKBlockLoop,
+    //           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    //           TailNumber TailNum = TailNumber::Odd>
+    // __device__ static void Run_2Lds(const ADataType* p_a_grid,
+    //                                 const BDataType* p_b_grid,
+    //                                 DsGridPointer& p_ds_grid,
+    //                                 CDataType* p_c_grid,
+    //                                 void* p_shared,
+    //                                 void* p_shared1,
+    //                                 const Problem& problem,
+    //                                 AElementwiseOperation a_element_op,
+    //                                 BElementwiseOperation b_element_op,
+    //                                 CElementwiseOperation c_element_op)
+    // {
+    //     // const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
+    //     // Run_2Lds<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+    //     //     p_a_grid,
+    //     //     p_b_grid,
+    //     //     p_ds_grid,
+    //     //     p_c_grid,
+    //     //     p_shared,
+    //     //     p_shared1,
+    //     //     problem,
+    //     //     a_element_op,
+    //     //     b_element_op,
+    //     //     c_element_op,
+    //     //     block_2_ctile_map);
+    // }
+
+    // template <typename Block2CTileMap,
+    //           bool HasMainKBlockLoop,
+    //           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    //           TailNumber TailNum = TailNumber::Odd>
+    // __device__ static void Run_2Lds(const ADataType* p_a_grid,
+    //                                 const BDataType* p_b_grid,
+    //                                 DsGridPointer& p_ds_grid,
+    //                                 CDataType* p_c_grid,
+    //                                 void* p_shared,
+    //                                 void* p_shared1,
+    //                                 const Problem& problem,
+    //                                 AElementwiseOperation a_element_op,
+    //                                 BElementwiseOperation b_element_op,
+    //                                 CElementwiseOperation c_element_op,
+    //                                 const Block2CTileMap& block_2_ctile_map)
+    // {
+    // }
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#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/block/thread_group_tensor_slice_transfer_v7r3_scatter.hpp"
+
+#define DEBUG_LOG 0
+
+namespace ck {
+
+// Currently we do not have a elegant way to put single lds buffer & double lds buffer pipe in same
+// kernel function Blockers:
+// 1. Two separted declaration of __shared__ pointer is the key to make sure data access operate on
+// two lds chunks.
+// 2. Occupied __shared__ won't release until whole shader end, a.k.a AB and C may not use same lds
+// buffer when we declare __shared__ inside blkgemmpipe
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t MinimumOccupancy = 1,
+          TailNumber TailNum       = TailNumber::Even>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+#endif
+    // __attribute__((amdgpu_waves_per_eu(1, 1)))
+    kernel_moe_gemm_scatter(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+        karg.p_sorted_token_ids,
+        karg.p_sorted_expert_ids,
+        karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+        karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+        karg.p_ds_grid,
+        karg.p_c_grid,
+        p_shared,
+        karg,
+        karg.a_element_op,
+        karg.b_element_op,
+        karg.c_element_op);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx9__))
+}
+
+// template <typename GridwiseGemm,
+//           bool HasMainKBlockLoop,
+//           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+//           index_t MinimumOccupancy = 1,
+//           TailNumber TailNum       = TailNumber::Even>
+// __global__ void
+// #if CK_USE_LAUNCH_BOUNDS
+//     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
+// #endif
+//     // __attribute__((amdgpu_waves_per_eu(1, 1)))
+//     kernel_moe_gemm_scatter_2lds(typename GridwiseGemm::Argument karg)
+// {
+// #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
+//     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+//     __shared__ char p_shared1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+//     auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
+
+//     GridwiseGemm::template Run_2Lds<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+//         karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
+//         karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
+//         karg.p_ds_grid,
+//         karg.p_c_grid,
+//         p_shared,
+//         p_shared1,
+//         karg,
+//         karg.a_element_op,
+//         karg.b_element_op,
+//         karg.c_element_op);
+// #else
+//     ignore = karg;
+// #endif // end of if (defined(__gfx9__))
+// }
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA                       = CDataType,
+          typename ComputeTypeB                       = ComputeTypeA,
+          typename LDSTypeA                           = ADataType,
+          typename LDSTypeB                           = BDataType>
+struct GridwiseMoeGemmScatter
+{
+    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>{};
+
+    static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock =
+        CDEShuffleBlockTransferScalarPerVectors{}[I0];
+    // K1 should be Number<...>
+    static constexpr auto AK0Number       = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number       = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number       = Number<AK1Value>{};
+    static constexpr auto BK1Number       = Number<BK1Value>{};
+    static constexpr auto BlockSizeNumber = Number<BlockSize>{};
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>;
+    static constexpr index_t KPack =
+        math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
+    static constexpr index_t KLane =
+        mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
+    static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
+    static constexpr index_t NLane   = NPerXdl;
+    static constexpr index_t NWave   = NPerBlock / NPerXdl / NXdlPerWave;
+    static_assert(NWave * warpSize == BlockSize);
+    // static constexpr index_t NumTokens = 1;
+    static constexpr index_t SortedTileSize   = MPerBlock;
+
+
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(math::integer_divide_ceil(N, NPerBlock),
+                               math::integer_divide_ceil(M, MPerBlock),
+                               1);
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateBN0Shuffled(index_t N)
+    {
+        return math::integer_divide_ceil(N, NLane);
+    }
+    __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
+    {
+        return math::integer_divide_ceil(K, KLane * KPack);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / AK1Value);
+    }
+
+    __host__ __device__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * (KPerBlock / BK1Value);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K_t = K_Batch * KPerBlock;
+        return (K + K_t - 1) / K_t * KPerBlock;
+    }
+
+    __host__ __device__ static auto CalculateKRead(index_t K, index_t K_Batch = 1)
+    {
+        constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+        auto K_t                = K_Batch * KReadVec;
+        return (K + K_t - 1) / K_t * KReadVec;
+    }
+
+    __host__ __device__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock);
+    }
+
+    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
+    {
+        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+        return transform_tensor_descriptor(
+            TileDesc_K0_MN_K1{},
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                       make_unmerge_transform(make_tuple(
+                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
+    {
+        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack>{};
+        return make_naive_tensor_descriptor(
+            make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber),
+            make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWave, NPerXdl>(BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ELayout>
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    template <typename DLayout>
+    __host__ __device__ static auto
+    MakeDGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I0));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I0, StrideC));
+            }
+        }();
+
+        // pad M and N
+        return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto MakeDsGridDescriptor_M_N(
+        index_t M, index_t MPad, index_t N, index_t NPad, std::array<index_t, NumDTensor> StrideDs)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                return MakeDGridDescriptor_M_N<DLayout>(M, MPad, N, NPad, StrideDs[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    template <typename DsGridDesc>
+    __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n[i], MBlock, NBlock);
+            },
+            Number<NumDTensor>{});
+    }
+
+    struct Problem
+    {
+        __host__ __device__ Problem(index_t NumTokens_,
+                                    index_t M_,
+                                    index_t N_,
+                                    index_t K_,
+                                    index_t StrideA_,
+                                    index_t StrideB_,
+                                    std::array<index_t, NumDTensor> StrideDs_,
+                                    index_t StrideC_,
+                                    index_t KBatch_)
+            : 
+              NumTokens{NumTokens_},
+              M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideDs{StrideDs_},
+              StrideC{StrideC_},
+              KBatch{KBatch_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KRead{CalculateKRead(K_, KBatch_)},
+              KPadded{CalculateKPadded(K_, KBatch_)},
+              AK0{CalculateAK0Padded(K_, KBatch_)},
+              BK0{CalculateBK0Padded(K_, KBatch_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)},
+              BN0Shuffled{CalculateBN0Shuffled(N_)},
+              BK0Shuffled{CalculateBK0Shuffled(K_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "NumTokens:" << NumTokens << ", "
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KRead:" << KRead << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t NumTokens;
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        std::array<index_t, NumDTensor> StrideDs;
+        index_t StrideC;
+        index_t KBatch;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KRead;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+        // FOR PRESHUFFLE ONLY
+        index_t BN0Shuffled;
+        index_t BK0Shuffled;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(
+                          const index_t* p_sorted_token_ids_,
+                          const index_t* p_sorted_expert_ids_,
+                          const ADataType* p_a_grid_,
+                          const BDataType* p_b_grid_,
+                          std::array<const void*, NumDTensor> p_ds_grid_,
+                          CDataType* p_c_grid_,
+                          index_t NumTokens_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          std::array<index_t, NumDTensor> StrideDs_,
+                          index_t StrideC_,
+                          index_t k_batch_,
+                          AElementwiseOperation a_element_op_,
+                          BElementwiseOperation b_element_op_,
+                          CElementwiseOperation c_element_op_)
+            : Problem{NumTokens_, M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
+            
+        p_sorted_token_ids{p_sorted_token_ids_},
+        p_sorted_expert_ids{p_sorted_expert_ids_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_ds_grid{},
+              p_c_grid{p_c_grid_},
+              a_element_op{a_element_op_},
+              b_element_op{b_element_op_},
+              c_element_op{c_element_op_}
+        {
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType_ = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid(i) = static_cast<const DDataType_*>(p_ds_grid_[i]);
+            });
+        }
+
+        const index_t * p_sorted_token_ids;
+        const index_t * p_sorted_expert_ids;
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        DsGridPointer p_ds_grid;
+        CDataType* p_c_grid;
+
+        const AElementwiseOperation a_element_op;
+        const BElementwiseOperation b_element_op;
+        const CElementwiseOperation c_element_op;
+    };
+
+    struct SplitKBatchOffset
+    {
+        __device__ SplitKBatchOffset(Argument& karg, index_t k_id)
+        {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                a_k_split_offset = k_id * karg.KRead * karg.StrideA;
+            }
+
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                b_k_split_offset = k_id * karg.KRead * karg.StrideB;
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                // KPack * NLane * KLane * K0 * N0
+                b_k_split_offset = k_id * karg.KRead * NLane;
+            }
+
+            if(k_id < karg.KBatch - 1)
+            {
+                karg.K = karg.KRead;
+            }
+            else
+            {
+                karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
+            }
+        }
+
+        index_t a_k_split_offset;
+        index_t b_k_split_offset;
+    };
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        if constexpr(ABlockLdsExtraM)
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock + ABlockLdsExtraM>{}, I1));
+        }
+        // xor tensor transformation request more unnecessary vgpr usage, would cause register spill
+        // in some cases.
+        else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1
+                                                  ? 1
+                                                  : 32 * 4 / KPerBlock / sizeof(LDSTypeA);
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(
+                    AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_with_modulo_transform(make_tuple(
+                               Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
+                make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
+
+            constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
+                           make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_ak0_mldslayer_m_ak1,
+                make_tuple(make_pass_through_transform(AK0Number),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
+            constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / MPerXdl;
+            constexpr auto K0PerThreadRead  = AK0Number / KThreadRead;
+
+            constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : 128 / (AK1Number * M0 * sizeof(LDSTypeA));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=n0
+            constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128)
+                                       ? 1
+                                       : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0
+                                              ? M0
+                                              : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           Number<K0PerThreadWrite>{},
+                           Number<KThreadReadPerm * M1>{},
+                           Number<kfold * M0 / mpair>{},
+                           Number<mpair>{},
+                           AK1Number));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_xor_with_modulo_transform(
+                        make_tuple(Number<KThreadReadPerm * M1>{}, Number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(Number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(Number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(Number<KThreadReadPerm>{}, Number<M1>{})),
+                    make_unmerge_transform(make_tuple(Number<kfold>{}, Number<M0 / mpair>{})),
+                    make_pass_through_transform(Number<mpair>{}),
+                    make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<0, 3>{},
+                           Sequence<4, 5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<KThreadReadPerm>{},
+                                          Number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          Number<kfold>{},
+                                          Number<K0PerThreadWrite>{})),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(Number<M0 / mpair>{}, Number<mpair>{}, Number<M1>{})),
+                           make_pass_through_transform(AK1Number)),
+                make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            return a_lds_block_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}));
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    using BlockwiseGemmPipe =
+        remove_cvref_t<decltype(BlockGemmBPreshufflePipeline_Selector<
+                                BlkGemmPipelineVer,
+                                BlkGemmPipeSched,
+                                BlockSize,
+                                LDSTypeA,
+                                LDSTypeB,
+                                ComputeTypeA,
+                                AccDataType,
+                                decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()),
+                                decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()),
+                                decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(
+                                    GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())),
+                                decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(
+                                    GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())),
+                                ABlockTransferSrcScalarPerVector,
+                                BBlockTransferSrcScalarPerVector,
+                                MPerBlock,
+                                NPerBlock,
+                                KPerBlock,
+                                MPerXdl,
+                                NPerXdl,
+                                MXdlPerWave,
+                                NXdlPerWave,
+                                KPack>())>;
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(LDSTypeA),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     !(is_same<tensor_layout::gemm::RowMajor, ALayout>::value))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding) &&
+                     (is_same<tensor_layout::gemm::RowMajor, BLayout>::value))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+
+            auto K_t = karg.KBatch * KPerBlock;
+            if(!(karg.K % K_t == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: "
+                          << karg.K << " " << __FILE__ << ":" << __LINE__
+                          << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            constexpr auto KReadVec = math::lcm(AK1Number, BK1Number);
+            auto K_t                = karg.KBatch * KReadVec;
+            auto KReadPadSplited    = math::integer_divide_ceil(karg.K, K_t) * KReadVec;
+            if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of "
+                             "CShuffleBlockTransferScalarPerVector_NPerBlock ("
+                          << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__
+                          << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+#if 1
+        const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value);
+
+        if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages)
+        {
+            return false;
+        }
+#endif
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockHasHotloop(num_loop);
+    }
+
+    __host__ __device__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return BlockwiseGemmPipe::BlockLoopTailNum(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    // if arch = gfx942
+    // using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              TailNumber TailNum = TailNumber::Odd>
+    __device__ static void Run(
+                               const index_t* p_sorted_token_ids,
+                               const index_t* p_sorted_expert_ids,
+                               const ADataType* p_a_grid,
+                               const BDataType* p_b_grid,
+                               DsGridPointer& p_ds_grid,
+                               CDataType* p_c_grid,
+                               void* p_shared,
+                               const Problem& problem,
+                               AElementwiseOperation a_element_op,
+                               BElementwiseOperation b_element_op,
+                               CElementwiseOperation c_element_op)
+    {
+        ignore                           = b_element_op;
+        const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+
+        const auto b_grid_desc_bpreshuffled =
+            MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
+            problem.NumTokens, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(blockIdx.x);
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(blockIdx.y);
+        const index_t expert_id = __builtin_amdgcn_readfirstlane(p_sorted_expert_ids[block_m_id]);
+        
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+        const index_t expert_stride = __builtin_amdgcn_readfirstlane(problem.N * problem.K);
+                
+        const index_t t0 = (p_sorted_token_ids[block_m_id * MPerBlock] & 0xffffff);
+        if(t0 >= problem.NumTokens)
+            return;
+        // N0, K0, Blocksize*KPack
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid + expert_id * expert_stride, b_grid_desc_bpreshuffled.GetElementSpaceSize());
+        // if(threadIdx.x==0)
+        // printf("tid %d eid %d expert_stride %d bufsize %d\n",
+        // threadIdx.x, expert_id, expert_stride, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        // dummy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                LDSTypeA,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                BlockwiseGemmPipe::GlobalBufferNum>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // Thread-wise copy
+        // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
+        auto b_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            BDataType,
+            BDataType,
+            decltype(b_grid_desc_bpreshuffled),
+            decltype(b_block_desc_bk0_n_bk1),
+            Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
+            Sequence<0, 1, 2, 3>,
+            3,
+            BBlockTransferSrcScalarPerVector,
+            BThreadTransferSrcResetCoordinateAfterRun,
+            true>(b_grid_desc_bpreshuffled,
+                  make_multi_index(n_block_data_idx_on_grid,
+                                   get_warp_local_1d_id(),
+                                   0,
+                                   KPack * (get_thread_local_1d_id() % warpSize)));
+
+        // LDS allocation for A and B: be careful of alignment
+        // Cast after lds
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0);
+
+        // Blockwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
+        auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
+        auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
+                                                                         a_block_desc_ak0_m_ak1,
+                                                                         a_blockwise_copy,
+                                                                         a_grid_buf,
+                                                                         a_block_buf,
+                                                                         a_block_slice_copy_step,
+                                                                         b_grid_desc_bpreshuffled,
+                                                                         b_blockwise_copy,
+                                                                         b_grid_buf,
+                                                                         b_block_buf,
+                                                                         b_block_slice_copy_step,
+                                                                         c_thread_buf,
+                                                                         num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_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_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_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_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            using EDataType = CDataType;
+
+            const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
+                problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
+
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+            const auto ds_grid_buf = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                    const DDataType *ptr_ = p_ds_grid[i];
+                    // hack logic here to support different kind of strides. todo fix it.
+                    // ascale M, 1; bscale E, N, 1, move ptr to E
+                    if (i.value == 1) 
+                    {
+                        ptr_ += expert_id * (problem.StrideDs[1]? problem.StrideDs[1] * problem.N : 1);
+                        // if ( threadIdx.x  ==0)
+                        //     printf("bid %d eid %d b eoff %d %f\n", blockIdx.y, expert_id, expert_id * (problem.StrideDs[1]? problem.StrideDs[1] * problem.N : 1), ptr_[0]);
+                    }
+                    return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                        ptr_, ds_grid_desc_m_n[i].GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_m_id, 0, block_n_id, 0);
+                        // return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                c_grid_desc_mblock_mperblock_nblock_nperblock;
+
+            using CDEBlockTransferCluster =
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
+            const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
+
+            constexpr auto EMThreads = CDEBlockTransferCluster{}.At(I0) * CDEBlockTransferCluster{}.At(I1);
+            constexpr auto EMRepeats = MPerBlock / EMThreads;
+            constexpr auto ENThreads = CDEBlockTransferCluster{}.At(I2) * CDEBlockTransferCluster{}.At(I3);
+            const index_t c_token_pos = block_m_id * MPerBlock + threadIdx.x / ENThreads * EMRepeats;
+            StaticallyIndexedArray<index_t, EMRepeats> scatter_offsets; //= p_sorted_token_ids[c_token_pos];
+            StaticallyIndexedArray<float, EMRepeats> scatter_weights; //= for topk
+            // too hack here, 2 specific for topk weights, fixme
+            const float *p_sorted_weights = p_ds_grid[I2];
+            static_for<0, EMRepeats, 1>{}([&](auto m0) {
+                scatter_offsets(m0) = (p_sorted_token_ids[c_token_pos + m0] & 0xffffff) * problem.N;
+                scatter_weights(m0) = p_sorted_weights[c_token_pos + m0];
+                // printf("init off bid %d tid %d m %d off %d\n", blockIdx.y, threadIdx.x, m0(), scatter_offsets(m0));
+            });
+
+            // printf("tid %d pos %d offset %d size %d\n", threadIdx.x, token_pos, scatter_offsets(I0), c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3_scatter<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferCluster,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
+                Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
+                3,                    // index_t SrcVectorDim,
+                3,                    // index_t DstVectorDim,
+                CDEShuffleBlockTransferScalarPerVectors,
+                CShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(0, 0, block_n_id, 0)),
+                 c_element_op,
+                 scatter_offsets,
+                 scatter_weights};
+            // if(threadIdx.x== 0)
+            auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(c_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+
+    // template <bool HasMainKBlockLoop,
+    //           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    //           TailNumber TailNum = TailNumber::Odd>
+    // __device__ static void Run_2Lds(const ADataType* p_a_grid,
+    //                                 const BDataType* p_b_grid,
+    //                                 DsGridPointer& p_ds_grid,
+    //                                 CDataType* p_c_grid,
+    //                                 void* p_shared,
+    //                                 void* p_shared1,
+    //                                 const Problem& problem,
+    //                                 AElementwiseOperation a_element_op,
+    //                                 BElementwiseOperation b_element_op,
+    //                                 CElementwiseOperation c_element_op)
+    // {
+    //     // const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
+    //     // Run_2Lds<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
+    //     //     p_a_grid,
+    //     //     p_b_grid,
+    //     //     p_ds_grid,
+    //     //     p_c_grid,
+    //     //     p_shared,
+    //     //     p_shared1,
+    //     //     problem,
+    //     //     a_element_op,
+    //     //     b_element_op,
+    //     //     c_element_op,
+    //     //     block_2_ctile_map);
+    // }
+
+    // template <typename Block2CTileMap,
+    //           bool HasMainKBlockLoop,
+    //           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    //           TailNumber TailNum = TailNumber::Odd>
+    // __device__ static void Run_2Lds(const ADataType* p_a_grid,
+    //                                 const BDataType* p_b_grid,
+    //                                 DsGridPointer& p_ds_grid,
+    //                                 CDataType* p_c_grid,
+    //                                 void* p_shared,
+    //                                 void* p_shared1,
+    //                                 const Problem& problem,
+    //                                 AElementwiseOperation a_element_op,
+    //                                 BElementwiseOperation b_element_op,
+    //                                 CElementwiseOperation c_element_op,
+    //                                 const Block2CTileMap& block_2_ctile_map)
+    // {
+    // }
+};
+
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_moe_gemm.hpp

@@ -83,21 +83,30 @@ struct ReferenceMoeGemm : public device::BaseOperator
                 if(t < token_cnt) {
                     for(int k = 0; k < K; ++k)
                     {
-                        // use PassThrough instead of ConvertBF16RTN for reference calculation
-                        if constexpr(is_same_v<AElementwiseOperation,
-                                            ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                        if constexpr(is_same_v<ADataType, pk_i4_t>)
                         {
-                            ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_t_k_(t, k));
+                            uint8_t i4x2 = arg.a_t_k_(t, k).data;
+                            uint8_t i4    = 0;
+                            if(k % 2 == 1)
+                                i4 = (i4x2 >> 0) & 0xf;
+                            else
+                                i4 = (i4x2 >> 4) & 0xf;
+                            v_a = i4_to_f32_gfx9(i4);
                         }
                         else
                         {
                             arg.a_element_op_(v_a, arg.a_t_k_(t, k));
                         }
                         // same for B matrix
-                        if constexpr(is_same_v<BElementwiseOperation,
-                                            ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                        if constexpr(is_same_v<BDataType, pk_i4_t>)
                         {
-                            ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_e_n_k_(e, n, k));
+                            uint8_t i4x2 = arg.b_e_n_k_(e, k, n).data;
+                            uint8_t i4    = 0;
+                            if(k % 2 == 1)
+                                i4 = (i4x2 >> 0) & 0xf;
+                            else
+                                i4 = (i4x2 >> 4) & 0xf;
+                            v_b = i4_to_f32_gfx9(i4);
                         }
                         else
                         {
@@ -170,6 +179,28 @@ struct ReferenceMoeGemm : public device::BaseOperator
 
         return str.str();
     }
+
+    static float i4_to_f32_gfx9(uint8_t i4)
+    {
+        static std::unordered_map<uint8_t, float> u = {{0b1000, -0.5000f},
+                                                       {0b1001, -0.4375f},
+                                                       {0b1010, -0.3750f},
+                                                       {0b1011, -0.3125f},
+                                                       {0b1100, -0.2500f},
+                                                       {0b1101, -0.1875f},
+                                                       {0b1110, -0.1250f},
+                                                       {0b1111, -0.0625f},
+                                                       {0b0, +0.0000f},
+                                                       {0b1, +0.0625f},
+                                                       {0b10, +0.1250f},
+                                                       {0b11, +0.1875f},
+                                                       {0b100, +0.2500f},
+                                                       {0b101, +0.3125f},
+                                                       {0b110, +0.3750f},
+                                                       {0b111, +0.4375f}};
+
+        return u[i4];
+    }
 };
 
 } // namespace host