merge upstream

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

@@ -500,6 +500,7 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
 
             for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
             {
+#if DEBUG_LOG
                 std::cout << "group: " << i << " arg.a_grid_desc_ak0_m_ak1_{"
                           << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0)
                           << ", "
@@ -520,6 +521,7 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
                           << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
                           << std::endl;
+#endif
 
                 if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_m_k_,
                                                 arg.gemm_desc_kernel_arg_[i].b_grid_desc_n_k_,

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/philox_rand.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_forward_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmAccDataType,
+          typename GroupKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool HasMainKBlockLoop,
+          bool IsDropout,
+          bool IsLseStoring>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v2(
+            const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const ushort p_dropout_in_16bits,
+            const GemmAccDataType p_dropout_rescale,
+            const unsigned long long seed,
+            const unsigned long long offset)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id         = get_block_1d_id();
+    const index_t global_thread_id = get_thread_global_1d_id();
+
+    ck::philox ph(seed, global_thread_id, offset);
+
+    const auto arg_ptr = reinterpret_cast<const GroupKernelArg*>(
+        cast_pointer_to_generic_address_space(group_kernel_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while(
+        (!(block_id >= arg_ptr[group_id].block_start_ && block_id < arg_ptr[group_id].block_end_)))
+    {
+        if(block_id < arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    // per-group batch offset
+    const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
+    const index_t g_idx                = __builtin_amdgcn_readfirstlane(
+        (block_id - arg_ptr[group_id].block_start_) / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
+        arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset  = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+    const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetZBasePtr(g_idx)));
+    const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
+        arg_ptr[group_id].compute_base_ptr_of_batch_.GetLSEBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>(
+        arg_ptr[group_id].p_a_grid_ + a_batch_offset,
+        arg_ptr[group_id].p_b_grid_ + b_batch_offset,
+        arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
+        arg_ptr[group_id].p_c_grid_ + c_batch_offset,
+        arg_ptr[group_id].p_z_grid_ == nullptr ? nullptr
+                                               : arg_ptr[group_id].p_z_grid_ + z_batch_offset,
+        arg_ptr[group_id].p_lse_grid_ == nullptr ? nullptr
+                                                 : arg_ptr[group_id].p_lse_grid_ + lse_batch_offset,
+        // arg_ptr[group_id].p_lse_grid_ + lse_batch_offset,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        acc_element_op,
+        b1_element_op,
+        c_element_op,
+        arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
+        arg_ptr[group_id].z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
+        arg_ptr[group_id].lse_grid_desc_m_,
+        arg_ptr[group_id].block_2_ctile_map_,
+        arg_ptr[group_id].c0_matrix_mask_,
+        p_dropout_in_16bits,
+        p_dropout_rescale,
+        ph);
+#else
+    ignore = group_kernel_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ZDataType,
+          typename LSEDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          MaskingSpecialization MaskingSpec,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle
+    : public DeviceGroupedMultiheadAttentionForward<NumDimG,
+                                                    NumDimM,
+                                                    NumDimN,
+                                                    NumDimK,
+                                                    NumDimO,
+                                                    ADataType,
+                                                    BDataType,
+                                                    B1DataType,
+                                                    CDataType,
+                                                    ZDataType,
+                                                    LSEDataType,
+                                                    Acc0BiasDataType,
+                                                    Acc1BiasDataType,
+                                                    AElementwiseOperation,
+                                                    BElementwiseOperation,
+                                                    AccElementwiseOperation,
+                                                    B1ElementwiseOperation,
+                                                    CElementwiseOperation,
+                                                    MaskingSpec>
+{
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp    = DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle;
+    using ProblemDesc = typename DeviceGroupedMultiheadAttentionForward<NumDimG,
+                                                                        NumDimM,
+                                                                        NumDimN,
+                                                                        NumDimK,
+                                                                        NumDimO,
+                                                                        ADataType,
+                                                                        BDataType,
+                                                                        B1DataType,
+                                                                        CDataType,
+                                                                        ZDataType,
+                                                                        LSEDataType,
+                                                                        Acc0BiasDataType,
+                                                                        Acc1BiasDataType,
+                                                                        AElementwiseOperation,
+                                                                        BElementwiseOperation,
+                                                                        AccElementwiseOperation,
+                                                                        B1ElementwiseOperation,
+                                                                        CElementwiseOperation,
+                                                                        MaskingSpec>::ProblemDesc;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
+    }
+
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
+    {
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
+    }
+
+    static auto MakeZGridDescriptor_M_N(const std::vector<index_t>& z_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& z_gs_ms_ns_strides_vec)
+    {
+        return Transform::MakeCGridDescriptor_M_N(z_gs_ms_ns_lengths_vec, z_gs_ms_ns_strides_vec);
+    }
+
+    static auto MakeLSEGridDescriptor_M(index_t MRaw)
+    {
+        const auto lse_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(lse_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return lse_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using LSEGridDesc_M        = decltype(MakeLSEGridDescriptor_M(1));
+    using ZGridDesc_M_N        = decltype(MakeZGridDescriptor_M_N({}, {}));
+
+    using AGridDesc_G_M_K  = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K  = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N  = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+    using ZGridDesc_G_M_N  = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+
+    constexpr static auto make_MaskOutPredicate()
+    {
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
+        {
+            return MaskDisabledPredicate{};
+        }
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+        {
+            return MaskOutUpperTrianglePredicate{};
+        }
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n,
+                                     const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
+                                     index_t BatchStrideLSE)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
+              z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
+              BatchStrideLSE_(BatchStrideLSE)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetZBasePtr(index_t g_idx) const
+        {
+            return z_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetLSEBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideLSE_);
+        }
+
+        private:
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+        ZGridDesc_G_M_N z_grid_desc_g_m_n_;
+
+        index_t BatchStrideLSE_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        LSEDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        ZGridDesc_M_N,
+        LSEGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
+
+    using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
+
+    struct GroupKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        ZDataType* p_z_grid_;
+        LSEDataType* p_lse_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
+        ZGridDesc_M_N z_grid_desc_m_n_;
+        LSEGridDesc_M lse_grid_desc_m_;
+
+        // batch & stride
+        index_t num_blocks_per_batch_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // block-to-c-tile map
+        Block2CTileMap block_2_ctile_map_;
+
+        index_t block_start_, block_end_;
+    };
+
+    struct GroupDeviceArg
+    {
+        // lengths for the last dimensions of overall problem for sanity check of vector load/store
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+
+        // strides for the last dimensions of each tensor for sanity check of vector load/store
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+
+        // for gridwise gemm check
+        CGridDesc_M_N c_grid_desc_m_n_;
+    };
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<const void*> p_b1_vec,
+                 std::vector<void*> p_c_vec,
+                 std::vector<void*> p_z_vec,
+                 std::vector<void*> p_lse_vec,
+                 std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                 std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                 std::vector<ProblemDesc> problem_desc_vec,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op,
+                 float p_dropout,
+                 std::tuple<unsigned long long, unsigned long long> seeds)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op}
+        {
+            // TODO ANT: implement bias addition
+            group_count_ = problem_desc_vec.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
+            }
+
+            if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
+            {
+                throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
+            }
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid   = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid   = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_b1_grid  = static_cast<const B1DataType*>(p_b1_vec[i]);
+                const auto p_c_grid   = static_cast<CDataType*>(p_c_vec[i]);
+                const auto p_z_grid   = static_cast<ZDataType*>(p_z_vec[i]);
+                const auto p_lse_grid = static_cast<LSEDataType*>(p_lse_vec[i]);
+
+                if(p_lse_grid == nullptr)
+                {
+                    is_lse_storing_ = false;
+                }
+
+                const auto& problem_desc = problem_desc_vec[i];
+
+                const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+                const auto z_grid_desc_m_n = MakeZGridDescriptor_M_N(
+                    problem_desc.z_gs_ms_ns_lengths, problem_desc.z_gs_ms_ns_strides);
+                const auto lse_grid_desc_m =
+                    DeviceOp::MakeLSEGridDescriptor_M(problem_desc.lse_gs_ms_lengths[NumDimG]);
+
+                const auto a_grid_desc_g_m_k = Transform::MakeAGridDescriptor_G_M_K(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+                const auto z_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
+                    problem_desc.z_gs_ms_ns_lengths, problem_desc.z_gs_ms_ns_strides);
+
+                const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n);
+
+                // typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                //    z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
+
+                const auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+                        z_grid_desc_m_n);
+
+                const index_t BlockStart     = grid_size_;
+                const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
+                const index_t batch_count    = c_grid_desc_g_m_n.GetLength(I0);
+                const index_t grid_size_grp =
+                    block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) * batch_count;
+                const index_t BlockEnd = grid_size_ + grid_size_grp;
+
+                // batch stride
+                const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
+                    a_grid_desc_g_m_k,
+                    b_grid_desc_g_n_k,
+                    b1_grid_desc_g_n_k,
+                    c_grid_desc_g_m_n,
+                    z_grid_desc_g_m_n,
+                    type_convert<index_t>(lse_grid_desc_m.GetElementSpaceSize()));
+
+                // C0 mask
+                const auto c0_matrix_mask = C0MatrixMask(b_grid_desc_g_n_k.GetLength(I1));
+
+                grid_size_ += grid_size_grp;
+
+                // for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
+                // so on
+                if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
+                     problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
+                {
+                    throw std::runtime_error(
+                        "wrong! number of biases in function argument does not "
+                        "match that in template argument");
+                }
+
+                group_kernel_args_.push_back({p_a_grid,
+                                              p_b_grid,
+                                              p_b1_grid,
+                                              p_c_grid,
+                                              p_z_grid,
+                                              p_lse_grid,
+                                              a_grid_desc_ak0_m_ak1,
+                                              b_grid_desc_bk0_n_bk1,
+                                              b1_grid_desc_bk0_n_bk1,
+                                              c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                              z_grid_desc_m_n,
+                                              lse_grid_desc_m,
+                                              block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n),
+                                              compute_base_ptr_of_batch,
+                                              c0_matrix_mask,
+                                              block_2_ctile_map,
+                                              BlockStart,
+                                              BlockEnd});
+
+                group_device_args_.push_back(
+                    {{problem_desc.a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                      problem_desc.b1_gs_os_ns_lengths[NumDimG + NumDimO - 1]},
+                     {problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                      problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+                     {problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+                     {problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO - 1],
+                      problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
+                     {problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
+                      problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
+                     c_grid_desc_m_n});
+            }
+
+            is_dropout_          = p_dropout > 0.0; //
+            p_dropout_           = 1.f - p_dropout;
+            p_dropout_in_16bits_ = uint16_t(std::floor(p_dropout_ * 65535.0));
+            p_dropout_           = 1.f / p_dropout_;
+            p_dropout_rescale_   = type_convert<GemmAccDataType>(p_dropout_);
+
+            seed_   = std::get<0>(seeds);
+            offset_ = std::get<1>(seeds);
+        }
+
+        std::vector<GroupKernelArg> group_kernel_args_;
+        std::vector<GroupDeviceArg> group_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        float p_dropout_;
+        ushort p_dropout_in_16bits_;
+        unsigned long long seed_;
+        unsigned long long offset_;
+        GemmAccDataType p_dropout_rescale_;
+        bool is_dropout_;
+
+        bool is_lse_storing_ = true;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            bool all_has_main_k_block_loop  = true;
+            bool some_has_main_k_block_loop = false;
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K = arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+                const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+                all_has_main_k_block_loop &= y;
+                some_has_main_k_block_loop |= y;
+            }
+
+            hipGetErrorString(hipMemcpy(arg.p_workspace_,
+                                        arg.group_kernel_args_.data(),
+                                        arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
+                                        hipMemcpyHostToDevice));
+
+            float ave_time = 0;
+
+            auto launch_kernel =
+                [&](auto has_main_k_block_loop_, auto is_dropout_, auto is_lse_storing_) {
+                    const auto kernel =
+                        kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v2<GridwiseGemm,
+                                                                         GemmAccDataType,
+                                                                         GroupKernelArg,
+                                                                         AElementwiseOperation,
+                                                                         BElementwiseOperation,
+                                                                         AccElementwiseOperation,
+                                                                         B1ElementwiseOperation,
+                                                                         CElementwiseOperation,
+                                                                         has_main_k_block_loop_,
+                                                                         is_dropout_,
+                                                                         is_lse_storing_>;
+
+                    return launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(arg.grid_size_),
+                        dim3(BlockSize),
+                        0,
+                        cast_pointer_to_constant_address_space(arg.p_workspace_),
+                        arg.group_count_,
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.acc_element_op_,
+                        arg.b1_element_op_,
+                        arg.c_element_op_,
+                        arg.p_dropout_in_16bits_,
+                        arg.p_dropout_rescale_,
+                        arg.seed_,
+                        arg.offset_);
+                };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(all_has_main_k_block_loop)
+            {
+                if(arg.is_dropout_)
+                {
+                    if(arg.is_lse_storing_)
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, true>{},
+                                                 integral_constant<bool, true>{},
+                                                 integral_constant<bool, true>{});
+                    }
+                    else
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, true>{},
+                                                 integral_constant<bool, true>{},
+                                                 integral_constant<bool, false>{});
+                    }
+                }
+                else
+                {
+                    if(arg.is_lse_storing_)
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, true>{},
+                                                 integral_constant<bool, false>{},
+                                                 integral_constant<bool, true>{});
+                    }
+                    else
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, true>{},
+                                                 integral_constant<bool, false>{},
+                                                 integral_constant<bool, false>{});
+                    }
+                }
+            }
+            else if(!some_has_main_k_block_loop)
+            {
+                if(arg.is_dropout_)
+                {
+                    if(arg.is_lse_storing_)
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, false>{},
+                                                 integral_constant<bool, true>{},
+                                                 integral_constant<bool, true>{});
+                    }
+                    else
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, false>{},
+                                                 integral_constant<bool, true>{},
+                                                 integral_constant<bool, false>{});
+                    }
+                }
+                else
+                {
+                    if(arg.is_lse_storing_)
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, false>{},
+                                                 integral_constant<bool, false>{},
+                                                 integral_constant<bool, true>{});
+                    }
+                    else
+                    {
+                        ave_time = launch_kernel(integral_constant<bool, false>{},
+                                                 integral_constant<bool, false>{},
+                                                 integral_constant<bool, false>{});
+                    }
+                }
+            }
+            else
+            {
+                throw std::runtime_error("wrong! all gemm problems have to simultaneously meet "
+                                         "has_main_k_block_loop or no_main_k_block_loop");
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
+        bool all_has_main_k_block_loop  = true;
+        bool some_has_main_k_block_loop = false;
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto& kernel_arg = arg.group_kernel_args_[i];
+            const auto& device_arg = arg.group_device_args_[i];
+
+            // Check if C permute dimension matches GEMM + GEMM shape
+            const index_t c_m       = device_arg.c_grid_desc_m_n_.GetLength(I0);
+            const index_t c_gemm1n  = device_arg.c_grid_desc_m_n_.GetLength(I1);
+            const index_t a_m       = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+            const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+            if(!(c_m == a_m && c_gemm1n == b1_gemm1n))
+            {
+                return false;
+            }
+
+            // Check if having main loop
+            const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                           kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+            const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+            all_has_main_k_block_loop &= y;
+            some_has_main_k_block_loop |= y;
+
+            // Note: we need raw lengths since threadwise copy can not handle vector load when
+            // part of vector is out of bounds
+            const auto MzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+            const auto NzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+            const auto KzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+            const auto Gemm1NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
+
+            // Check scalar per vector requirement
+            const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+            const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+            const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+            const auto c_extent_lowest  = Gemm1NzRaw;
+
+            if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+                 b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+                 b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+                 c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+
+            // Check vector load/store requirement
+            const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
+                                             ? device_arg.a_mz_kz_strides_[1]
+                                             : device_arg.a_mz_kz_strides_[0];
+            const auto b_stride_lowest = BBlockTransferSrcVectorDim == 2
+                                             ? device_arg.b_nz_kz_strides_[1]
+                                             : device_arg.b_nz_kz_strides_[0];
+            const auto b1_stride_lowest = B1BlockTransferSrcVectorDim == 2
+                                              ? device_arg.b1_nz_kz_strides_[1]
+                                              : device_arg.b1_nz_kz_strides_[0];
+            const auto c_stride_lowest =
+                device_arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be
+                                                     // contiguous
+
+            if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+                 c_stride_lowest == 1))
+            {
+                return false;
+            }
+
+            if(!GridwiseGemm::CheckValidity(kernel_arg.a_grid_desc_ak0_m_ak1_,
+                                            kernel_arg.b_grid_desc_bk0_n_bk1_,
+                                            kernel_arg.b1_grid_desc_bk0_n_bk1_,
+                                            device_arg.c_grid_desc_m_n_,
+                                            kernel_arg.block_2_ctile_map_))
+            {
+                return false;
+            }
+        }
+
+        // all gemm problems have to simultaneously meet has_main_k_block_loop or
+        // no_main_k_block_loop
+        if(!(all_has_main_k_block_loop || !some_has_main_k_block_loop))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<const void*> p_b1_vec,
+                             std::vector<void*> p_c_vec,
+                             std::vector<void*> p_z_vec,
+                             std::vector<void*> p_lse_vec,
+                             std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                             std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                             std::vector<ProblemDesc> problem_desc_vec,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op,
+                             float p_dropout,
+                             std::tuple<unsigned long long, unsigned long long> seeds)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_b1_vec,
+                        p_c_vec,
+                        p_z_vec,
+                        p_lse_vec,
+                        p_acc0_biases_vec,
+                        p_acc1_biases_vec,
+                        problem_desc_vec,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op,
+                        p_dropout,
+                        seeds};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<void*> p_z_vec,
+                        std::vector<void*> p_lse_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        AccElementwiseOperation acc_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op,
+                        float p_dropout,
+                        std::tuple<unsigned long long, unsigned long long> seeds) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_b1_vec,
+                                          p_c_vec,
+                                          p_z_vec,
+                                          p_lse_vec,
+                                          p_acc0_biases_vec,
+                                          p_acc1_biases_vec,
+                                          problem_desc_vec,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op,
+                                          p_dropout,
+                                          seeds);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GroupKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -95,6 +95,8 @@ struct Scale
         y = scale_ * x;
     };
 
+    __host__ __device__ auto Value() const { return scale_; }
+
     float scale_;
 };
 

include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/philox_rand.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/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.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/blockwise_softmax.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_dropout.hpp"
+
+namespace ck {
+
+template <typename DataType,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatLSE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename SElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename QGridDesc_K0_M_K1,
+          typename KGridDesc_K0_N_K1,
+          typename ZGridDesc_M_N,
+          typename VGridDesc_O0_N_O1,
+          typename CGridDesc_M_N,
+          typename LSEGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          index_t Gemm2NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    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 I8 = Number<8>{};
+    static constexpr auto I9 = Number<9>{};
+
+    static constexpr auto WaveSize = 64;
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    // C desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const ZGridDesc_M_N& z_grid_desc_m_n)
+    {
+        const auto M = z_grid_desc_m_n.GetLength(I0);
+        const auto N = z_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            z_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    template <typename AccThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4>
+    __host__ __device__ static constexpr auto GetA1SrcThreadDescriptor_AK0PerBlock_MPerBlock_AK1(
+        const AccThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4& acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4)
+    {
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to a_src_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        const auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        const auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        const auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        const auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        const auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        const auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        const auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        const auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        return transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        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;
+    }
+
+    template <typename Gemm2Param>
+    __host__ __device__ static constexpr auto GetA2BlockDescriptor_M0_N_M1()
+    {
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<Gemm2Param::A_M0>{},
+                       Number<Gemm2Param::Free0_N>{},
+                       Number<Gemm2Param::A_M1>{}),
+            make_tuple(Number<Gemm2Param::Free0_N + Gemm2Param::A_LdsPad>{} *
+                           Number<Gemm2Param::A_M1>{},
+                       Number<Gemm2Param::A_M1>{},
+                       I1));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
+                  const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
+                  const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = q_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N      = k_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K      = q_grid_desc_k0_m_k1.GetLength(I0) * q_grid_desc_k0_m_k1.GetLength(I2);
+        const auto Gemm1N = v_grid_desc_o0_n_o1.GetLength(I0) * v_grid_desc_o0_n_o1.GetLength(I2);
+
+        // This assumption reduces implemention complexity by categorizing 6 separate GEMMs into 3
+        // types of GEMM operations, therefore some code body can be reused accordingly
+        // P_MNK / dP_MNO Gemm (Gemm0 rcr)
+        // Y_MON / dQ_MKN Gemm (Gemm1 rrr)
+        // dV_NOM / dK_NKM Gemm (Gemm2 crr)
+        if(Gemm1N != K)
+        {
+            std::cerr << "SizeK must be equal to SizeO (equal attention head size)" << '\n';
+            return false;
+        }
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // 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 GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    // }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto y_grid_desc_mblock_mperblock_oblock_operblock = 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<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return y_grid_desc_mblock_mperblock_oblock_operblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeLSEGridDescriptor_MBlock_MRepeat_NWave_MPerXdl(const LSEGridDesc_M& lse_grid_desc_m)
+    {
+        const index_t M         = lse_grid_desc_m.GetLength(I0);
+        const index_t MBlock    = M / MPerBlock;
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        const auto lse_grid_desc_mblock_mrepeat_mwave_mperxdl = transform_tensor_descriptor(
+            lse_grid_desc_m,
+            make_tuple(make_unmerge_transform(
+                make_tuple(MBlock, Number<MXdlPerWave>{}, MWave, Number<MPerXdl>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2, 3>{}));
+
+        return lse_grid_desc_mblock_mrepeat_mwave_mperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>;
+
+    // Q / K / V / dY
+    struct GemmBlockwiseCopy
+    {
+        // Q matrix in LDS memory, dst of blockwise copy
+        static constexpr auto q_block_desc_k0_m_k1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // K matrix in LDS memory, dst of blockwise copy
+        static constexpr auto k_block_desc_k0_n_k1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // V matrix in LDS memory, dst of blockwise copy
+        static constexpr auto v_block_desc_k0_n_k1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // dY matrix in LDS memory, dst of blockwise copy
+        static constexpr auto ygrad_block_desc_k0_m_k1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // // A matrix in LDS memory, dst of blockwise copy
+        // static constexpr auto a_block_desc_ak0_m_ak1 =
+        //     GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // // B matrix in LDS memory, dst of blockwise copy
+        // static constexpr auto b_block_desc_bk0_n_bk1 =
+        //     GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        template <typename GridDesc_K0_M_K1>
+        using QBlockwiseCopy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                tensor_operation::element_wise::PassThrough,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                DataType,
+                                                DataType,
+                                                GridDesc_K0_M_K1,
+                                                decltype(q_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>;
+
+        template <typename GridDesc_K0_N_K1>
+        using KBlockwiseCopy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                tensor_operation::element_wise::PassThrough,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                DataType,
+                                                DataType,
+                                                GridDesc_K0_N_K1,
+                                                decltype(k_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>;
+
+        template <typename GridDesc_K0_N_K1>
+        using VBlockwiseCopy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                tensor_operation::element_wise::PassThrough,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                DataType,
+                                                DataType,
+                                                GridDesc_K0_N_K1,
+                                                decltype(v_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>;
+
+        template <typename GridDesc_K0_M_K1>
+        using YGradBlockwiseCopy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                tensor_operation::element_wise::PassThrough,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                DataType,
+                                                DataType,
+                                                GridDesc_K0_M_K1,
+                                                decltype(ygrad_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>;
+
+        static constexpr auto gemm_tile_k_block_slice_copy_step = make_multi_index(0, NPerBlock, 0);
+        static constexpr auto gemm_tile_v_block_slice_copy_step = make_multi_index(0, NPerBlock, 0);
+    };
+
+    // S / dP Gemm (type 1 rcr)
+    struct Gemm0
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        template <typename ABlockDesc_AK0_M_AK1>
+        __host__ __device__ static constexpr auto
+        MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+        {
+            constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+                ABlockDesc_AK0_M_AK1{});
+        }
+
+        template <typename BBlockDesc_BK0_N_BK1>
+        __host__ __device__ static constexpr auto
+        MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+        {
+            constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+                BBlockDesc_BK0_N_BK1{});
+        }
+
+        static constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        // Blockwise gemm with transposed XDL output
+        using BlockwiseGemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            DataType,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>; // TransposeC
+    };
+
+    // Y / dQ Gemm (type 2 rrr)
+    template <typename ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4,
+              typename ASrcBlockDesc_M0_N0_M1_N1_M2_N2_N3_N4>
+    struct Gemm1
+    {
+        private:
+        static constexpr auto m0 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I0);
+        static constexpr auto n0 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I1);
+        static constexpr auto m1 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I2);
+        static constexpr auto n1 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I3);
+        static constexpr auto m2 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I4);
+        static constexpr auto n2 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I5);
+        static constexpr auto n3 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I6);
+        static constexpr auto n4 = ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I7);
+
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        static constexpr auto N3 = ASrcBlockDesc_M0_N0_M1_N1_M2_N2_N3_N4{}.GetLength(I6);
+
+        public:
+        static constexpr auto AThreadSliceLength_K0 = Number<Gemm1KPerBlock / n4 / N3>{};
+        static constexpr auto AThreadSliceLength_M  = Number<m0 * m1 * m2>{};
+        static constexpr auto AThreadSliceLength_K1 = Number<n4>{};
+
+        // A source matrix layout in AccVGPR
+        static constexpr auto a_src_thread_desc_k0_m_k1 =
+            GetA1SrcThreadDescriptor_AK0PerBlock_MPerBlock_AK1(
+                ASrcThreadDesc_M0_N0_M1_N1_M2_N2_N3_N4{});
+
+        // A matrix in VGPR memory, dst of AccVGPR-to-VGPR copy
+        static constexpr auto a_thread_desc_k0_m_k1 = make_naive_tensor_descriptor_packed(
+            make_tuple(AThreadSliceLength_K0, AThreadSliceLength_M, AThreadSliceLength_K1));
+
+        // B matrix in LDS memory, dst of blockwise copy
+        static constexpr auto b_block_desc_bn0_k_bn1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        template <typename ABlockDesc_AK0_M_AK1>
+        __host__ __device__ static constexpr auto
+        MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+        {
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(
+                ABlockDesc_AK0_M_AK1{});
+        }
+
+        template <typename BBlockDesc_BK0_N_BK1>
+        __host__ __device__ static constexpr auto
+        MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+        {
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, 1, 1>(
+                BBlockDesc_BK0_N_BK1{});
+        }
+
+        static constexpr auto ASrcScalarPerVector = n4;
+
+        using AThreadSliceLengths_K0_M_K1 = decltype(a_thread_desc_k0_m_k1.GetLengths());
+
+        using ABlockwiseCopy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            DataType,
+            decltype(a_src_thread_desc_k0_m_k1),
+            decltype(a_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            AThreadSliceLengths_K0_M_K1,
+            Sequence<1, 0, 2>,
+            2,
+            ASrcScalarPerVector>;
+
+        // for a_block_slice_copy_step to be able to address static buffers, it MUST be a
+        // tuple-based container as well as containing ONLY integral constants
+        static constexpr auto a_block_slice_copy_step = make_tuple(AThreadSliceLength_K0, I0, I0);
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        static constexpr index_t GemmKPack =
+            MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        static constexpr index_t GemmMWave   = Gemm0MWaves;
+        static constexpr index_t GemmNWave   = Gemm0NWaves;
+        static constexpr index_t GemmMRepeat = MXdlPerWave;
+        static constexpr index_t GemmNRepeat = Gemm1NXdlPerWave;
+        static constexpr index_t GemmKLoop   = NPerBlock / Gemm1KPerBlock;   // 128/32=4
+        static constexpr index_t B_K3        = GemmKPack;                    // 4
+        static constexpr index_t B_K2        = N3;                           // 2
+        static constexpr index_t B_K1        = Gemm1KPerBlock / B_K2 / B_K3; // 4
+        static constexpr index_t B_K0        = GemmKLoop;                    // 4
+
+        __host__ __device__ static constexpr auto MakeBBlockDesc_N0_N1_N2_K0_K1_K2_K3()
+        {
+            const auto N0_ = b_block_desc_bn0_k_bn1.GetLength(I0);
+            const auto K_  = b_block_desc_bn0_k_bn1.GetLength(I1);
+            const auto N1_ = b_block_desc_bn0_k_bn1.GetLength(I2);
+
+            constexpr auto b_block_desc_n_k = transform_tensor_descriptor( //(64, 128)
+                b_block_desc_bn0_k_bn1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(N0_, N1_)), //(8, 8)
+                           make_pass_through_transform(K_)),                           // 128
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_block_desc_n_k,
+                make_tuple(
+                    make_unmerge_transform(
+                        make_tuple(GemmNRepeat, GemmNWave, NPerXdl)),            //(2, 1, 32)
+                    make_unmerge_transform(make_tuple(B_K0, B_K1, B_K2, B_K3))), //(4, 4, 2, 4)
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}));
+        }
+
+        static constexpr auto b_block_desc_n0_n1_n2_k0_k1_k2_k3 =
+            MakeBBlockDesc_N0_N1_N2_K0_K1_K2_K3();
+
+        using BThreadSlice_N0_N1_N2_K0_K1_K2_K3 = Sequence<GemmNRepeat, 1, 1, 1, B_K1, 1, B_K3>;
+
+        static constexpr auto b_thread_desc_n0_n1_n2_k0_k1_k2_k3 =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(Number<GemmNRepeat>{}, I1, I1, I1, Number<B_K1>{}, I1, Number<B_K3>{}));
+
+        __host__ __device__ static constexpr auto MakeBThreadDesc_K0_N_K1()
+        {
+            constexpr auto b_thread_desc_n_k = transform_tensor_descriptor(
+                b_thread_desc_n0_n1_n2_k0_k1_k2_k3,
+                make_tuple(
+                    make_merge_transform_v3_division_mod(make_tuple(Number<GemmNRepeat>{}, I1, I1)),
+                    make_merge_transform_v3_division_mod(
+                        make_tuple(I1, Number<B_K1>{}, I1, Number<B_K3>{}))),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_thread_desc_n_k,
+                make_tuple(make_pass_through_transform(Number<GemmNRepeat>{}),
+                           make_unmerge_transform(make_tuple(Number<B_K1>{}, Number<B_K3>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+        }
+
+        static constexpr auto b_thread_desc_k0_n_k1 = MakeBThreadDesc_K0_N_K1();
+
+        using BBlockwiseCopy =
+            ThreadwiseTensorSliceTransfer_v2<DataType,
+                                             DataType,
+                                             decltype(b_block_desc_n0_n1_n2_k0_k1_k2_k3),
+                                             decltype(b_thread_desc_n0_n1_n2_k0_k1_k2_k3),
+                                             BThreadSlice_N0_N1_N2_K0_K1_K2_K3,
+                                             Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                             6,
+                                             1,
+                                             1,
+                                             true>;
+
+        static constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, 0, 1, 0, 0, 0);
+        static constexpr auto b_block_reset_copy_step = make_multi_index(0, 0, 0, -B_K0, 0, 0, 0);
+
+        using BlockwiseGemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            DataType,
+            FloatGemmAcc,
+            decltype(a_thread_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_n_k1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b_thread_desc_k0_n_k1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            GemmKPack,
+            true,      // TransposeC
+            GemmKPack, // AMmaKStride
+            GemmKPack>;
+    };
+
+    // dV / dK Gemm (type 3 crr)
+    // Describes tuning parameter for C2_n_o = A2_n_m * B2_m_o
+    template <index_t Sum_M_ = MPerXdl * 2>
+    struct Gemm2Params_N_O_M_
+    {
+        static constexpr index_t Free0_N = NPerBlock;
+        static constexpr index_t Free1_M = MPerBlock;
+        static constexpr index_t Free1_O = Gemm1NPerBlock;
+        static constexpr index_t Sum_M   = Sum_M_;
+
+        static constexpr index_t A_M1     = 8; // P will be row-major
+        static constexpr index_t A_M0     = Sum_M / A_M1;
+        static constexpr index_t A_LdsPad = 0; // how many multiples of M1 per N * M1 elements
+
+        static_assert(Sum_M % MPerXdl == 0, "");
+
+        static constexpr index_t GemmNWave   = Free0_N / Gemm2NXdlPerWave / MPerXdl;
+        static constexpr index_t GemmOWave   = BlockSize / get_warp_size() / GemmNWave;
+        static constexpr index_t GemmNRepeat = Gemm2NXdlPerWave;
+        static constexpr index_t GemmORepeat = Free1_O / GemmOWave / NPerXdl;
+        static constexpr index_t GemmMLoop   = Free1_M / Sum_M;
+        static constexpr index_t GemmMPack =
+            math::max(A_M1, MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        static constexpr index_t B_M3 = GemmMPack; // 8
+        static constexpr index_t B_M2 =
+            XdlopsGemm<DataType, MPerXdl, NPerXdl, GemmMPack, false>{}.K0PerXdlops; // 2
+        static constexpr index_t B_M1 = Sum_M / B_M2 / B_M3;                        // 4
+        static constexpr index_t B_M0 = GemmMLoop;                                  // 2
+
+        __host__ __device__ static constexpr auto GetABlockSliceLengths_M0_N0_M1_N1_M2_N2()
+        {
+            // perform manual unmerge: m -> m_repeat, m_waves, m_per_xdl
+            constexpr index_t m  = Gemm2Params_N_O_M::Sum_M - 1;
+            constexpr index_t m2 = m % MPerXdl;
+            constexpr index_t m1 = m / MPerXdl % Gemm0MWaves;
+            constexpr index_t m0 = m / MPerXdl / Gemm0MWaves % MXdlPerWave;
+
+            // perform manual unmerge: n -> n_repeat, n_waves, n_per_xdl
+            constexpr index_t n  = Gemm2Params_N_O_M::Free0_N - 1;
+            constexpr index_t n2 = n % NPerXdl;
+            constexpr index_t n1 = n / NPerXdl % Gemm0NWaves;
+            constexpr index_t n0 = n / NPerXdl / Gemm0NWaves % NXdlPerWave;
+
+            // assume 256 decomposed into 2 x 4 x 32
+            // 1d idx ( 32 - 1) -> 3d idx 0, 0, 31 -> 3d dim 1 x 1 x 32
+            // 1d idx (256 - 1) -> 3d idx 1, 3, 31 -> 3d dim 2 x 4 x 32
+            return Sequence<m0, n0, m1, n1, m2, n2>{} + Sequence<1, 1, 1, 1, 1, 1>{};
+        }
+
+        __host__ __device__ static constexpr auto GetABlockSliceLengths_M0_N0_M1_N1()
+        {
+            return generate_sequence_v2(
+                [](auto I) { return GetABlockSliceLengths_M0_N0_M1_N1_M2_N2().At(I); },
+                Number<4>{});
+        }
+
+        using ABlockSliceLengths_M0_N0_M1_N1 = decltype(GetABlockSliceLengths_M0_N0_M1_N1());
+    };
+    using Gemm2Params_N_O_M = Gemm2Params_N_O_M_<>; // tune later
+
+    // dV / dK Gemm (type 3 crr)
+    template <typename Gemm2Params_N_O_M, typename ASrcBlockwiseGemm>
+    struct Gemm2
+    {
+        private:
+        static constexpr auto a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            ASrcBlockwiseGemm::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+        static constexpr auto M0 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0); // repeat
+        static constexpr auto N0 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        static constexpr auto M1 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2); // wave
+        static constexpr auto N1 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        static constexpr auto M2 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4); // xdl
+        static constexpr auto N2 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        static constexpr auto N3 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        static constexpr auto N4 = a_src_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        public:
+        // A source matrix layout in VGPR, src of VGPR-to-LDS copy
+        static constexpr auto a_src_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            ASrcBlockwiseGemm::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        // A matrix in LDS memory, dst of blockwise copy
+        static constexpr auto a_block_desc_m0_n_m1 =
+            GetA2BlockDescriptor_M0_N_M1<Gemm2Params_N_O_M>();
+
+        // // B matrix in LDS memory, dst of blockwise copy
+        static constexpr auto b_block_desc_o0_m_o1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        template <typename ABlockDesc_M0_N_M1>
+        __host__ __device__ static constexpr auto
+        MakeGemm2AMmaTileDescriptor_N0_N1_N2_M(const ABlockDesc_M0_N_M1&)
+        {
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm2Params_N_O_M::GemmNRepeat,
+                                                           Gemm2Params_N_O_M::GemmNWave,
+                                                           MPerXdl>(ABlockDesc_M0_N_M1{});
+        }
+
+        template <typename BBlockDesc_M0_O_M1>
+        __host__ __device__ static constexpr auto
+        MakeGemm2BMmaTileDescriptor_O0_O1_O2_M(const BBlockDesc_M0_O_M1&)
+        {
+            return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm2Params_N_O_M::GemmORepeat, 1, 1>(
+                BBlockDesc_M0_O_M1{});
+        }
+
+        __host__ __device__ static constexpr auto MakeABlockDesc_M0_N0_M1_N1_M2_N2_N3_N4()
+        {
+            const auto M0_ = a_block_desc_m0_n_m1.GetLength(I0);
+            const auto N_  = a_block_desc_m0_n_m1.GetLength(I1);
+            const auto M1_ = a_block_desc_m0_n_m1.GetLength(I2);
+
+            const auto a_block_desc_m_n = transform_tensor_descriptor(
+                a_block_desc_m0_n_m1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(M0_, M1_)),
+                           make_pass_through_transform(N_)),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // HACK: for unmerge transform, the length of highest dim is irrelevant so we put dummy
+            // variable I1 there
+            return transform_tensor_descriptor(
+                a_block_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(I1, M1, M2)),
+                           make_unmerge_transform(make_tuple(I1, N1, N2, N3, N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+        }
+
+        // Note: we will perform sub-workgroup VGPR-to-LDS copy to save LDS space, therefore the
+        // destination coordinate can overlap between wavefronts in a workgroup as seen in the mod
+        // operation before returning the values
+        __host__ __device__ static auto MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4()
+        {
+            const auto a_thread_origin_on_block_idx =
+                ASrcBlockwiseGemm::CalculateCThreadOriginDataIndex8D(I0, I0, I0, I0);
+
+            constexpr auto c_block_slice_lengths_m0_n0_m1_n1 =
+                typename Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1{}; // mrepeat, nrepeat,
+                                                                              // mwaves, nwaves,
+
+            return make_tuple(
+                a_thread_origin_on_block_idx[I0],                                         // mrepeat
+                a_thread_origin_on_block_idx[I1],                                         // nrepeat
+                a_thread_origin_on_block_idx[I2] % c_block_slice_lengths_m0_n0_m1_n1[I2], // mwave
+                a_thread_origin_on_block_idx[I3] % c_block_slice_lengths_m0_n0_m1_n1[I3], // nwave
+                a_thread_origin_on_block_idx[I4],                                         // xdlops
+                a_thread_origin_on_block_idx[I5],
+                a_thread_origin_on_block_idx[I6],
+                a_thread_origin_on_block_idx[I7]);
+        }
+
+        static constexpr auto a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            MakeABlockDesc_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        using ASrcBlockSliceWindowIterator =
+            SpaceFillingCurve<Sequence<M0, N0, M1, N1>,
+                              Sequence<0, 1, 2, 3>,
+                              typename Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1,
+                              false>;
+
+        template <typename ElementwiseOp = tensor_operation::element_wise::PassThrough>
+        using ABlockwiseCopy = ThreadwiseTensorSliceTransfer_v1r3<
+            FloatGemmAcc,
+            DataType,
+            decltype(a_src_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+            decltype(a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+            ElementwiseOp,
+            Sequence<Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(
+                         I0), // ThreadSliceLengths
+                     Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I1),
+                     I1,
+                     I1,
+                     I1,
+                     N2,
+                     I1,
+                     N4>,
+            Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+            7, // DstVectorDim
+            1, // DstScalarPerVector
+            InMemoryDataOperationEnum::Set,
+            1, // DstScalarStrideInVector
+            true>;
+
+        __host__ __device__ static constexpr auto MakeBBlockDesc_O0_O1_O2_M0_M1_M2_M3()
+        {
+            const auto O0_ = b_block_desc_o0_m_o1.GetLength(I0);
+            const auto M_  = b_block_desc_o0_m_o1.GetLength(I1);
+            const auto O1_ = b_block_desc_o0_m_o1.GetLength(I2);
+
+            constexpr auto b_block_desc_o_m = transform_tensor_descriptor( //(64, 128)
+                b_block_desc_o0_m_o1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(O0_, O1_)), //(8, 8)
+                           make_pass_through_transform(M_)),                           // 128
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_block_desc_o_m,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(Gemm2Params_N_O_M::GemmORepeat,
+                                                      Gemm2Params_N_O_M::GemmOWave,
+                                                      NPerXdl)), //(1, 2, 32)
+                    make_unmerge_transform(make_tuple(Gemm2Params_N_O_M::B_M0,
+                                                      Gemm2Params_N_O_M::B_M1,
+                                                      Gemm2Params_N_O_M::B_M2,
+                                                      Gemm2Params_N_O_M::B_M3))), //(2, 4, 2, 8)
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}));
+        }
+
+        static constexpr auto b_block_desc_o0_o1_o2_m0_m1_m2_m3 =
+            MakeBBlockDesc_O0_O1_O2_M0_M1_M2_M3();
+
+        using BThreadSlice_O0_O1_O2_M0_M1_M2_M3 = Sequence<Gemm2Params_N_O_M::GemmORepeat,
+                                                           1,
+                                                           1,
+                                                           1,
+                                                           Gemm2Params_N_O_M::B_M1,
+                                                           1,
+                                                           Gemm2Params_N_O_M::B_M3>;
+
+        static constexpr auto b_thread_desc_o0_o1_o2_m0_m1_m2_m3 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<Gemm2Params_N_O_M::GemmORepeat>{},
+                                                           I1,
+                                                           I1,
+                                                           I1,
+                                                           Number<Gemm2Params_N_O_M::B_M1>{},
+                                                           I1,
+                                                           Number<Gemm2Params_N_O_M::B_M3>{}));
+
+        __host__ __device__ static constexpr auto MakeBThreadDesc_M0_O_M1()
+        {
+            constexpr auto b_thread_desc_o_m = transform_tensor_descriptor(
+                b_thread_desc_o0_o1_o2_m0_m1_m2_m3,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(Number<Gemm2Params_N_O_M::GemmORepeat>{}, I1, I1)),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(I1,
+                                          Number<Gemm2Params_N_O_M::B_M1>{},
+                                          I1,
+                                          Number<Gemm2Params_N_O_M::B_M3>{}))),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_thread_desc_o_m,
+                make_tuple(make_pass_through_transform(Number<Gemm2Params_N_O_M::GemmORepeat>{}),
+                           make_unmerge_transform(make_tuple(Number<Gemm2Params_N_O_M::B_M1>{},
+                                                             Number<Gemm2Params_N_O_M::B_M3>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+        }
+
+        static constexpr auto b_thread_desc_m0_o_m1 = MakeBThreadDesc_M0_O_M1();
+
+        using BBlockwiseCopy =
+            ThreadwiseTensorSliceTransfer_v2<DataType,
+                                             DataType,
+                                             decltype(b_block_desc_o0_o1_o2_m0_m1_m2_m3),
+                                             decltype(b_thread_desc_o0_o1_o2_m0_m1_m2_m3),
+                                             BThreadSlice_O0_O1_O2_M0_M1_M2_M3,
+                                             Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                             6,
+                                             1,
+                                             1,
+                                             true>;
+
+        static constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, 0, 1, 0, 0, 0);
+        static constexpr auto b_block_reset_copy_step =
+            make_multi_index(0, 0, 0, -Gemm2Params_N_O_M::B_M0, 0, 0, 0);
+
+        using BlockwiseGemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            DataType,
+            FloatGemmAcc,
+            decltype(a_block_desc_m0_n_m1),
+            decltype(b_thread_desc_m0_o_m1),
+            decltype(MakeGemm2AMmaTileDescriptor_N0_N1_N2_M(a_block_desc_m0_n_m1)),
+            decltype(MakeGemm2BMmaTileDescriptor_O0_O1_O2_M(b_thread_desc_m0_o_m1)),
+            NPerBlock,
+            Gemm1NPerBlock,
+            Gemm2Params_N_O_M::Sum_M,
+            MPerXdl,
+            NPerXdl,
+            Gemm2Params_N_O_M::GemmNRepeat,
+            Gemm2Params_N_O_M::GemmORepeat,
+            Gemm2Params_N_O_M::GemmMPack,
+            true, // TransposeC
+            Gemm2Params_N_O_M::GemmMPack *
+                XdlopsGemm<DataType, MPerXdl, NPerXdl, Gemm2Params_N_O_M::GemmMPack, false>{}
+                    .K0PerXdlops,
+            Gemm2Params_N_O_M::GemmMPack>;
+
+        static constexpr auto c_block_slice_copy_step =
+            make_multi_index(Gemm2Params_N_O_M::GemmNRepeat, 0, 0, 0, 0, 0, 0, 0);
+
+        template <typename CGradDesc_N_O>
+        __host__ __device__ static auto
+        MakeCGridDesc_N0_O0_N1_O1_N2_O2_O3_O4(const CGradDesc_N_O& c_grid_desc_n_o)
+        {
+            // HACK: for unmerge transform, the length of highest dim is irrelevant so we put dummy
+            // variable I1 there
+            const auto c_grid_desc_n0_o0_n1_o1_n2_o2 = transform_tensor_descriptor(
+                c_grid_desc_n_o,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(I1, Gemm2Params_N_O_M::GemmNWave, MPerXdl)),
+                    make_unmerge_transform(make_tuple(I1, Gemm2Params_N_O_M::GemmOWave, NPerXdl))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+            const auto c_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4 =
+                BlockwiseGemm{}.xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(
+                    c_grid_desc_n0_o0_n1_o1_n2_o2);
+
+            return c_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4;
+        }
+
+        static constexpr auto c_thread_desc_n0_o0_n1_o1_n2_o2_o3_o4 =
+            BlockwiseGemm::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        __host__ __device__ static auto GetCThreadOriginOnBlock_N0_O0_N1_O1_N2_O2_O3_O4()
+        {
+            return to_multi_index(BlockwiseGemm::CalculateCThreadOriginDataIndex8D(I0, I0, I0, I0));
+        }
+
+        template <typename CGridDesc_N0_O0_N1_O1_N2_O2_O3_O4,
+                  typename ElementwiseOp = tensor_operation::element_wise::PassThrough>
+        using CBlockwiseCopy = ThreadwiseTensorSliceTransfer_v1r3<
+            FloatGemmAcc,
+            DataType,
+            decltype(c_thread_desc_n0_o0_n1_o1_n2_o2_o3_o4),
+            CGridDesc_N0_O0_N1_O1_N2_O2_O3_O4,
+            ElementwiseOp,                                                // CElementwiseOperation
+            decltype(c_thread_desc_n0_o0_n1_o1_n2_o2_o3_o4.GetLengths()), // SliceLengths
+            Sequence<0, 1, 2, 3, 4, 5, 6, 7>,                             // AccessOrder
+            7,                                                            // VectorDim
+            2,                                                            // ScalarPerVector
+            InMemoryDataOperationEnum::AtomicAdd, // GlobalMemoryDataOperation
+            1,                                    // DstScalarStrideInVector
+            true>;
+    };
+
+    template <index_t BlockSize_, index_t BlockSliceLength_M_, index_t BlockSliceLength_O_>
+    struct YDotYGrad_M_O_
+    {
+        static constexpr index_t SrcScalarPerVector = 16 / sizeof(DataType);
+        static constexpr auto ThreadClusterLength_O =
+            Number<BlockSliceLength_O_ / SrcScalarPerVector>{};
+        static constexpr auto ThreadClusterLength_M = Number<BlockSize_ / ThreadClusterLength_O>{};
+        static constexpr auto ThreadSliceLength_O   = Number<SrcScalarPerVector>{};
+        static constexpr auto ThreadSliceLength_M =
+            Number<BlockSliceLength_M_ * ThreadClusterLength_O / BlockSize_>{};
+
+        // dY matrix in LDS memory, dst of blockwise copy
+        static constexpr auto ygrad_block_desc_o0_m_o1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        __host__ __device__ static constexpr auto MakeYGradBlockDesc_M_O()
+        {
+            const auto O0_ = ygrad_block_desc_o0_m_o1.GetLength(I0);
+            const auto M_  = ygrad_block_desc_o0_m_o1.GetLength(I1);
+            const auto O1_ = ygrad_block_desc_o0_m_o1.GetLength(I2);
+
+            static_assert(O0_ * O1_ == BlockSliceLength_O_, "");
+            static_assert(M_ == BlockSliceLength_M_, "");
+
+            return transform_tensor_descriptor( //(128, 64)
+                ygrad_block_desc_o0_m_o1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(O0_, O1_)), //(8, 8)
+                           make_pass_through_transform(M_)),                           // 128
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+        }
+
+        static constexpr auto ygrad_block_desc_m_o = MakeYGradBlockDesc_M_O();
+
+        static_assert(ThreadClusterLength_O * ThreadSliceLength_O == BlockSliceLength_O_, "");
+        static_assert(ThreadClusterLength_M * ThreadSliceLength_M == BlockSliceLength_M_, "");
+
+        using SrcBufType = StaticBuffer<AddressSpaceEnum::Vgpr,
+                                        DataType,
+                                        ThreadSliceLength_M * ThreadSliceLength_O,
+                                        true>;
+
+        using DstBufType =
+            StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, ThreadSliceLength_M, true>;
+    };
+    using YDotYGrad_M_O = YDotYGrad_M_O_<BlockSize, MPerBlock, Gemm1NPerBlock>;
+
+    // QGrad Gemm has the same layout as Y = P * V Gemm (A in acc B row-major)
+    struct QGradGemmTile_M_K_N
+    {
+        template <typename QGridDesc_K0_M_K1_>
+        __device__ static auto MakeQGradGridDesc_MBlock_MPerBlock_KBlock_KPerBlock(
+            const QGridDesc_K0_M_K1_& q_grid_desc_k0_m_k1)
+        {
+            const auto K0 = q_grid_desc_k0_m_k1.GetLength(I0);
+            const auto M  = q_grid_desc_k0_m_k1.GetLength(I1);
+            const auto K1 = q_grid_desc_k0_m_k1.GetLength(I2);
+            const auto K  = K0 * K1;
+
+            const auto MBlock = M / MPerBlock;
+            const auto KBlock = K / Gemm1NPerBlock; // NOTE: QGrad gemm is similar to Y gemm
+
+            const auto q_grid_desc_m_k = transform_tensor_descriptor(
+                q_grid_desc_k0_m_k1,
+                make_tuple(make_pass_through_transform(M),
+                           make_merge_transform_v3_division_mod(make_tuple(K0, K1))),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                q_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                           make_unmerge_transform(make_tuple(KBlock, Number<Gemm1NPerBlock>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+        }
+    };
+
+    struct KGradGemmTile_N_K_M
+    {
+        // C position
+        template <typename KGridDesc_K0_N_K1_>
+        __device__ static auto MakeKGradGridDesc_N_K(const KGridDesc_K0_N_K1_& k_grid_desc_k0_n_k1)
+        {
+            const auto K_K0 = k_grid_desc_k0_n_k1.GetLength(I0);
+            const auto N    = k_grid_desc_k0_n_k1.GetLength(I1);
+            const auto K_K1 = k_grid_desc_k0_n_k1.GetLength(I2);
+
+            return transform_tensor_descriptor(
+                k_grid_desc_k0_n_k1,
+                make_tuple(make_pass_through_transform(N),
+                           make_merge_transform_v3_division_mod(make_tuple(K_K0, K_K1))),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    };
+
+    struct SharedMemTrait
+    {
+        // // LDS allocation for A and B: be careful of alignment
+        static constexpr auto q_block_desc_k0_m_k1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto k_block_desc_k0_n_k1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto v_block_desc_k0_n_k1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto ygrad_block_desc_k0_m_k1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto p_slash_sgrad_block_desc_m0_n_m1 =
+            GetA2BlockDescriptor_M0_N_M1<Gemm2Params_N_O_M>();
+
+        static constexpr auto max_lds_align = Number<16 / sizeof(DataType)>{};
+
+        static constexpr auto q_block_space_size_aligned =
+            math::integer_least_multiple(q_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto k_block_space_size_aligned =
+            math::integer_least_multiple(k_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto v_block_space_size_aligned =
+            math::integer_least_multiple(v_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto ygrad_block_space_size_aligned = math::integer_least_multiple(
+            ygrad_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto p_slash_sgrad_block_space_size_aligned = math::integer_least_multiple(
+            p_slash_sgrad_block_desc_m0_n_m1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto ygrad_block_space_offset = 0;
+        static constexpr auto q_block_space_offset     = ygrad_block_space_size_aligned.value;
+        static constexpr auto k_block_space_offset =
+            ygrad_block_space_size_aligned.value + q_block_space_size_aligned.value;
+        static constexpr auto v_block_space_offset =
+            ygrad_block_space_size_aligned.value + q_block_space_size_aligned.value;
+        static constexpr auto p_slash_sgrad_block_space_offset =
+            ygrad_block_space_size_aligned.value + q_block_space_size_aligned.value;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset =
+            (ygrad_block_space_size_aligned.value + q_block_space_size_aligned.value) *
+            sizeof(DataType) / sizeof(FloatGemmAcc);
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t k_bytes_end =
+            (SharedMemTrait::k_block_space_offset + SharedMemTrait::k_block_space_size_aligned) *
+            sizeof(DataType);
+        const index_t v_bytes_end =
+            (SharedMemTrait::v_block_space_offset + SharedMemTrait::v_block_space_size_aligned) *
+            sizeof(DataType);
+        const index_t p_slash_sgrad_bytes_end =
+            (SharedMemTrait::p_slash_sgrad_block_space_offset +
+             SharedMemTrait::p_slash_sgrad_block_space_size_aligned) *
+            sizeof(DataType);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(k_bytes_end,
+                         v_bytes_end,
+                         p_slash_sgrad_bytes_end,
+                         softmax_bytes_end,
+                         c_block_bytes_end);
+    }
+
+    template <bool HasMainKBlockLoop,
+              typename Block2CTileMap,
+              typename C0MatrixMask,
+              typename VGradGridDescriptor_N_O,
+              typename YGradGridDesc_O0_M_O1>
+    __device__ static void Run(const DataType* __restrict__ p_q_grid,
+                               const DataType* __restrict__ p_k_grid,
+                               unsigned short* __restrict__ p_z_grid,
+                               const DataType* __restrict__ p_v_grid,
+                               const DataType* __restrict__ p_y_grid,
+                               const FloatLSE* __restrict__ p_lse_grid,
+                               const DataType* __restrict__ p_ygrad_grid,
+                               DataType* __restrict__ p_qgrad_grid,
+                               DataType* __restrict__ p_kgrad_grid,
+                               DataType* __restrict__ p_vgrad_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const SElementwiseOperation& s_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
+                               const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
+                               const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1,
+                               const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock&
+                                   y_grid_desc_mblock_mperblock_oblock_operblock,
+                               const LSEGridDesc_M& lse_grid_desc_m,
+                               const VGradGridDescriptor_N_O& vgrad_grid_desc_n_o,
+                               const YGradGridDesc_O0_M_O1& ygrad_grid_desc_o0_m_o1,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask,
+                               const float p_drop,
+                               ck::philox& ph)
+    {
+        const FloatGemmAcc p_dropout     = type_convert<FloatGemmAcc>(1.0f - p_drop);
+        const FloatGemmAcc rp_dropout    = type_convert<FloatGemmAcc>(1.0f / p_dropout);
+        const ushort p_dropout_in_16bits = uint16_t(std::floor(p_dropout * 65535.0));
+        const tensor_operation::element_wise::Scale scale_rp_dropout(s_element_op.Value() *
+                                                                     rp_dropout);
+
+        const auto q_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_q_grid, q_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto k_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_k_grid, k_grid_desc_k0_n_k1.GetElementSpaceSize());
+        const auto v_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_v_grid, v_grid_desc_o0_n_o1.GetElementSpaceSize());
+        const auto y_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_grid, y_grid_desc_mblock_mperblock_oblock_operblock.GetElementSpaceSize());
+        const auto lse_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_lse_grid, lse_grid_desc_m.GetElementSpaceSize());
+        const auto ygrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_ygrad_grid, ygrad_grid_desc_o0_m_o1.GetElementSpaceSize());
+        auto vgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_vgrad_grid, vgrad_grid_desc_n_o.GetElementSpaceSize());
+        auto qgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_qgrad_grid, q_grid_desc_k0_m_k1.GetElementSpaceSize());
+        auto kgrad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_kgrad_grid, k_grid_desc_k0_n_k1.GetElementSpaceSize());
+
+        // divide block work by [M, O]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(y_grid_desc_mblock_mperblock_oblock_operblock.GetLength(I0),
+                          y_grid_desc_mblock_mperblock_oblock_operblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/o_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        // const index_t o_block_data_idx_on_grid =
+        //     __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // 6 GEMM operations are categorized into 3 buckets. SizeK == SizeO == head_dim
+        // S_MNK / dP_MNO Gemm (Gemm0 rcr)
+        // Y_MON / dQ_MKN Gemm (Gemm1 rrr)
+        // dV_NOM / dK_NKM Gemm (Gemm2 crr)
+
+        // LDS allocation for Q / K / V / dY
+        auto q_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<DataType*>(p_shared) + SharedMemTrait::q_block_space_offset,
+            GemmBlockwiseCopy::q_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto k_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<DataType*>(p_shared) + SharedMemTrait::k_block_space_offset,
+            GemmBlockwiseCopy::k_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        auto v_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<DataType*>(p_shared) + SharedMemTrait::v_block_space_offset,
+            GemmBlockwiseCopy::v_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        auto ygrad_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<DataType*>(p_shared) + SharedMemTrait::ygrad_block_space_offset,
+            GemmBlockwiseCopy::ygrad_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        // Q matrix blockwise copy
+        auto gemm_tile_q_blockwise_copy =
+            typename GemmBlockwiseCopy::template QBlockwiseCopy<decltype(q_grid_desc_k0_m_k1)>(
+                q_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                GemmBlockwiseCopy::q_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // K matrix blockwise copy
+        auto gemm_tile_k_blockwise_copy =
+            typename GemmBlockwiseCopy::template KBlockwiseCopy<decltype(k_grid_desc_k0_n_k1)>(
+                k_grid_desc_k0_n_k1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                GemmBlockwiseCopy::k_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // V matrix blockwise copy
+        auto gemm_tile_v_blockwise_copy =
+            typename GemmBlockwiseCopy::template VBlockwiseCopy<decltype(v_grid_desc_o0_n_o1)>(
+                v_grid_desc_o0_n_o1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b1_element_op,
+                GemmBlockwiseCopy::v_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // dY matrix blockwise copy
+        auto gemm_tile_ygrad_blockwise_copy =
+            typename GemmBlockwiseCopy::template YGradBlockwiseCopy<decltype(
+                ygrad_grid_desc_o0_m_o1)>(ygrad_grid_desc_o0_m_o1,
+                                          make_multi_index(0, m_block_data_idx_on_grid, 0),
+                                          a_element_op,
+                                          GemmBlockwiseCopy::ygrad_block_desc_k0_m_k1,
+                                          make_multi_index(0, 0, 0),
+                                          tensor_operation::element_wise::PassThrough{});
+
+        //
+        // set up S / dP Gemm (type 1 rcr)
+        //
+
+        // S: blockwise gemm
+        auto s_blockwise_gemm = typename Gemm0::BlockwiseGemm{}; // TransposeC
+
+        auto s_slash_p_thread_buf = s_blockwise_gemm.GetCThreadBuffer();
+
+        // dP: blockwise gemm
+        // we need separate blockwise gemm object because we need separate thread buffer
+        auto pgrad_blockwise_gemm = typename Gemm0::BlockwiseGemm{};
+
+        auto pgrad_thread_buf = pgrad_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // set up Y / dQ Gemm (type 2 rrr)
+        //
+        // Note: Y is pre-calculated in forward pass and loaded to backward pass kernel
+        using Gemm1 =
+            Gemm1<decltype(s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()),
+                  decltype(s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4())>;
+
+        // Gemm1: VGPR allocation for A and B
+        auto gemm1_a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, DataType>(
+            Gemm1::a_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto gemm1_b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, DataType>(
+            Gemm1::b_thread_desc_n0_n1_n2_k0_k1_k2_k3.GetElementSpaceSize());
+
+        // dQ: transform input and output tensor descriptors
+        auto qgrad_grid_desc_mblock_mperblock_kblock_kperblock =
+            QGradGemmTile_M_K_N::MakeQGradGridDesc_MBlock_MPerBlock_KBlock_KPerBlock(
+                q_grid_desc_k0_m_k1);
+
+        // dQ: A matrix blockwise copy
+        auto qgrad_gemm_tile_sgrad_blockwise_copy =
+            typename Gemm1::ABlockwiseCopy{tensor_operation::element_wise::PassThrough{}};
+
+        // dQ: blockwise gemm
+        auto qgrad_blockwise_gemm =
+            typename Gemm1::BlockwiseGemm{make_tuple(0, 0, 0, 0), make_tuple(0, 0, 0, 0)};
+
+        // dQ: B matrix blockwise copy
+        auto k_thread_origin = qgrad_blockwise_gemm.CalculateBThreadOriginDataIndex();
+
+        auto qgrad_gemm_tile_k_blockwise_copy = typename Gemm1::BBlockwiseCopy{
+            Gemm1::b_block_desc_n0_n1_n2_k0_k1_k2_k3,
+            make_multi_index(0,                                      // nrepeat
+                             k_thread_origin[I1],                    // nwave
+                             k_thread_origin[I2],                    // nperxdl
+                             0,                                      // k0
+                             0,                                      // k1
+                             k_thread_origin[I3] / Gemm1::GemmKPack, // k2
+                             0)};                                    // k3
+
+        auto qgrad_thread_buf = qgrad_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+        constexpr auto m0 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        auto blockwise_dropout = BlockwiseDropout<FloatGemmAcc, decltype(thread_slice_desc_m_n)>{
+            p_dropout_in_16bits, rp_dropout};
+
+        auto lse_grid_desc_mblock_mrepeat_mwave_mperxdl =
+            MakeLSEGridDescriptor_MBlock_MRepeat_NWave_MPerXdl(lse_grid_desc_m);
+
+        constexpr auto lse_thread_desc_mblock_mrepeat_mwave_mperxdl =
+            make_naive_tensor_descriptor_packed(make_tuple(I1, m0, m1, m2));
+
+        auto lse_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatLSE>(
+            lse_thread_desc_mblock_mrepeat_mwave_mperxdl.GetElementSpaceSize());
+
+        auto acc0_thread_origin = s_blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+            Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+        auto lse_thread_copy_global_to_vgpr =
+            ThreadwiseTensorSliceTransfer_v2<FloatLSE,
+                                             FloatLSE,
+                                             decltype(lse_grid_desc_mblock_mrepeat_mwave_mperxdl),
+                                             decltype(lse_thread_desc_mblock_mrepeat_mwave_mperxdl),
+                                             Sequence<1, m0, m1, m2>,
+                                             Sequence<0, 1, 2, 3>,
+                                             3,
+                                             m2,
+                                             1,
+                                             false>{
+                lse_grid_desc_mblock_mrepeat_mwave_mperxdl,
+                make_multi_index(block_work_idx[I0],       // mblock
+                                 acc0_thread_origin[I0],   // mrepeat
+                                 acc0_thread_origin[I2],   // mwave
+                                 acc0_thread_origin[I4])}; // mperxdl
+
+        //
+        // z vgpr copy to global
+        //
+        // z matrix threadwise desc
+        constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
+                                                           n4)); // registerNum
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     unsigned short,
+                     z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                     true>
+            z_tenor_buffer;
+        z_tenor_buffer.Clear();
+        // z matrix global desc
+        /*const auto M = q_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N = k_grid_desc_k0_n_k1.GetLength(I1);
+
+        auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(M, N);*/
+
+        auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize());
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+            ushort,
+            ushort,
+            decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<I1, // MBlockId
+                     I1, // NBlockID
+                     m0, // MRepeat
+                     n0, // NRepeat
+                     m1, // MWaveId
+                     n1, // NWaveId
+                     m2, // MPerXdl
+                     n2, // NGroupNum
+                     n3, // NInputNum
+                     n4>,
+            Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+            9, // DstVectorDim
+            1, // DstScalarPerVector
+            InMemoryDataOperationEnum::Set,
+            1, // DstScalarStrideInVector
+            true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                  make_multi_index(block_work_idx[I0], // MBlockId
+                                   0,                  // NBlockId
+                                   0,                  // mrepeat
+                                   0,                  // nrepeat
+                                   wave_id[I0],        // MWaveId
+                                   wave_id[I1],        // NWaveId
+                                   wave_m_n_id[I1],    // MPerXdl
+                                   0,                  // group
+                                   wave_m_n_id[I0],    // NInputIndex
+                                   0),
+                  tensor_operation::element_wise::PassThrough{}};
+
+        //
+        // set up dV / dK Gemm (type 3 crr)
+        //
+        using Gemm2 = Gemm2<Gemm2Params_N_O_M, decltype(s_blockwise_gemm)>;
+
+        // Gemm2: LDS allocation for A and B: be careful of alignment
+        auto gemm2_a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<DataType*>(p_shared) + SharedMemTrait::p_slash_sgrad_block_space_offset,
+            Gemm2::a_block_desc_m0_n_m1.GetElementSpaceSize());
+
+        auto gemm2_b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, DataType>(
+            Gemm2::b_thread_desc_o0_o1_o2_m0_m1_m2_m3.GetElementSpaceSize());
+
+        // dV: transform input and output tensor descriptors
+        const auto vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4 =
+            Gemm2::MakeCGridDesc_N0_O0_N1_O1_N2_O2_O3_O4(vgrad_grid_desc_n_o);
+
+        // dV: A matrix VGPR-to-LDS blockwise copy
+        auto vgrad_gemm_tile_p_thread_copy_vgpr_to_lds =
+            typename Gemm2::template ABlockwiseCopy<tensor_operation::element_wise::Relu>{
+                Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
+                tensor_operation::element_wise::Relu{}}; // relu(P-dropped)
+
+        // dV: blockwise gemm
+        auto v_slash_k_grad_blockwise_gemm = typename Gemm2::BlockwiseGemm{};
+        v_slash_k_grad_blockwise_gemm.SetBBlockStartWindow(make_tuple(0, 0, 0, 0));
+
+        auto q_slash_ygrad_thread_origin =
+            v_slash_k_grad_blockwise_gemm.CalculateBThreadOriginDataIndex();
+
+        // dV: B matrix LDS-to-VGPR blockwise copy
+        auto vgrad_gemm_tile_ygrad_blockwise_copy = typename Gemm2::BBlockwiseCopy{
+            Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+            make_multi_index(0,                               // orepeat
+                             q_slash_ygrad_thread_origin[I1], // owave
+                             q_slash_ygrad_thread_origin[I2], // nperxdl
+                             0,                               // m0
+                             0,                               // m1
+                             q_slash_ygrad_thread_origin[I3] / Gemm2Params_N_O_M::GemmMPack, // m2
+                             0)};                                                            // m3
+
+        auto v_slash_k_grad_thread_buf = v_slash_k_grad_blockwise_gemm.GetCThreadBuffer();
+
+        // dV: C VGPR-to-global copy
+        const auto vgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4 =
+            Gemm2::GetCThreadOriginOnBlock_N0_O0_N1_O1_N2_O2_O3_O4() +
+            make_multi_index(
+                I0, block_work_idx[I1] * Gemm2Params_N_O_M::GemmORepeat, I0, I0, I0, I0, I0, I0);
+
+        auto vgrad_thread_copy_vgpr_to_global =
+            typename Gemm2::template CBlockwiseCopy<decltype(
+                                                        vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4),
+                                                    tensor_operation::element_wise::Scale>(
+                vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                vgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
+                tensor_operation::element_wise::Scale{rp_dropout});
+
+        // dK: transform output tensor descriptors
+        const auto kgrad_grid_desc_n_k =
+            KGradGemmTile_N_K_M::MakeKGradGridDesc_N_K(k_grid_desc_k0_n_k1);
+        const auto kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4 =
+            Gemm2::MakeCGridDesc_N0_O0_N1_O1_N2_O2_O3_O4(kgrad_grid_desc_n_k);
+
+        // dK: A matrix VGPR-to-LDS blockwise copy
+        auto kgrad_gemm_tile_sgrad_thread_copy_vgpr_to_lds =
+            typename Gemm2::template ABlockwiseCopy<tensor_operation::element_wise::PassThrough>{
+                Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
+                tensor_operation::element_wise::PassThrough{}};
+
+        // dK: B matrix LDS-to-VGPR blockwise copy
+        auto kgrad_gemm_tile_q_blockwise_copy = typename Gemm2::BBlockwiseCopy{
+            Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+            make_multi_index(0,                               // orepeat
+                             q_slash_ygrad_thread_origin[I1], // owave
+                             q_slash_ygrad_thread_origin[I2], // nperxdl
+                             0,                               // m0
+                             0,                               // m1
+                             q_slash_ygrad_thread_origin[I3] / Gemm2Params_N_O_M::GemmMPack, // m2
+                             0)};                                                            // m3
+
+        // dK: blockwise gemm
+        /* reuse v_slash_k_grad_blockwise_gemm, v_slash_k_grad_thread_buf */
+
+        // dK: C VGPR-to-global copy
+        const auto kgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4 =
+            Gemm2::GetCThreadOriginOnBlock_N0_O0_N1_O1_N2_O2_O3_O4() +
+            make_multi_index(
+                I0, block_work_idx[I1] * Gemm2Params_N_O_M::GemmORepeat, I0, I0, I0, I0, I0, I0);
+
+        auto kgrad_thread_copy_vgpr_to_global = typename Gemm2::template CBlockwiseCopy<
+            decltype(kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4),
+            decltype(scale_rp_dropout)>(kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                        kgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
+                                        scale_rp_dropout);
+
+        //
+        // set up Y dot dY
+        //
+
+        // m0, n0 are m/n repeat per wave
+        // m1, n1 are number of waves
+        constexpr auto p_block_lengths =
+            s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto P_M0 = p_block_lengths[I0]; // repeats
+        constexpr auto P_M1 = p_block_lengths[I2]; // waves
+        constexpr auto P_M2 = p_block_lengths[I4]; // xdl
+
+        constexpr auto y_thread_desc_m0_m1_o0_o1 = make_naive_tensor_descriptor_packed(make_tuple(
+            I1, YDotYGrad_M_O::ThreadSliceLength_M, I1, YDotYGrad_M_O::ThreadSliceLength_O));
+        constexpr auto ygrad_thread_desc_m_o     = make_naive_tensor_descriptor_packed(
+            make_tuple(YDotYGrad_M_O::ThreadSliceLength_M, YDotYGrad_M_O::ThreadSliceLength_O));
+        constexpr auto y_thread_cluster_desc =
+            make_cluster_descriptor(Sequence<I1,
+                                             YDotYGrad_M_O::ThreadClusterLength_M,
+                                             I1,
+                                             YDotYGrad_M_O::ThreadClusterLength_O>{},
+                                    Sequence<0, 1, 2, 3>{});
+        const auto y_thread_cluster_idx =
+            y_thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        constexpr auto ygrad_thread_cluster_desc = make_cluster_descriptor(
+            Sequence<YDotYGrad_M_O::ThreadClusterLength_M, YDotYGrad_M_O::ThreadClusterLength_O>{},
+            Sequence<0, 1>{});
+        const auto ygrad_thread_cluster_idx = ygrad_thread_cluster_desc.CalculateBottomIndex(
+            make_multi_index(get_thread_local_1d_id()));
+
+        const auto y_thread_data_on_block_idx =
+            y_thread_cluster_idx * y_thread_desc_m0_m1_o0_o1.GetLengths();
+        const auto ygrad_thread_data_on_block_idx =
+            ygrad_thread_cluster_idx * ygrad_thread_desc_m_o.GetLengths();
+        const auto y_thread_data_on_grid_idx =
+            make_multi_index(
+                block_work_idx[I0], I0, I0 /* all WGs start from o_block_idx = 0 */, I0) +
+            y_thread_data_on_block_idx;
+
+        // performs for y
+        auto y_threadwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            DataType,
+            DataType,
+            YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
+            decltype(y_thread_desc_m0_m1_o0_o1),
+            decltype(y_thread_desc_m0_m1_o0_o1.GetLengths()),
+            Sequence<0, 1, 2, 3>,
+            3,                                 // SrcVectorDim
+            YDotYGrad_M_O::SrcScalarPerVector, // SrcScalarPerVector
+            1,                                 // SrcScalarStrideInVector
+            true /* ResetCoordAfterRun */>(y_grid_desc_mblock_mperblock_oblock_operblock,
+                                           y_thread_data_on_grid_idx);
+
+        // performs for ygrad
+        auto ygrad_threadwise_copy = ThreadwiseTensorSliceTransfer_v2<
+            DataType,
+            DataType,
+            decltype(YDotYGrad_M_O::ygrad_block_desc_m_o),
+            decltype(ygrad_thread_desc_m_o),
+            decltype(ygrad_thread_desc_m_o.GetLengths()),
+            Sequence<0, 1>,
+            1,                                 // SrcVectorDim
+            YDotYGrad_M_O::SrcScalarPerVector, // SrcScalarPerVector
+            1,                                 // SrcScalarStrideInVector
+            true /* ResetCoordAfterRun */>(YDotYGrad_M_O::ygrad_block_desc_m_o,
+                                           ygrad_thread_data_on_block_idx);
+
+        auto y_thread_buf                 = typename YDotYGrad_M_O::SrcBufType{};
+        auto ygrad_thread_buf             = typename YDotYGrad_M_O::SrcBufType{};
+        auto y_dot_ygrad_thread_accum_buf = typename YDotYGrad_M_O::DstBufType{};
+        auto y_dot_ygrad_block_accum_buf  = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            MPerBlock);
+
+        constexpr auto y_dot_ygrad_block_desc_mblock_mrepeat_mwave_mperxdl =
+            make_naive_tensor_descriptor(make_tuple(I1, P_M0, P_M1, P_M2),
+                                         make_tuple(P_M0 * P_M1 * P_M2, P_M1 * P_M2, P_M2, I1));
+        constexpr auto y_dot_ygrad_thread_desc_mblock_mrepeat_mwave_mperxdl =
+            lse_thread_desc_mblock_mrepeat_mwave_mperxdl; // reuse LSE thread descriptor because
+                                                          // per-thread LSE data and y_dot_ygrad is
+                                                          // tiled the same way
+
+        auto y_dot_ygrad_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+            FloatGemmAcc,
+            FloatGemmAcc,
+            decltype(y_dot_ygrad_block_desc_mblock_mrepeat_mwave_mperxdl),
+            decltype(y_dot_ygrad_thread_desc_mblock_mrepeat_mwave_mperxdl),
+            Sequence<1, m0, m1, m2>,
+            Sequence<0, 1, 2, 3>,
+            3,
+            m2,
+            1,
+            false>{y_dot_ygrad_block_desc_mblock_mrepeat_mwave_mperxdl,
+                   make_multi_index(I0,                       // mblock
+                                    acc0_thread_origin[I0],   // mrepeat
+                                    acc0_thread_origin[I2],   // mwave
+                                    acc0_thread_origin[I4])}; // mperxdl
+
+        auto y_dot_ygrad_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatGemmAcc>(
+            y_dot_ygrad_thread_desc_mblock_mrepeat_mwave_mperxdl.GetElementSpaceSize());
+
+        // load ygrad
+        gemm_tile_ygrad_blockwise_copy.Run(ygrad_grid_desc_o0_m_o1,
+                                           ygrad_grid_buf,
+                                           GemmBlockwiseCopy::ygrad_block_desc_k0_m_k1,
+                                           ygrad_block_buf,
+                                           I0);
+
+        block_sync_lds();
+
+        //
+        // calculate Y dot dY
+        //
+
+        // clear accum buffers
+        y_dot_ygrad_thread_accum_buf.Clear();
+        y_dot_ygrad_block_accum_buf.Clear();
+
+        y_threadwise_copy.Run(y_grid_desc_mblock_mperblock_oblock_operblock,
+                              y_grid_buf,
+                              y_thread_desc_m0_m1_o0_o1,
+                              make_tuple(I0, I0, I0, I0),
+                              y_thread_buf);
+        ygrad_threadwise_copy.Run(YDotYGrad_M_O::ygrad_block_desc_m_o,
+                                  ygrad_block_buf,
+                                  ygrad_thread_desc_m_o,
+                                  make_tuple(I0, I0),
+                                  ygrad_thread_buf);
+
+        static_for<0, YDotYGrad_M_O::ThreadSliceLength_M, 1>{}([&](auto iM) {
+            static_for<0, YDotYGrad_M_O::ThreadSliceLength_O, 1>{}([&](auto iO) {
+                constexpr auto y_offset =
+                    y_thread_desc_m0_m1_o0_o1.CalculateOffset(make_multi_index(I0, iM, I0, iO));
+                constexpr auto ygrad_offset =
+                    ygrad_thread_desc_m_o.CalculateOffset(make_multi_index(iM, iO));
+                y_dot_ygrad_thread_accum_buf(iM) +=
+                    y_thread_buf[Number<y_offset>{}] * ygrad_thread_buf[Number<ygrad_offset>{}];
+            });
+        });
+
+        // blockwise reduction using atomic_add
+        block_sync_lds();
+        static_for<0, YDotYGrad_M_O::ThreadSliceLength_M, 1>{}([&](auto iM) {
+            const auto idx_on_block = y_thread_data_on_block_idx[I1] + iM;
+            y_dot_ygrad_block_accum_buf.AtomicAdd(
+                idx_on_block, true, y_dot_ygrad_thread_accum_buf[iM] * p_dropout); // p_dropoutD1
+        });
+        block_sync_lds();
+
+        // distribute y_dot_ygrad to threads; LDS accum buffer can be safely reused after barrier
+        y_dot_ygrad_thread_copy_lds_to_vgpr.Run(
+            y_dot_ygrad_block_desc_mblock_mrepeat_mwave_mperxdl,
+            y_dot_ygrad_block_accum_buf,
+            y_dot_ygrad_thread_desc_mblock_mrepeat_mwave_mperxdl,
+            make_tuple(I0, I0, I0, I0),
+            y_dot_ygrad_thread_buf);
+
+        lse_thread_copy_global_to_vgpr.Run(lse_grid_desc_mblock_mrepeat_mwave_mperxdl,
+                                           lse_grid_buf,
+                                           lse_thread_desc_mblock_mrepeat_mwave_mperxdl,
+                                           make_tuple(I0, I0, I0, I0),
+                                           lse_thread_buf);
+
+        const index_t num_gemm1_k_block_outer_loop = k_grid_desc_k0_n_k1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize dQ
+        qgrad_thread_buf.Clear();
+
+        // load q
+        gemm_tile_q_blockwise_copy.Run(q_grid_desc_k0_m_k1,
+                                       q_grid_buf,
+                                       GemmBlockwiseCopy::q_block_desc_k0_m_k1,
+                                       q_block_buf,
+                                       I0);
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+
+            // gemm dP
+            // dP = dY * V^T
+            pgrad_thread_buf.Clear();
+            gemm_tile_v_blockwise_copy.Run(v_grid_desc_o0_n_o1,
+                                           v_grid_buf,
+                                           GemmBlockwiseCopy::v_block_desc_k0_n_k1,
+                                           v_block_buf,
+                                           I0);
+            block_sync_lds();
+            pgrad_blockwise_gemm.Run(ygrad_block_buf, v_block_buf, pgrad_thread_buf);
+
+            // gemm S
+            // S = Q * K^T
+            s_slash_p_thread_buf.Clear();
+            gemm_tile_k_blockwise_copy.Run(k_grid_desc_k0_n_k1,
+                                           k_grid_buf,
+                                           GemmBlockwiseCopy::k_block_desc_k0_n_k1,
+                                           k_block_buf,
+                                           I0);
+            block_sync_lds();
+            s_blockwise_gemm.Run(q_block_buf, k_block_buf, s_slash_p_thread_buf);
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(c_thread_lengths),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        s_slash_p_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                    else
+                    {
+                        s_element_op(s_slash_p_thread_buf(i), s_slash_p_thread_buf[i]);
+                    }
+                });
+            }
+            else
+            {
+                static_for<0, s_slash_p_thread_buf.Size(), 1>{}([&](auto i) {
+                    s_element_op(s_slash_p_thread_buf(i), s_slash_p_thread_buf[i]);
+                });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // P_i: = softmax(scalar * S_i:)
+            // scaling is already performed in the preceding statements with s_element_op
+            blockwise_softmax.RunWithPreCalcStats(s_slash_p_thread_buf, lse_thread_buf);
+
+            // save z to global
+            if(p_z_grid)
+            {
+                // P_dropped
+                blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf),
+                                                        decltype(z_tenor_buffer),
+                                                        true>(
+                    s_slash_p_thread_buf, ph, z_tenor_buffer);
+
+                z_thread_copy_vgpr_to_global.Run(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                                 z_tenor_buffer,
+                                                 z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                 z_grid_buf);
+            }
+            else
+            {
+                // P_dropped
+                blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>(
+                    s_slash_p_thread_buf, ph);
+            }
+
+            block_sync_lds(); // wait for gemm1 LDS read
+
+            // dS = P * (dP - Y_dot_dY)
+            auto& sgrad_thread_buf = pgrad_thread_buf;
+            constexpr auto pgrad_thread_tile_iterator =
+                pgrad_blockwise_gemm.MakeCThreadTileIterator();
+            constexpr auto pgrad_thread_idx_to_m_n_adaptor =
+                pgrad_blockwise_gemm.MakeCThreadIndexAdaptor8DTo2D();
+            static_for<0, pgrad_thread_tile_iterator.GetNumOfAccess(), 1>{}([&](auto i) {
+                constexpr auto pgrad_thread_idx = pgrad_thread_tile_iterator.GetIndex(i);
+                constexpr auto m =
+                    pgrad_thread_idx_to_m_n_adaptor.CalculateBottomIndex(pgrad_thread_idx)[I0];
+                // dS and P has same thread buf layout
+                if(s_slash_p_thread_buf[i] >= 0)
+                {
+                    sgrad_thread_buf(i) =
+                        s_slash_p_thread_buf[i] *
+                        (pgrad_thread_buf[i] - y_dot_ygrad_thread_buf[Number<m>{}]);
+                }
+                else
+                {
+                    sgrad_thread_buf(i) =
+                        s_slash_p_thread_buf[i] * y_dot_ygrad_thread_buf[Number<m>{}];
+                }
+            });
+
+            // gemm dQ
+            // dQ = scalar * dS * K
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // main body
+                static_for<0, num_gemm1_k_block_inner_loop, 1>{}([&](auto i) {
+                    qgrad_gemm_tile_sgrad_blockwise_copy.Run(Gemm1::a_src_thread_desc_k0_m_k1,
+                                                             Gemm1::a_block_slice_copy_step * i,
+                                                             sgrad_thread_buf,
+                                                             Gemm1::a_thread_desc_k0_m_k1,
+                                                             make_tuple(I0, I0, I0),
+                                                             gemm1_a_thread_buf);
+
+                    qgrad_gemm_tile_k_blockwise_copy.Run(Gemm1::b_block_desc_n0_n1_n2_k0_k1_k2_k3,
+                                                         k_block_buf,
+                                                         Gemm1::b_thread_desc_n0_n1_n2_k0_k1_k2_k3,
+                                                         make_tuple(I0, I0, I0, I0, I0, I0, I0),
+                                                         gemm1_b_thread_buf);
+
+                    qgrad_gemm_tile_k_blockwise_copy.MoveSrcSliceWindow(
+                        Gemm1::b_block_desc_n0_n1_n2_k0_k1_k2_k3, Gemm1::b_block_slice_copy_step);
+
+                    block_sync_lds();
+
+                    qgrad_blockwise_gemm.Run(
+                        gemm1_a_thread_buf, gemm1_b_thread_buf, qgrad_thread_buf);
+
+                    // block_sync_lds();
+                });
+            } // end gemm dQ
+
+            SubThreadBlock<BlockSize> gemm2_a_copy_subgroup(s_blockwise_gemm.GetWaveIdx()[I0],
+                                                            s_blockwise_gemm.GetWaveIdx()[I1]);
+
+            constexpr index_t num_gemm2_loop = MPerBlock / Gemm2Params_N_O_M::Sum_M;
+            static_assert(Gemm2::ASrcBlockSliceWindowIterator::GetNumOfAccess() == num_gemm2_loop,
+                          "");
+
+            // TODO: tune gemm2 pipeline
+            // dV = P_drop^T * dY
+            v_slash_k_grad_thread_buf.Clear();
+            static_for<0, num_gemm2_loop, 1>{}([&](auto gemm2_loop_idx) { // gemm dV
+                // load VGrad Gemm A
+                const auto p_slice_idx =
+                    Gemm2::ASrcBlockSliceWindowIterator::GetIndexTupleOfNumber(gemm2_loop_idx);
+                constexpr auto mwave_range = make_tuple(
+                    p_slice_idx[I2],
+                    p_slice_idx[I2] + Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I2));
+                constexpr auto nwave_range = make_tuple(
+                    p_slice_idx[I3],
+                    p_slice_idx[I3] + Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I3));
+
+                if(gemm2_a_copy_subgroup.IsBelong(mwave_range, nwave_range))
+                {
+                    vgrad_gemm_tile_p_thread_copy_vgpr_to_lds.Run(
+                        Gemm2::a_src_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                        make_tuple(p_slice_idx[I0], p_slice_idx[I1], I0, I0, I0, I0, I0, I0),
+                        s_slash_p_thread_buf,
+                        Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                        gemm2_a_block_buf);
+                }
+
+                // block_sync_lds(); // sync before write
+
+                vgrad_gemm_tile_ygrad_blockwise_copy.Run(Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+                                                         ygrad_block_buf,
+                                                         Gemm2::b_thread_desc_o0_o1_o2_m0_m1_m2_m3,
+                                                         make_tuple(I0, I0, I0, I0, I0, I0, I0),
+                                                         gemm2_b_thread_buf);
+
+                vgrad_gemm_tile_ygrad_blockwise_copy.MoveSrcSliceWindow(
+                    Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3, Gemm2::b_block_slice_copy_step);
+
+                block_sync_lds(); // sync before read
+                v_slash_k_grad_blockwise_gemm.Run(
+                    gemm2_a_block_buf, gemm2_b_thread_buf, v_slash_k_grad_thread_buf);
+
+            }); // end gemm dV
+            // atomic_add dV
+            vgrad_thread_copy_vgpr_to_global.Run(Gemm2::c_thread_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                                 v_slash_k_grad_thread_buf,
+                                                 vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                                 vgrad_grid_buf);
+
+            // dK = scalar * dS^T * dQ
+            v_slash_k_grad_thread_buf.Clear();
+            // gemm2_b_thread_buf.Clear();
+            static_for<0, num_gemm2_loop, 1>{}([&](auto gemm2_loop_idx) { // gemm dK
+                // load KGrad Gemm A
+                const auto sgrad_slice_idx =
+                    Gemm2::ASrcBlockSliceWindowIterator::GetIndexTupleOfNumber(gemm2_loop_idx);
+                constexpr auto mwave_range =
+                    make_tuple(sgrad_slice_idx[I2],
+                               sgrad_slice_idx[I2] +
+                                   Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I2));
+                constexpr auto nwave_range =
+                    make_tuple(sgrad_slice_idx[I3],
+                               sgrad_slice_idx[I3] +
+                                   Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I3));
+
+                if(gemm2_a_copy_subgroup.IsBelong(mwave_range, nwave_range))
+                {
+                    kgrad_gemm_tile_sgrad_thread_copy_vgpr_to_lds.Run(
+                        Gemm2::a_src_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                        make_tuple(
+                            sgrad_slice_idx[I0], sgrad_slice_idx[I1], I0, I0, I0, I0, I0, I0),
+                        sgrad_thread_buf,
+                        Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                        gemm2_a_block_buf);
+                }
+
+                kgrad_gemm_tile_q_blockwise_copy.Run(Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+                                                     q_block_buf,
+                                                     Gemm2::b_thread_desc_o0_o1_o2_m0_m1_m2_m3,
+                                                     make_tuple(I0, I0, I0, I0, I0, I0, I0),
+                                                     gemm2_b_thread_buf);
+
+                kgrad_gemm_tile_q_blockwise_copy.MoveSrcSliceWindow(
+                    Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3, Gemm2::b_block_slice_copy_step);
+
+                block_sync_lds(); // sync before read
+                v_slash_k_grad_blockwise_gemm.Run(
+                    gemm2_a_block_buf, gemm2_b_thread_buf, v_slash_k_grad_thread_buf);
+
+            }); // end gemm dK
+            // atomic_add dK
+            kgrad_thread_copy_vgpr_to_global.Run(Gemm2::c_thread_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                                 v_slash_k_grad_thread_buf,
+                                                 kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                                 kgrad_grid_buf);
+
+            // move slice window
+            gemm_tile_k_blockwise_copy.MoveSrcSliceWindow(
+                k_grid_desc_k0_n_k1,
+                GemmBlockwiseCopy::gemm_tile_k_block_slice_copy_step); // step N
+            gemm_tile_v_blockwise_copy.MoveSrcSliceWindow(
+                v_grid_desc_o0_n_o1,
+                GemmBlockwiseCopy::gemm_tile_v_block_slice_copy_step); // step N
+            vgrad_gemm_tile_ygrad_blockwise_copy.MoveSrcSliceWindow(
+                Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+                Gemm2::b_block_reset_copy_step); // rewind M
+            vgrad_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4, Gemm2::c_block_slice_copy_step); // step N
+            qgrad_gemm_tile_k_blockwise_copy.MoveSrcSliceWindow(
+                Gemm1::b_block_desc_n0_n1_n2_k0_k1_k2_k3,
+                Gemm1::b_block_reset_copy_step); // rewind K
+            kgrad_gemm_tile_q_blockwise_copy.MoveSrcSliceWindow(
+                Gemm2::b_block_desc_o0_o1_o2_m0_m1_m2_m3,
+                Gemm2::b_block_reset_copy_step); // rewind M
+            kgrad_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4, Gemm2::c_block_slice_copy_step); // step N
+            z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle dQ and write
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                qgrad_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                qgrad_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_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<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = 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 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                qgrad_blockwise_gemm.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_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   SElementwiseOperation,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    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],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    scale_rp_dropout};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                DataType,             // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(qgrad_grid_desc_mblock_mperblock_kblock_kperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 qgrad_grid_desc_mblock_mperblock_kblock_kperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.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_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              qgrad_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              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
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    qgrad_grid_desc_mblock_mperblock_kblock_kperblock,
+                    qgrad_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        qgrad_grid_desc_mblock_mperblock_kblock_kperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_v1.hpp → include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_v2.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/utility/common_header.hpp"
+#include "ck/utility/philox_rand.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"
@@ -15,6 +16,7 @@
 #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/blockwise_softmax.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_dropout.hpp"
 
 namespace ck {
 
@@ -30,6 +32,7 @@ template <typename DataType,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
           typename QGridDesc_K0_M_K1,
           typename KGridDesc_K0_N_K1,
+          typename ZGridDesc_M_N,
           typename VGridDesc_N0_O_N1,
           typename CGridDesc_M_N,
           typename LSEGridDesc_M,
@@ -80,8 +83,23 @@ template <typename DataType,
           bool PadN,
           bool MaskOutUpperTriangle,
           PipelineVersion PipelineVer = PipelineVersion::v1>
-struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
+struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
 {
+    template <typename T>
+    struct TypeMap
+    {
+        using type = T;
+    };
+#if defined(__gfx90a__)
+    template <>
+    struct TypeMap<ck::half_t>
+    {
+        using type = ck::bhalf_t;
+    };
+#endif
+
+    using LDSDataType = typename TypeMap<DataType>::type;
+
     static_assert(LoopSched == LoopScheduler::Default,
                   "Non-default loop scheduler is currently not supported");
 
@@ -93,7 +111,10 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
     static constexpr auto I5 = Number<5>{};
     static constexpr auto I6 = Number<6>{};
     static constexpr auto I7 = Number<7>{};
+    static constexpr auto I8 = Number<8>{};
+    static constexpr auto I9 = Number<9>{};
 
+    static constexpr auto WaveSize = 64;
     // K1 should be Number<...>
     // Gemm0
     static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
@@ -113,6 +134,65 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
     using GridwiseGemmPipe = remove_cvref_t<decltype(
         GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
+    // C desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const ZGridDesc_M_N& z_grid_desc_m_n)
+    {
+        const auto M = z_grid_desc_m_n.GetLength(I0);
+        const auto N = z_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<LDSDataType, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            z_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+    __host__ __device__ static constexpr auto
+    MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const index_t M, const index_t N)
+    {
+        constexpr auto mfma = MfmaSelector<LDSDataType, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            make_naive_tensor_descriptor_packed(make_tuple(M, N)),
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
     template <typename ABlockDesc_AK0_M_AK1>
     __host__ __device__ static constexpr auto
     MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
@@ -347,6 +427,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
 
+    using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>;
+
     // S / dP Gemm (type 1 rcr)
     struct Gemm0
     {
@@ -388,7 +471,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
                                                 DataType,
-                                                DataType,
+                                                LDSDataType,
                                                 GridDesc_K0_M_K1,
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -413,7 +496,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
                                                 DataType,
-                                                DataType,
+                                                LDSDataType,
                                                 GridDesc_K0_N_K1,
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -428,13 +511,14 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                                 true, // DstResetCoord
                                                 NumGemmKPrefetchStage>;
 
-        static constexpr index_t KPack = math::max(
-            math::lcm(AK1, BK1), MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        static constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<LDSDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         // Blockwise gemm with transposed XDL output
         using BlockwiseGemm = BlockwiseGemmXdlops_v2<
             BlockSize,
-            DataType,
+            LDSDataType,
             FloatGemmAcc,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -496,7 +580,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
 
         using ABlockwiseCopy = ThreadwiseTensorSliceTransfer_StaticToStatic<
             FloatGemmAcc,
-            DataType,
+            LDSDataType,
             decltype(a_src_thread_desc_k0_m_k1),
             decltype(a_thread_desc_k0_m_k1),
             tensor_operation::element_wise::PassThrough,
@@ -515,7 +599,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                                 B1BlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 B1BlockTransferThreadClusterArrangeOrder,
                                                 DataType,
-                                                DataType,
+                                                LDSDataType,
                                                 GridDesc_K0_N_K1,
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 B1BlockTransferSrcAccessOrder,
@@ -546,11 +630,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
         // therefore we may just as well assign Gemm1KPack = group_size
         static constexpr index_t GemmKPack =
-            MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.group_size;
+            MfmaSelector<LDSDataType, MPerXdl, NPerXdl>::selected_mfma.group_size;
 
         using BlockwiseGemm = BlockwiseGemmXdlops_v2<
             BlockSize,
-            DataType,
+            LDSDataType,
             FloatGemmAcc,
             decltype(a_thread_desc_k0_m_k1),
             decltype(b_block_desc_bk0_n_bk1),
@@ -566,7 +650,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             GemmKPack,
             true,      // TransposeC
             GemmKPack, // AMmaKStride
-            GemmKPack * XdlopsGemm<DataType, MPerXdl, NPerXdl, GemmKPack, false>{}
+            GemmKPack * XdlopsGemm<LDSDataType, MPerXdl, NPerXdl, GemmKPack, false>{}
                             .K0PerXdlops /* BMmaKStride */>;
     };
 
@@ -598,7 +682,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         static constexpr index_t GemmORepeat = Free1_O / GemmOWave / NPerXdl;
         static constexpr index_t GemmMPack =
             math::max(math::lcm(A_M1, B_M1),
-                      MfmaSelector<DataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+                      MfmaSelector<LDSDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         using BBlockSliceLengths = Sequence<B_M0, Free1_O, B_M1>;
         using BThreadClusterLengths =
@@ -720,12 +804,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                               typename Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1,
                               false>;
 
+        template <typename ElementwiseOp = tensor_operation::element_wise::PassThrough>
         using ABlockwiseCopy = ThreadwiseTensorSliceTransfer_v1r3<
             FloatGemmAcc,
-            DataType,
+            LDSDataType,
             decltype(a_src_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
             decltype(a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
-            tensor_operation::element_wise::PassThrough,
+            ElementwiseOp,
             Sequence<Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(
                          I0), // ThreadSliceLengths
                      Gemm2Params_N_O_M::ABlockSliceLengths_M0_N0_M1_N1::At(I1),
@@ -752,7 +837,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             typename Gemm2Params_N_O_M::BThreadClusterLengths,
             typename Gemm2Params_N_O_M::BThreadClusterArrangeOrder,
             DataType,
-            DataType,
+            LDSDataType,
             GridDesc_M0_O_M1,
             decltype(b_block_desc_m0_o_m1),
             typename Gemm2Params_N_O_M::BThreadClusterArrangeOrder, // access order == thread order
@@ -769,7 +854,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
 
         using BlockwiseGemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
-                                                                DataType,
+                                                                LDSDataType,
                                                                 FloatGemmAcc,
                                                                 decltype(a_block_desc_m0_n_m1),
                                                                 decltype(b_block_desc_m0_o_m1),
@@ -836,7 +921,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
     template <index_t BlockSize_, index_t BlockSliceLength_M_, index_t BlockSliceLength_O_>
     struct YDotYGrad_M_O_
     {
-        static constexpr index_t SrcScalarPerVector = 16 / sizeof(DataType);
+        static constexpr index_t SrcScalarPerVector = 16 / sizeof(FloatGemmAcc);
         static constexpr auto ThreadClusterLength_O =
             Number<BlockSliceLength_O_ / SrcScalarPerVector>{};
         static constexpr auto ThreadClusterLength_M = Number<BlockSize_ / ThreadClusterLength_O>{};
@@ -848,7 +933,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         static_assert(ThreadClusterLength_M * ThreadSliceLength_M == BlockSliceLength_M_, "");
 
         using SrcBufType = StaticBuffer<AddressSpaceEnum::Vgpr,
-                                        DataType,
+                                        FloatGemmAcc,
                                         ThreadSliceLength_M * ThreadSliceLength_O,
                                         true>;
 
@@ -1010,7 +1095,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         static constexpr auto b2_block_desc_m0_o_m1 =
             GetB2BlockDescriptor_M0_O_M1<Gemm2Params_N_O_M>();
 
-        static constexpr auto max_lds_align = Number<16 / sizeof(DataType)>{};
+        static constexpr auto max_lds_align = Number<16 / sizeof(LDSDataType)>{};
 
         static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
@@ -1046,13 +1131,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
     {
         const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
                                          SharedMemTrait::b_block_space_size_aligned) *
-                                        sizeof(DataType);
+                                        sizeof(LDSDataType);
         const index_t gemm1_bytes_end =
             (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
-            sizeof(DataType);
+            sizeof(LDSDataType);
         const index_t vgrad_gemm_bytes_end = (SharedMemTrait::p_block_space_size_aligned +
                                               SharedMemTrait::ygrad_block_space_size_aligned) *
-                                             sizeof(DataType);
+                                             sizeof(LDSDataType);
 
         const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
                                            SharedMemTrait::reduction_space_size_aligned) *
@@ -1074,6 +1159,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
               typename YGradGridDesc_M0_O_M1>
     __device__ static void Run(const DataType* __restrict__ p_q_grid,
                                const DataType* __restrict__ p_k_grid,
+                               unsigned short* __restrict__ p_z_grid,
                                const DataType* __restrict__ p_v_grid,
                                const DataType* __restrict__ p_y_grid,
                                const FloatLSE* __restrict__ p_lse_grid,
@@ -1089,6 +1175,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                const CElementwiseOperation& c_element_op,
                                const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
                                const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
+                               const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
                                const VGridDesc_N0_O_N1& v_grid_desc_n0_o_n1,
                                const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock&
                                    y_grid_desc_mblock_mperblock_oblock_operblock,
@@ -1096,8 +1184,16 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                const VGradGridDescriptor_N_O& vgrad_grid_desc_n_o,
                                const YGradGridDesc_M0_O_M1& ygrad_grid_desc_m0_o_m1,
                                const Block2CTileMap& block_2_ctile_map,
-                               const C0MatrixMask& c0_matrix_mask)
+                               const C0MatrixMask& c0_matrix_mask,
+                               const float p_drop,
+                               ck::philox& ph)
     {
+        const FloatGemmAcc p_dropout     = type_convert<FloatGemmAcc>(1.0f - p_drop);
+        const FloatGemmAcc rp_dropout    = type_convert<FloatGemmAcc>(1.0f / p_dropout);
+        const ushort p_dropout_in_16bits = uint16_t(std::floor(p_dropout * 65535.0));
+        const tensor_operation::element_wise::Scale scale_rp_dropout(s_element_op.Value() *
+                                                                     rp_dropout);
+
         const auto q_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_q_grid, q_grid_desc_k0_m_k1.GetElementSpaceSize());
         const auto k_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
@@ -1147,11 +1243,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
 
         // Gemm0: LDS allocation for A and B: be careful of alignment
         auto gemm0_a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<DataType*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            static_cast<LDSDataType*>(p_shared) + SharedMemTrait::a_block_space_offset,
             Gemm0::a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto gemm0_b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<DataType*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            static_cast<LDSDataType*>(p_shared) + SharedMemTrait::b_block_space_offset,
             Gemm0::b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         // Gemm0: gridwise GEMM pipeline
@@ -1243,11 +1339,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                   decltype(s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4())>;
 
         // Gemm1: VGPR allocation for A and LDS allocation for B
-        auto gemm1_a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, DataType>(
+        auto gemm1_a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, LDSDataType>(
             Gemm1::a_thread_desc_k0_m_k1.GetElementSpaceSize());
 
         auto gemm1_b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<DataType*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            static_cast<LDSDataType*>(p_shared) + SharedMemTrait::b1_block_space_offset,
             Gemm1::b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         // dQ: transform input and output tensor descriptors
@@ -1331,6 +1427,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                                   decltype(thread_cluster_desc_m_n),
                                                   decltype(thread_slice_desc_m_n)>{};
 
+        auto blockwise_dropout = BlockwiseDropout<FloatGemmAcc, decltype(thread_slice_desc_m_n)>{
+            p_dropout_in_16bits, rp_dropout};
+
         auto lse_grid_desc_mblock_mrepeat_mwave_mperxdl =
             MakeLSEGridDescriptor_MBlock_MRepeat_NWave_MPerXdl(lse_grid_desc_m);
 
@@ -1360,6 +1459,75 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                  acc0_thread_origin[I2],   // mwave
                                  acc0_thread_origin[I4])}; // mperxdl
 
+        //
+        // z vgpr copy to global
+        //
+        // z matrix threadwise desc
+        constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
+                                                           n4)); // registerNum
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     unsigned short,
+                     z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                     true>
+            z_tenor_buffer;
+        z_tenor_buffer.Clear();
+        // z matrix global desc
+        /*const auto M = q_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N = k_grid_desc_k0_n_k1.GetLength(I1);
+
+        auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(M, N);*/
+
+        auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize());
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+            ushort,
+            ushort,
+            decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<I1, // MBlockId
+                     I1, // NBlockID
+                     m0, // MRepeat
+                     n0, // NRepeat
+                     m1, // MWaveId
+                     n1, // NWaveId
+                     m2, // MPerXdl
+                     n2, // NGroupNum
+                     n3, // NInputNum
+                     n4>,
+            Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+            9, // DstVectorDim
+            1, // DstScalarPerVector
+            InMemoryDataOperationEnum::Set,
+            1, // DstScalarStrideInVector
+            true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                  make_multi_index(block_work_idx[I0], // MBlockId
+                                   0,                  // NBlockId
+                                   0,                  // mrepeat
+                                   0,                  // nrepeat
+                                   wave_id[I0],        // MWaveId
+                                   wave_id[I1],        // NWaveId
+                                   wave_m_n_id[I1],    // MPerXdl
+                                   0,                  // group
+                                   wave_m_n_id[I0],    // NInputIndex
+                                   0),
+                  tensor_operation::element_wise::PassThrough{}};
+
         //
         // set up dV / dK Gemm (type 3 crr)
         //
@@ -1367,11 +1535,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
 
         // Gemm2: LDS allocation for A and B: be careful of alignment
         auto gemm2_a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<DataType*>(p_shared) + SharedMemTrait::a2_block_space_offset,
+            static_cast<LDSDataType*>(p_shared) + SharedMemTrait::a2_block_space_offset,
             Gemm2::a_block_desc_m0_n_m1.GetElementSpaceSize());
 
         auto gemm2_b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<DataType*>(p_shared) + SharedMemTrait::b2_block_space_offset,
+            static_cast<LDSDataType*>(p_shared) + SharedMemTrait::b2_block_space_offset,
             Gemm2::b_block_desc_m0_o_m1.GetElementSpaceSize());
 
         // dV: transform input and output tensor descriptors
@@ -1379,10 +1547,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             Gemm2::MakeCGridDesc_N0_O0_N1_O1_N2_O2_O3_O4(vgrad_grid_desc_n_o);
 
         // dV: A matrix VGPR-to-LDS blockwise copy
-        auto vgrad_gemm_tile_p_thread_copy_vgpr_to_lds = typename Gemm2::ABlockwiseCopy{
-            Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
-            Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
-            tensor_operation::element_wise::PassThrough{}};
+        auto vgrad_gemm_tile_p_thread_copy_vgpr_to_lds =
+            typename Gemm2::template ABlockwiseCopy<tensor_operation::element_wise::Relu>{
+                Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
+                tensor_operation::element_wise::Relu{}}; // relu(P-dropped)
 
         // dV: B matrix global-to-LDS blockwise copy
         auto vgrad_gemm_tile_ygrad_blockwise_copy =
@@ -1407,11 +1576,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             make_multi_index(
                 I0, block_work_idx[I1] * Gemm2Params_N_O_M::GemmORepeat, I0, I0, I0, I0, I0, I0);
 
-        auto vgrad_thread_copy_vgpr_to_global = typename Gemm2::template CBlockwiseCopy<decltype(
-            vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4)>(
-            vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
-            vgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
-            tensor_operation::element_wise::PassThrough{});
+        auto vgrad_thread_copy_vgpr_to_global =
+            typename Gemm2::template CBlockwiseCopy<decltype(
+                                                        vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4),
+                                                    tensor_operation::element_wise::Scale>(
+                vgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                vgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
+                tensor_operation::element_wise::Scale{rp_dropout});
 
         // dK: transform input and output tensor descriptors
         const auto q_grid_desc_m0_k_m1 =
@@ -1422,10 +1593,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             Gemm2::MakeCGridDesc_N0_O0_N1_O1_N2_O2_O3_O4(kgrad_grid_desc_n_k);
 
         // dK: A matrix VGPR-to-LDS blockwise copy
-        auto kgrad_gemm_tile_sgrad_thread_copy_vgpr_to_lds = typename Gemm2::ABlockwiseCopy{
-            Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
-            Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
-            tensor_operation::element_wise::PassThrough{}};
+        auto kgrad_gemm_tile_sgrad_thread_copy_vgpr_to_lds =
+            typename Gemm2::template ABlockwiseCopy<tensor_operation::element_wise::PassThrough>{
+                Gemm2::a_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                Gemm2::MakeAThreadOriginOnBlock_M0_N0_M1_N1_M2_N2_N3_N4(),
+                tensor_operation::element_wise::PassThrough{}};
 
         // dK: B matrix global-to-LDS blockwise copy
         auto kgrad_gemm_tile_q_blockwise_copy =
@@ -1450,9 +1622,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
 
         auto kgrad_thread_copy_vgpr_to_global = typename Gemm2::template CBlockwiseCopy<
             decltype(kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4),
-            decltype(s_element_op)>(kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
-                                    kgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
-                                    s_element_op);
+            decltype(scale_rp_dropout)>(kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4,
+                                        kgrad_thread_origin_on_grid_n0_o0_n1_o1_n2_o2_o3_o4,
+                                        scale_rp_dropout);
 
         //
         // set up Y dot dY
@@ -1487,7 +1659,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         // performs double duty for both y and ygrad
         auto yygrad_threadwise_copy = ThreadwiseTensorSliceTransfer_v2<
             DataType,
-            DataType,
+            FloatGemmAcc,
             YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
             decltype(y_thread_desc_m0_m1_o0_o1),
             decltype(y_thread_desc_m0_m1_o0_o1.GetLengths()),
@@ -1496,8 +1668,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             YDotYGrad_M_O::SrcScalarPerVector, // SrcScalarPerVector
             1,                                 // SrcScalarStrideInVector
             true /* ResetCoordAfterRun */,
-            true /* InvalidElementAsNaN */>(y_grid_desc_mblock_mperblock_oblock_operblock,
-                                            y_thread_data_on_grid_idx);
+            false /* InvalidElementAsNaN */>(y_grid_desc_mblock_mperblock_oblock_operblock,
+                                             y_thread_data_on_grid_idx);
 
         auto y_thread_buf                 = typename YDotYGrad_M_O::SrcBufType{};
         auto ygrad_thread_buf             = typename YDotYGrad_M_O::SrcBufType{};
@@ -1574,7 +1746,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
         static_for<0, YDotYGrad_M_O::ThreadSliceLength_M, 1>{}([&](auto iM) {
             const auto idx_on_block = y_thread_data_on_block_idx[I1] + iM;
             y_dot_ygrad_block_accum_buf.AtomicAdd(
-                idx_on_block, true, y_dot_ygrad_thread_accum_buf[iM]);
+                idx_on_block, true, y_dot_ygrad_thread_accum_buf[iM] * p_dropout); // p_dropoutD1
         });
         block_sync_lds();
 
@@ -1681,8 +1853,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             }
             else
             {
-                static_for<0, s_slash_p_thread_buf.Size(), 1>{}(
-                    [&](auto i) { s_element_op(acc_thread_buf(i), s_slash_p_thread_buf[i]); });
+                static_for<0, s_slash_p_thread_buf.Size(), 1>{}([&](auto i) {
+                    s_element_op(s_slash_p_thread_buf(i), s_slash_p_thread_buf[i]);
+                });
             }
 
             block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
@@ -1691,6 +1864,28 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             // scaling is already performed in the preceding statements with s_element_op
             blockwise_softmax.RunWithPreCalcStats(s_slash_p_thread_buf, lse_thread_buf);
 
+            // save z to global
+            if(p_z_grid)
+            {
+                // P_dropped
+                blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf),
+                                                        decltype(z_tenor_buffer),
+                                                        true>(
+                    s_slash_p_thread_buf, ph, z_tenor_buffer);
+
+                z_thread_copy_vgpr_to_global.Run(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                 make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                                 z_tenor_buffer,
+                                                 z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                 z_grid_buf);
+            }
+            else
+            {
+                // P_dropped
+                blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>(
+                    s_slash_p_thread_buf, ph);
+            }
+
             block_sync_lds(); // wait for gemm1 LDS read
 
             SubThreadBlock<BlockSize> gemm2_a_copy_subgroup(s_blockwise_gemm.GetWaveIdx()[I0],
@@ -1701,7 +1896,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                           "");
 
             // TODO: tune gemm2 pipeline
-            // dV = P^T * dY
+            // dV = P_drop^T * dY
             v_slash_k_grad_thread_buf.Clear();
             static_for<0, num_gemm2_loop, 1>{}([&](auto gemm2_loop_idx) { // gemm dV
                 // load VGrad Gemm B
@@ -1781,8 +1976,17 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                 constexpr auto m =
                     pgrad_thread_idx_to_m_n_adaptor.CalculateBottomIndex(pgrad_thread_idx)[I0];
                 // dS and P has same thread buf layout
-                sgrad_thread_buf(i) = s_slash_p_thread_buf[i] *
-                                      (pgrad_thread_buf[i] - y_dot_ygrad_thread_buf[Number<m>{}]);
+                if(s_slash_p_thread_buf[i] >= 0)
+                {
+                    sgrad_thread_buf(i) =
+                        s_slash_p_thread_buf[i] *
+                        (pgrad_thread_buf[i] - y_dot_ygrad_thread_buf[Number<m>{}]);
+                }
+                else
+                {
+                    sgrad_thread_buf(i) =
+                        s_slash_p_thread_buf[i] * y_dot_ygrad_thread_buf[Number<m>{}];
+                }
             });
 
             // gemm dQ
@@ -1922,6 +2126,10 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
             kgrad_thread_copy_vgpr_to_global.MoveDstSliceWindow(
                 kgrad_grid_desc_n0_o0_n1_o1_n2_o2_o3_o4, Gemm2::c_block_slice_copy_step); // step N
 
+            z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+
         } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
 
         // shuffle dQ and write
@@ -2035,7 +2243,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle
                                      n_thread_data_on_block_idx[I2],
                                      n_thread_data_on_block_idx[I3],
                                      n_thread_data_on_block_idx[I4]),
-                    s_element_op};
+                    scale_rp_dropout};
 
             // shuffle: blockwise copy C from LDS to global
             auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/philox_rand.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/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.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/blockwise_softmax.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_dropout.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatLSE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename ZGridDesc_M_N,
+          typename LSEGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    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 WaveSize = 64;
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<decltype(
+        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    // C desc for source in blockwise copy
+    __host__ __device__ static constexpr auto MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+        const ZGridDesc_M_N& z_grid_desc_m_n) ////=> for z use
+    {
+        const auto M = z_grid_desc_m_n.GetLength(I0);
+        const auto N = z_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            z_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        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;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(Gemm1N != K)
+        {
+            std::cout << "SizeK must be equal to SizeO (equal attention head size)" << '\n';
+            return false;
+        }
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // 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 GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        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<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeLSEGridDescriptor_MBlock_MRepeat_NWave_MPerXdl(const LSEGridDesc_M& lse_grid_desc_m)
+    {
+        const index_t M         = lse_grid_desc_m.GetLength(I0);
+        const index_t MBlock    = M / MPerBlock;
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        const auto lse_grid_desc_mblock_mrepeat_mwave_mperxdl = transform_tensor_descriptor(
+            lse_grid_desc_m,
+            make_tuple(make_unmerge_transform(
+                make_tuple(MBlock, Number<MXdlPerWave>{}, MWave, Number<MPerXdl>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2, 3>{}));
+
+        return lse_grid_desc_mblock_mrepeat_mwave_mperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
+        MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop,
+              bool IsDropout,
+              bool IsLseStoring,
+              typename Block2CTileMap,
+              typename C0MatrixMask>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               unsigned short* __restrict__ p_z_grid,
+                               FloatLSE* __restrict__ p_lse_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const LSEGridDesc_M& lse_grid_desc_m,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask,
+                               const ushort p_dropout_in_16bits,
+                               FloatGemmAcc p_dropout_rescale,
+                               ck::philox ph)
+    {
+        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, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto lse_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_lse_grid, lse_grid_desc_m.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // 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
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                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),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        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);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+        constexpr auto AccM2 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I4);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        auto blockwise_dropout = BlockwiseDropout<FloatGemmAcc, decltype(thread_slice_desc_m_n)>{
+            p_dropout_in_16bits, p_dropout_rescale};
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        auto lse_grid_desc_mblock_mrepeat_mwave_mperxdl =
+            MakeLSEGridDescriptor_MBlock_MRepeat_NWave_MPerXdl(lse_grid_desc_m);
+
+        constexpr auto lse_thread_desc_mblock_mrepeat_mwave_mperxdl =
+            make_naive_tensor_descriptor_packed(make_tuple(I1, m0, m1, m2));
+
+        auto lse_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatLSE>(
+            lse_thread_desc_mblock_mrepeat_mwave_mperxdl.GetElementSpaceSize());
+
+        auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+            Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+        auto lse_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+            FloatGemmAcc,
+            FloatLSE,
+            decltype(lse_thread_desc_mblock_mrepeat_mwave_mperxdl),
+            decltype(lse_grid_desc_mblock_mrepeat_mwave_mperxdl),
+            ck::tensor_operation::element_wise::PassThrough,
+            Sequence<1, 1, 1, 1>,
+            Sequence<0, 1, 2, 3>,
+            3,
+            1,
+            InMemoryDataOperationEnum::Set,
+            1,
+            false>{lse_grid_desc_mblock_mrepeat_mwave_mperxdl,
+                   make_multi_index(block_work_idx[I0],      // mblock
+                                    0,                       // mrepeat
+                                    acc0_thread_origin[I2],  // mwave
+                                    acc0_thread_origin[I4]), // mperxdl
+                   ck::tensor_operation::element_wise::PassThrough{}};
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+
+        // z is random number matrix for dropout verify
+        //
+        // z vgpr copy to global
+        //
+        // z matrix threadwise desc
+        constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
+                                                           n4)); // registerNum
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     unsigned short,
+                     z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                     true>
+            z_tenor_buffer;
+        z_tenor_buffer.Clear();
+        // z matrix global desc
+
+        auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize());
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+            ushort,
+            ushort,
+            decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<I1, // MBlockId
+                     I1, // NBlockID
+                     m0, // MRepeat
+                     n0, // NRepeat
+                     m1, // MWaveId
+                     n1, // NWaveId
+                     m2, // MPerXdl
+                     n2, // NGroupNum
+                     n3, // NInputNum
+                     n4>,
+            Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+            9,  // DstVectorDim
+            n4, // DstScalarPerVector
+            InMemoryDataOperationEnum::Set,
+            1, // DstScalarStrideInVector
+            true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                  make_multi_index(block_work_idx[I0], // MBlockId
+                                   0,                  // NBlockId
+                                   0,                  // mrepeat
+                                   0,                  // nrepeat
+                                   wave_id[I0],        // MWaveId
+                                   wave_id[I1],        // NWaveId
+                                   wave_m_n_id[I1],    // MPerXdl
+                                   0,                  // group
+                                   wave_m_n_id[I0],    // NInputIndex
+                                   0),
+                  tensor_operation::element_wise::PassThrough{}};
+
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(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_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(c_thread_lengths),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                    else
+                    {
+                        acc_element_op(acc_thread_buf(i), acc_thread_buf[i]);
+                    }
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            if constexpr(IsDropout) // dropout
+            {
+
+                // save z to global
+                if(p_z_grid)
+                {
+                    // P_dropped
+                    blockwise_dropout.template ApplyDropout<decltype(acc_thread_buf),
+                                                            decltype(z_tenor_buffer),
+                                                            false>(
+                        acc_thread_buf, ph, z_tenor_buffer);
+
+                    z_thread_copy_vgpr_to_global.Run(
+                        z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                        make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                        z_tenor_buffer,
+                        z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                        z_grid_buf);
+
+                    z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                        make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+                }
+                else
+                {
+                    // P_dropped
+                    blockwise_dropout.template ApplyDropout<decltype(acc_thread_buf), false>(
+                        acc_thread_buf, ph);
+                }
+            }
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for reduction LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            // workaround compiler issue; see ck/ck.hpp
+            if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
+                         is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
+                         Gemm1NPerBlock == 128)
+            {
+                __builtin_amdgcn_sched_barrier(0);
+            }
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // Calculate max + ln(sum) and write out
+
+        if constexpr(IsLseStoring)
+        {
+            static_for<0, MXdlPerWave, 1>{}(
+                [&](auto I) { lse_thread_buf(I) = running_max(I) + math::log(running_sum(I)); });
+
+            if(get_warp_local_1d_id() < AccM2)
+            {
+                static_for<0, MXdlPerWave, 1>{}([&](auto I) {
+                    // copy from VGPR to Global
+                    lse_thread_copy_vgpr_to_global.Run(lse_thread_desc_mblock_mrepeat_mwave_mperxdl,
+                                                       make_tuple(I0, Number<I>{}, I0, I0),
+                                                       lse_thread_buf,
+                                                       lse_grid_desc_mblock_mrepeat_mwave_mperxdl,
+                                                       lse_grid_buf);
+
+                    lse_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        lse_grid_desc_mblock_mrepeat_mwave_mperxdl, make_multi_index(0, 1, 0, 0));
+                });
+            }
+        }
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_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<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = 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 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.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_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_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<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    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],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.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_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              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
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck

include/ck/utility/data_type.hpp

@@ -1010,6 +1010,42 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float
     return uint16_t(u.int32 >> 16);
 }
 
+// convert fp16 to bf16
+template <>
+inline __host__ __device__ bhalf_t type_convert<bhalf_t, half_t>(half_t x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {static_cast<float>(x)};
+
+    return uint16_t(u.int32 >> 16);
+}
+
+template <>
+inline __host__ __device__ bhalf2_t type_convert<bhalf2_t, half2_t>(half2_t x)
+{
+    float y0{0}, y1{0};
+    bhalf2_t y{0};
+    asm volatile("\n \
+            v_cvt_f32_f16 %0, %1 \n \
+            "
+                 : "=v"(y0)
+                 : "v"(x));
+    asm volatile("\n \
+            v_cvt_f32_f16 %0, %1 src0_sel:WORD_1\n \
+            "
+                 : "=v"(y1)
+                 : "v"(x));
+    asm volatile("\n \
+            v_pack_b32_f16 %0, %1, %2 op_sel:[1, 1] \n \
+            "
+                 : "=v"(y)
+                 : "v"(y0), "v"(y1));
+    return y;
+}
+
 template <typename T>
 struct NumericLimits
 {

include/ck/utility/math.hpp

@@ -168,6 +168,22 @@ __device__ double exp<double>(double x)
     return exp(x);
 }
 
+// disallow implicit type casting
+template <typename T>
+__device__ T log(T x);
+
+template <>
+__device__ float log<float>(float x)
+{
+    return __logf(x);
+}
+
+template <>
+__device__ double log<double>(double x)
+{
+    return log(x);
+}
+
 // greatest common divisor, aka highest common factor
 __host__ __device__ constexpr index_t gcd(index_t x, index_t y)
 {

include/ck/utility/math_v2.hpp

@@ -3,7 +3,9 @@
 
 #pragma once
 
+#ifndef __HIP_DEVICE_COMPILE__
 #include <cmath>
+#endif
 
 #include "ck/utility/data_type.hpp"
 #include "ck/utility/type.hpp"

include/ck/utility/philox_rand.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+class philox
+{
+    public:
+    __device__ inline philox(unsigned long long seed,
+                             unsigned long long subsequence,
+                             unsigned long long offset)
+        : h_seed(reinterpret_cast<const uint2&>(seed))
+    {
+
+        ull2* tmp = reinterpret_cast<ull2*>(&counter);
+        tmp->x    = offset / 4;
+        tmp->y    = subsequence;
+    }
+
+    __device__ inline uint4 get_philox_4x32()
+    {
+
+        uint4 counter_ = counter;
+        uint2 key_     = h_seed;
+// 7-round philox
+#pragma unroll
+        for(int i = 0; i < 6; i++)
+        {
+            counter_ = single_loop(counter_, key_);
+            key_.x += kPhilox10A;
+            key_.y += kPhilox10B;
+        }
+        uint4 output = single_loop(counter_, key_);
+        incr();
+
+        return output;
+    }
+
+    __device__ inline uint4 get_philox_4x32(const unsigned long long subsequence)
+    {
+
+        uint4 counter_ = counter;
+        ull2* tmp      = reinterpret_cast<ull2*>(&counter_);
+        tmp->y         = subsequence;
+
+        uint2 key_ = h_seed;
+// 7-round philox
+#pragma unroll
+        for(int i = 0; i < 6; i++)
+        {
+            counter_ = single_loop(counter_, key_);
+            key_.x += kPhilox10A;
+            key_.y += kPhilox10B;
+        }
+        uint4 output = single_loop(counter_, key_);
+        return output;
+    }
+
+    __device__ void get_random_8x16(ushort* out)
+    {
+        uint4 tmp_ph;
+        tmp_ph = get_philox_4x32();
+
+        uint32_t* out_tmp = reinterpret_cast<uint32_t*>(&out[0]);
+
+        out_tmp[0] = tmp_ph.x;
+        out_tmp[1] = tmp_ph.y;
+        out_tmp[2] = tmp_ph.z;
+        out_tmp[3] = tmp_ph.w;
+    }
+
+    __device__ void get_random_8x16(ushort* out, const unsigned long long subsequence)
+    {
+        uint4 tmp_ph;
+        tmp_ph = get_philox_4x32(subsequence);
+
+        uint32_t* out_tmp = reinterpret_cast<uint32_t*>(&out[0]);
+
+        out_tmp[0] = tmp_ph.x;
+        out_tmp[1] = tmp_ph.y;
+        out_tmp[2] = tmp_ph.z;
+        out_tmp[3] = tmp_ph.w;
+    }
+
+    private:
+    struct ull2
+    {
+        uint64_t x;
+        uint64_t y;
+    };
+    uint4 counter;
+    const uint2 h_seed;
+
+    __device__ uint4 incr(uint4 ctr)
+    {
+
+        uint4 res;
+        res.x = ctr.x + 1;
+        res.y = ctr.y;
+        res.z = ctr.z;
+        res.w = ctr.w;
+        return res;
+    }
+
+    __device__ inline void incr() { counter = incr(counter); }
+
+    __device__ uint2 u32_high_low_multi(const unsigned int a, const unsigned int b)
+    {
+        uint2* res;
+        unsigned long long tmp;
+        tmp = static_cast<unsigned long long>(a) * b;
+        res = reinterpret_cast<uint2*>(&tmp);
+        return *res;
+    }
+
+    __device__ inline uint4 single_loop(const uint4 ctr, const uint2 i_key)
+    {
+
+        uint2 res0 = u32_high_low_multi(kPhiloxSA, ctr.x);
+        uint2 res1 = u32_high_low_multi(kPhiloxSB, ctr.z);
+        uint4 ret  = {res1.y ^ ctr.y ^ i_key.x, res1.x, res0.y ^ ctr.w ^ i_key.y, res0.x};
+        return ret;
+    }
+
+    static const unsigned long kPhilox10A = 0x9E3779B9;
+    static const unsigned long kPhilox10B = 0xBB67AE85;
+    static const unsigned long kPhiloxSA  = 0xD2511F53;
+    static const unsigned long kPhiloxSB  = 0xCD9E8D57;
+};
+
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename RefDataType, typename InDataType, typename OutDataType>
+struct ReferenceDropout : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<RefDataType>& ref,
+                 const Tensor<InDataType>& in,
+                 Tensor<OutDataType>& out,
+                 RefDataType p_dropout_in_16bits,
+                 float rp_dropout)
+            : ref_(ref),
+              in_(in),
+              out_(out),
+              p_dropout_in_16bits_(p_dropout_in_16bits),
+              rp_dropout_(rp_dropout)
+        {
+        }
+        const Tensor<RefDataType>& ref_;
+        const Tensor<InDataType>& in_;
+        Tensor<OutDataType>& out_;
+        RefDataType p_dropout_in_16bits_;
+        float rp_dropout_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            arg.out_.ForEach([&](auto& self, auto idx) {
+                self(idx) =
+                    arg.ref_(idx) <= arg.p_dropout_in_16bits_
+                        ? ck::type_convert<OutDataType>(ck::type_convert<float>(arg.in_(idx)) *
+                                                        ck::type_convert<float>(arg.rp_dropout_))
+                        : 0;
+            });
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<RefDataType>& ref,
+                             const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             RefDataType p_dropout_in_16bits,
+                             float rp_dropout)
+    {
+        return Argument{ref, in, out, p_dropout_in_16bits, rp_dropout};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceDropout"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

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

@@ -149,13 +149,6 @@ struct ReferenceSoftmax : public device::BaseOperator
                                                   ck::type_convert<AccDataType>(
                                                       arg.sm_stats_ptr_[0](to_sm_stats_idx(idx)))) +
                             arg.beta_ * self(idx);
-                // printf(
-                //     "exponent %f, exp() = %f\n",
-                //     ck::type_convert<AccDataType>(arg.in_(idx)) -
-                //         ck::type_convert<AccDataType>(arg.sm_stats_ptr_[0](to_sm_stats_idx(idx))),
-                //     std::exp(
-                //         ck::type_convert<AccDataType>(arg.in_(idx)) -
-                //         ck::type_convert<AccDataType>(arg.sm_stats_ptr_[0](to_sm_stats_idx(idx)))));
             });
 
             return 0;