attn bwd prototype1 bf16

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

@@ -11,6 +11,7 @@ add_example_executable(example_grouped_multihead_attention_forward_bf16 grouped_
 add_example_executable(example_batched_multihead_attention_forward_bf16 batched_multihead_attention_forward_bf16.cpp)
 add_example_executable(example_batched_multihead_attention_backward_fp16 batched_multihead_attention_backward_fp16.cpp)
 add_example_executable(example_batched_multihead_attention_backward_pt1_fp16 batched_multihead_attention_backward_pt1_fp16.cpp)
+add_example_executable(example_batched_multihead_attention_backward_pt1_bf16 batched_multihead_attention_backward_pt1_bf16.cpp)
 
 add_custom_target(example_gemm_scale_softmax_gemm)
 add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)

example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_fp16.cpp

@@ -24,7 +24,7 @@ Kernel outputs:
 */
 
 #define PRINT_HOST 0
-#define USING_MASK 1
+#define USING_MASK 0
 
 #include <iostream>
 #include <numeric>
@@ -268,8 +268,8 @@ int run(int argc, char* argv[])
     ck::index_t N  = 512;
     ck::index_t K  = 128;
     ck::index_t O  = 128;
-    ck::index_t G0 = 3;
-    ck::index_t G1 = 2;
+    ck::index_t G0 = 54;
+    ck::index_t G1 = 16;
 
     float alpha = 1.f / std::sqrt(K);
 

example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_pt1_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+/*
+Backprop for Gemm + Softmax + Gemm fused operation, where forward prop is defined as:
+
+  Y_g_m_o = Softmax(alpha * Q_g_m_k * K_g_k_n) * V_g_n_o
+
+Computation graph:
+
+          K^T                   V
+          |                     |
+          |                     |
+    Q --- * ----- Softmax ----- * --> Y
+              S             P
+
+Kernel inputs:
+
+    Q, K, V, Y, dY, per-row softmax stats (LSE)
+
+Kernel outputs:
+
+    dQ, dK, dV
+
+*/
+
+#define PRINT_HOST 0
+#define USING_MASK 0
+#define USING_HD32 0
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_pt1_bf16v1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_dropout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32 = float;
+using U16 = unsigned short;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using QKVElementOp = PassThrough;
+using YElementOp   = PassThrough;
+
+using DataType         = BF16;
+using AccDataType      = F32;
+using ShuffleDataType  = F32;
+using LSEDataType      = F32;
+using ZDataType        = U16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+#if USING_MASK
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+#else
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+#endif
+
+static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// Headdim/K/O should be a multiple of 8, and it's only supported up to 64 in prototype1.
+// If Headdim/K/O <= 32, ues 1st template.
+// If 32 < Headdim/K/O <= 64, ues 2nd template.
+
+#if USING_HD32
+// 1st template
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        DataType,
+        ZDataType,
+        LSEDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        ShuffleDataType,
+        QKVElementOp,
+        QKVElementOp,
+        Scale,
+        QKVElementOp,
+        YElementOp,
+        GemmSpec,
+        TensorSpecQ,
+        TensorSpecK,
+        TensorSpecV,
+        TensorSpecY,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        32,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        1,           // Gemm1NXdlPerWave
+        1,           // Gemm2NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        1,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+#else
+//2nd template
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        DataType,
+        ZDataType,
+        LSEDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        ShuffleDataType,
+        QKVElementOp,
+        QKVElementOp,
+        Scale,
+        QKVElementOp,
+        YElementOp,
+        GemmSpec,
+        TensorSpecQ,
+        TensorSpecK,
+        TensorSpecV,
+        TensorSpecY,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        64,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        2,           // Gemm2NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+#endif
+
+// Ref Gemm0: S = alpha * Q * K^T
+// fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType,
+                                                                                DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                Scale>;
+
+// Ref Softmax: P = Softmax(S)
+// fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, DataType, AccDataType>;
+
+// Ref Gemm1: Y = P * V
+// fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType,
+                                                                                DataType,
+                                                                                DataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+
+// Ref Gemm for backward pass
+// fp16 in, fp16 out
+using ReferenceGemmGradInstance = ck::tensor_operation::host::ReferenceBatchedGemm<DataType,
+                                                                                   DataType,
+                                                                                   DataType,
+                                                                                   AccDataType,
+                                                                                   PassThrough,
+                                                                                   PassThrough,
+                                                                                   Scale>;
+// Ref dropout
+using ReferenceDropoutInstance =
+    ck::tensor_operation::host::ReferenceDropout<ushort, DataType, DataType>;
+
+template <typename TensorQ,
+          typename TensorK,
+          typename TensorV,
+          typename TensorS,
+          typename TensorP,
+          typename TensorZ,
+          typename TensorY,
+          typename TensorLSE = TensorP>
+void run_attention_fwd_host(const TensorQ& q_g_m_k,
+                            const TensorK& k_g_n_k,
+                            const TensorV& v_g_n_o,
+                            const float alpha,
+                            TensorS& s_g_m_n,
+                            TensorP& p_g_m_n,
+                            TensorY& y_g_m_o,
+                            TensorLSE& lse_g_m,
+                            TensorP& p_drop_g_m_n,
+                            TensorZ& z_g_m_n,
+                            ushort p_dropout_in_16bits,
+                            float rp_dropout)
+{
+    // S = alpha * Q * K^T
+    auto k_g_k_n            = k_g_n_k.Transpose({0, 2, 1});
+    auto ref_gemm0          = ReferenceGemm0Instance{};
+    auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+    auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+        q_g_m_k, k_g_k_n, s_g_m_n, PassThrough{}, PassThrough{}, Scale{alpha});
+
+    ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+    // masking
+#if USING_MASK
+    auto N          = s_g_m_n.GetLengths()[2];
+    const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+    s_g_m_n.ForEach([&](auto& self, auto idx) {
+        if(mask.IsMaskedElement(idx[1], idx[2]))
+            self(idx) = -ck::NumericLimits<float>::Infinity();
+    });
+#endif
+
+    // P = Softmax(S)
+    auto ref_softmax          = ReferenceSoftmaxInstance{};
+    auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+    auto ref_softmax_argument = ref_softmax.MakeArgument(s_g_m_n, p_g_m_n, 1, 0, {2}, &lse_g_m);
+
+    ref_softmax_invoker.Run(ref_softmax_argument);
+
+    // P_dropped
+    auto ref_dropout         = ReferenceDropoutInstance{};
+    auto ref_dropout_invoker = ref_dropout.MakeInvoker();
+    auto ref_dropout_argment =
+        ref_dropout.MakeArgument(z_g_m_n, p_g_m_n, p_drop_g_m_n, p_dropout_in_16bits, rp_dropout);
+    ref_dropout_invoker.Run(ref_dropout_argment);
+
+    // Y = P_dropout * V
+    auto ref_gemm1          = ReferenceGemm1Instance{};
+    auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+    auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+        p_drop_g_m_n, v_g_n_o, y_g_m_o, PassThrough{}, PassThrough{}, PassThrough{});
+
+    ref_gemm1_invoker.Run(ref_gemm1_argument);
+}
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 2; // method 1 will have slightly higher error; TODO: to investigate
+    bool time_kernel     = true;
+
+    // Overall QKV matrices shape
+    // y_g_m_o = Softmax(alpha * Q_g_m_k * K_g_k_n) * V_g_n_o
+    // y_g0_g1_m_o = reshape(y_g_m_o, [G0, G1, M, O])
+    // y_g0_m_g1_o = permute(y_g0_g1_m_o, [0, 2, 1, 3])
+    ck::index_t M  = 512; // 512
+    ck::index_t N  = 512; // 512
+    ck::index_t K  = 64;
+    ck::index_t O  = 64;
+    ck::index_t G0 = 54; // 54
+    ck::index_t G1 = 16; // 16
+
+    float alpha = 1.f / std::sqrt(K);
+
+    bool input_permute  = false;
+    bool output_permute = false;
+
+    float p_drop                    = 0.2;
+    float p_dropout                 = 1 - p_drop;
+    uint16_t p_dropout_in_16bits    = uint16_t(std::floor(p_dropout * 65535.0));
+    float rp_dropout                = 1.0 / p_dropout;
+    const unsigned long long seed   = 1;
+    const unsigned long long offset = 0;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+
+        input_permute  = std::stoi(argv[11]);
+        output_permute = std::stoi(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        printf("arg11 to 12: input / output permute\n");
+        exit(0);
+    }
+
+    const ck::index_t BatchCount = G0 * G1;
+
+    std::vector<ck::index_t> q_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> q_gs_ms_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // Q layout [G0, M, G1, K]
+            : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // Q layout [G0, G1, M, K]
+
+    std::vector<ck::index_t> k_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> k_gs_ns_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // K layout [G0, N, G1, K]
+            : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // K layout [G0, G1, N, K]
+
+    std::vector<ck::index_t> v_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> v_gs_os_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // V layout [G0, N, G1, O]
+            : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // V layout [G0, G1, N, O]
+
+    std::vector<ck::index_t> y_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> y_gs_ms_os_strides =
+        output_permute
+            ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // Y layout [G0, M, G1, O]
+            : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // Y layout [G0, G1, M, O]
+
+    std::vector<ck::index_t> z_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> z_gs_ms_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
+            : std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
+    // The softmax stat log-sum-exp (LSE) is used to speed up softmax calculation in backward pass
+    // Pi = exp(Si) / sum(exp(S0) + exp(S1) + ...)
+    //    = exp(Si) / exp(log(sum(exp() + ...)))
+    //    = exp(Si - log(sum(exp() + ...)))
+    //               ^^^^^^^^^^^^^^^^^^^^^
+    //                       LSE
+    std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M};
+    std::vector<ck::index_t> lse_gs_ms_strides{G1 * M, M, 1}; // LSE layout [G0, G1, M]
+
+    Tensor<DataType> q_gs_ms_ks(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
+    Tensor<DataType> k_gs_ns_ks(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
+    Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides);
+    Tensor<DataType> v_gs_os_ns(v_gs_os_ns_lengths, v_gs_os_ns_strides);
+    Tensor<DataType> y_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides);
+    Tensor<DataType> ygrad_gs_ms_os(y_gs_ms_os_lengths, y_gs_ms_os_strides);
+    Tensor<LSEDataType> lse_gs_ms(lse_gs_ms_lengths, lse_gs_ms_strides);
+
+    std::cout << "q_gs_ms_ks: " << q_gs_ms_ks.mDesc << std::endl;
+    std::cout << "k_gs_ns_ks: " << k_gs_ns_ks.mDesc << std::endl;
+    std::cout << "z_gs_ms_ks: " << z_gs_ms_ns.mDesc << std::endl;
+    std::cout << "v_gs_os_ns: " << v_gs_os_ns.mDesc << std::endl;
+    std::cout << "y_gs_ms_os: " << y_gs_ms_os.mDesc << std::endl;
+    std::cout << "lse_gs_ms_os: " << lse_gs_ms.mDesc << std::endl;
+
+    z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DataType>{0});
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_2<DataType>{-2, 2});
+        break;
+    case 2:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        break;
+    case 4:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        break;
+    case 5:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<2>{}); // dy[g0, g1, m, o]
+        // dO dot O = [0; 1; 2; ...]
+        break;
+    case 6:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_Sequential<3>{}); // dy[g0, g1, m, o]
+        // assume mnko = 256
+        // P = softmax(QK) = 0.0039 * ones
+        // O = P V = 0.0039 * ones
+        // dP = dO V = [0, 1, 2, ...; 0, 1, 2, ...; ...]
+        // dO dot O = [127.5; ...]
+        // dS = P * (dP - dO dot O)
+        //
+        break;
+    default:
+        q_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<DataType>{1});
+        k_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        v_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<DataType>{});
+        ygrad_gs_ms_os.GenerateTensorValue(GeneratorTensor_1<DataType>{1}); // dy[g0, g1, m, o]
+        // assume mnko = 256
+        // P = softmax(QK) = 0.0039 * ones
+        // O = P V = 0.0039 * ones
+        // dP = dO V = ones
+        // dS = P * (dP - (dO dot O))
+        //    = 0.0039 * ones * (ones - 0.0039*256)
+        //    = 0.0039 * ones * (ones - 1)
+        //    = 0
+    }
+
+    // calculate y & log-sum-exp beforehand
+    Tensor<DataType> q_g_m_k({BatchCount, M, K});
+    Tensor<DataType> k_g_n_k({BatchCount, N, K});
+    Tensor<ZDataType> z_g_m_n({BatchCount, M, N});
+    Tensor<DataType> v_g_n_o({BatchCount, N, O});
+    Tensor<AccDataType> s_g_m_n({BatchCount, M, N});
+    Tensor<DataType> p_g_m_n({BatchCount, M, N});
+    Tensor<DataType> p_drop_g_m_n({BatchCount, M, N});
+    Tensor<DataType> y_g_m_o({BatchCount, M, O});
+    Tensor<LSEDataType> lse_g_m({BatchCount, M});
+
+    q_gs_ms_ks.ForEach(
+        [&](auto& self, auto idx) { q_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); });
+    k_gs_ns_ks.ForEach(
+        [&](auto& self, auto idx) { k_g_n_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); });
+    z_gs_ms_ns.ForEach(
+        [&](auto& self, auto idx) { z_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx); });
+    v_gs_os_ns.ForEach(
+        [&](auto& self, auto idx) { v_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx); });
+    lse_gs_ms.ForEach(
+        [&](auto& self, auto idx) { lse_g_m(idx[0] * G1 + idx[1], idx[2]) = self(idx); });
+
+    run_attention_fwd_host(q_g_m_k,
+                           k_g_n_k,
+                           v_g_n_o,
+                           alpha,
+                           s_g_m_n,
+                           p_g_m_n,
+                           y_g_m_o,
+                           lse_g_m,
+                           p_drop_g_m_n,
+                           z_g_m_n,
+                           p_dropout_in_16bits,
+                           rp_dropout);
+
+    y_gs_ms_os.ForEach(
+        [&](auto& self, auto idx) { self(idx) = y_g_m_o(idx[0] * G1 + idx[1], idx[2], idx[3]); });
+    lse_gs_ms.ForEach(
+        [&](auto& self, auto idx) { self(idx) = lse_g_m(idx[0] * G1 + idx[1], idx[2]); });
+
+    // qkv gradients have the same descriptor as with qkv
+    DeviceMem q_device_buf(sizeof(DataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem k_device_buf(sizeof(DataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem z_device_buf(sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem v_device_buf(sizeof(DataType) * v_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem y_device_buf(sizeof(DataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
+    DeviceMem lse_device_buf(sizeof(LSEDataType) * lse_gs_ms.mDesc.GetElementSpaceSize());
+    DeviceMem qgrad_device_buf(sizeof(DataType) * q_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem kgrad_device_buf(sizeof(DataType) * k_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem vgrad_device_buf(sizeof(DataType) * v_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem ygrad_device_buf(sizeof(DataType) * y_gs_ms_os.mDesc.GetElementSpaceSize());
+
+    q_device_buf.ToDevice(q_gs_ms_ks.mData.data());
+    k_device_buf.ToDevice(k_gs_ns_ks.mData.data());
+    z_device_buf.ToDevice(z_gs_ms_ns.mData.data());
+    v_device_buf.ToDevice(v_gs_os_ns.mData.data());
+    y_device_buf.ToDevice(y_gs_ms_os.mData.data());
+    lse_device_buf.ToDevice(lse_gs_ms.mData.data());
+    ygrad_device_buf.ToDevice(ygrad_gs_ms_os.mData.data());
+    kgrad_device_buf.SetZero();
+    vgrad_device_buf.SetZero();
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    // get z matrix
+    {
+        auto argument = gemm.MakeArgument(
+            static_cast<DataType*>(q_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(k_device_buf.GetDeviceBuffer()),
+            static_cast<ZDataType*>(z_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(v_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(y_device_buf.GetDeviceBuffer()),
+            static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(ygrad_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(qgrad_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(kgrad_device_buf.GetDeviceBuffer()),
+            static_cast<DataType*>(vgrad_device_buf.GetDeviceBuffer()),
+            {}, // std::array<void*, 1> p_acc0_biases;
+            {}, // std::array<void*, 1> p_acc1_biases;
+            q_gs_ms_ks_lengths,
+            q_gs_ms_ks_strides,
+            k_gs_ns_ks_lengths,
+            k_gs_ns_ks_strides,
+            z_gs_ms_ns_lengths,
+            z_gs_ms_ns_strides,
+            v_gs_os_ns_lengths,
+            v_gs_os_ns_strides,
+            y_gs_ms_os_lengths,
+            y_gs_ms_os_strides,
+            lse_gs_ms_lengths,
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            QKVElementOp{},
+            QKVElementOp{},
+            Scale{alpha},
+            QKVElementOp{},
+            YElementOp{},
+            p_drop,
+            std::tuple<unsigned long long, unsigned long long>(seed, offset));
+
+        if(!gemm.IsSupportedArgument(argument))
+        {
+            std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+            return 0;
+        }
+        invoker.Run(argument, StreamConfig{nullptr, false});
+    }
+    // not need output z matrix
+    auto argument = gemm.MakeArgument(
+        static_cast<DataType*>(q_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(k_device_buf.GetDeviceBuffer()),
+        static_cast<ZDataType*>(nullptr), // set to nullptr
+        static_cast<DataType*>(v_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(y_device_buf.GetDeviceBuffer()),
+        static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(ygrad_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(qgrad_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(kgrad_device_buf.GetDeviceBuffer()),
+        static_cast<DataType*>(vgrad_device_buf.GetDeviceBuffer()),
+        {}, // std::array<void*, 1> p_acc0_biases;
+        {}, // std::array<void*, 1> p_acc1_biases;
+        q_gs_ms_ks_lengths,
+        q_gs_ms_ks_strides,
+        k_gs_ns_ks_lengths,
+        k_gs_ns_ks_strides,
+        z_gs_ms_ns_lengths,
+        z_gs_ms_ns_strides,
+        v_gs_os_ns_lengths,
+        v_gs_os_ns_strides,
+        y_gs_ms_os_lengths,
+        y_gs_ms_os_strides,
+        lse_gs_ms_lengths,
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+        QKVElementOp{},
+        QKVElementOp{},
+        Scale{alpha},
+        QKVElementOp{},
+        YElementOp{},
+        p_drop,
+        std::tuple<unsigned long long, unsigned long long>(seed, offset));
+    kgrad_device_buf.SetZero(); // reset global accum buffer and rerun
+    vgrad_device_buf.SetZero();
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    // 5 GEMM ops in total:
+    // S_MNK / dP_MNO Gemm (Gemm0 rcr)
+    // dQ_MKN Gemm (Gemm1 rrr)
+    // dV_NOM / dK_NKM Gemm (Gemm2 crr)
+    // 3x MNK + 2x MNO
+    std::size_t flop = (size_t(3) * M * N * K + size_t(2) * M * N * O) * 2 * BatchCount;
+    // Q/K/V/Y, dQ/dK/dV/dY, LSE
+    std::size_t num_btype = (sizeof(DataType) * M * K + sizeof(DataType) * K * N +
+                             sizeof(DataType) * N * O + sizeof(DataType) * M * O) *
+                                size_t(2) * BatchCount +
+                            sizeof(LSEDataType) * M * BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    // copy z matirx data form device
+    z_device_buf.FromDevice(z_g_m_n.mData.data());
+
+    //       std::cout << "z_g_m_n ref:\n" << z_g_m_n;
+    bool pass = true;
+    if(do_verification)
+    {
+        // run fwd again for y, cause z_g_m_n update
+        run_attention_fwd_host(q_g_m_k,
+                               k_g_n_k,
+                               v_g_n_o,
+                               alpha,
+                               s_g_m_n,
+                               p_g_m_n,
+                               y_g_m_o,
+                               lse_g_m,
+                               p_drop_g_m_n,
+                               z_g_m_n,
+                               p_dropout_in_16bits,
+                               rp_dropout);
+        y_gs_ms_os.ForEach([&](auto& self, auto idx) {
+            self(idx) = y_g_m_o(idx[0] * G1 + idx[1], idx[2], idx[3]);
+        });
+        y_device_buf.ToDevice(y_gs_ms_os.mData.data());
+
+        // call kernel again
+        kgrad_device_buf.SetZero(); // reset global accum buffer and rerun
+        vgrad_device_buf.SetZero();
+        invoker.Run(argument, StreamConfig{nullptr, false});
+
+        Tensor<DataType> qgrad_g_m_k({BatchCount, M, K});
+        Tensor<DataType> kgrad_g_n_k({BatchCount, N, K});
+        Tensor<DataType> vgrad_g_n_o({BatchCount, N, O});
+        Tensor<DataType> sgrad_g_m_n({BatchCount, M, N});
+        Tensor<DataType> pgrad_g_m_n({BatchCount, M, N});
+        Tensor<DataType> pgrad_drop_g_m_n({BatchCount, M, N});
+        Tensor<DataType> ygrad_g_m_o({BatchCount, M, O});
+        Tensor<DataType> ygrad_dot_y_g_m({BatchCount, M});
+
+        ygrad_gs_ms_os.ForEach([&](auto& self, auto idx) {
+            ygrad_g_m_o(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+#if PRINT_HOST
+        {
+            std::cout << "q_g_m_k ref:\n" << q_g_m_k;
+            std::cout << "k_g_n_k ref:\n" << k_g_n_k;
+            std::cout << "v_g_n_o ref:\n" << v_g_n_o;
+            std::cout << "ygrad_g_m_o ref:\n" << ygrad_g_m_o;
+        }
+#endif
+
+        // Gradients
+        auto ref_gemm_grad         = ReferenceGemmGradInstance{};
+        auto ref_gemm_grad_invoker = ref_gemm_grad.MakeInvoker();
+        using RefGemmGradArg       = ReferenceGemmGradInstance::Argument;
+
+        // dP_dropout = dY * V^T
+        auto v_g_o_n = v_g_n_o.Transpose({0, 2, 1});
+        ref_gemm_grad_invoker.Run(RefGemmGradArg{
+            ygrad_g_m_o, v_g_o_n, pgrad_drop_g_m_n, PassThrough{}, PassThrough{}, Scale{1.f}});
+#if PRINT_HOST
+        {
+            std::cout << "===== dP = dY * V^T\n";
+            std::cout << "ygrad_g_m_o ref:\n" << ygrad_g_m_o;
+            std::cout << "v_g_o_n ref:\n" << v_g_o_n;
+            std::cout << "pgrad_drop_g_m_n ref:\n" << pgrad_drop_g_m_n;
+        }
+#endif
+        // dP = dP_dropout x Z
+        auto ref_dropout         = ReferenceDropoutInstance{};
+        auto ref_dropout_invoker = ref_dropout.MakeInvoker();
+        auto ref_dropout_argment = ref_dropout.MakeArgument(
+            z_g_m_n, pgrad_drop_g_m_n, pgrad_g_m_n, p_dropout_in_16bits, rp_dropout);
+        ref_dropout_invoker.Run(ref_dropout_argment);
+
+        // dS_i_j = P_i_j .* (dP_i_j - dY_i dot Y_i)
+        sgrad_g_m_n.ForEach([&](auto& self, auto idx_gmn) {
+            float ygrad_dot_y = 0;
+            for(int o = 0; o < O; o++)
+            {
+                auto idx_gmo = idx_gmn;
+                idx_gmo[2]   = o;
+                ygrad_dot_y += ck::type_convert<AccDataType>(ygrad_g_m_o(idx_gmo)) * ck::type_convert<AccDataType>(y_g_m_o(idx_gmo));
+            }
+            self(idx_gmn) = ck::type_convert<DataType>(ck::type_convert<AccDataType>(p_g_m_n(idx_gmn)) * (ck::type_convert<AccDataType>(pgrad_g_m_n(idx_gmn)) - ygrad_dot_y));
+        });
+#if PRINT_HOST
+        {
+            std::cout << "===== dS_i_j = P_i_j .* (dP_i_j - dY_i dot Y_i)\n";
+            std::cout << "p_g_m_n ref:\n" << p_g_m_n;
+            std::cout << "pgrad_g_m_n ref:\n" << pgrad_g_m_n;
+            std::cout << "y_g_m_o ref:\n" << y_g_m_o;
+            std::cout << "ygrad_g_m_o ref:\n" << ygrad_g_m_o;
+            std::cout << "sgrad_g_m_n ref:\n" << sgrad_g_m_n;
+        }
+#endif
+        // dV = P_drop^T * dY
+        auto p_drop_g_n_m = p_drop_g_m_n.Transpose({0, 2, 1});
+        ref_gemm_grad_invoker.Run(RefGemmGradArg{
+            p_drop_g_n_m, ygrad_g_m_o, vgrad_g_n_o, PassThrough{}, PassThrough{}, Scale{1.0f}});
+#if PRINT_HOST
+        {
+            std::cout << "===== dV = P^T * dY\n";
+            std::cout << "p_drop_g_n_m ref:\n" << p_drop_g_n_m;
+            std::cout << "ygrad_g_m_o ref:\n" << ygrad_g_m_o;
+            std::cout << "vgrad_g_n_o ref:\n" << vgrad_g_n_o;
+        }
+#endif
+
+        // dQ = alpha * dS * K
+        ref_gemm_grad_invoker.Run(RefGemmGradArg{
+            sgrad_g_m_n, k_g_n_k, qgrad_g_m_k, PassThrough{}, PassThrough{}, Scale{alpha}});
+#if PRINT_HOST
+        {
+            std::cout << "===== dQ = alpha * dS * K\n";
+            std::cout << "sgrad_g_m_n ref:\n" << sgrad_g_m_n;
+            std::cout << "k_g_n_k ref:\n" << k_g_n_k;
+            std::cout << "qgrad_g_m_k ref:\n" << qgrad_g_m_k;
+        }
+#endif
+
+        // dK = alpha * dS^T * Q
+        auto sgrad_g_n_m = sgrad_g_m_n.Transpose({0, 2, 1});
+        ref_gemm_grad_invoker.Run(RefGemmGradArg{
+            sgrad_g_n_m, q_g_m_k, kgrad_g_n_k, PassThrough{}, PassThrough{}, Scale{alpha}});
+#if PRINT_HOST
+        {
+            std::cout << "===== dK = alpha * dS^T * Q\n";
+            std::cout << "sgrad_g_n_m ref:\n" << sgrad_g_n_m;
+            std::cout << "q_g_m_k ref:\n" << q_g_m_k;
+            std::cout << "kgrad_g_n_k ref:\n" << kgrad_g_n_k;
+        }
+#endif
+
+        Tensor<DataType> qgrad_gs_ms_ks_host_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
+        Tensor<DataType> kgrad_gs_ns_ks_host_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
+        Tensor<DataType> vgrad_gs_os_ns_host_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
+
+        Tensor<DataType> qgrad_gs_ms_ks_device_result(q_gs_ms_ks_lengths, q_gs_ms_ks_strides);
+        Tensor<DataType> kgrad_gs_ns_ks_device_result(k_gs_ns_ks_lengths, k_gs_ns_ks_strides);
+        Tensor<DataType> vgrad_gs_os_ns_device_result(v_gs_os_ns_lengths, v_gs_os_ns_strides);
+
+        qgrad_device_buf.FromDevice(qgrad_gs_ms_ks_device_result.mData.data());
+        kgrad_device_buf.FromDevice(kgrad_gs_ns_ks_device_result.mData.data());
+        vgrad_device_buf.FromDevice(vgrad_gs_os_ns_device_result.mData.data());
+
+        // permute
+        qgrad_gs_ms_ks_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = qgrad_g_m_k(g, idx[2], idx[3]);
+        });
+        kgrad_gs_ns_ks_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = kgrad_g_n_k(g, idx[2], idx[3]);
+        });
+        vgrad_gs_os_ns_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = vgrad_g_n_o(g, idx[3], idx[2]);
+        });
+
+        std::cout << "Checking qgrad:\n";
+        pass &= ck::utils::check_err(qgrad_gs_ms_ks_device_result.mData,
+                                     qgrad_gs_ms_ks_host_result.mData,
+                                     "error",
+                                     1e-2,
+                                     1e-2);
+        std::cout << "Checking kgrad:\n";
+        pass &= ck::utils::check_err(kgrad_gs_ns_ks_device_result.mData,
+                                     kgrad_gs_ns_ks_host_result.mData,
+                                     "error",
+                                     1e-2,
+                                     1e-2);
+        std::cout << "Checking vgrad:\n";
+        pass &= ck::utils::check_err(vgrad_gs_os_ns_device_result.mData,
+                                     vgrad_gs_os_ns_host_result.mData,
+                                     "error",
+                                     1e-2,
+                                     1e-2);
+    }
+
+    return pass ? ((void)(std::cout << "pass\n"), 0) : ((void)(std::cout << "fail\n"), 1);
+}
+
+int main(int argc, char* argv[]) { return run(argc, argv); }

example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_pt1_fp16.cpp

@@ -25,6 +25,7 @@ Kernel outputs:
 
 #define PRINT_HOST 0
 #define USING_MASK 0
+#define USING_HD32 0
 
 #include <iostream>
 #include <numeric>
@@ -86,6 +87,80 @@ static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpeciali
 static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
 static constexpr auto TensorSpecY = ck::tensor_operation::device::TensorSpecialization::Default;
 
+// Headdim/K/O should be a multiple of 8, and it's only supported up to 64 in prototype1.
+// If Headdim/K/O <= 32, ues 1st template.
+// If 32 < Headdim/K/O <= 64, ues 2nd template.
+
+#if USING_HD32
+// 1st template
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        DataType,
+        ZDataType,
+        LSEDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        ShuffleDataType,
+        QKVElementOp,
+        QKVElementOp,
+        Scale,
+        QKVElementOp,
+        YElementOp,
+        GemmSpec,
+        TensorSpecQ,
+        TensorSpecK,
+        TensorSpecV,
+        TensorSpecY,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        32,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        1,           // Gemm1NXdlPerWave
+        1,           // Gemm2NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        1,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+#else
+//2nd template
 using DeviceGemmInstance =
     ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1<
         NumDimG,
@@ -125,6 +200,7 @@ using DeviceGemmInstance =
         1,           // MXdlPerWave
         4,           // NXdlPerWave
         2,           // Gemm1NXdlPerWave
+        2,           // Gemm2NXdlPerWave
         S<4, 64, 1>, // ABlockTransfer
         S<1, 0, 2>,
         S<1, 0, 2>,
@@ -151,6 +227,7 @@ using DeviceGemmInstance =
         S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
         8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
         MaskingSpec>;   // MaskingSpecialization
+#endif
 
 // Ref Gemm0: S = alpha * Q * K^T
 // fp16 in, fp32 out

include/ck/ck.hpp

@@ -118,7 +118,7 @@
 #define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
 #endif
 #define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
-#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 0
 #define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
 
 // experimental feature: in-regsiter sub-dword transpose

include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_pt1.hpp

@@ -204,6 +204,7 @@ template <index_t NumDimG,
           index_t MXdlPerWave,
           index_t NXdlPerWave,
           index_t Gemm1NXdlPerWave,
+          index_t Gemm2NXdlPerWave,
           typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
           typename ABlockTransferThreadClusterArrangeOrder,
           typename ABlockTransferSrcAccessOrder,
@@ -627,6 +628,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
         MXdlPerWave,
         NXdlPerWave,
         Gemm1NXdlPerWave,
+        Gemm2NXdlPerWave,
         ABlockTransferThreadClusterLengths_AK0_M_AK1,
         ABlockTransferThreadClusterArrangeOrder,
         ABlockTransferSrcAccessOrder,

include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_pt1_bf16v1.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/device_batched_multihead_attention_backward.hpp" // TODO
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt1_bf16v1.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"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename DataType,
+          typename ZDataType,
+          typename LSEDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+          typename B1GridDesc_BK0_N_BK1,
+          typename YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
+          typename LSEGridDescriptor_M,
+          typename VGradGridDescriptor_N_O,
+          typename YGradGridDesc_O0_M_O1,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          typename C0MatrixMask,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    // __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, 1)
+#endif
+        kernel_batched_multihead_attention_backward_xdl_cshuffle_pt1(
+            const DataType* __restrict__ p_a_grid,
+            const DataType* __restrict__ p_b_grid,
+            ZDataType* __restrict__ p_z_grid,
+            const DataType* __restrict__ p_b1_grid,
+            const DataType* __restrict__ p_c_grid,
+            const LSEDataType* __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,
+            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 ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const LSEGridDescriptor_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 index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
+            const C0MatrixMask c0_matrix_mask,
+            const float p_drop,
+            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 num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    // NOTE: assumes QKVY has the same layout as dQ/dK/dV/dY therefore being able to reuse batch
+    // offsets
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetZBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+    const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetLSEBasePtr(g_idx)));
+
+    const index_t global_thread_id = get_thread_global_1d_id();
+    ck::philox ph(seed, global_thread_id, offset);
+    ZDataType* z_matrix_ptr = (p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset);
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  z_matrix_ptr,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_lse_grid + lse_batch_offset,
+                                                  p_ygrad_grid + c_batch_offset,
+                                                  p_qgrad_grid + a_batch_offset,
+                                                  p_kgrad_grid + b_batch_offset,
+                                                  p_vgrad_grid + b1_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  lse_grid_desc_m,
+                                                  vgrad_grid_desc_n_o,
+                                                  ygrad_grid_desc_o0_m_o1,
+                                                  block_2_ctile_map,
+                                                  c0_matrix_mask,
+                                                  p_drop,
+                                                  ph);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+    ignore = c0_matrix_mask;
+    ignore = p_drop;
+    ignore = seed;
+    ignore = offset;
+#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 DataType,
+          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,
+          index_t Gemm2NXdlPerWave,
+          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 DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
+    : public BaseOperator // TODO inherit atten bwd op once API stablizes
+{
+    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: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO: 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 = DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t Q_K1 = 8;
+    static constexpr index_t K_K1 = 8;
+    static constexpr index_t V_N1 = 2;
+
+    static constexpr index_t Q_M1 = 2;
+    static constexpr index_t K_N1 = 2;
+    static constexpr index_t V_O1 = 8;
+    static constexpr index_t Y_O1 = 8;
+    static constexpr index_t Y_M1 = 2;
+
+    static constexpr auto padder = GemmGemmPadder<GemmSpec,
+                                                  Number<MPerBlock>,
+                                                  Number<NPerBlock>,
+                                                  Number<KPerBlock>,
+                                                  Number<Gemm1NPerBlock>>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    /*
+    Descriptors for inputs:
+
+      Q, K, V, Y, dY, per-row softmax stats
+
+    Descriptors for outputs:
+
+      dQ, dK, dV
+
+    */
+
+    // Q in Gemm A position
+    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>{});
+    }
+
+    // K in Gemm B0 position
+    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>{});
+    }
+
+    // V in Gemm B1 position
+    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>{});
+    }
+
+    //
+    // dV = P^T * dY
+    //
+
+    // VGrad in Gemm C position
+    static auto MakeVGradGridDescriptor_N_O(const std::vector<index_t>& v_gs_os_ns_lengths_vec,
+                                            const std::vector<index_t>& v_gs_os_ns_strides_vec)
+    {
+        // v_gs_os_ns -> vgrad_gs_ns_os. O dims last because output is row-major.
+        // Here directly rearrange lengths/strides before constructing tensor descriptor to reduce
+        // transformation overhead
+        // TODO: This will be much easier when inputs are Gs, Ms, Ns, Os. So there's no need to
+        // extract subsequence and shuffle them.
+        const index_t num_dims = NumDimG + NumDimN + NumDimO;
+
+        // 0, 1, .. NumDimG - 1
+        std::vector<index_t> gs_ids(NumDimG);
+        std::iota(gs_ids.begin(), gs_ids.end(), 0);
+
+        // NumDimG, NumDimG + 1, ... NumDimG + NumDimO - 1
+        std::vector<index_t> os_ids(NumDimO);
+        std::iota(os_ids.begin(), os_ids.end(), NumDimG);
+
+        // NumDimG + NumDimO, NumDimG + NumDimO + 1, ... NumDimG + NumDimO + NumDimN - 1
+        std::vector<index_t> ns_ids(NumDimN);
+        std::iota(ns_ids.begin(), ns_ids.end(), NumDimG + NumDimO);
+
+        std::vector<index_t> ids_old2new;
+        ids_old2new.insert(ids_old2new.end(), gs_ids.begin(), gs_ids.end());
+        ids_old2new.insert(ids_old2new.end(), ns_ids.begin(), ns_ids.end());
+        ids_old2new.insert(ids_old2new.end(), os_ids.begin(), os_ids.end());
+
+        std::vector<index_t> v_gs_ns_os_lengths_vec(num_dims), v_gs_ns_os_strides_vec(num_dims);
+        for(int i = 0; i < num_dims; i++)
+        {
+            index_t id_new            = ids_old2new[i];
+            v_gs_ns_os_lengths_vec[i] = v_gs_os_ns_lengths_vec[id_new];
+            v_gs_ns_os_strides_vec[i] = v_gs_os_ns_strides_vec[id_new];
+        }
+
+        const auto vgrad_desc_nraw_oraw =
+            MakeGridDescriptorPair<NumDimG, NumDimN, NumDimO, TensorSpecialization::Default>(
+                v_gs_ns_os_lengths_vec, v_gs_ns_os_strides_vec)
+                .second;
+
+        return PadTensorDescriptor(vgrad_desc_nraw_oraw,
+                                   make_tuple(NPerBlock, Gemm1NPerBlock),
+                                   Sequence<padder.PadN, padder.PadO>{});
+    }
+
+    template <typename YGridDesc_M_O>
+    static auto MakeYGradGridDescriptor_M0_O_M1(const YGridDesc_M_O& ygrad_grid_desc_m_o)
+    {
+        const auto M = ygrad_grid_desc_m_o.GetLength(I0);
+        const auto O = ygrad_grid_desc_m_o.GetLength(I1);
+
+        const auto Y_M0 = M / Y_M1;
+
+        return transform_tensor_descriptor(
+            ygrad_grid_desc_m_o,
+            make_tuple(make_unmerge_transform(make_tuple(Y_M0, Y_M1)),
+                       make_pass_through_transform(O)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // dP = dY * V^T
+    //
+
+    // YGrad in Gemm A position
+    static auto MakeYGradGridDescriptor_O0_M_O1(const std::vector<index_t>& y_gs_ms_os_lengths_vec,
+                                                const std::vector<index_t>& y_gs_ms_os_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(y_gs_ms_os_lengths_vec, y_gs_ms_os_strides_vec),
+            Number<Y_O1>{});
+    }
+
+    // V in Gemm B position
+    static auto MakeVGridDescriptor_O0_N_O1(const std::vector<index_t>& v_gs_os_ns_lengths_vec,
+                                            const std::vector<index_t>& v_gs_os_ns_strides_vec)
+    {
+        // v_gs_os_ns -> vgrad_gs_ns_os. O dims last because output is row-major.
+        // Here directly rearrange lengths/strides before constructing tensor descriptor to reduce
+        // transformation overhead
+        // TODO: This will be much easier when inputs are Gs, Ms, Ns, Os. So there's no need to
+        // extract subsequence and shuffle them.
+        const index_t num_dims = NumDimG + NumDimN + NumDimO;
+
+        // 0, 1, .. NumDimG - 1
+        std::vector<index_t> gs_ids(NumDimG);
+        std::iota(gs_ids.begin(), gs_ids.end(), 0);
+
+        // NumDimG, NumDimG + 1, ... NumDimG + NumDimO - 1
+        std::vector<index_t> os_ids(NumDimO);
+        std::iota(os_ids.begin(), os_ids.end(), NumDimG);
+
+        // NumDimG + NumDimO, NumDimG + NumDimO + 1, ... NumDimG + NumDimO + NumDimN - 1
+        std::vector<index_t> ns_ids(NumDimN);
+        std::iota(ns_ids.begin(), ns_ids.end(), NumDimG + NumDimO);
+
+        std::vector<index_t> ids_old2new;
+        ids_old2new.insert(ids_old2new.end(), gs_ids.begin(), gs_ids.end());
+        ids_old2new.insert(ids_old2new.end(), ns_ids.begin(), ns_ids.end());
+        ids_old2new.insert(ids_old2new.end(), os_ids.begin(), os_ids.end());
+
+        std::vector<index_t> v_gs_ns_os_lengths_vec(num_dims), v_gs_ns_os_strides_vec(num_dims);
+        for(int i = 0; i < num_dims; i++)
+        {
+            index_t id_new            = ids_old2new[i];
+            v_gs_ns_os_lengths_vec[i] = v_gs_os_ns_lengths_vec[id_new];
+            v_gs_ns_os_strides_vec[i] = v_gs_os_ns_strides_vec[id_new];
+        }
+
+        const auto v_grid_desc_nraw_oraw =
+            MakeGridDescriptorPair<NumDimG, NumDimN, NumDimO, TensorSpecialization::Default>(
+                v_gs_ns_os_lengths_vec, v_gs_ns_os_strides_vec)
+                .second;
+
+        const auto v_grid_desc_n_o = PadTensorDescriptor(v_grid_desc_nraw_oraw,
+                                                         make_tuple(NPerBlock, Gemm1NPerBlock),
+                                                         Sequence<padder.PadN, padder.PadO>{});
+
+        // N_O to O0_N_O1; to refactor
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(v_grid_desc_n_o, Number<V_O1>{});
+    }
+
+    // Z in Gemm0 C position
+    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);
+    }
+    //
+    // dS_i_j = P_i_j .* (dP_i_j - dY_i dot Y_i)
+    //
+
+    //
+    // dQ = alpha * dS * K
+    //
+
+    // QGrad in Gemm C position
+    static auto MakeQGradGridDescriptor_M_K(const std::vector<index_t>& q_gs_ms_ks_lengths_vec,
+                                            const std::vector<index_t>& q_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeCGridDescriptor_M_N(q_gs_ms_ks_lengths_vec, q_gs_ms_ks_strides_vec);
+    }
+
+    //
+    // dK = alpha * dS^T * Q
+    //
+
+    // KGrad in Gemm C position
+    static auto MakeKGradGridDescriptor_N_K(const std::vector<index_t>& k_gs_ns_ks_lengths_vec,
+                                            const std::vector<index_t>& k_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeCGridDescriptor_M_N(k_gs_ns_ks_lengths_vec, k_gs_ns_ks_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(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using YGridDesc_M_O        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using LSEGridDesc_M        = decltype(MakeLSEGridDescriptor_M(1));
+    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({}, {}));
+
+    using VGradGridDesc_N_O     = decltype(MakeVGradGridDescriptor_N_O({}, {}));
+    using YGradGridDesc_O0_M_O1 = decltype(MakeYGradGridDescriptor_O0_M_O1({}, {}));
+    using ZGridDesc_M_N         = decltype(MakeZGridDescriptor_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 ZGridDesc_G_M_N& z_grid_desc_g_m_n,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_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),
+              z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c_grid_desc_g_m_n_(c_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 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 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 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_;
+        ZGridDesc_G_M_N z_grid_desc_g_m_n_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+
+        index_t BatchStrideLSE_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1<
+        DataType, // TODO: distinguish A/B datatype
+        LSEDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        ZGridDesc_M_N,
+        B1GridDesc_BK0_N_BK1,
+        YGridDesc_M_O,
+        LSEGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        Gemm2NXdlPerWave,
+        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>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(
+            const DataType* p_a_grid,
+            const DataType* p_b_grid,
+            ZDataType* p_z_grid,
+            const DataType* p_b1_grid,
+            const DataType* p_c_grid, // for dS
+            const LSEDataType* p_lse_grid,
+            const DataType* p_ygrad_grid,
+            DataType* p_qgrad_grid,
+            DataType* p_kgrad_grid,
+            DataType* p_vgrad_grid,
+            const std::array<void*, NumAcc0Bias> p_acc0_biases,
+            const std::array<void*, NumAcc1Bias> p_acc1_biases,
+            const std::vector<index_t>& a_gs_ms_ks_lengths,
+            const std::vector<index_t>& a_gs_ms_ks_strides,
+            const std::vector<index_t>& b_gs_ns_ks_lengths,
+            const std::vector<index_t>& b_gs_ns_ks_strides,
+            const std::vector<index_t>& z_gs_ms_ns_lengths,
+            const std::vector<index_t>& z_gs_ms_ns_strides,
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+            const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+            const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+            const std::vector<index_t>& lse_gs_ms_lengths,
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+            const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+            const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+                acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+            AElementwiseOperation a_element_op,
+            BElementwiseOperation b_element_op,
+            AccElementwiseOperation acc_element_op,
+            B1ElementwiseOperation b1_element_op,
+            CElementwiseOperation c_element_op,
+            float p_drop,
+            std::tuple<unsigned long long, unsigned long long> seeds)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_z_grid_{p_z_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              p_lse_grid_{p_lse_grid},
+              p_ygrad_grid_{p_ygrad_grid},
+              p_qgrad_grid_{p_qgrad_grid},
+              p_kgrad_grid_{p_kgrad_grid},
+              p_vgrad_grid_{p_vgrad_grid},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              z_grid_desc_m_n_{MakeZGridDescriptor_M_N(z_gs_ms_ns_lengths, z_gs_ms_ns_strides)},
+              b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeVGridDescriptor_O0_N_O1(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              y_grid_desc_m_o_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                  c_gs_ms_gemm1ns_strides)},
+              lse_grid_desc_m_{DeviceOp::MakeLSEGridDescriptor_M(lse_gs_ms_lengths[NumDimG])},
+              vgrad_grid_desc_n_o_{DeviceOp::MakeVGradGridDescriptor_N_O(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              ygrad_grid_desc_o0_m_o1_{DeviceOp::MakeYGradGridDescriptor_O0_M_O1(
+                  c_gs_ms_gemm1ns_lengths, c_gs_ms_gemm1ns_strides)},
+              // batch offsets
+              a_grid_desc_g_m_k_{
+                  Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_g_n_k_{
+                  Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                      c_gs_ms_gemm1ns_strides)},
+              z_grid_desc_g_m_n_{
+                  Transform::MakeCGridDescriptor_G_M_N(z_gs_ms_ns_lengths, z_gs_ms_ns_strides)},
+              y_grid_desc_mblock_mperblock_oblock_operblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(y_grid_desc_m_o_)},
+              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},
+              c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
+              raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                                            b1_gs_gemm1ns_gemm1ks_lengths[NumDimG + NumDimO - 1]},
+              a_mz_kz_strides_{a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                               a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+              b_nz_kz_strides_{b_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                               b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+              b1_nz_kz_strides_{b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO - 1],
+                                b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
+              c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
+                                    c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
+              batch_count_{c_grid_desc_g_m_n_.GetLength(I0)},
+              compute_base_ptr_of_batch_{
+                  a_grid_desc_g_m_k_,
+                  b_grid_desc_g_n_k_,
+                  z_grid_desc_g_m_n_,
+                  b1_grid_desc_g_n_k_,
+                  c_grid_desc_g_m_n_,
+                  type_convert<index_t>(lse_grid_desc_m_.GetElementSpaceSize())},
+              p_drop_{p_drop}
+        {
+            // TODO: implement bias addition
+            ignore = p_acc0_biases;
+            ignore = p_acc1_biases;
+            ignore = acc0_biases_gs_ms_ns_lengths;
+            ignore = acc0_biases_gs_ms_ns_strides;
+            ignore = acc1_biases_gs_ms_gemm1ns_lengths;
+            ignore = acc1_biases_gs_ms_gemm1ns_strides;
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           y_grid_desc_m_o_,
+                                           block_2_ctile_map_))
+            {
+                y_grid_desc_mblock_mperblock_oblock_operblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        y_grid_desc_m_o_);
+            }
+
+            seed_   = std::get<0>(seeds);
+            offset_ = std::get<1>(seeds);
+
+            c_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_);
+            // Print();
+        }
+
+        void Print() const
+        {
+            std::cout << "a_grid_desc_g_m_k_: " << a_grid_desc_g_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I1) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I2) << '\n';
+            // a_grid_desc_g_m_k_.Print();
+            std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b_grid_desc_g_n_k_.Print();
+            std::cout << "b1_grid_desc_g_o_n_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            // b1_grid_desc_g_n_k_.Print();
+            std::cout << "c_grid_desc_g_m_o_: " << c_grid_desc_g_m_n_.GetLength(I0) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I1) << ", "
+                      << c_grid_desc_g_m_n_.GetLength(I2) << '\n';
+            // c_grid_desc_g_m_n_.Print();
+            std::cout << "vgrad_grid_desc_n_o_: " << vgrad_grid_desc_n_o_.GetLength(I0) << ", "
+                      << vgrad_grid_desc_n_o_.GetLength(I1) << '\n';
+            std::cout << "ygrad_grid_desc_o0_m_o1_: " << ygrad_grid_desc_o0_m_o1_.GetLength(I0)
+                      << ", " << ygrad_grid_desc_o0_m_o1_.GetLength(I1) << ", "
+                      << ygrad_grid_desc_o0_m_o1_.GetLength(I2) << '\n';
+        }
+
+        // pointers
+        const DataType* p_a_grid_;
+        const DataType* p_b_grid_;
+        ZDataType* p_z_grid_;
+        const DataType* p_b1_grid_;
+        const DataType* p_c_grid_;
+        const LSEDataType* p_lse_grid_;
+        const DataType* p_ygrad_grid_;
+        DataType* p_qgrad_grid_;
+        DataType* p_kgrad_grid_;
+        DataType* p_vgrad_grid_;
+
+        // tensor descriptor
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        ZGridDesc_M_N z_grid_desc_m_n_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        YGridDesc_M_O y_grid_desc_m_o_;
+        LSEGridDesc_M lse_grid_desc_m_;
+        VGradGridDesc_N_O vgrad_grid_desc_n_o_;
+        YGradGridDesc_O0_M_O1 ygrad_grid_desc_o0_m_o1_;
+
+        // batch offsets
+        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_;
+        typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock
+            y_grid_desc_mblock_mperblock_oblock_operblock_;
+
+        typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
+
+        // block-to-c-tile map
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+        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_;
+
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        float p_drop_;
+        unsigned long long seed_;
+        unsigned long long offset_;
+    };
+
+    // 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");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.y_grid_desc_m_o_) * arg.batch_count_;
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_multihead_attention_backward_xdl_cshuffle_pt1<
+                    GridwiseGemm,
+                    DataType,
+                    ZDataType,
+                    LSEDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
+                    DeviceOp::LSEGridDesc_M,
+                    DeviceOp::VGradGridDesc_N_O,
+                    DeviceOp::YGradGridDesc_O0_M_O1,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    C0MatrixMask,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_z_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.p_lse_grid_,
+                                              arg.p_ygrad_grid_,
+                                              arg.p_qgrad_grid_,
+                                              arg.p_kgrad_grid_,
+                                              arg.p_vgrad_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.y_grid_desc_mblock_mperblock_oblock_operblock_,
+                                              arg.lse_grid_desc_m_,
+                                              arg.vgrad_grid_desc_n_o_,
+                                              arg.ygrad_grid_desc_o0_m_o1_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_,
+                                              arg.c0_matrix_mask_,
+                                              arg.p_drop_,
+                                              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 1
+            // if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            // {
+            //     ave_time = launch_kernel(integral_constant<bool, true>{});
+            // }
+            // else
+            // {
+            //     ave_time = launch_kernel(integral_constant<bool, false>{});
+            // }
+            ave_time = launch_kernel(integral_constant<bool, false>{});
+#endif
+            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 0
+        arg.Print();
+#endif
+
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        {
+            return false;
+        }
+
+        // TODO: Check if tensor specialization & strides mismatch
+
+        // Check if C permute dimension matches GEMM + GEMM shape
+        const index_t c_g      = arg.c_grid_desc_g_m_n_.GetLength(I0); // unpadded
+        const index_t c_m      = arg.y_grid_desc_m_o_.GetLength(I0);
+        const index_t c_gemm1n = arg.y_grid_desc_m_o_.GetLength(I1);
+        const index_t a_m      = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+        const index_t b1_gemm1n =
+            arg.b1_grid_desc_bk0_n_bk1_.GetLength(I0) * arg.b1_grid_desc_bk0_n_bk1_.GetLength(I2);
+
+        if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        // Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O
+        const auto MzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+        const auto NzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+        const auto KzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+        const auto Gemm1NzRaw = 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 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];
+        const auto b_stride_lowest =
+            BBlockTransferSrcVectorDim == 2 ? arg.b_nz_kz_strides_[1] : arg.b_nz_kz_strides_[0];
+        const auto b1_stride_lowest =
+            B1BlockTransferSrcVectorDim == 2 ? arg.b1_nz_kz_strides_[1] : arg.b1_nz_kz_strides_[0];
+        const auto c_stride_lowest =
+            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;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.y_grid_desc_m_o_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const DataType* p_a,
+        const DataType* p_b,
+        ZDataType* p_z,
+        const DataType* p_b1,
+        const DataType* p_c,
+        const LSEDataType* p_lse,
+        const DataType* p_ygrad_grid,
+        DataType* p_qgrad_grid,
+        DataType* p_kgrad_grid,
+        DataType* p_vgrad_grid,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& z_gs_ms_ns_lengths,
+        const std::vector<index_t>& z_gs_ms_ns_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::vector<index_t>& lse_gs_ms_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op,
+        float p_drop,
+        std::tuple<unsigned long long, unsigned long long> seeds)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_z,
+                        p_b1,
+                        p_c,
+                        p_lse,
+                        p_ygrad_grid,
+                        p_qgrad_grid,
+                        p_kgrad_grid,
+                        p_vgrad_grid,
+                        p_acc0_biases,
+                        p_acc1_biases,
+                        a_gs_ms_ks_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        z_gs_ms_ns_lengths,
+                        z_gs_ms_ns_strides,
+                        b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                        b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                        c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                        c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                        lse_gs_ms_lengths,
+                        acc0_biases_gs_ms_ns_lengths,
+                        acc0_biases_gs_ms_ns_strides,
+                        acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+                        acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op,
+                        p_drop,
+                        seeds};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    // FIXME: constness
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        void* p_z,
+        const void* p_b1,
+        const void* p_c,
+        const void* p_lse,
+        const void* p_ygrad_grid,
+        void* p_qgrad_grid,
+        void* p_kgrad_grid,
+        void* p_vgrad_grid,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& z_gs_ms_ns_lengths,
+        const std::vector<index_t>& z_gs_ms_ns_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::vector<index_t>& lse_gs_ms_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        AccElementwiseOperation acc_element_op,
+        B1ElementwiseOperation b1_element_op,
+        CElementwiseOperation c_element_op,
+        float p_drop,
+        std::tuple<unsigned long long, unsigned long long> seeds) // override
+    {
+        return std::make_unique<Argument>(static_cast<const DataType*>(p_a),
+                                          static_cast<const DataType*>(p_b),
+                                          static_cast<ZDataType*>(p_z),
+                                          static_cast<const DataType*>(p_b1),
+                                          static_cast<const DataType*>(p_c),
+                                          static_cast<const LSEDataType*>(p_lse),
+                                          static_cast<const DataType*>(p_ygrad_grid),
+                                          static_cast<DataType*>(p_qgrad_grid),
+                                          static_cast<DataType*>(p_kgrad_grid),
+                                          static_cast<DataType*>(p_vgrad_grid),
+                                          p_acc0_biases, // cast in struct Argument
+                                          p_acc1_biases, // cast in struct Argument
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          z_gs_ms_ns_lengths,
+                                          z_gs_ms_ns_strides,
+                                          b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                                          b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                                          c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                                          c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                                          lse_gs_ms_lengths,
+                                          acc0_biases_gs_ms_ns_lengths,
+                                          acc0_biases_gs_ms_ns_strides,
+                                          acc1_biases_gs_ms_gemm1ns_lengths,
+                                          acc1_biases_gs_ms_gemm1ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op,
+                                          p_drop,
+                                          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 << "DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1"
+            << "<"
+            << 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();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -51,6 +51,7 @@ template <typename DataType,
           index_t MXdlPerWave,
           index_t NXdlPerWave,
           index_t Gemm1NXdlPerWave,
+          index_t Gemm2NXdlPerWave,
           typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
           typename ABlockTransferThreadClusterArrangeOrder,
           typename ABlockTransferSrcAccessOrder,
@@ -726,9 +727,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
 
         static_assert(Sum_M % MPerXdl == 0, "");
 
-        static constexpr index_t GemmNWave   = 2;
+        static constexpr index_t GemmNWave   = Free0_N / Gemm2NXdlPerWave / MPerXdl;
         static constexpr index_t GemmOWave   = BlockSize / get_warp_size() / GemmNWave;
-        static constexpr index_t GemmNRepeat = Free0_N / GemmNWave / MPerXdl;
+        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 =
@@ -1563,8 +1564,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
                      n3, // NInputNum
                      n4>,
             Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
-            9,  // DstVectorDim
-            1,  // DstScalarPerVector
+            9, // DstVectorDim
+            1, // DstScalarPerVector
             InMemoryDataOperationEnum::Set,
             1, // DstScalarStrideInVector
             true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
@@ -1937,7 +1938,6 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_PT1
                     else
                     {
                         s_element_op(s_slash_p_thread_buf(i), s_slash_p_thread_buf[i]);
-
                     }
                 });
             }

include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt1_bf16v1.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,
+                                        FloatGemmAcc,
+                                        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,
+            FloatGemmAcc,
+            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,
+            FloatGemmAcc,
+            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/utility/generic_memory_space_atomic.hpp

@@ -71,6 +71,75 @@ __device__ double2_t atomic_add<double2_t>(double2_t* p_dst, const double2_t& x)
     return vy.template AsType<double2_t>()[I0];
 }
 
+inline __host__ __device__ half2_t add_fp16x2_t(const half2_t& a, const half2_t& b)
+{
+    half2_t rtn;
+    rtn[0] = a[0] + b[0];
+    rtn[1] = a[1] + b[1];
+    return rtn;
+}
+
+template <>
+__device__ half2_t atomic_add<half2_t>(half2_t* p_dst, const half2_t& x)
+{
+    uint32_t * dword_addr = reinterpret_cast<uint32_t*>(p_dst);
+    uint32_t cur_v = *dword_addr;
+    uint32_t old_v, new_v;
+
+    do {
+        old_v = cur_v;
+        half2_t new_ = add_fp16x2_t(*reinterpret_cast<half2_t*>(&cur_v), x);
+        new_v = *reinterpret_cast<uint32_t*>(&new_);
+        cur_v = atomicCAS(dword_addr, old_v, new_v);
+    }while(cur_v != old_v);
+
+    return x;
+}
+
+// union U16BF16 {
+//     uint16_t u16;
+//     bhalf_t bf16;
+// };
+ 
+// inline __host__ __device__ bhalf_t add_bf16_t(const bhalf_t& a, const bhalf_t& b){
+//     U16BF16 xa {.bf16 = a};
+//     U16BF16 xb {.bf16 = b};
+    
+//     U16BF16 xr;
+//     xr.u16 = xa.u16 + xb.u16;
+//     return xr.bf16;
+// }
+
+inline __host__ __device__ bhalf_t add_bf16_t(const bhalf_t& a, const bhalf_t& b)
+{
+    return type_convert<bhalf_t>(type_convert<float>(a) + type_convert<float>(b));
+}
+
+inline __host__ __device__ bhalf2_t add_bf16x2_t(const bhalf2_t& a, const bhalf2_t& b)
+{
+    bhalf2_t rtn;
+    rtn[0] = add_bf16_t(a[0], b[0]);
+    rtn[1] = add_bf16_t(a[1], b[1]);
+    return rtn;
+}
+
+template <>
+__device__ bhalf2_t atomic_add<bhalf2_t>(bhalf2_t* p_dst, const bhalf2_t& x)
+{
+    uint32_t * dword_addr = reinterpret_cast<uint32_t*>(p_dst);
+    uint32_t cur_v = *dword_addr;
+    uint32_t old_v, new_v;
+
+    do {
+        old_v = cur_v;
+        bhalf2_t new_ = add_bf16x2_t(*reinterpret_cast<bhalf2_t*>(&cur_v), x);
+        new_v = *reinterpret_cast<uint32_t*>(&new_);
+        cur_v = atomicCAS(dword_addr, old_v, new_v);
+    }while(cur_v != old_v);
+
+    return x;
+}
+
 // Caution: DO NOT REMOVE
 // intentionally have only declaration but no definition to cause compilation failure when trying to
 // instantiate this template. The purpose is to make the implementation of atomic_max explicit for