bf16A_Int8B with fastgelu/bias (#1264)

* changed the copy function to v7r2

* adding multi_abd

* in-progress

* add post-load oob check

* debugging

* adjust instances

* add run_lds

* add elemntwise_op

* replace multi_abd_device with v3

* clean up

* clean

* clean

* Added LDSType

* profiling

* adjust oobcheck

* add missing file

* refactor

* clean

* add examples

client_example/30_gemm_multi_abd/CMakeLists.txt → client_example/30_gemm_bf16Aint8B/CMakeLists.txt

@@ -10,4 +10,7 @@ if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf
 
 	add_executable(client_gemm_bf16_i8_bf16 gemm_xdl_bf16_i8.cpp)
 	target_link_libraries(client_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_gemm_multiply_bf16_i8_bf16 gemm_xdl_multiply_bf16_i8.cpp)
+	target_link_libraries(client_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
 endif()

client_example/30_gemm_multi_abd/gemm_bias_fastgelu_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_bias_fastgelu_xdl_bf16_i8.cpp

@@ -38,19 +38,19 @@ using EDataType        = BF16;
 
 using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Col;
+using B0Layout = Row;
 using B1Layout = B0Layout;
 using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 using DsLayout = ck::Tuple<D0Layout>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = AddFastGelu;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

client_example/30_gemm_multi_abd/gemm_bias_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_bias_xdl_bf16_i8.cpp

@@ -36,7 +36,7 @@ using D0DataType       = BF16;
 using DsDataType       = ck::Tuple<D0DataType>;
 using EDataType        = BF16;
 
-using A0Layout = Col;
+using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
 using B0Layout = Row;
 using B1Layout = B0Layout;
@@ -45,12 +45,12 @@ using D0Layout = Row;
 using DsLayout = ck::Tuple<D0Layout>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using Add         = ck::tensor_operation::element_wise::Add;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = Add;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

client_example/30_gemm_multi_abd/gemm_xdl_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_xdl_bf16_i8.cpp

@@ -37,19 +37,19 @@ using EDataType        = BF16;
 
 using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Col;
+using B0Layout = Row;
 using B1Layout = B0Layout;
 using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 using DsLayout = ck::Tuple<>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using Add         = ck::tensor_operation::element_wise::Add;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = PassThrough;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
@@ -74,12 +74,12 @@ struct SimpleDeviceMem
 int main(int argc, char* argv[])
 {
     // GEMM shape
-    ck::index_t M = 64;
-    ck::index_t N = 1024;
-    ck::index_t K = 512;
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
 
     ck::index_t StrideA = K;
-    ck::index_t StrideB = N;
+    ck::index_t StrideB = K;
     ck::index_t StrideE = N;
 
     if(argc == 1)

client_example/30_gemm_multi_abd/gemm_xdl_gelu_bf16_i8.cpp → client_example/30_gemm_bf16Aint8B/gemm_xdl_gelu_bf16_i8.cpp

@@ -37,19 +37,19 @@ using EDataType        = BF16;
 
 using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Col;
+using B0Layout = Row;
 using B1Layout = B0Layout;
 using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 using DsLayout = ck::Tuple<>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = FastGelu;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

client_example/30_gemm_bf16Aint8B/gemm_xdl_multiply_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <iomanip>
+#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/device_gemm_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<B1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Multiply;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// clang-format on
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, Row>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a0_device_buf(sizeof(A0DataType) *
+                                  f_matrix_space_size(M, K, StrideA, A0Layout{}));
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) *
+                                  f_matrix_space_size(K, N, StrideB, B0Layout{}));
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * f_matrix_space_size(K, N, 0, B1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 1;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
+                                                                         BsLayout,
+                                                                         DsLayout,
+                                                                         Row,
+                                                                         AsDataType,
+                                                                         BsDataType,
+                                                                         DsDataType,
+                                                                         BF16,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CDEElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
+            std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            std::array<ck::index_t, NumATensor>{StrideA},
+            std::array<ck::index_t, NumBTensor>{StrideB},
+            std::array<ck::index_t, NumDTensor>{0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

client_example/31_grouped_gemm_multi_abd/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp → client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_bias_fastgelu_xdl_bf16_i8.cpp

@@ -38,19 +38,19 @@ using EDataType        = BF16;
 
 using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Col;
+using B0Layout = Row;
 using B1Layout = B0Layout;
 using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 using DsLayout = ck::Tuple<D0Layout>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = AddFastGelu;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

client_example/31_grouped_gemm_multi_abd/grouped_gemm_fastgelu_xdl_bf16_i8.cpp → client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_fastgelu_xdl_bf16_i8.cpp

@@ -36,7 +36,7 @@ using D0DataType       = BF16;
 using DsDataType       = ck::Tuple<>;
 using EDataType        = BF16;
 
-using A0Layout = Col;
+using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
 using B0Layout = Row;
 using B1Layout = B0Layout;
@@ -45,12 +45,12 @@ using D0Layout = Row;
 using DsLayout = ck::Tuple<>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = FastGelu;
 
 static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp

@@ -52,12 +52,12 @@ using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using DsLayout = ck::Tuple<Row>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = AddFastGelu;
 
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;

example/60_gemm_multi_ABD/CMakeLists.txt

 add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
-add_example_executable(example_gemm_multi_ABD_xdl_bf16_i8 gemm_multi_ABD_xdl_bf16_i8.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp)
\ No newline at end of file
+add_example_executable(example_gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_fastgelu_bf16_i8 gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp)

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bf16_i8.cpp → example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp

@@ -18,9 +18,12 @@
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 #include "ck/library/utility/check_err.hpp"
 
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
+using F16  = ck::half_t;
 using BF16 = ck::bhalf_t;
 using I8   = int8_t;
 using F32  = float;
@@ -41,22 +44,22 @@ using EDataType        = BF16;
 
 using A0Layout = Row;
 using AsLayout = ck::Tuple<A0Layout>;
-using B0Layout = Col;
+using B0Layout = Row;
 using B1Layout = B0Layout;
 using BsLayout = ck::Tuple<B0Layout, B1Layout>;
 using D0Layout = Row;
 using DsLayout = ck::Tuple<D0Layout>;
 using ELayout  = Row;
 
-using Scales      = ck::tensor_operation::element_wise::Scales;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
 
 using AElementOp   = PassThrough;
-using BElementOp   = Scales;
+using BElementOp   = Multiply;
 using CDEElementOp = AddFastGelu;
 
-static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
 
 using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
     // clang-format off
@@ -64,9 +67,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl
 ///######|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
 ///######|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
 ///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1>;
-
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
 // clang-format on
+
 int main(int argc, char* argv[])
 {
     bool do_verification = true;
@@ -74,13 +77,13 @@ int main(int argc, char* argv[])
     bool time_kernel     = false;
 
     // GEMM shape
-    ck::index_t M = 64;
-    ck::index_t N = 1024;
-    ck::index_t K = 512;
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
 
     ck::index_t StrideA = K;
-    ck::index_t StrideB = K;
-    ck::index_t StrideD = N;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
     ck::index_t StrideE = N;
 
     if(argc == 1)

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = FastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    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 == 11)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                                   b1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB, 0},
+                               std::array<ck::index_t, NumDTensor>{},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B1DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType, D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout, D0Layout>;
+using ELayout  = Row;
+
+using PassThrough         = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyAddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    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 == 11)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 1;
+    constexpr ck::index_t NumDTensor = 2;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer(),
+                                                                   d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB},
+                               std::array<ck::index_t, NumDTensor>{0, StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+#if 0
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+#endif
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), b1_k_n(0, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp

@@ -41,7 +41,8 @@ template <typename ThreadGroup,
           index_t SrcScalarPerVector,
           index_t DstScalarPerVector,
           typename ThreadTransferSrcResetCoordinateAfterRunFlags,
-          typename ThreadTransferDstResetCoordinateAfterRunFlags>
+          typename ThreadTransferDstResetCoordinateAfterRunFlags,
+          index_t NumThreadScratch = 1>
 struct ThreadGroupTensorSliceTransfer_v7r2
 {
     static constexpr index_t nDim =
@@ -100,7 +101,7 @@ struct ThreadGroupTensorSliceTransfer_v7r2
            ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
         {
             const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
-                make_multi_index(get_thread_local_1d_id()));
+                make_multi_index(ThreadGroup::GetThreadId()));
 
             const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
 
@@ -117,29 +118,33 @@ struct ThreadGroupTensorSliceTransfer_v7r2
         }
     }
 
-    template <typename SrcBuffers>
-    __device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
+    template <typename SrcBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDescs& src_descs,
+                            const SrcBuffers& src_bufs,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
     {
         if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
            ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
         {
-            threadwise_transfer_.RunRead(src_descs, src_bufs);
+            threadwise_transfer_.RunRead(src_descs, src_bufs, thread_scratch_id);
         }
     }
 
     template <typename T>
     using is_tuple = decltype(std::declval<T&>().IsTuple());
 
-    template <typename DstBuffers>
-    __device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
+    template <typename DstBuffers, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDescs& dst_descs,
+                             DstBuffers dst_bufs,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
     {
         if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
            ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
         {
             if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
-                threadwise_transfer_.RunWrite(dst_descs, dst_bufs);
+                threadwise_transfer_.RunWrite(dst_descs, dst_bufs, thread_scratch_id);
             else
-                threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs));
+                threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs), thread_scratch_id);
         }
     }
 
@@ -206,7 +211,8 @@ struct ThreadGroupTensorSliceTransfer_v7r2
                                            SrcScalarPerVector,
                                            DstScalarPerVector,
                                            ThreadTransferSrcResetCoordinateAfterRunFlags,
-                                           ThreadTransferDstResetCoordinateAfterRunFlags>;
+                                           ThreadTransferDstResetCoordinateAfterRunFlags,
+                                           NumThreadScratch>;
 
     ThreadwiseTransfer threadwise_transfer_;
 };

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

@@ -92,15 +92,6 @@ struct Add
     };
 };
 
-struct Scales
-{
-    template <typename Y, typename X0, typename X1>
-    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const
-    {
-        y = ck::type_convert<Y>(ck::type_convert<float>(x0) * ck::type_convert<float>(x1));
-    }
-};
-
 struct Max
 {
     template <typename Y, typename X0, typename X1>
@@ -188,6 +179,16 @@ struct Multiply
         y                  = ck::type_convert<bhalf_t>(y_tmp);
     }
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const int8_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp * x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
+    }
+
     template <>
     __host__ __device__ constexpr void
     operator()<bhalf_t>(bhalf_t& y, const float& x0, const bhalf_t& x1) const
@@ -521,6 +522,71 @@ struct AddFastGelu
     }
 };
 
+// E = MultiplyFastGelu(C + D)
+struct MultiplyFastGelu
+{
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& e, const float& c, const float& d) const
+    {
+        const float x = c * d;
+
+        FastGelu{}.template operator()<float, float>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
+    {
+        const half_t x = c * d;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
+    {
+        const float x0_f = c * d;
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<half_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, bhalf_t, bhalf_t>(bhalf_t& e, const bhalf_t& c, const bhalf_t& d) const
+    {
+        const float x0_f = type_convert<float>(c) * type_convert<float>(d);
+
+        float x1_f = 0;
+
+        FastGelu{}.template operator()<float, float>(x1_f, x0_f);
+
+        e = type_convert<bhalf_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, float, bhalf_t>(bhalf_t& e, const float& c, const bhalf_t& d) const
+    {
+        const float x0_f = c * type_convert<float>(d);
+
+        float x1_f = 0;
+
+        FastGelu{}.template operator()<float, float>(x1_f, x0_f);
+
+        e = type_convert<bhalf_t>(x1_f);
+    }
+};
+
 // E = Silu(C + D)
 struct AddSilu
 {

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

@@ -221,6 +221,15 @@ struct MultiplyAdd
         e              = y;
     }
     template <>
+    __host__ __device__ void operator()<bhalf_t, float, bhalf_t, bhalf_t>(bhalf_t& e,
+                                                                          const float& c,
+                                                                          const bhalf_t& d0,
+                                                                          const bhalf_t& d1) const
+    {
+        const bhalf_t y = type_convert<bhalf_t>(c) * d0 + d1;
+        e               = y;
+    }
+    template <>
     __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
                                                                       const float& c,
                                                                       const half_t& d0,
@@ -240,6 +249,26 @@ struct MultiplyAdd
     }
 };
 
+struct MultiplyAddFastGelu
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::bhalf_t, float, ck::bhalf_t, ck::bhalf_t>(
+        ck::bhalf_t& e, const float& c, const ck::bhalf_t& d0, const ck::bhalf_t& d1) const
+    {
+        const float x0_f = c * ck::type_convert<float>(d0) + ck::type_convert<float>(d1);
+
+        float x1_f = 0;
+
+        FastGelu{}.template operator()<float, float>(x1_f, x0_f);
+
+        e = ck::type_convert<ck::bhalf_t>(x1_f);
+    }
+};
+
 // E = FastGelu(C + D0 + D1)
 struct AddAddFastGelu
 {

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

@@ -504,6 +504,16 @@ struct FastGelu
         y = type_convert<half_t>(y_f);
     }
 
+    template <>
+    __host__ void operator()<bhalf_t, float>(bhalf_t& y, const float& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, x);
+
+        y = type_convert<bhalf_t>(y_f);
+    }
+
     template <>
     __device__ void operator()<bhalf_t, float>(bhalf_t& y, const float& x) const
     {

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

@@ -594,11 +594,6 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             generate_tuple([&](auto) { return make_multi_index(0, m_block_data_idx_on_grid, 0); },
                            Number<NumATensor>{});
 
-#if 0
-        static_assert(ABlockTransferSrcScalarPerVector == ABlockTransferDstScalarPerVector_AK1,
-                      "Src and Dst ScalarPerVector must be the same");
-#endif
-
         auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
             ThisThreadBlock,
             AsDataType,
@@ -616,7 +611,7 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             2,
             ABlockTransferSrcScalarPerVector,
             ABlockTransferDstScalarPerVector_AK1,
-            uniform_sequence_gen_t<NumATensor, false>,
+            uniform_sequence_gen_t<NumATensor, AThreadTransferSrcResetCoordinateAfterRun>,
             Sequence<true>>{as_grid_desc_ak0_m_ak1,
                             idx_as_block_begin,
                             tie(a_block_desc_ak0_m_ak1),
@@ -627,11 +622,6 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             generate_tuple([&](auto) { return make_multi_index(0, n_block_data_idx_on_grid, 0); },
                            Number<NumBTensor>{});
 
-#if 0
-        static_assert(BBlockTransferSrcScalarPerVector == BBlockTransferDstScalarPerVector_BK1,
-                      "Src and Dst ScalarPerVector must be the same");
-#endif
-
         auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v7r2<
             ThisThreadBlock,
             BsDataType,
@@ -649,7 +639,7 @@ struct GridwiseGemmMultipleABD_xdl_cshuffle
             2,
             BBlockTransferSrcScalarPerVector,
             BBlockTransferDstScalarPerVector_BK1,
-            uniform_sequence_gen_t<NumBTensor, false>,
+            uniform_sequence_gen_t<NumBTensor, BThreadTransferSrcResetCoordinateAfterRun>,
             Sequence<true>>{bs_grid_desc_bk0_n_bk1,
                             idx_bs_block_begin,
                             tie(b_block_desc_bk0_n_bk1),