Merge branch 'develop' into amd-develop

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/CMakeLists.txt → client_example/31_grouped_gemm_bf16Aint8B/CMakeLists.txt

@@ -4,4 +4,13 @@ if(GPU_TARGETS MATCHES "gfx9" AND ((DTYPES MATCHES "int8" AND DTYPES MATCHES "bf
 
 	add_executable(client_grouped_gemm_fastgelu_bf16_i8_bf16 grouped_gemm_fastgelu_xdl_bf16_i8.cpp)
 	target_link_libraries(client_grouped_gemm_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_multiply_bf16_i8_bf16 grouped_gemm_multiply_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_multiply_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_multiply_bias_fastgelu_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
+
+	add_executable(client_grouped_gemm_bf16_i8_bf16 grouped_gemm_xdl_bf16_i8.cpp)
+	target_link_libraries(client_grouped_gemm_bf16_i8_bf16 PRIVATE composable_kernel::device_gemm_operations)
 endif()

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

@@ -15,6 +15,8 @@
 
 #include "ck/library/tensor_operation_instance/gpu/grouped_gemm_multi_abd_fixed_nk.hpp"
 
+#include "ck/host_utility/hip_check_error.hpp"
+
 template <ck::index_t... Is>
 using S = ck::Sequence<Is...>;
 
@@ -36,7 +38,7 @@ 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 +47,12 @@ 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;

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_bias_fastgelu_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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_grouped_gemm_multi_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/grouped_gemm_tile_loop_multply.hpp"
+
+#include "ck/host_utility/hip_check_error.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 B0DataType       = I8;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType, D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B0Layout, 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::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_;
+};
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+    }
+
+    constexpr ck::index_t NumDTensor = 2;
+
+    using GroupedGemmKernelArgument =
+        ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        gemm_descs.push_back({problem_size.Ms[i],
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              problem_size.stride_As[i],
+                              problem_size.stride_Bs[i],
+                              problem_size.stride_Cs[i],
+                              {0, 0}});
+
+        grouped_gemm_kernel_args_.push_back(
+            {a0_tensors_device[i]->GetDeviceBuffer(),
+             b0_tensors_device[i]->GetDeviceBuffer(),
+             {b1_tensors_device[i]->GetDeviceBuffer(), d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             problem_size.stride_As[i],
+             problem_size.stride_Bs[i],
+             {0, 0},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
+                                                                             B0Layout,
+                                                                             DsLayout,
+                                                                             ELayout,
+                                                                             A0DataType,
+                                                                             B0DataType,
+                                                                             DsDataType,
+                                                                             EDataType,
+                                                                             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];
+
+        std::vector<const void*> p_As                         = {};
+        std::vector<const void*> p_Bs                         = {};
+        std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+        std::vector<void*> p_Cs                               = {};
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_As, p_Bs, p_Ds, p_Cs, gemm_descs, 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()))
+        {
+            SimpleDeviceMem gemm_kernel_args_dev(
+                op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+            hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                                      grouped_gemm_kernel_args_.data(),
+                                      op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                      hipMemcpyHostToDevice));
+
+            op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
+
+            std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
+
+            std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
+                                    sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
+                                    sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
+
+            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 true;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(1 + rand() % 1024);
+        problem_size.Ns.push_back(6144);
+        problem_size.Ks.push_back(4096);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_multiply_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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_grouped_gemm_multi_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/grouped_gemm_tile_loop_multply.hpp"
+
+#include "ck/host_utility/hip_check_error.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 B0DataType       = I8;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout>;
+using ELayout  = Row;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+
+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_;
+};
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+    }
+
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument =
+        ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(A0DataType) * problem_size.Ms[i] * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
+
+        gemm_descs.push_back({problem_size.Ms[i],
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              problem_size.stride_As[i],
+                              problem_size.stride_Bs[i],
+                              problem_size.stride_Cs[i],
+                              {0}});
+
+        grouped_gemm_kernel_args_.push_back({a0_tensors_device[i]->GetDeviceBuffer(),
+                                             b0_tensors_device[i]->GetDeviceBuffer(),
+                                             {b1_tensors_device[i]->GetDeviceBuffer()},
+                                             c_tensors_device[i]->GetDeviceBuffer(),
+                                             problem_size.Ms[i],
+                                             problem_size.Ns[i],
+                                             problem_size.Ks[i],
+                                             problem_size.stride_As[i],
+                                             problem_size.stride_Bs[i],
+                                             {0},
+                                             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmTileLoop<A0Layout,
+                                                                             B0Layout,
+                                                                             DsLayout,
+                                                                             ELayout,
+                                                                             A0DataType,
+                                                                             B0DataType,
+                                                                             DsDataType,
+                                                                             EDataType,
+                                                                             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];
+
+        std::vector<const void*> p_As                         = {};
+        std::vector<const void*> p_Bs                         = {};
+        std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+        std::vector<void*> p_Cs                               = {};
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_As, p_Bs, p_Ds, p_Cs, gemm_descs, 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()))
+        {
+            SimpleDeviceMem gemm_kernel_args_dev(
+                op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+            hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                                      grouped_gemm_kernel_args_.data(),
+                                      op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                      hipMemcpyHostToDevice));
+
+            op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
+
+            std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
+
+            std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
+                                    sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
+                                    sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
+
+            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 true;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(1 + rand() % 1024);
+        problem_size.Ns.push_back(4096);
+        problem_size.Ks.push_back(4096);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

client_example/31_grouped_gemm_bf16Aint8B/grouped_gemm_xdl_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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_grouped_gemm_multi_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/grouped_gemm_multi_abd_fixed_nk.hpp"
+
+#include "ck/host_utility/hip_check_error.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, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+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 AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = PassThrough;
+
+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_;
+};
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmMultiABDDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    using DeviceMemPtr = std::unique_ptr<SimpleDeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+    }
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<SimpleDeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<SimpleDeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        gemm_descs.push_back(
+            {sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {}, 1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
+                                                 b1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
+             std::array<ck::index_t, NumDTensor>{},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmMultiABDFixedNK<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];
+
+        std::vector<std::array<const void*, NumATensor>> p_As = {};
+        std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+        std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+        std::vector<void*> p_Cs                               = {};
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            SimpleDeviceMem gemm_kernel_args_dev(
+                op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+            hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                                      grouped_gemm_kernel_args_.data(),
+                                      op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                      hipMemcpyHostToDevice));
+
+            op_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kernel_args_dev.GetDeviceBuffer());
+
+            op_ptr->SetElementwiseOps(
+                argument_ptr.get(), a_element_op, b_element_op, cde_element_op);
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, 0, 20, 50});
+
+            std::size_t flop = std::size_t(2) * sum_of_m * problem_size.Ns[0] * problem_size.Ks[0];
+
+            std::size_t num_btype = sizeof(A0DataType) * sum_of_m * problem_size.Ks[0] +
+                                    sizeof(B0DataType) * problem_size.Ks[0] * problem_size.Ns[0] +
+                                    sizeof(EDataType) * sum_of_m * problem_size.Ns[0];
+
+            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 true;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(1 + rand() % 1024);
+        problem_size.Ns.push_back(4096);
+        problem_size.Ks.push_back(4096);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+
+        std::cout << " M = " << problem_size.Ms[i] << " N = " << problem_size.Ns[i] << " K "
+                  << problem_size.Ks[i] << std::endl;
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/01_gemm/CMakeLists.txt

@@ -28,6 +28,8 @@ add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
 add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
+add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
 
 add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)

example/01_gemm/gemm_xdl_bf16_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType,   BDataType,  CDataType,  AccDataType,  CShuffleDataType, 
+        PassThrough, PassThrough, PassThrough, GemmDefault, 
+        256,
+        128, 128, 
+        64, 8, 8,
+        16,   16,
+        4,    4,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        1, 2, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example_v2.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/01_gemm/run_gemm_example_v2.inc

@@ -3,6 +3,88 @@
 
 #pragma once
 
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {
@@ -180,7 +262,11 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 #else
         c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-        pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+        pass &= ck::utils::check_err(c_m_n_device_result,
+                                     c_m_n_host_result,
+                                     "Error: Incorrect results!",
+                                     get_rtol<CDataType>(),
+                                     get_atol<CDataType>());
 #endif
     }
 

example/15_grouped_gemm/CMakeLists.txt

@@ -26,6 +26,9 @@ add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
 add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16_fp8 grouped_gemm_xdl_fixed_nk_fp16_fp8.cpp)
 add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp16_fp8)
 
+add_example_executable(example_grouped_gemm_multiple_d_xdl_fp16 grouped_gemm_multiple_d_xdl_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_xdl_fp16)
+
 if(USE_BITINT_EXTENSION_INT4)
     add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
     add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)

example/15_grouped_gemm/grouped_gemm_multiple_d_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_tile_loop.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include <ck/utility/data_type.hpp>
+#include <ck/utility/tuple.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/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm_multiple_d.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddAdd      = ck::tensor_operation::element_wise::AddAdd;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType, DDataType>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DLayout  = Row;
+using DsLayout = ck::Tuple<DLayout, DLayout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+static constexpr int NumDs           = 2;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmMultipleDXdlCShuffleTileLoop
+    // 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|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,    S<4, 64, 1>,     S<1, 0, 2>,      S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>;
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<std::vector<ck::index_t>> stride_Ds;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDs>;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<KernelArguments> ggemm_kargs;
+    std::vector<void*> p_Cs;
+    std::vector<const void*> p_As;
+    std::vector<const void*> p_Bs;
+    std::vector<std::array<const void*, NumDs>> p_Ds = {};
+
+    gemm_descs.reserve(group_count);
+    ggemm_kargs.reserve(group_count);
+    p_As.reserve(group_count);
+    p_Bs.reserve(group_count);
+    p_Ds.reserve(group_count);
+
+    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});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<std::array<Tensor<DDataType>, NumDs>> d_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_result_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_result_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+    std::vector<std::vector<DeviceMemPtr>> d_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
+
+        auto d0_tensor = Tensor<DDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], DLayout{}));
+        auto d1_tensor = Tensor<DDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], DLayout{}));
+
+        std::array<Tensor<DDataType>, NumDs> d_tens = {d0_tensor, d1_tensor};
+        d_tensors.push_back(d_tens);
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_result_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc
+                  << " c_m_n: " << c_device_result_tensors[i].mDesc << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(ADataType) * a_tensors[i].GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].GetElementSize() +
+                     sizeof(DDataType) * d_tensors[i][0].GetElementSize() * NumDs +
+                     sizeof(EDataType) * c_device_result_tensors[i].GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            for(int j = 0; j < NumDs; ++j)
+            {
+                d_tensors[i][j].GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+            }
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            for(int j = 0; j < NumDs; ++j)
+            {
+                d_tensors[i][j].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            }
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            for(int j = 0; j < NumDs; ++j)
+            {
+                d_tensors[i][j].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            }
+        }
+    }
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(a_tensors[i].GetElementSpaceSize() * sizeof(ADataType)));
+        b_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(b_tensors[i].GetElementSpaceSize() * sizeof(BDataType)));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            c_device_result_tensors[i].GetElementSpaceSize() * sizeof(EDataType)));
+
+        for(int j = 0; j < NumDs; ++j)
+        {
+            d_tensors_device[i].emplace_back(std::make_unique<DeviceMem>(
+                d_tensors[i][j].GetElementSpaceSize() * sizeof(DDataType)));
+        }
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+        for(int j = 0; j < NumDs; ++j)
+        {
+            d_tensors_device[i][j]->ToDevice(d_tensors[i][j].mData.data());
+        }
+        c_tensors_device[i]->SetZero();
+
+        p_As.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_Bs.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_Ds.push_back(
+            {d_tensors_device[i][0]->GetDeviceBuffer(), d_tensors_device[i][1]->GetDeviceBuffer()});
+        p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
+
+        // The device op does not have to know M problem size at lunch time.
+        gemm_descs.push_back({0,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              problem_size.stride_As[i],
+                              problem_size.stride_Bs[i],
+                              problem_size.stride_Cs[i],
+                              {problem_size.stride_Cs[i], problem_size.stride_Cs[i]}});
+        ggemm_kargs.push_back(
+            {a_tensors_device[i]->GetDeviceBuffer(),
+             b_tensors_device[i]->GetDeviceBuffer(),
+             {d_tensors_device[i][0]->GetDeviceBuffer(), d_tensors_device[i][1]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             problem_size.stride_As[i],
+             problem_size.stride_Bs[i],
+             {problem_size.stride_Cs[i], problem_size.stride_Cs[i]},
+             problem_size.stride_Cs[i]});
+    }
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
+                              ggemm_kargs.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+    gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
+
+    invoker.Run(argument, StreamConfig{nullptr, false, 1});
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceGemmMultipleD<ADataType,
+                                                               BDataType,
+                                                               DsDataType,
+                                                               EDataType,
+                                                               AccDataType,
+                                                               AElementOp,
+                                                               BElementOp,
+                                                               CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            auto karg = ggemm_kargs[i];
+            auto dev_res_tensor =
+                Tensor<float>(f_host_tensor_descriptor(karg.M, karg.N, karg.StrideE, ELayout{}));
+            c_tensors_device[i]->FromDevice(c_device_result_tensors[i].mData.data());
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      d_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      cde_element_op);
+
+            ref_invoker.Run(ref_argument);
+            pass &= ck::utils::check_err(c_device_result_tensors[i], c_host_tensors[i]);
+        }
+
+        std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << "!" << std::endl;
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        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;
+    }
+
+    return pass;
+}
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+    std::istringstream in(input);
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+    return out;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    if(argc < 10)
+    {
+        std::vector<ck::index_t> Ms{64, 127, 255, 129, 260, 190, 77};
+        problem_size.group_count = Ms.size();
+
+        for(int i = 0; i < problem_size.group_count; i++)
+        {
+            problem_size.Ms.push_back(Ms[i]);
+            problem_size.Ns.push_back(252);
+            problem_size.Ks.push_back(4608);
+
+            problem_size.stride_As.push_back(problem_size.Ks[i]);
+            problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+            problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+
+            problem_size.stride_Ds.push_back({});
+            for(int j = 0; j < NumDs; ++j)
+            {
+                problem_size.stride_Ds[i].push_back(problem_size.Ns[i]);
+            }
+        }
+
+        std::cout
+            << "Usage:\n"
+            << "arg1: verification (0=no, 1=yes)\n"
+            << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+            << "arg3: time kernel (0=n0, 1=yes)\n"
+            << "arg4 to 9: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n"
+            << "... setting default values." << std::endl;
+    }
+    else
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.Ms = argToIntArray(argv[4]);
+        problem_size.Ns = argToIntArray(argv[5]);
+        problem_size.Ks = argToIntArray(argv[6]);
+
+        problem_size.stride_As = argToIntArray(argv[7]);
+        problem_size.stride_Bs = argToIntArray(argv[8]);
+        problem_size.stride_Cs = argToIntArray(argv[9]);
+
+        for(int j = 0; j < NumDs; ++j)
+        {
+            problem_size.stride_Ds.push_back(problem_size.stride_Cs);
+        }
+
+        problem_size.group_count = problem_size.Ms.size();
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

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)