Merge branch 'amd-develop-0605' into amd-master

example/01_gemm/README.md

@@ -7,17 +7,3 @@
 #arg3: run kernel # of times (>1)
 ./bin/example_gemm_xdl 0 1 5
 ```
-
-Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
-a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
-b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
-c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
-arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
-arg.c_grid_desc_m_n_{ 3840, 4096}
-launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
-Warm up
-Start running 5 times...
-Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
-```

example/01_gemm/common.hpp

@@ -46,6 +46,19 @@ struct ProblemSizeStreamK final
     ck::index_t NumSKBlocks = -1;
 };
 
+struct ProblemSizeSplitK final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    ck::index_t KBatch = 1;
+};
+
 struct ExecutionConfig final
 {
     bool do_verification = true;
@@ -158,3 +171,52 @@ bool parse_cmd_args<ProblemSizeStreamK>(int argc,
 
     return true;
 }
+
+template <>
+bool parse_cmd_args<ProblemSizeSplitK>(int argc,
+                                       char* argv[],
+                                       ProblemSizeSplitK& problem_size,
+                                       ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc >= 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+
+        if(argc >= 11)
+        {
+            problem_size.KBatch = std::stoi(argv[10]);
+        }
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
+                  << "arg10: KBatch" << std::endl;
+        return false;
+    }
+
+    return true;
+}

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/gemm_xdl_fp16_fp8_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::f8_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_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, 
+        AElementOp, BElementOp, CElementOp, GemmDefault, 
+        64,
+        16, 16, 
+        64, 16, 8,
+        16,   16,
+        1,    1, 
+        S<4, 16, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 16, 16, 0,
+        S<8, 8, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        1, 1, S<1, 16, 1, 4>, 4,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+
+#include "run_gemm_example_v2.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

example/01_gemm/gemm_xdl_fp16_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::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// 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,
+        224, 256, 
+        64, 8, 2,
+        16,   16,
+        7,    8,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        S<8, 32, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
+        1, 8, 2, 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/gemm_xdl_fp8_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::f8_t;
+using BDataType        = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_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, 256, 
+        128, 16, 16,
+        16,   16,
+        4,    8, 
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 16, 16, 1,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 16, 16, 1,
+        1, 2, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3, ck::f8_t>;
+// 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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#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)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    using namespace ck::literals;
+
+    auto M       = problem_size.M;
+    auto N       = problem_size.N;
+    auto K       = problem_size.K;
+    auto StrideA = problem_size.StrideA;
+    auto StrideB = problem_size.StrideB;
+    auto StrideC = problem_size.StrideC;
+    auto KBatch  = problem_size.KBatch;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+#if 0
+    printf("B matrix:\n");
+    for (int in = 0; in < N; in++)
+    {
+        for (int ik = 0; ik < K; ik++)
+        {
+            printf("%02x ", *(reinterpret_cast<uint8_t*>(&b_k_n(ik,in))));
+            if(ik%8==7) printf("|");
+        }
+        printf("\n");
+    }
+#endif
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
+                               c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+    DeviceMem workspace;
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm      = DeviceGemmV2Instance{};
+    auto invoker   = gemm.MakeInvoker();
+    float ave_time = 0;
+
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#endif
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        StrideC,
+        KBatch,
+        a_element_op,
+        b_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return true;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 1});
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
+
+        c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
+
+        c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
+
+        return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
+#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,
+                                     "Error: Incorrect results!",
+                                     get_rtol<CDataType>(),
+                                     get_atol<CDataType>());
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop = 2_uz * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, " << gemm.GetTypeString() << std::endl;
+    }
+    return pass;
+}
+
+bool run_gemm_splitk_example(int argc, char* argv[])
+{
+    ProblemSizeSplitK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}

example/02_gemm_bilinear/README.md

@@ -9,20 +9,3 @@
 #arg11 to 12: alpha, beta
 ./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
 ```
-Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
-```
-a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
-b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
-c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
-arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
-arg.c0_grid_desc_m_n_{ 3840, 4096}
-arg.c_grid_desc_m_n_{ 3840, 4096}
-launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
-Warm up
-Start running 1 times...
-Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
-error: 0
-max_diff: 0, 558.5, 558.5
-```

example/04_gemm_add_add_fastgelu/README.md

@@ -8,16 +8,3 @@
 #arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
 ./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
 ```
-
-Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
-a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
-b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
-d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
-d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
-launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
-Warm up 1 time
-Start running 10 times...
-Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
-```

example/09_convnd_fwd/README.md

@@ -16,17 +16,3 @@
 # <right padding>, (ie RightPy, RightPx for 2D)
 ./bin/example_convnd_fwd_xdl 0 1 100
 ```
-
-Result (MI100 @ 1087Mhz, 33.4TFlops peak FP32)
-```
-input: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
-weights: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
-output: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
-arg.a_grid_desc_k0_m_k1_{432, 165888, 4}
-arg.b_grid_desc_k0_n_k1_{432, 256, 4}
-arg.c_grid_desc_m_n_{ 165888, 256}
-launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
-Warm up
-Start running 100 times...
-Perf: 4.43736 ms, 33.0753 TFlops, 150.357 GB/s
-```

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/README.md

@@ -7,19 +7,3 @@
 #arg3: run kernel # of times (>1)
 ./bin/example_grouped_gemm_xdl_fp16 0 1 5
 ```
-
-Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
-gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
-gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
-gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
-gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
-group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
-group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
-group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
-group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
-launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1} 
-Warm up
-Start running 5 times...
-Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
-```

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>;
+    using GemmDesc        = ck::tensor_operation::device::GemmDesc;
+
+    // GEMM shape
+    std::vector<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/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -27,7 +27,12 @@ using DeviceElementwiseAddInstance =
                                                         ck::Tuple<CDataType>,
                                                         Add,
                                                         2,
+                                                        64,
+                                                        64,
+                                                        64,
                                                         8,
+                                                        8,
+                                                        ck::Sequence<1, 0>,
                                                         ck::Sequence<8, 8>,
                                                         ck::Sequence<8>>;
 

example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -27,9 +27,14 @@ using DeviceElementwiseAddInstance =
                                                         ck::Tuple<CDataType>,
                                                         Add,
                                                         3,
-                                                        8,
-                                                        ck::Sequence<1, 8>,
-                                                        ck::Sequence<8>>;
+                                                        64,
+                                                        16,
+                                                        16,
+                                                        2,
+                                                        2,
+                                                        ck::Sequence<1, 0>,
+                                                        ck::Sequence<1, 2>,
+                                                        ck::Sequence<2>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_broadcast3D_am_bmnk(HostTensorC& C,

example/19_binary_elementwise/elementwise_add_1d.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <cstdlib>
 
 #include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -25,9 +25,14 @@ using DeviceElementwiseAddInstance =
                                                         ck::Tuple<CDataType>,
                                                         Add,
                                                         1,
-                                                        8,
-                                                        ck::Sequence<8, 8>,
-                                                        ck::Sequence<8>>;
+                                                        64,
+                                                        16,
+                                                        16,
+                                                        2,
+                                                        2,
+                                                        ck::Sequence<1, 0>,
+                                                        ck::Sequence<2, 2>,
+                                                        ck::Sequence<2>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_elementwise1D(

example/19_binary_elementwise/elementwise_add_4d.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
 
 #include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -27,9 +27,14 @@ using DeviceElementwiseAddInstance =
                                                         ck::Tuple<CDataType>,
                                                         Add,
                                                         4,
-                                                        8,
-                                                        ck::Sequence<8, 8>,
-                                                        ck::Sequence<8>>;
+                                                        64,
+                                                        2,
+                                                        128,
+                                                        2,
+                                                        2,
+                                                        ck::Sequence<1, 0>,
+                                                        ck::Sequence<2, 2>,
+                                                        ck::Sequence<2>>;
 
 template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
 void host_elementwise4D(HostTensorC& C,

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -78,6 +78,9 @@ using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWe
                                                                                      InElementOp,
                                                                                      WeiElementOp,
                                                                                      OutElementOp,
+                                                                                     0,
+                                                                                     0,
+                                                                                     0,
                                                                                      ComputeTypeA,
                                                                                      ComputeTypeB>;
 

example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc

@@ -119,7 +119,10 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
                                                   conv_param.input_right_pads_,
                                                   InElementOp{},
                                                   WeiElementOp{},
-                                                  OutElementOp{});
+                                                  OutElementOp{},
+                                                  {},
+                                                  {},
+                                                  {});
 
         ref_invoker.Run(ref_argument);
 

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <numeric>
@@ -9,7 +9,7 @@
 #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_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/device_memory.hpp"
@@ -103,9 +103,14 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseI
     ck::Tuple<LayerNormOutDataType>, // y
     NormalizeFunctor,
     2,
-    8,                           // MPerthread
-    ck::Sequence<8, 1, 1, 8, 8>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
-    ck::Sequence<8>>;            // scalarPerVector: y(layerNorm_out)
+    64,                          // BlockSize
+    16,                          // MPerBlock
+    16,                          // NPerBlock
+    2,                           // MPerthread
+    2,                           // NPerthread
+    ck::Sequence<1, 0>,          // ThreadClusterArrangeOrder
+    ck::Sequence<2, 1, 1, 2, 2>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
+    ck::Sequence<2>>;            // scalarPerVector: y(layerNorm_out)
 
 auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
     return HostTensorDescriptor({len}, {stride});