DeviceOp + GridwiseGemm Draft GroupedGEMM+SplitK+TileLoop

* First Part: accumulation across tiles in CThreadBuffer

example/15_grouped_gemm/CMakeLists.txt

@@ -24,7 +24,11 @@ add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
 add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
 
 add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
-add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
+add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
+
+add_example_executable(example_grouped_gemm_multiple_d_splitk_xdl_fp16 grouped_gemm_multiple_d_splitk_xdl_fp16.cpp)
+add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_multiple_d_splitk_xdl_fp16)
+
 
 if(USE_BITINT_EXTENSION_INT4)
     add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)

example/15_grouped_gemm/grouped_gemm_multiple_d_splitk_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_splitk_xdl_cshuffle_tile_loop.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+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 ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmMultipleDSplitKXdlCShuffle
+    // 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| AThreadTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BThreadTransfer| 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|        SrcReset| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|        SrcReset| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1| CoordinateAfter|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1| CoordinateAfter|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |             Run|          |                |               |               |              |               |               |             Run|          |            |            |                             |                |
+        < 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<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,           false,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,           false,         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<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    int k_batch          = 1;
+    bool time_kernel     = false;
+};
+
+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;
+    std::vector<void*> p_Cs;
+    std::vector<const void*> p_As;
+    std::vector<const void*> p_Bs;
+
+    gemm_descs.reserve(group_count);
+    p_As.reserve(group_count);
+    p_Bs.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<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_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{})));
+        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_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_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].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.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});
+            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});
+            break;
+        default:
+            // a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            // b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            ck::utils::FillConstant<ADataType>{1.f}(a_tensors[i]);
+            ck::utils::FillConstant<BDataType>{1.f}(b_tensors[i]);
+        }
+    }
+
+    using GroupedGemmKernelArgument =
+        ck::tensor_operation::device::GroupedGemmMultipleDKernelArguments<>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * problem_size.Ms[i] * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * problem_size.Ms[i] * problem_size.Ns[i]));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
+                                      a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
+        c_tensors_device[i]->SetZero();
+
+        p_As.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_Bs.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
+
+        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],
+                              {}});
+
+        grouped_gemm_kernel_args_.push_back({a_tensors_device[i]->GetDeviceBuffer(),
+                                             b_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],
+                                             {},
+                                             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
+    gemm.SetKBatchSize(argument, config.k_batch);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    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;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    std::vector<ck::index_t> Ms{64};
+
+    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(128);
+        problem_size.Ks.push_back(128);
+
+        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]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (> 0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

include/ck/tensor_operation/gpu/device/device_grouped_gemm_multiple_d_splitk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <iostream>
+#include <vector>
+#include <sstream>
+
+#include "device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+///
+/// @brief      Structure representing single GEMM problem arguments.
+///
+///             The pointer to the vector of those structures is passed to the GroupedGEMM entry
+///             point kernel.
+///
+/// @tparam     NumDTensor  The number of D input tensors.
+///
+template <index_t NumDTensor = 0>
+struct GroupedGemmMultipleDKernelArguments
+{
+    __host__ __device__
+    GroupedGemmMultipleDKernelArguments(const void* p_a_grid_,
+                                        const void* p_b_grid_,
+                                        std::array<const void*, NumDTensor> p_ds_grid_,
+                                        void* p_e_grid_,
+                                        index_t M_,
+                                        index_t N_,
+                                        index_t K_,
+                                        index_t StrideA_,
+                                        index_t StrideB_,
+                                        std::array<index_t, NumDTensor> StrideDs_,
+                                        index_t StrideE_)
+        : p_a_grid{p_a_grid_},
+          p_b_grid{p_b_grid_},
+          p_ds_grid{p_ds_grid_},
+          p_e_grid{p_e_grid_},
+          M{M_},
+          N{N_},
+          K{K_},
+          StrideA{StrideA_},
+          StrideB{StrideB_},
+          StrideDs{StrideDs_},
+          StrideE{StrideE_}
+    {
+    }
+
+    const void* p_a_grid;
+    const void* p_b_grid;
+    std::array<const void*, NumDTensor> p_ds_grid;
+    void* p_e_grid;
+    index_t M;
+    index_t N;
+    index_t K;
+    index_t StrideA;
+    index_t StrideB;
+    std::array<index_t, NumDTensor> StrideDs;
+    index_t StrideE;
+
+    void Print() const
+    {
+        std::stringstream str;
+        for(auto sd : StrideDs)
+            str << sd << ",";
+
+        std::cout << "arg {"
+                  << "M:" << M << ", "
+                  << "N:" << N << ", "
+                  << "K:" << K << ", "
+                  << "SA:" << StrideA << ", "
+                  << "SB:" << StrideB << ", "
+                  << "SE:" << StrideE << ", "
+                  << "SDs: {" << str.str() << "}"
+                  << "}" << std::endl;
+    }
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedGemmMultipleDSplitK : public DeviceGroupedGemm<ALayout,
+                                                                   BLayout,
+                                                                   DsLayout,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   DsDataType,
+                                                                   EDataType,
+                                                                   AElementwiseOperation,
+                                                                   BElementwiseOperation,
+                                                                   CDEElementwiseOperation>
+{
+    //----------------------------------------------------------------------------------------------
+    /// @brief      Sets the k batch size.
+    ///
+    /// @param      p_arg   Pointer to the Argument we're going to change.
+    /// @param[in]  kbatch  The kbatch value.
+    ///
+    virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
+
+    //----------------------------------------------------------------------------------------------
+    /// @brief      Sets the device kernel arguments pointer.
+    ///
+    /// @param      p_arg              The pointer to the Argument we're going to update.
+    /// @param[in]  p_dev_kernel_args  The pointer to the device memory which contains kernel
+    ///                                arguments.
+    ///
+    virtual void SetDeviceKernelArgs(BaseArgument* p_arg, const void* p_dev_kernel_args) const = 0;
+
+    //----------------------------------------------------------------------------------------------
+    /// @brief      Gets the device kernel argument size.
+    ///
+    /// @param[in]  p_arg  The pointer to the Device op Argument.
+    ///
+    /// @return     The device kernel argument size.
+    ///
+    virtual size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
 #pragma once
 #include <iostream>
 #include <vector>