AccCpuThreads.hpp

Go to the documentation of this file.

/* Copyright 2026 Axel Huebl, Benjamin Worpitz, René Widera, Jan Stephan, Bernhard Manfred Gruber, Andrea Bocci
 * SPDX-License-Identifier: MPL-2.0
 */
 
#pragma once
 
// Base classes.
#include "alpaka/atomic/AtomicCpu.hpp"
#include "alpaka/atomic/AtomicHierarchy.hpp"
#include "alpaka/block/shared/dyn/BlockSharedMemDynMember.hpp"
#include "alpaka/block/shared/st/BlockSharedMemStMemberMasterSync.hpp"
#include "alpaka/block/sync/BlockSyncBarrierThread.hpp"
#include "alpaka/core/DemangleTypeNames.hpp"
#include "alpaka/grid/GridSyncBarrierCpuThread.hpp"
#include "alpaka/idx/bt/IdxBtRefThreadIdMap.hpp"
#include "alpaka/idx/gb/IdxGbRef.hpp"
#include "alpaka/intrinsic/IntrinsicCpu.hpp"
#include "alpaka/math/MathStdLib.hpp"
#include "alpaka/mem/fence/MemFenceCpu.hpp"
#include "alpaka/rand/RandDefault.hpp"
#include "alpaka/rand/RandStdLib.hpp"
#include "alpaka/warp/WarpSingleThread.hpp"
#include "alpaka/workdiv/WorkDivMembers.hpp"
 
// Specialized traits.
#include "alpaka/acc/Traits.hpp"
#include "alpaka/dev/Traits.hpp"
#include "alpaka/idx/Traits.hpp"
#include "alpaka/kernel/Traits.hpp"
#include "alpaka/platform/Traits.hpp"
 
// Implementation details.
#include "alpaka/acc/Tag.hpp"
#include "alpaka/core/ClipCast.hpp"
#include "alpaka/core/Config.hpp"
#include "alpaka/core/Interface.hpp"
#include "alpaka/dev/DevCpu.hpp"
 
#ifdef __cpp_lib_format
#    include <format>
#endif
#include <memory>
#include <string>
#include <thread>
 
#ifdef ALPAKA_ACC_CPU_B_SEQ_T_THREADS_ENABLED
 
namespace alpaka
{
    template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
    class TaskKernelCpuThreads;
 
    //! The CPU threads accelerator.
    //!
    //! This accelerator allows parallel kernel execution on a CPU device.
    //! It uses std::thread to implement the parallelism.
    template<typename TDim, typename TIdx>
    class AccCpuThreads final
        : public WorkDivMembers<TDim, TIdx>
        , public gb::IdxGbRef<TDim, TIdx>
        , public bt::IdxBtRefThreadIdMap<TDim, TIdx>
        , public AtomicHierarchy<
              AtomicCpu, // grid atomics
              AtomicCpu, // block atomics
              AtomicCpu> // thread atomics
        , public math::MathStdLib
        , public BlockSharedMemDynMember<>
        , public BlockSharedMemStMemberMasterSync<>
        , public BlockSyncBarrierThread<TIdx>
        , public GridSyncBarrierThread<TIdx>
        , public IntrinsicCpu
        , public MemFenceCpu
#    ifdef ALPAKA_DISABLE_VENDOR_RNG
        , public rand::RandDefault
#    else
        , public rand::RandStdLib
#    endif
        , public warp::WarpSingleThread
        , public interface::Implements<InterfaceAcc, AccCpuThreads<TDim, TIdx>>
    {
        static_assert(
            sizeof(TIdx) >= sizeof(int),
            "Index type is not supported, consider using int or a larger type.");
 
    public:
        // Partial specialization with the correct TDim and TIdx is not allowed.
        template<typename TDim2, typename TIdx2, typename TKernelFnObj, typename... TArgs>
        friend class ::alpaka::TaskKernelCpuThreads;
 
        AccCpuThreads(AccCpuThreads const&) = delete;
        AccCpuThreads(AccCpuThreads&&) = delete;
        auto operator=(AccCpuThreads const&) -> AccCpuThreads& = delete;
        auto operator=(AccCpuThreads&&) -> AccCpuThreads& = delete;
 
    private:
        template<typename TWorkDiv>
        ALPAKA_FN_HOST AccCpuThreads(TWorkDiv const& workDiv, std::size_t const& blockSharedMemDynSizeBytes)
            : WorkDivMembers<TDim, TIdx>(workDiv)
            , gb::IdxGbRef<TDim, TIdx>(m_gridBlockIdx)
            , bt::IdxBtRefThreadIdMap<TDim, TIdx>(m_threadToIndexMap)
            , BlockSharedMemDynMember<>(blockSharedMemDynSizeBytes)
            , BlockSharedMemStMemberMasterSync<>(
                  staticMemBegin(),
                  staticMemCapacity(),
                  [this]() { syncBlockThreads(*this); },
                  [this]() noexcept { return (m_idMasterThread == std::this_thread::get_id()); })
            , BlockSyncBarrierThread<TIdx>(getWorkDiv<Block, Threads>(workDiv).prod())
            , GridSyncBarrierThread<TIdx>(getWorkDiv<Block, Threads>(workDiv).prod())
            , m_gridBlockIdx(Vec<TDim, TIdx>::zeros())
        {
        }
 
    private:
        // getIdx
        std::mutex mutable m_mtxMapInsert; //!< The mutex used to secure insertion into the ThreadIdToIdxMap.
        typename bt::IdxBtRefThreadIdMap<TDim, TIdx>::
            ThreadIdToIdxMap mutable m_threadToIndexMap; //!< The mapping of thread id's to indices.
        Vec<TDim, TIdx> mutable m_gridBlockIdx; //!< The index of the currently executed block.
 
        // allocBlockSharedArr
        std::thread::id mutable m_idMasterThread; //!< The id of the master thread.
    };
 
    namespace trait
    {
        //! The CPU threads accelerator accelerator type trait specialization.
        template<typename TDim, typename TIdx>
        struct AccType<AccCpuThreads<TDim, TIdx>>
        {
            using type = AccCpuThreads<TDim, TIdx>;
        };
 
        //! The CPU threads single thread accelerator type trait specialization.
        template<typename TDim, typename TIdx>
        struct IsSingleThreadAcc<AccCpuThreads<TDim, TIdx>> : std::false_type
        {
        };
 
        //! The CPU threads multi thread accelerator type trait specialization.
        template<typename TDim, typename TIdx>
        struct IsMultiThreadAcc<AccCpuThreads<TDim, TIdx>> : std::true_type
        {
        };
 
        //! The CPU threads accelerator device properties get trait specialization.
        template<typename TDim, typename TIdx>
        struct GetAccDevProps<AccCpuThreads<TDim, TIdx>>
        {
            ALPAKA_FN_HOST static auto getAccDevProps(DevCpu const& dev) -> AccDevProps<TDim, TIdx>
            {
#    ifdef ALPAKA_CI
                auto const blockThreadCountMax = static_cast<TIdx>(8);
#    else
                // \TODO: Magic number. What is the maximum? Just set a reasonable value? There is a implementation
                // defined maximum where the creation of a new thread crashes. std::thread::hardware_concurrency can
                // return 0, so 1 is the default case?
                auto const blockThreadCountMax = std::max(
                    static_cast<TIdx>(1),
                    alpaka::core::clipCast<TIdx>(std::thread::hardware_concurrency() * 8));
#    endif
                auto const memBytes = getMemBytes(dev);
                return {// m_multiProcessorCount
                        static_cast<TIdx>(1),
                        // m_gridBlockExtentMax
                        Vec<TDim, TIdx>::all(std::numeric_limits<TIdx>::max()),
                        // m_gridBlockCountMax
                        std::numeric_limits<TIdx>::max(),
                        // m_blockThreadExtentMax
                        Vec<TDim, TIdx>::all(blockThreadCountMax),
                        // m_blockThreadCountMax
                        blockThreadCountMax,
                        // m_threadElemExtentMax
                        Vec<TDim, TIdx>::all(std::numeric_limits<TIdx>::max()),
                        // m_threadElemCountMax
                        std::numeric_limits<TIdx>::max(),
                        // m_sharedMemSizeBytes
                        memBytes,
                        // m_globalMemSizeBytes
                        memBytes,
                        // m_cooperativeLaunch
                        true};
            }
        };
 
        //! The CPU threads accelerator name trait specialization.
        template<typename TDim, typename TIdx>
        struct GetAccName<AccCpuThreads<TDim, TIdx>>
        {
            ALPAKA_FN_HOST static auto getAccName() -> std::string
            {
#    if ALPAKA_COMP_CLANG
#        pragma clang diagnostic push
#        pragma clang diagnostic ignored "-Wexit-time-destructors"
#    endif
                using namespace std::literals;
                static std::string const accName =
#    ifdef __cpp_lib_format
                    std::format("AccCpuThreads<{},{}>", TDim::value, core::demangled<TIdx>);
#    else
                    "AccCpuThreads<"s + std::to_string(TDim::value) + ","s + std::string(core::demangled<TIdx>) + ">"s;
#    endif
                return accName;
#    if ALPAKA_COMP_CLANG
#        pragma clang diagnostic pop
#    endif
            }
        };
 
        //! The CPU threads accelerator device type trait specialization.
        template<typename TDim, typename TIdx>
        struct DevType<AccCpuThreads<TDim, TIdx>>
        {
            using type = DevCpu;
        };
 
        //! The CPU threads accelerator dimension getter trait specialization.
        template<typename TDim, typename TIdx>
        struct DimType<AccCpuThreads<TDim, TIdx>>
        {
            using type = TDim;
        };
 
        //! The CPU threads accelerator execution task type trait specialization.
        template<typename TDim, typename TIdx, typename TWorkDiv, typename TKernelFnObj, typename... TArgs>
        struct CreateTaskKernel<AccCpuThreads<TDim, TIdx>, TWorkDiv, TKernelFnObj, TArgs...>
        {
            ALPAKA_FN_HOST static auto createTaskKernel(
                TWorkDiv const& workDiv,
                TKernelFnObj const& kernelFnObj,
                TArgs&&... args)
            {
                return TaskKernelCpuThreads<TDim, TIdx, TKernelFnObj, TArgs...>(
                    workDiv,
                    kernelFnObj,
                    std::forward<TArgs>(args)...);
            }
        };
 
        //! The CPU threads thread accelerator execution cooperative task type trait specialization.
        template<typename TDim, typename TIdx, typename TWorkDiv, typename TKernelFnObj, typename... TArgs>
        struct CreateTaskCooperativeKernel<AccCpuThreads<TDim, TIdx>, TWorkDiv, TKernelFnObj, TArgs...>
        {
            ALPAKA_FN_HOST static auto createTaskCooperativeKernel(
                TWorkDiv const& workDiv,
                TKernelFnObj const& kernelFnObj,
                TArgs&&... args)
            {
                auto const gridBlockExtent = getWorkDiv<Grid, Blocks>(workDiv);
                if(gridBlockExtent.prod() != static_cast<TIdx>(1u))
                {
                    throw std::runtime_error(
                        "The std::thread accelerator supports only a single block operation with cooperative kernel!\n"
                        "Consider useing a different CPU accelerator.");
                }
 
                return TaskKernelCpuThreads<TDim, TIdx, TKernelFnObj, TArgs...>(
                    workDiv,
                    kernelFnObj,
                    std::forward<TArgs>(args)...);
            }
        };
 
        //! The CPU threads execution task platform type trait specialization.
        template<typename TDim, typename TIdx>
        struct PlatformType<AccCpuThreads<TDim, TIdx>>
        {
            using type = PlatformCpu;
        };
 
        //! The CPU threads accelerator idx type trait specialization.
        template<typename TDim, typename TIdx>
        struct IdxType<AccCpuThreads<TDim, TIdx>>
        {
            using type = TIdx;
        };
 
        template<typename TDim, typename TIdx>
        struct AccToTag<alpaka::AccCpuThreads<TDim, TIdx>>
        {
            using type = alpaka::TagCpuThreads;
        };
 
        template<typename TDim, typename TIdx>
        struct TagToAcc<alpaka::TagCpuThreads, TDim, TIdx>
        {
            using type = alpaka::AccCpuThreads<TDim, TIdx>;
        };
    } // namespace trait
} // namespace alpaka
 
#endif