TaskKernelCpuOmp2Threads.hpp

Go to the documentation of this file.

/* Copyright 2022 Axel Huebl, Benjamin Worpitz, Bert Wesarg, René Widera, Jan Stephan, Bernhard Manfred Gruber
 * SPDX-License-Identifier: MPL-2.0
 */
 
#pragma once
 
// Specialized traits.
#include "alpaka/acc/Traits.hpp"
#include "alpaka/dev/Traits.hpp"
#include "alpaka/dim/Traits.hpp"
#include "alpaka/idx/Traits.hpp"
#include "alpaka/platform/Traits.hpp"
 
// Implementation details.
#include "alpaka/acc/AccCpuOmp2Threads.hpp"
#include "alpaka/core/Decay.hpp"
#include "alpaka/dev/DevCpu.hpp"
#include "alpaka/kernel/KernelFunctionAttributes.hpp"
#include "alpaka/kernel/Traits.hpp"
#include "alpaka/meta/NdLoop.hpp"
#include "alpaka/platform/PlatformCpu.hpp"
#include "alpaka/workdiv/WorkDivMembers.hpp"
 
#include <functional>
#include <stdexcept>
#include <tuple>
#include <type_traits>
#if ALPAKA_DEBUG >= ALPAKA_DEBUG_MINIMAL
#    include <iostream>
#endif
 
#ifdef ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED
 
#    if _OPENMP < 200203
#        error If ALPAKA_ACC_CPU_B_SEQ_T_OMP2_ENABLED is set, the compiler has to support OpenMP 2.0 or higher!
#    endif
 
#    include <omp.h>
 
namespace alpaka
{
    //! The CPU OpenMP 2.0 thread accelerator execution task.
    template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
    class TaskKernelCpuOmp2Threads final : public WorkDivMembers<TDim, TIdx>
    {
    public:
        template<typename TWorkDiv>
        ALPAKA_FN_HOST TaskKernelCpuOmp2Threads(TWorkDiv&& workDiv, TKernelFnObj const& kernelFnObj, TArgs&&... args)
            : WorkDivMembers<TDim, TIdx>(std::forward<TWorkDiv>(workDiv))
            , m_kernelFnObj(kernelFnObj)
            , m_args(std::forward<TArgs>(args)...)
        {
            static_assert(
                Dim<std::decay_t<TWorkDiv>>::value == TDim::value,
                "The work division and the execution task have to be of the same dimensionality!");
        }
 
        //! Executes the kernel function object.
        ALPAKA_FN_HOST auto operator()() const -> void
        {
            ALPAKA_DEBUG_MINIMAL_LOG_SCOPE;
 
            auto const gridBlockExtent = getWorkDiv<Grid, Blocks>(*this);
            auto const blockThreadExtent = getWorkDiv<Block, Threads>(*this);
            auto const threadElemExtent = getWorkDiv<Thread, Elems>(*this);
 
            // Get the size of the block shared dynamic memory.
            auto const blockSharedMemDynSizeBytes = std::apply(
                [&](std::decay_t<TArgs> const&... args)
                {
                    return getBlockSharedMemDynSizeBytes<AccCpuOmp2Threads<TDim, TIdx>>(
                        m_kernelFnObj,
                        blockThreadExtent,
                        threadElemExtent,
                        args...);
                },
                m_args);
 
#    if ALPAKA_DEBUG >= ALPAKA_DEBUG_FULL
            std::cout << __func__ << " blockSharedMemDynSizeBytes: " << blockSharedMemDynSizeBytes << " B"
                      << std::endl;
#    endif
 
            AccCpuOmp2Threads<TDim, TIdx> acc(
                *static_cast<WorkDivMembers<TDim, TIdx> const*>(this),
                blockSharedMemDynSizeBytes);
 
            // The number of threads in this block.
            TIdx const blockThreadCount(blockThreadExtent.prod());
            [[maybe_unused]] int const iBlockThreadCount(static_cast<int>(blockThreadCount));
 
            if(::omp_in_parallel() != 0)
            {
                throw std::runtime_error(
                    "The OpenMP 2.0 thread backend can not be used within an existing parallel region!");
            }
 
            // Force the environment to use the given number of threads.
            int const ompIsDynamic(::omp_get_dynamic());
            ::omp_set_dynamic(0);
 
            // Execute the blocks serially.
            meta::ndLoopIncIdx(
                gridBlockExtent,
                [&](Vec<TDim, TIdx> const& gridBlockIdx)
                {
                    acc.m_gridBlockIdx = gridBlockIdx;
 
// Execute the threads in parallel.
 
// Parallel execution of the threads in a block is required because when syncBlockThreads is called all of them have to
// be done with their work up to this line. So we have to spawn one OS thread per thread in a block. 'omp for' is not
// useful because it is meant for cases where multiple iterations are executed by one thread but in our case a 1:1
// mapping is required. Therefore we use 'omp parallel' with the specified number of threads in a block.
#    pragma omp parallel num_threads(iBlockThreadCount)
                    {
                        // The guard is for gcc internal compiler error, as discussed in #735
                        if constexpr((!BOOST_COMP_GNUC) || (BOOST_COMP_GNUC >= BOOST_VERSION_NUMBER(8, 1, 0)))
                        {
#    pragma omp single nowait
                            {
                                // The OpenMP runtime does not create a parallel region when only one thread is
                                // required in the num_threads clause. In all other cases we expect to be in a parallel
                                // region now.
                                if((iBlockThreadCount > 1) && (::omp_in_parallel() == 0))
                                {
                                    throw std::runtime_error(
                                        "The OpenMP 2.0 runtime did not create a parallel region!");
                                }
 
                                int const numThreads = ::omp_get_num_threads();
                                if(numThreads != iBlockThreadCount)
                                {
                                    throw std::runtime_error(
                                        "The OpenMP 2.0 runtime did not use the number of threads "
                                        "that had been required!");
                                }
                            }
                        }
 
                        std::apply(m_kernelFnObj, std::tuple_cat(std::tie(acc), m_args));
 
                        // Wait for all threads to finish before deleting the shared memory.
                        // This is done by default if the omp 'nowait' clause is missing on the omp parallel directive
                        // syncBlockThreads(acc);
                    }
 
                    // After a block has been processed, the shared memory has to be deleted.
                    freeSharedVars(acc);
                });
 
            // Reset the dynamic thread number setting.
            ::omp_set_dynamic(ompIsDynamic);
        }
 
    private:
        TKernelFnObj m_kernelFnObj;
        std::tuple<std::decay_t<TArgs>...> m_args;
    };
 
    namespace trait
    {
        //! The CPU OpenMP 2.0 block thread execution task accelerator type trait specialization.
        template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
        struct AccType<TaskKernelCpuOmp2Threads<TDim, TIdx, TKernelFnObj, TArgs...>>
        {
            using type = AccCpuOmp2Threads<TDim, TIdx>;
        };
 
        //! The CPU OpenMP 2.0 block thread execution task device type trait specialization.
        template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
        struct DevType<TaskKernelCpuOmp2Threads<TDim, TIdx, TKernelFnObj, TArgs...>>
        {
            using type = DevCpu;
        };
 
        //! The CPU OpenMP 2.0 block thread execution task dimension getter trait specialization.
        template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
        struct DimType<TaskKernelCpuOmp2Threads<TDim, TIdx, TKernelFnObj, TArgs...>>
        {
            using type = TDim;
        };
 
        //! The CPU OpenMP 2.0 block thread execution task platform type trait specialization.
        template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
        struct PlatformType<TaskKernelCpuOmp2Threads<TDim, TIdx, TKernelFnObj, TArgs...>>
        {
            using type = PlatformCpu;
        };
 
        //! The CPU OpenMP 2.0 block thread execution task idx type trait specialization.
        template<typename TDim, typename TIdx, typename TKernelFnObj, typename... TArgs>
        struct IdxType<TaskKernelCpuOmp2Threads<TDim, TIdx, TKernelFnObj, TArgs...>>
        {
            using type = TIdx;
        };
 
        //! \brief Specialisation of the class template FunctionAttributes
        //! \tparam TDev The device type.
        //! \tparam TDim The dimensionality of the accelerator device properties.
        //! \tparam TIdx The idx type of the accelerator device properties.
        //! \tparam TKernelFn Kernel function object type.
        //! \tparam TArgs Kernel function object argument types as a parameter pack.
        template<typename TDev, typename TDim, typename TIdx, typename TKernelFn, typename... TArgs>
        struct FunctionAttributes<AccCpuOmp2Threads<TDim, TIdx>, TDev, TKernelFn, TArgs...>
        {
            //! \param dev The device instance
            //! \param kernelFn The kernel function object which should be executed.
            //! \param args The kernel invocation arguments.
            //! \return KernelFunctionAttributes instance. The default version always returns an instance with zero
            //! fields. For CPU, the field of max threads allowed by kernel function for the block is 1.
            ALPAKA_FN_HOST static auto getFunctionAttributes(
                TDev const& dev,
                [[maybe_unused]] TKernelFn const& kernelFn,
                [[maybe_unused]] TArgs&&... args) -> alpaka::KernelFunctionAttributes
            {
                alpaka::KernelFunctionAttributes kernelFunctionAttributes;
 
                // set function properties for maxThreadsPerBlock to device properties, since API doesn't have function
                // properties function.
                auto const& props = alpaka::getAccDevProps<AccCpuOmp2Threads<TDim, TIdx>>(dev);
                kernelFunctionAttributes.maxThreadsPerBlock = static_cast<int>(props.m_blockThreadCountMax);
                kernelFunctionAttributes.maxDynamicSharedSizeBytes
                    = static_cast<int>(alpaka::BlockSharedDynMemberAllocKiB * 1024);
                return kernelFunctionAttributes;
            }
        };
 
    } // namespace trait
} // namespace alpaka
 
#endif