AMReX_GpuLaunch.H

Go to the documentation of this file.

#ifndef AMREX_GPU_LAUNCH_H_
#define AMREX_GPU_LAUNCH_H_
#include <AMReX_Config.H>
 
#include <AMReX_GpuQualifiers.H>
#include <AMReX_GpuKernelInfo.H>
#include <AMReX_GpuControl.H>
#include <AMReX_GpuTypes.H>
#include <AMReX_GpuError.H>
#include <AMReX_GpuRange.H>
#include <AMReX_GpuDevice.H>
#include <AMReX_GpuMemory.H>
#include <AMReX_GpuReduce.H>
#include <AMReX_Arena.H>
#include <AMReX_Tuple.H>
#include <AMReX_Box.H>
#include <AMReX_Loop.H>
#include <AMReX_Extension.H>
#include <AMReX_BLassert.H>
#include <AMReX_TypeTraits.H>
#include <AMReX_GpuLaunchGlobal.H>
#include <AMReX_RandomEngine.H>
#include <AMReX_Algorithm.H>
#include <AMReX_Math.H>
#include <AMReX_Vector.H>
#include <concepts>
#include <cstddef>
#include <limits>
#include <algorithm>
#include <utility>
 
#define AMREX_GPU_NCELLS_PER_THREAD 3
#define AMREX_GPU_Y_STRIDE 1
#define AMREX_GPU_Z_STRIDE 1
 
#ifdef AMREX_USE_CUDA
#  define AMREX_LAUNCH_KERNEL(MT, blocks, threads, sharedMem, stream, ... ) \
        amrex::launch_global<MT><<<blocks, threads, sharedMem, stream>>>(__VA_ARGS__)
#  define AMREX_LAUNCH_KERNEL_NOBOUND(blocks, threads, sharedMem, stream, ... ) \
        amrex::launch_global    <<<blocks, threads, sharedMem, stream>>>(__VA_ARGS__)
#elif defined(AMREX_USE_HIP)
#  define AMREX_LAUNCH_KERNEL(MT, blocks, threads, sharedMem, stream, ... ) \
        hipLaunchKernelGGL(amrex::launch_global<MT>, blocks, threads, sharedMem, stream, __VA_ARGS__)
#  define AMREX_LAUNCH_KERNEL_NOBOUND(blocks, threads, sharedMem, stream, ... ) \
        hipLaunchKernelGGL(amrex::launch_global    , blocks, threads, sharedMem, stream, __VA_ARGS__)
#endif
 
 
namespace amrex {
 
// We cannot take rvalue lambdas.
// ************************************************
//  Variadic lambda function wrappers for C++ CUDA/HIP Kernel calls.
 
#if defined(AMREX_USE_CUDA) || defined(AMREX_USE_HIP)
 
 
    template <typename... Lambdas>
    AMREX_GPU_DEVICE void call_device (Lambdas&&... fs) noexcept
    {
        (std::forward<Lambdas>(fs)(), ...);
    }
 
 
    template<class L, class... Lambdas>
    AMREX_GPU_GLOBAL void launch_global (L f0, Lambdas... fs) { f0(); call_device(fs...); }
 
#endif
 
// CPU variation
 
    template<class L>
    void launch_host (L&& f0) noexcept { std::forward<L>(f0)(); }
 
    template<class L, class... Lambdas>
    void launch_host (L&& f0, Lambdas&&... fs) noexcept {
        std::forward<L>(f0)();
        launch_host(std::forward<Lambdas>(fs)...);
    }
 
 
    template <class T> class LayoutData;
    class FabArrayBase;
 
namespace Gpu {
 
#ifdef AMREX_USE_GPU
    constexpr std::size_t numThreadsPerBlockParallelFor () {
        return AMREX_GPU_MAX_THREADS;
    }
#else
    constexpr std::size_t numThreadsPerBlockParallelFor () { return 0; }
#endif
 
    AMREX_GPU_HOST_DEVICE
    inline
    Box getThreadBox (const Box& bx, Long offset) noexcept
    {
        AMREX_IF_ON_DEVICE((
            const auto len = bx.length3d();
            Long k = offset / (len[0]*len[1]);
            Long j = (offset - k*(len[0]*len[1])) / len[0];
            Long i = (offset - k*(len[0]*len[1])) - j*len[0];
            IntVect iv{AMREX_D_DECL(static_cast<int>(i),
                                    static_cast<int>(j),
                                    static_cast<int>(k))};
            iv += bx.smallEnd();
            return (bx & Box(iv,iv,bx.type()));
        ))
        AMREX_IF_ON_HOST((
            amrex::ignore_unused(offset);
            return bx;
        ))
    }
 
// ************************************************
 
#ifdef AMREX_USE_GPU
    struct ExecutionConfig {
        ExecutionConfig () noexcept {
            Gpu::Device::grid_stride_threads_and_blocks(numBlocks,numThreads);
        }
        ExecutionConfig (const Box& box) noexcept {
            // If we change this, we must make sure it doesn't break FabArrayUtility Reduce*,
            // which assumes the decomposition is 1D.
            Gpu::Device::n_threads_and_blocks( ((box.numPts()+AMREX_GPU_NCELLS_PER_THREAD-1)/AMREX_GPU_NCELLS_PER_THREAD), numBlocks, numThreads );
#if 0
            Box b = amrex::surroundingNodes(box);
            b -= box.smallEnd();
            b.coarsen(IntVect(AMREX_D_DECL(1,AMREX_GPU_Y_STRIDE,AMREX_GPU_Z_STRIDE)));
            Gpu::Device::c_threads_and_blocks(b.loVect(), b.hiVect(), numBlocks, numThreads);
#endif
        }
        ExecutionConfig (const Box& box, int comps) noexcept {
            const Box& b = amrex::surroundingNodes(box);
            Gpu::Device::c_comps_threads_and_blocks(b.loVect(), b.hiVect(), comps, numBlocks, numThreads);
        }
        ExecutionConfig (Long N) noexcept {
            Gpu::Device::n_threads_and_blocks(N, numBlocks, numThreads);
        }
        ExecutionConfig (dim3 nb, dim3 nt, std::size_t sm=0) noexcept
            : numBlocks(nb), numThreads(nt), sharedMem(sm) {}
 
        dim3 numBlocks;
        dim3 numThreads;
        std::size_t sharedMem = 0;
    };
 
    template <int MT>
    ExecutionConfig
    makeExecutionConfig (Long N) noexcept
    {
        ExecutionConfig ec(dim3{}, dim3{});
        Long numBlocks = (std::max(N,Long(1)) + MT - 1) / MT;
        // ensure that blockDim.x*gridDim.x does not overflow
        numBlocks = std::min(numBlocks, Long(std::numeric_limits<unsigned int>::max()/MT));
        // ensure that the maximum grid size of 2^31-1 won't be exceeded
        numBlocks = std::min(numBlocks, Long(std::numeric_limits<int>::max()));
        ec.numBlocks.x = numBlocks;
        ec.numThreads.x = MT;
        AMREX_ASSERT(MT % Gpu::Device::warp_size == 0);
        return ec;
    }
 
    template <int MT>
    ExecutionConfig
    makeExecutionConfig (const Box& box) noexcept
    {
        return makeExecutionConfig<MT>(box.numPts());
    }
 
    struct ExecConfig
    {
        Long start_idx;
        int nblocks;
    };
 
    template <int MT>
    Vector<ExecConfig> makeNExecutionConfigs (Long N) noexcept
    {
        // Max # of blocks in a kernel launch
        int numblocks_max = std::numeric_limits<int>::max();
        // Max # of threads in a kernel launch
        Long nmax = Long(MT) * numblocks_max;
        // # of launches needed for N elements without using grid-stride
        // loops inside GPU kernels.
        auto nlaunches = int((N+nmax-1)/nmax);
        Vector<ExecConfig> r(nlaunches);
        Long ndone = 0;
        for (int i = 0; i < nlaunches; ++i) {
            int nblocks;
            if (N > nmax) {
                nblocks = numblocks_max;
                N -= nmax;
            } else {
                nblocks = int((N+MT-1)/MT);
            }
            // At which element ID the kernel should start
            r[i].start_idx = ndone;
            ndone += Long(nblocks) * MT;
            // # of blocks in this launch
            r[i].nblocks = nblocks;
        }
        return r;
    }
 
    template <int MT, int dim>
    Vector<ExecConfig> makeNExecutionConfigs (BoxND<dim> const& box) noexcept
    {
        return makeNExecutionConfigs<MT>(box.numPts());
    }
#endif
 
}
}
 
#ifdef AMREX_USE_SYCL
namespace amrex::detail {
 
    template <typename... L> constexpr bool is_big_kernel () {
        return (sizeof(L) + ... + 0) > 1792;
    }
 
    template <typename... L>
    struct SyclKernelDevPtr
    {
        using Ls = GpuTuple<L...>;
        Ls* m_dp = nullptr;
        Ls* m_hp = nullptr;
        gpuStream_t m_stream;
 
        SyclKernelDevPtr (L const&... f, gpuStream_t const& stream)
            : m_stream(stream)
        {
            if constexpr (is_big_kernel<L...>()) {
                std::size_t sz = sizeof(Ls);
                m_dp = (Ls*)The_Arena()->alloc(sz);
                m_hp = (Ls*)The_Pinned_Arena()->alloc(sz);
                new (m_hp) Ls(f...);
                auto* l_hp = (void const*)m_hp;
                auto* l_dp = (void*)m_dp;
                stream.queue->submit([&] (sycl::handler& h) {
                    h.memcpy(l_dp, l_hp, sz);
                });
            } else {
                amrex::ignore_unused(f...);
            }
        }
 
        ~SyclKernelDevPtr ()
        {
            if constexpr (is_big_kernel<Ls>()) {
                try {
                    m_stream.queue->wait_and_throw();
                } catch (sycl::exception const& ex) {
                    if (m_dp) {
                        The_Arena()->free((void*)m_dp);
                        m_dp = nullptr;
                    }
                    if (m_hp) {
                        m_hp->~Ls();
                        The_Pinned_Arena()->free((void*)m_hp);
                        m_hp = nullptr;
                    }
                    amrex::Abort(std::string("~SyclKernelDevPtr: ")+ex.what());
                }
                if (m_dp) {
                    The_Arena()->free((void*)m_dp);
                    m_dp = nullptr;
                }
                if (m_hp) {
                    m_hp->~Ls();
                    The_Pinned_Arena()->free((void*)m_hp);
                    m_hp = nullptr;
                }
            }
        }
 
        template <int N>
        auto get () const
        {
            using pointer_t = std::add_pointer_t<
                std::add_const_t<typename std::tuple_element<N, Ls>::type>>;
            if (m_hp && m_dp) {
                std::ptrdiff_t offset = (char*)(&(amrex::get<N>(*m_hp))) - (char*)m_hp;
                return pointer_t((char const*)m_dp + offset);
            } else {
                return pointer_t(nullptr);
            }
        }
    };
 
}
#endif
 
 
#ifdef AMREX_USE_GPU
#include <AMReX_GpuLaunchMacrosG.H>
#include <AMReX_GpuLaunchFunctsG.H>
#else
#include <AMReX_GpuLaunchMacrosC.H>
#include <AMReX_GpuLaunchFunctsC.H>
#include <AMReX_SIMD.H>
#endif
#include <AMReX_GpuLaunchFunctsSIMD.H>
 
#include <AMReX_GpuLaunch.nolint.H>
 
#include <AMReX_CTOParallelForImpl.H>
 
namespace amrex {
 
#if defined(AMREX_USE_GPU) || !defined(AMREX_USE_OMP)
 
template <std::integral T, typename L >
void ParallelForOMP (T n, L const& f) noexcept
{
    ParallelFor(n, f);
}
 
template <typename L>
void ParallelForOMP (Box const& box, L const& f) noexcept
{
    ParallelFor(box, f);
}
 
template <std::integral T, typename L >
void ParallelForOMP (Box const& box, T ncomp, L const& f) noexcept
{
    ParallelFor(box, ncomp, f);
}
 
#else /* !defined(AMREX_USE_GPU) && defined(AMREX_USE_OMP) */
 
template <std::integral T, typename L >
AMREX_ATTRIBUTE_FLATTEN_FOR
void ParallelForOMP (T n, L const& f) noexcept
{
#pragma omp parallel for
    for (T i = 0; i < n; ++i) {
        f(i);
    }
}
 
template <typename L>
AMREX_ATTRIBUTE_FLATTEN_FOR
void ParallelForOMP (Box const& box, L const& f) noexcept
{
    auto lo = amrex::lbound(box);
    auto hi = amrex::ubound(box);
#if (AMREX_SPACEDIM == 1)
#pragma omp parallel for
    for (int i = lo.x; i <= hi.x; ++i) {
        f(i,0,0);
    }
#elif (AMREX_SPACEDIM == 2)
#pragma omp parallel for
    for     (int j = lo.y; j <= hi.y; ++j) {
        AMREX_PRAGMA_SIMD
        for (int i = lo.x; i <= hi.x; ++i) {
            f(i,j,0);
        }
    }
#else
#pragma omp parallel for collapse(2)
    for         (int k = lo.z; k <= hi.z; ++k) {
        for     (int j = lo.y; j <= hi.y; ++j) {
            AMREX_PRAGMA_SIMD
            for (int i = lo.x; i <= hi.x; ++i) {
                f(i,j,k);
            }
        }
    }
#endif
}
 
template <std::integral T, typename L >
AMREX_ATTRIBUTE_FLATTEN_FOR
void ParallelForOMP (Box const& box, T ncomp, L const& f) noexcept
{
    auto lo = amrex::lbound(box);
    auto hi = amrex::ubound(box);
#if (AMREX_SPACEDIM == 1)
#pragma omp parallel for collapse(2)
    for (T n = 0; n < ncomp; ++n) {
        for (int i = lo.x; i <= hi.x; ++i) {
            f(i,0,0,n);
        }
    }
#elif (AMREX_SPACEDIM == 2)
#pragma omp parallel for collapse(2)
    for (T n = 0; n < ncomp; ++n) {
        for     (int j = lo.y; j <= hi.y; ++j) {
            AMREX_PRAGMA_SIMD
            for (int i = lo.x; i <= hi.x; ++i) {
                f(i,j,0,n);
            }
        }
    }
#else
#pragma omp parallel for collapse(3)
    for (T n = 0; n < ncomp; ++n) {
        for         (int k = lo.z; k <= hi.z; ++k) {
            for     (int j = lo.y; j <= hi.y; ++j) {
                AMREX_PRAGMA_SIMD
                for (int i = lo.x; i <= hi.x; ++i) {
                    f(i,j,k,n);
                }
            }
        }
    }
#endif
}
 
#endif
 
}
 
#endif