AMReX_GpuControl.H

Go to the documentation of this file.

#ifndef AMREX_GPU_CONTROL_H_
#define AMREX_GPU_CONTROL_H_
#include <AMReX_Config.H>
 
#include <AMReX_GpuQualifiers.H>
#include <AMReX_GpuTypes.H>
 
#include <utility>
 
#if defined(AMREX_USE_CUDA) && (__CUDACC_VER_MAJOR__ > 11 || ((__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ >= 2)))
#define AMREX_CUDA_GE_11_2 1
#endif
 
#if defined(AMREX_USE_HIP) || defined(AMREX_CUDA_GE_11_2)
#define AMREX_GPU_STREAM_ALLOC_SUPPORT 1
#endif
 
#if defined(AMREX_USE_HIP)
#define AMREX_HIP_OR_CUDA(a,b) a
#elif defined(AMREX_USE_CUDA)
#define AMREX_HIP_OR_CUDA(a,b) b
#else
#define AMREX_HIP_OR_CUDA(a,b) ((void)0);
#endif
 
#if defined(AMREX_USE_HIP)
#define AMREX_HIP_OR_CUDA_OR_SYCL(a,b,c) a
#elif defined(AMREX_USE_CUDA)
#define AMREX_HIP_OR_CUDA_OR_SYCL(a,b,c) b
#elif defined(AMREX_USE_SYCL)
#define AMREX_HIP_OR_CUDA_OR_SYCL(a,b,c) c
#else
#define AMREX_HIP_OR_CUDA_OR_SYCL(a,b,c) ((void)0);
#endif
 
#ifdef AMREX_USE_GPU
#define AMREX_GPU_OR_CPU(a,b) a
#else
#define AMREX_GPU_OR_CPU(a,b) b
#endif
 
#ifdef AMREX_USE_SYCL
#define AMREX_SYCL_ONLY(a) a
#else
#define AMREX_SYCL_ONLY(a) ((void)0)
#endif
 
#ifdef AMREX_USE_SYCL
#if (AMREX_SPACEDIM == 1)
#  define AMREX_SYCL_1D_ONLY(a) a
#  define AMREX_SYCL_2D_ONLY(a) ((void)0)
#  define AMREX_SYCL_3D_ONLY(a) ((void)0)
#elif (AMREX_SPACEDIM == 2)
#  define AMREX_SYCL_1D_ONLY(a) ((void)0)
#  define AMREX_SYCL_2D_ONLY(a) a
#  define AMREX_SYCL_3D_ONLY(a) ((void)0)
#elif (AMREX_SPACEDIM == 3)
#  define AMREX_SYCL_1D_ONLY(a) ((void)0)
#  define AMREX_SYCL_2D_ONLY(a) ((void)0)
#  define AMREX_SYCL_3D_ONLY(a) a
#endif
#else
#  define AMREX_SYCL_1D_ONLY(a) ((void)0)
#  define AMREX_SYCL_2D_ONLY(a) ((void)0)
#  define AMREX_SYCL_3D_ONLY(a) ((void)0)
#endif
 
namespace amrex {
    enum struct RunOn { Gpu, Cpu, Device=Gpu, Host=Cpu };
}
 
namespace amrex { // NOLINT(modernize-concat-nested-namespaces)
 
#ifdef AMREX_USE_HIP
using gpuStream_t = hipStream_t;
#elif defined(AMREX_USE_CUDA)
using gpuStream_t = cudaStream_t;
#endif
 
namespace Gpu {
 
#if defined(AMREX_USE_GPU)
 
    extern bool in_launch_region;
 
    [[nodiscard]] inline bool inLaunchRegion () noexcept { return in_launch_region; }
    [[nodiscard]] inline bool notInLaunchRegion () noexcept { return !in_launch_region; }
 
    inline bool setLaunchRegion (bool launch) noexcept {
        bool r = in_launch_region;
        in_launch_region = launch;
        return r;
    }
 
    extern bool in_graph_region;
    [[nodiscard]] inline bool inGraphRegion() { return (in_graph_region && in_launch_region); }
    [[nodiscard]] inline bool notInGraphRegion() { return (!in_graph_region || !in_launch_region); }
 
    inline bool setGraphRegion (bool graph) {
        bool r = in_graph_region;
        in_graph_region = graph;
        return r;
    }
 
    struct [[nodiscard]] LaunchSafeGuard
    {
        explicit LaunchSafeGuard (bool flag) noexcept
            : m_old(setLaunchRegion(flag)) {}
        ~LaunchSafeGuard () { setLaunchRegion(m_old); }
    private:
        bool m_old;
    };
 
    struct [[nodiscard]] GraphSafeGuard
    {
        explicit GraphSafeGuard (bool flag) noexcept
            : m_old(setGraphRegion(flag)) {}
        ~GraphSafeGuard () { setGraphRegion(m_old); }
    private:
        bool m_old;
    };
 
    extern bool in_single_stream_region;
    extern bool in_nosync_region;
 
    [[nodiscard]] inline bool inSingleStreamRegion () noexcept { return in_single_stream_region; }
    [[nodiscard]] inline bool inNoSyncRegion () noexcept { return in_nosync_region; }
 
    inline bool setSingleStreamRegion (bool b) noexcept {
        return std::exchange(in_single_stream_region, b);
    }
 
    inline bool setNoSyncRegion (bool b) noexcept {
        return std::exchange(in_nosync_region, b);
    }
 
    struct [[nodiscard]] SingleStreamRegion
    {
        SingleStreamRegion () noexcept
            : m_prev_flag(std::exchange(in_single_stream_region,true))
        {}
 
        ~SingleStreamRegion () { in_single_stream_region = m_prev_flag; }
 
    private:
        bool m_prev_flag;
    };
 
    struct [[nodiscard]] NoSyncRegion
    {
        NoSyncRegion () noexcept
            : m_prev_flag(std::exchange(in_nosync_region,true))
        {}
 
        ~NoSyncRegion () { in_nosync_region = m_prev_flag; }
 
    private:
        bool m_prev_flag;
    };
 
#else
 
    [[nodiscard]] inline constexpr bool inLaunchRegion () { return false; }
    [[nodiscard]] inline constexpr bool notInLaunchRegion () { return true; }
    [[nodiscard]] inline constexpr bool setLaunchRegion (bool) { return false; }
 
    [[nodiscard]] inline constexpr bool inGraphRegion () { return false; }
    [[nodiscard]] inline constexpr bool notInGraphRegion () { return true; }
    [[nodiscard]] inline constexpr bool setGraphRegion (bool) { return false; }
 
    struct [[nodiscard]] LaunchSafeGuard
    {
        explicit LaunchSafeGuard (bool) {}
    };
 
    struct [[nodiscard]] GraphSafeGuard
    {
        explicit GraphSafeGuard (bool) {}
    };
 
    [[nodiscard]] inline constexpr bool inSingleStreamRegion () { return false; }
    [[nodiscard]] inline constexpr bool inNoSyncRegion () { return true; }
    [[nodiscard]] inline constexpr bool setSingleStreamRegion (bool) { return false; }
    [[nodiscard]] inline constexpr bool setNoSyncRegion (bool) { return true; }
    struct [[nodiscard]] SingleStreamRegion {};
    struct [[nodiscard]] NoSyncRegion {};
 
#endif
 
}
}
 
#endif