AMReX_CudaGraph.H

Go to the documentation of this file.

#ifndef AMREX_CUDA_GRAPH_H_
#define AMREX_CUDA_GRAPH_H_
#include <AMReX_Config.H>
 
#if defined(__CUDACC__) && defined(AMREX_USE_CUDA)
 
#include <AMReX.H>
#include <AMReX_Array.H>
#include <AMReX_GpuDevice.H>
 
namespace amrex {
 
struct CopyMemory
{
    void* src = nullptr;
    void* dst = nullptr;
    Dim3 src_begin = {0,0,0};
    Dim3 src_end = {0,0,0};
    Dim3 dst_begin = {0,0,0};
    Dim3 dst_end = {0,0,0};
    int scomp = 0;
    int ncomp = 0;
 
    template <class T>
    AMREX_GPU_HOST_DEVICE
    Array4<T const> getSrc () { return Array4<T const>(static_cast<T const*>(src), src_begin, src_end, scomp+ncomp); }
 
    template <class T>
    AMREX_GPU_HOST_DEVICE
    Array4<T> getDst () { return Array4<T>(static_cast<T*>(dst), dst_begin, dst_end, scomp+ncomp); }
};
 
template <typename T, typename U>
CopyMemory
makeCopyMemory (Array4<T> const& src, Array4<U> const& dst, int scomp, int ncomp)
{
    return CopyMemory{ (void*)(src.p), (void*)(dst.p), IntVectShrink<3>(src.begin).dim3(), IntVectShrink<3>(src.end).dim3(),
        IntVectShrink<3>(dst.begin).dim3(), IntVectShrink<3>(dst.end).dim3(), scomp, ncomp };
}
 
// ======================================================================================
 
template <typename T>
struct CudaGraph
{
    cudaGraphExec_t m_graph;
    Vector<T> m_parms;
    T* m_parms_d = nullptr;
    bool graph_is_ready = false;
 
    CudaGraph()
        : m_parms(0)
    {
        static_assert(std::is_trivially_copyable<T>::value, "CudaGraph's T must be trivially copyable");
    }
    CudaGraph(int num)
        : m_parms(num)
    {
        static_assert(std::is_trivially_copyable<T>::value, "CudaGraph's T must be trivially copyable");
        m_parms_d = static_cast<T*>( The_Arena()->alloc(sizeof(T)*m_parms.size()) );
    }
    ~CudaGraph() {
        The_Arena()->free(m_parms_d);
 
        if (graph_is_ready)
        {
            AMREX_CUDA_SAFE_CALL(cudaGraphExecDestroy(m_graph));
        }
    }
 
    void resize(Long num) {
        m_parms.resize(num);
        if (m_parms_d != nullptr)
        {
            The_Arena()->free(m_parms_d);
        }
        m_parms_d = static_cast<T*>( The_Arena()->alloc(sizeof(T)*m_parms.size()) );
    }
    void setGraph(cudaGraphExec_t const& graph) {
        m_graph = graph;
        graph_is_ready = true;
    }
    void setParams(int idx, T const& a_parm) { m_parms[idx] = a_parm; }
 
    T* getHostPtr (int idx) { return (m_parms.data() + idx); }
    T* getDevicePtr (int idx) const { return (m_parms_d + idx); }
    bool ready() const { return graph_is_ready; }
 
    void executeGraph (bool synch = true) const {
        Gpu::Device::executeGraph(m_graph, synch);
    }
};
 
}
 
#endif
#endif