Function copyMatrixIfGpuCompiled

quest/src/gpu/gpu_config.cpp:517–555

void copyMatrixIfGpuCompiled(qcomp** cpuMatr, qcomp* gpuArr, qindex matrDim, enum CopyDirection direction) {
#if COMPILE_CUDA

    // NOTE:
    // this function copies a 2D CPU matrix into a 1D row-major GPU array,
    // although this is not actually needed by the QuEST backend which
    // maintains 1D row-major CPU memories merely aliased by 2D structures
    // for the user's benefit. As such, this is dead code, but preserved in
    // case it is ever needed (like if custom user 2D data was needed in GPU).
    error_gpuDeadCopyMatrixFunctionCalled();

    // for completeness, we permit copying from the 1D GPU memory to the 2D CPU memory,
    // although we never actually have the need to do this!
    auto flag = (direction == TO_HOST)? 
        cudaMemcpyDeviceToHost:
        cudaMemcpyHostToDevice;

    // copy each CPU row into flattened GPU memory. we make each memcpy asynch,
    // but it's unclear it helps, nor whether single-stream sync is necessary
    size_t numBytesPerRow = matrDim * sizeof(**cpuMatr);
    gpu_sync();

    for (qindex r=0; r<matrDim; r++) {
        qcomp* cpuRow = cpuMatr[r];
        qcomp* gpuSlice = &gpuArr[r*matrDim];

        auto src = (direction == TO_HOST)? gpuSlice : cpuRow;
        auto dst = (direction == TO_HOST)? cpuRow   : gpuSlice;

        CUDA_CHECK( cudaMemcpyAsync(dst, src, numBytesPerRow, flag) );
    }

    // wait for async copies to complete
    gpu_sync();

#else
    error_gpuCopyButGpuNotCompiled();
#endif
}


template <typename T>