Function copyArray

src/backend/cuda/copy.cpp:38–54

template<typename T>
Array<T> copyArray(const Array<T> &src) {
    Array<T> out = createEmptyArray<T>(src.dims());
    if (src.elements() > 0) {
        if (src.isReady()) {
            if (src.isLinear()) {
                CUDA_CHECK(cudaMemcpyAsync(
                    out.get(), src.get(), src.elements() * sizeof(T),
                    cudaMemcpyDeviceToDevice, getActiveStream()));
            } else {
                kernel::memcopy<T>(out, src, src.ndims());
            }
        } else {
            evalNodes<T>(out, src.getNode().get());
        }
    }
    return out;
}

template<typename T>
void multiply_inplace(Array<T> &src, double norm) {