Function morph

src/backend/cuda/kernel/morph.hpp:30–60

template<typename T>
void morph(Param<T> out, CParam<T> in, CParam<T> mask, bool isDilation) {
    const int windLen  = mask.dims[0];
    const int SeLength = (windLen <= 10 ? windLen : 0);

    auto morph = common::getKernel(
        "arrayfire::cuda::morph", {{morph_cuh_src}},
        TemplateArgs(TemplateTypename<T>(), TemplateArg(isDilation),
                     TemplateArg(SeLength)),
        {{DefineValue(MAX_MORPH_FILTER_LEN)}});

    morph.copyToReadOnly(morph.getDevPtr("cFilter"),
                         reinterpret_cast<CUdeviceptr>(mask.ptr),
                         mask.dims[0] * mask.dims[1] * sizeof(T));

    dim3 threads(kernel::THREADS_X, kernel::THREADS_Y);

    int blk_x = divup(in.dims[0], THREADS_X);
    int blk_y = divup(in.dims[1], THREADS_Y);
    // launch batch * blk_x blocks along x dimension
    dim3 blocks(blk_x * in.dims[2], blk_y * in.dims[3]);

    // calculate shared memory size
    int padding = (windLen % 2 == 0 ? (windLen - 1) : (2 * (windLen / 2)));
    int shrdLen =
        kernel::THREADS_X + padding + 1;  // +1 for to avoid bank conflicts
    int shrdSize = shrdLen * (kernel::THREADS_Y + padding) * sizeof(T);

    EnqueueArgs qArgs(blocks, threads, getActiveStream(), shrdSize);
    morph(qArgs, out, in, blk_x, blk_y, windLen);
    POST_LAUNCH_CHECK();
}

template<typename T>
void morph3d(Param<T> out, CParam<T> in, CParam<T> mask, bool isDilation) {