Function morph3d

src/backend/cuda/kernel/morph.hpp:63–97

template<typename T>
void morph3d(Param<T> out, CParam<T> in, CParam<T> mask, bool isDilation) {
    const int windLen = mask.dims[0];

    if (windLen > 7) {
        CUDA_NOT_SUPPORTED("Morph 3D does not support kernels larger than 7.");
    }

    auto morph3D = common::getKernel(
        "arrayfire::cuda::morph3D", {{morph_cuh_src}},
        TemplateArgs(TemplateTypename<T>(), TemplateArg(isDilation),
                     TemplateArg(windLen)),
        {{DefineValue(MAX_MORPH_FILTER_LEN)}});

    morph3D.copyToReadOnly(
        morph3D.getDevPtr("cFilter"), reinterpret_cast<CUdeviceptr>(mask.ptr),
        mask.dims[0] * mask.dims[1] * mask.dims[2] * sizeof(T));

    dim3 threads(kernel::CUBE_X, kernel::CUBE_Y, kernel::CUBE_Z);

    int blk_x = divup(in.dims[0], CUBE_X);
    int blk_y = divup(in.dims[1], CUBE_Y);
    int blk_z = divup(in.dims[2], CUBE_Z);
    dim3 blocks(blk_x * in.dims[3], blk_y, blk_z);

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

    EnqueueArgs qArgs(blocks, threads, getActiveStream(), shrdSize);
    morph3D(qArgs, out, in, blk_x);
    POST_LAUNCH_CHECK();
}

}  // namespace kernel
}  // namespace cuda