Function convolve_3d

src/backend/cuda/kernel/convolve.hpp:202–235

template<typename T, typename aT>
void convolve_3d(conv_kparam_t& p, Param<T> out, CParam<T> sig, CParam<aT> filt,
                 const bool expand) {
    auto convolve3 = common::getKernel(
        "arrayfire::cuda::convolve3", {{convolve3_cuh_src}},
        TemplateArgs(TemplateTypename<T>(), TemplateTypename<aT>(),
                     TemplateArg(expand)),
        {{DefineValue(MAX_CONV1_FILTER_LEN), DefineValue(CONV_THREADS),
          DefineValue(CONV3_CUBE_X), DefineValue(CONV3_CUBE_Y),
          DefineValue(CONV3_CUBE_Z)}});

    prepareKernelArgs<T>(p, out.dims, filt.dims, 3);

    size_t filterSize = filt.dims[0] * filt.dims[1] * filt.dims[2] * sizeof(aT);

    for (int b3 = 0; b3 < filt.dims[3]; ++b3) {
        int f3Off = b3 * filt.strides[3];

        const aT* fptr = filt.ptr + f3Off;

        // FIXME: case where filter array is strided
        auto constMemPtr = convolve3.getDevPtr(conv_c_name);
        convolve3.copyToReadOnly(
            constMemPtr, reinterpret_cast<CUdeviceptr>(fptr), filterSize);

        p.o[2] = (p.outHasNoOffset ? 0 : b3);
        p.s[2] = (p.inHasNoOffset ? 0 : b3);

        EnqueueArgs qArgs(p.mBlocks, p.mThreads, getActiveStream(),
                          p.mSharedSize);
        convolve3(qArgs, out, sig, filt.dims[0], filt.dims[1], filt.dims[2],
                  p.mBlk_x, p.o[2], p.s[2]);
        POST_LAUNCH_CHECK();
    }
}

template<typename T, typename aT>
void convolve_nd(Param<T> out, CParam<T> signal, CParam<aT> filt,