Method operator()

src/backend/oneapi/kernel/bilateral.hpp:55–136

        , nBBS0_(nBBS0)
        , nBBS1_(nBBS1) {}
    void operator()(sycl::nd_item<2> it) const {
        sycl::group g              = it.get_group();
        const int radius           = sycl::max((int)(sigma_space_ * 1.5f), 1);
        const int padding          = 2 * radius;
        const int window_size      = padding + 1;
        const int shrdLen          = g.get_local_range(0) + padding;
        const float variance_range = sigma_color_ * sigma_color_;
        const float variance_space = sigma_space_ * sigma_space_;
        const float variance_space_neg2     = -2.0 * variance_space;
        const float inv_variance_range_neg2 = -0.5 / (variance_range);

        // gfor batch offsets
        unsigned b2 = g.get_group_id(0) / nBBS0_;
        unsigned b3 = g.get_group_id(1) / nBBS1_;

        const inType* in =
            d_src_.get_pointer() +
            (b2 * iInfo_.strides[2] + b3 * iInfo_.strides[3] + iInfo_.offset);
        outType* out = d_dst_.get_pointer() +
                       (b2 * oInfo_.strides[2] + b3 * oInfo_.strides[3]);

        int lx = it.get_local_id(0);
        int ly = it.get_local_id(1);

        const int gx =
            g.get_local_range(0) * (g.get_group_id(0) - b2 * nBBS0_) + lx;
        const int gy =
            g.get_local_range(1) * (g.get_group_id(1) - b3 * nBBS1_) + ly;

        // generate gauss2d_ spatial variance values for block
        if (lx < window_size && ly < window_size) {
            int x = lx - radius;
            int y = ly - radius;
            gauss2d_[ly * window_size + lx] =
                exp_native_nonnative<outType, USE_NATIVE_EXP>(
                    ((x * x) + (y * y)) / variance_space_neg2);
        }

        int s0 = iInfo_.strides[0];
        int s1 = iInfo_.strides[1];
        int d0 = iInfo_.dims[0];
        int d1 = iInfo_.dims[1];
        // pull image to local memory
        for (int b = ly, gy2 = gy; b < shrdLen;
             b += g.get_local_range(1), gy2 += g.get_local_range(1)) {
            // move row_set g.get_local_range(1) along coloumns
            for (int a = lx, gx2 = gx; a < shrdLen;
                 a += g.get_local_range(0), gx2 += g.get_local_range(0)) {
                load2LocalMem(localMem_, in, a, b, shrdLen, d0, d1,
                              gx2 - radius, gy2 - radius, s1, s0);
            }
        }

        it.barrier();

        if (gx < iInfo_.dims[0] && gy < iInfo_.dims[1]) {
            lx += radius;
            ly += radius;