simd_conv_121()

static void fl::gfx::blur_detail::simd_conv_121 ( const u8 *FL_RESTRICT_PARAM a, const u8 *FL_RESTRICT_PARAM b, const u8 *FL_RESTRICT_PARAM c, u8 *FL_RESTRICT_PARAM out, int nbytes )

static

Definition at line 605 of file blur.cpp.hpp.

                                                                   {
    namespace fsimd = fl::simd; // ok bare using
    int i = 0;
    for (; i + 63 < nbytes; i += 64) {
        auto va0 = fsimd::load_u8_32(a+i), vb0 = fsimd::load_u8_32(b+i), vc0 = fsimd::load_u8_32(c+i);
        auto va1 = fsimd::load_u8_32(a+i+32), vb1 = fsimd::load_u8_32(b+i+32), vc1 = fsimd::load_u8_32(c+i+32);
        fsimd::store_u8_32(out+i, fsimd::avg_round_u8_32(fsimd::avg_round_u8_32(va0, vc0), vb0));
        fsimd::store_u8_32(out+i+32, fsimd::avg_round_u8_32(fsimd::avg_round_u8_32(va1, vc1), vb1));
    }
    for (; i + 31 < nbytes; i += 32) {
        auto va = fsimd::load_u8_32(a+i), vb = fsimd::load_u8_32(b+i), vc = fsimd::load_u8_32(c+i);
        fsimd::store_u8_32(out+i, fsimd::avg_round_u8_32(fsimd::avg_round_u8_32(va, vc), vb));
    }
    for (; i + 15 < nbytes; i += 16) {
        auto va = fsimd::load_u8_16(a+i), vb = fsimd::load_u8_16(b+i), vc = fsimd::load_u8_16(c+i);
        fsimd::store_u8_16(out+i, fsimd::avg_round_u8_16(fsimd::avg_round_u8_16(va, vc), vb));
    }
    for (; i < nbytes; ++i)
        out[i] = (u8)(((u16)a[i] + ((u16)b[i] << 1) + (u16)c[i]) >> 2);
}

References FL_RESTRICT_PARAM.

Referenced by fl::gfx::blur_detail::simd_hconv_dispatch< 1 >::apply(), and fl::gfx::blur_detail::simd_vconv_dispatch< 1 >::apply().

Here is the caller graph for this function: