ISPC is faster

[cloudhan]

struct Pixel { unsigned int8 r, g, b; };

unsigned int8 rgb2gray_kernel(Pixel p) {
    return (p.r *76 + p.g *150 + p.b *30) >> 8;
}

export void rgb2gray(uniform unsigned int8* uniform src,
                     uniform unsigned int8* uniform dst,
                     uniform size_t width,
                     uniform size_t height,
                     uniform size_t stride) {
    for (uniform int h=0; h<height; ++h) {
        uniform unsigned int8* uniform scanline_start = src + h * stride;
        uniform Pixel* scanline = (uniform Pixel* uniform)scanline_start;

        uniform unsigned int8* uniform out_scanline = dst + h * width;

        foreach (w= 0...width) {
            unsigned int8 v = rgb2gray_kernel(scanline[w]);
            out_scanline[w] = v;
        }
    }
}

compiled with

ispc rgb2gray.ispc -h rgb2gray.h -o rgb2gray.o -O3 --pic --target=avx2-i32x16