That's pretty cool. The compare_and_swap ops of the sorting network can be done in SIMD using _mm_min_epi32, _mm_max_epi32, and _mm_blend_epi32. For the 8-val

That's pretty cool. The compare_and_swap ops of the sorting network can be done in SIMD using _mm_min_epi32, _mm_max_epi32, and _mm_blend_epi32. For the 8-value network, they can operate on 4 pairs of the values and 8 pairs for 16-value network, as long as the pairs have no dependency with each other.