understanding SIMD intrinsics

[CeleritasCelery]

I read the blog post about fast byte counting and I was reading through the code to better understand the SIMD intrinsic. I was mostly able to follow along but I got confused at this part

bytecount/src/simd/x86_avx2.rs

Lines 58 to 61 in 5085205

	counts = _mm256_sub_epi8(
	counts,
	_mm256_cmpeq_epi8(mm256_from_offset(haystack, offset), needles)
	);

I don't understand why we are subtracting the results of the compare instead of adding them. It seems like this will give incorrect results, but obviously doesn't, and I can't explain why. Could someone explain it to me?

[llogiq]

The reason for that is that the result of _mm256_cmpeq_epi8 is either all 1s or all 0s. And an i8 with all 1s is -1. So we need to subtract that to add 1 wherever the byte is equal.

[CeleritasCelery]

Thanks, that makes sense now.