Use __m512i or __m256i instead of __m128i to increase the SIMD vector length

[eriksjolund]

The AVX2 call
_mm_blend_epi32
was introduced in the git commit
4f2b2c8

When starting to use AVX2 we could also double the SIMD vector length by using
__m256i
instead of the current
__m128i

Maybe we could even make the SIMD vector length four times bigger by starting to use
__m512i

Here is a list of the current intrinsic functions:

$ cd diagonalsw
$ git branch | grep \*
* changeblend
$ git rev-parse HEAD
4f2b2c8d66d595c74806e43e306541f53742ecdf
$ grep -oh "_mm_[^(]*" diagonalsw/src/c/smith_waterman_sse_byte.1load.xml | sort | uniq
_mm_adds_epu8
_mm_alignr_epi8
_mm_blend_epi16
_mm_blend_epi32
_mm_blendv_epi8
_mm_cmpeq_epi8 
_mm_extract_epi8
_mm_load_si128
_mm_max_epu8
_mm_set1_epi8
_mm_set_epi8
_mm_setzero_si128
_mm_slli_si128
_mm_store_si128
_mm_subs_epu8
_mm_testc_si128

Do they all have 512-bit and 256-bit counterparts in AVX2?

[martin-steinegger]

I wrote a small preprocessor abstraction for simd instructions (not all instructions are covered). Maybe it could be useful for this purpose. See https://github.com/soedinglab/mmseqs2/blob/master/lib/simd/simd.h

[eriksjolund]

Thanks for the tip, but it seems the software license is not compatible with the MIT license that diagonalsw is licensed under.