GCC Lands AVX-512 Totally-Masked Vectorization

Stemming from wanting on the generated x264 video encode binary and a few efficiency inefficiencies, SUSE engineers have labored out AVX-512 totally masked vectorization help for the GCC 14 growth code.

Again in January SUSe compiler engineer Jan Hubicka opened this bug across the x264 benchmark with the averaging loop not being properly optimized for AVX-512.

“x264 benchmark has a loop averaging two unsigned char arrays that’s executed with comparatively low journey counts that doesn’t play properly with our vectorized code. For AVX512 most time is spent in unvectorized variant for the reason that common variety of iterations is just too small to succeed in the vector code.

…

For sizes 12-16 128bit vectorization wins, 20-28 behaves funily. Nonetheless avx512 vectorization is a large loss for all sizes as much as 31 bytes. aocc appears to win for 16 bytes.

…

One subject is that we at most carry out one epilogue loop vectorization, so with AVX512 we vectorize the epilogue with AVX2 however its epilogue stays unvectorized. With AVX512 we would need to use a totally masked epilogue utilizing AVX512 as a substitute.

I began engaged on totally masked vectorization help for AVX512 however obtained distracted.”

Quick ahead almost six months, SUSE compiler engineer Richard Biener has landed an preliminary implementation of AVX-512 totally masked vectorization throughout the GNU Compiler Assortment codebase for serving to out the x264 check case and different less-than-full vector circumstances.

“This implements totally masked vectorization or a masked epilog for avx512 type masks which single themselves out by representing every lane with a single bit and by utilizing integer modes for the masks (each is very like gcn).

avx512 can also be particular in that it does not have any instruction to compute the masks from a scalar iv like sve has with while_ult. As a substitute the masks are produced by vector compares and the loop management retains the scalar iv (primarily to keep away from dependences on masks technology, an appropriate masks check instruction is obtainable).

like rvv code technology prefers a decrementing iv although ivopts messes issues up in some circumstances eradicating that iv to get rid of it with an incrementing one used for deal with technology.

one of many motivating testcases is from pr108410 which in flip is extracted from x264 the place massive dimension vectorization reveals points with small journey loops. Execution time there improves in comparison with traditional avx512 with avx2 epilogues for the circumstances of lower than 32 iterations.”

The AVX-512 totally masked vectorization help landed this morning in GCC 14 Git through this commit.