AVX-VNNI-INT8 & AVX-IFMA Land In GCC 13
AVX-VNNI-INT8 are akin to the existing AVX-VNNI found with Alder lake and Sapphire Rapids but for an INT8 type. In turn AVX-VNNI is the VEX variant of AVX512-VNNI. The Vector Neural Network Instructions (VNNI) with native INT8 should further help with the deep learning performance for Intel's new processors in 2024.
This patch now in GCC 13 provides all the plumbing for AVX-VNNI-INT8 and the new "-mavxvnniint8". The AVX-VNNI-INT8 support will be flipped on automatically once the -march=sierraforest target is added.
A follow-up patch merged at the same time will auto-vectorize sdot product and udot product with VNNI INT8 instructions.
Another patch merged overnight adds the AVX-IFMA infrastructure for the GNU Compiler Collection. AVX-IFMA is for non-AVX-512 Integer Fused Multiply Add.
It's good as always seeing Intel engineers being punctual with their compiler enablement support for new instructions and processor support. GCC 13 with these additions and much more will debut as stable around April.