ChaCha AVX-512 rework provides up to 99% throughput boost on Zen 5

Reworked ChaCha stream cipher code slated for Linux 7.5 can increase throughput by up to 99% on AMD Zen 5 processors, reports say. The current Linux implementation uses 256-bit ymm registers to process eight blocks per pass. Martin Willi, who wrote much of the original x86 ChaCha SIMD code, said this eight-block limit was a deliberate choice to avoid heavy core down-clocking on Intel Skylake-X processors.

Image: Hardware Busters

The new code uses 512-bit zmm registers to process 16 blocks per call. In a test using 1024-byte blocks on a Zen 5 machine, Willi measured 7.5 GB/s with the new code compared to 4.2 GB/s with the existing path. This optimization only activates for inputs longer than 512 bytes. It also requires support for AVX-512BW and AVX-512VL, and executes only if the kernel lacks the prefer-ymm processor hint.