Re: [PATCH v2 0/3] lib/crypto: x86/chacha: Add a 16-block AVX-512 function

[Eric Biggers <ebiggers@kernel.org>]

On Mon, Nov 02, 2026 at 12:39:23PM +0100, Martin Willi wrote:
> Here is v2 of the 16-block ChaCha function for 512-bit zmm registers.
> Patch 1 renames the assembly file, patch 2 wraps the 8-block code in
> a macro, and patch 3 adds the 16-block function on top.
> 
> I previously mentioned a larger series with more AVX refactorings to
> shared macros along with a 4-block zmm variant. That needs some
> more work, so here is just the 16-block variant.
> 
> v1 used embedded broadcasts for the final ChaCha state add. This comes
> with a minor regression on older Intel chips when used in the 8-block
> variant. Therefore I kept the state words in registers in the macro
> version used by both 8/16-block variants, see patch 3.
> 
> I don't have a Zen 4, which splits zmm instructions into two 256-bit
> halves, so I can't tell how the 16-block function performs there.
> If anybody has access to one, numbers from there would be great.
> 
> Changes in v2:
>  - Rename chacha-avx512vl-x86_64.S to chacha-avx512-x86_64.S
>  - Generate the 8- and 16-block functions from a shared macro
>  - Drop the cpu_has_xfeatures() and X86_FEATURE_AVX512F checks
>  - Keep the state words in registers for the final add
>  - Benchmark with the ChaCha KUnit suite instead of tcrypt
> 
> v1: https://lore.kernel.org/linux-crypto/20260722153247.630519-1-martin@strongswan.org/
> 
> Martin Willi (3):
>   lib/crypto: x86/chacha: Rename the AVX-512VL assembly file
>   lib/crypto: x86/chacha: Wrap the 8-block AVX-512 code in a macro
>   lib/crypto: x86/chacha: Add a 16-block AVX-512 function

Applied to https://git.kernel.org/pub/scm/linux/kernel/git/ebiggers/linux.git/log/?h=libcrypto-next

I updated the last commit message to add my results for Zen 4:

    Results on Zen 4 (specifically Genoa), which has a 256-bit datapath:
    
        Length      Before       After  Change
        ======  ==========  ==========  ======
           768   2674 MB/s   3417 MB/s  +27.8%
          1024   3241 MB/s   4514 MB/s  +39.3%
          1420   3242 MB/s   3845 MB/s  +18.6%
          2048   3587 MB/s   4829 MB/s  +34.6%
          4096   3743 MB/s   5040 MB/s  +34.7%
          8920   3606 MB/s   4846 MB/s  +34.4%
         16384   3739 MB/s   5105 MB/s  +36.5%

- Eric