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:
> 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.

chacha_kunit benchmark results from AMD Genoa (Zen 4 server processor):

(I added some extra lengths as well:)

     | Len   | Before    | After     | Improvement |
     +-------+-----------+-----------+-------------+
     | 1     | 6 MB/s    | 6 MB/s    |   0.00%     |
     | 16    | 106 MB/s  | 106 MB/s  |   0.00%     |
     | 64    | 437 MB/s  | 437 MB/s  |   0.00%     |
     | 127   | 1078 MB/s | 1072 MB/s |  -0.56%     |
     | 128   | 1106 MB/s | 1101 MB/s |  -0.45%     |
     | 200   | 1361 MB/s | 1316 MB/s |  -3.31%     |
     | 256   | 1674 MB/s | 1612 MB/s |  -3.70%     |
     | 511   | 2950 MB/s | 2970 MB/s |  +0.68%     |
     | 512   | 3052 MB/s | 2976 MB/s |  -2.49%     |
     | 768   | 2674 MB/s | 3417 MB/s | +27.79%     |
     | 1024  | 3241 MB/s | 4514 MB/s | +39.28%     |
     | 1420  | 3242 MB/s | 3845 MB/s | +18.60%     |
     | 2048  | 3587 MB/s | 4829 MB/s | +34.63%     |
     | 3173  | 3088 MB/s | 4439 MB/s | +43.75%     |
     | 4096  | 3743 MB/s | 5040 MB/s | +34.65%     |
     | 8920  | 3606 MB/s | 4846 MB/s | +34.39%     |
     | 16384 | 3739 MB/s | 5105 MB/s | +36.53%     |

Note that mobile Zen 5 processors have only 256-bit ALUs as well.  There
the results are similar.

It's still worth enabling AVX-512 code like this on these processors.
It's just not quite as helpful as it is on processors with 512-bit ALUs.

- Eric