On Fri, 31 Jan 2025 16:48:09 GMT, Jamil Nimeh <jni...@openjdk.org> wrote:
> This enhancement makes a change to the ChaCha20 block function intrinsic on > aarch64, moving away from the block parallel implementation and to the > quarter-round parallel implementation that was done on x86_64. Assembly > language profiling yielded an 11% improvement in throughput. When put > together as an intrinsic and hooked into the JCE ChaCha20 cipher, the gains > are more modest, somewhere in the 2-4% range depending on job size, but still > an improvement. This looks very nice, and I'm tempted to just approve it as it is. My only concern is that the algorithm changes aren't really explained, but I guess what you have done here is the _128-Bit Vectorization_ in `https://eprint.iacr.org/2013/759.pdf`. Is that right? ------------- PR Comment: https://git.openjdk.org/jdk/pull/23397#issuecomment-2630610061
