github aous72/OpenJPH 0.32.0

5 hours ago

Many updates/PRs have been merged into this release.
This is slightly rushed to go along with the next release of OpenEXR; in particular, I am adding an experimental non-linearity, NLT type 4 #356, which is not part of the standard -- if you do not need it, please do not use it.
There are a couple of PRs I wish to have added to this release; hopefully I will add them to the next release.
It worth saying that for #356, I wrote the whole generic code myself, and then I used AI as explained below.

  • I used on-premise Qwen3.8-Flash-Next (Unsloth Q4_K_XL) to generate SSE2 and AVX2 code paths from generic C code; the only difficulty here is the gather command for AVX2, which I never used before. It is worth saying that this models runs at 10-12 t/s decode at the start, dropping to 4-5 t/s at 50-75% of the context; I am using a context of 262144.
  • I used Cloud-based Deepseek-Flash APIs to generate and debug test code. However, the first merged PR failed on the s390x big-endian machine. To fix this problem, AI was a great help; it discovered the problem in like 30 minutes and fixed it.

What's Changed

New Contributors

Full Changelog: 0.31.0...0.32.0

Don't miss a new OpenJPH release

NewReleases is sending notifications on new releases.