On Tue, 29 Sep 2026 23:16:56 GMT, Mohamed Issa <[email protected]> wrote:
>> It turns out the 2-step FP16 to integral conversion process is usually more >> beneficial on platforms that support AVX10.2 instructions. This is mainly >> due to the reduced instruction count and automatic special case handling >> (e.g., NaN). With that in mind, the changes in this PR go back to the >> original approach when the AVX10.2 vectorized path is detected in the C2 >> compiler. There are also some updates to the JTREG tests and JMH benchmarks. >> >> The JTREG test listed below was used to verify correctness with >> `-XX:-UseSuperWord` and `-XX:+UseSuperWord` JVM options applied. All >> modifications and tests used [OpenJDK >> v28-b15](https://github.com/openjdk/jdk/releases/tag/jdk-28%2B15) as the >> baseline build. >> >> 1. >> `jtreg:test/hotspot/jtreg/compiler/vectorapi/TestFloat16ToIntegralConv.java` >> >> --------- >> - [x] I confirm that I make this contribution in accordance with the >> [OpenJDK Interim AI Policy](https://openjdk.org/legal/ai). > > Mohamed Issa has updated the pull request with a new target base due to a > merge or a rebase. The incremental webrev excludes the unrelated changes > brought in by the merge/rebase. The pull request contains five additional > commits since the last revision: > > - Merge branch 'master' into user/missa-prime/avx10_2 > - Use AVX512 direct conversion for scalar half-precision path and add new > classes to JMH source. > - Remove AVX10.2 fp16 to byte direct conversion vector instruction as it > causes correctness issues. > - Use AVX10.2 fp16 to byte direct conversion vector instruction whenever > possible. > - Disable AVX512 FP16 direct conversion on targets that support AVX10.2 @missa-prime Your change (at version 8bd8b7f9f5e8c9187138514910dacb6a25aec5c3) is now ready to be sponsored by a Committer. ------------- PR Comment: https://git.openjdk.org/jdk/pull/32957#issuecomment-5903502871
