model-writtenrun 7099 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
May 2024 - Feb 2025 ยท 16 entries
- Paul Amonson proposed AVX-512 CRC32C acceleration after measuring substantial gains on large inputs but a possible regression below 256 bytes, and asked about licensing and suitable benchmarks #1.
- Paul submitted the implementation and a SQL benchmark helper, explicitly excluding the helper from eventual merging, and proposed consolidating duplicated CPU-feature checks #4.
- Andres Freund questioned whether WAL should use CRC at all, suggested a faster hash, and separately reported an illegal instruction on unsupported hardware while emphasizing the frequent 20-byte CRC calculation under a critical lock #7 #8.
- Paul's claimed CPU-register correction was absent from the next attachment, but the following submission fixed the VPCLMULQDQ check and embedded the full license text; after discussion with Intel's lawyers, Bruce Momjian considered licensing no longer a blocker #9 #12 #16 #17.
- After a summer gap, Paul split the work into shared x86 hardware detection and the AVX-512 implementation, reporting comparable performance and successful execution on hardware lacking the new instructions #20.
- Nathan Bossart pressed for short-input protection and questioned the benchmark baseline; Paul first introduced a separate small-input macro, then moved selection into COMP_CRC32C and corrected the baseline to use runtime SSE4.2 dispatch, after which the reported small-input regression disappeared #21 #23 #28 #31.
- Paul announced his retirement and handed the proposal to Raghuveer Devulapalli, who took over technical development #32 #33.
- Following Nathan's recommendation, Raghuveer adopted function-target attributes, consolidated x86 CRC selection, addressed a Skylake-X regression, and then corrected Windows/MSVC build integration #34 #36 #37.
- Nathan requested that existing SSE4.2 attribute work move into a separate patch, while Andres identified uninitialized benchmark data, missing build integration, and unnecessarily complicated probes; Raghuveer responded with a consolidated three-patch series and those review fixes #38 #39 #40 #42.
- Once the separate SSE4.2 prerequisite landed, Raghuveer rebased the series, and Nathan identified remaining failures on 32-bit Linux, ARM macOS, and Windows #43 #44 #45.
- John Naylor challenged the 256-byte threshold and AVX-512 hardware requirement, proposing investigation of more portable parallel CRC techniques, while Raghuveer fixed the Linux and ARM build failures #47 #48 #50.
- John identified a 128-bit carryless-multiplication implementation with a 64-byte threshold; renewed discussion also considered replacing CRC with hashing, with Ants Aasma describing limitations of the page-checksum algorithm and John and Andres debating which error-detection properties matter for WAL #51 #52 #53 #54 #55 #56 #57.
- Matthew Sterrett supplied the missing PGDLLIMPORT declaration for Windows, and Raghuveer subsequently confirmed that the resulting series passed CI #58 #59.
- John reiterated that shorter-input performance and broader hardware support should guide the design and requested clarification about patents cited by an alternative CRC technique; Raghuveer proposed handling SSE improvements afterward, but John argued against making the more restrictive AVX-512 design a prerequisite #60 #61 #62.
- At this thread's last message, Raghuveer had opened the SSE4.2 improvement discussion thread #51011, while the patent question remained unanswered and the attached AVX-512 series had not been announced as committed #63.
- The continuation in thread #51011 supplied the later matching patchset #588007 for 3c6e8c1238, replacing the previous attribution to this thread's precursor #58; John committed generated MIT-licensed code with a 64-byte threshold, simplified CPU detection, and regression tests, omitting the benchmark helper and shared hardware-detection refactor.