Pipeline cache persistence uses full-size temporary heap buffers #2623

Closed

opened 2026-11-12 02:53:19 +00:00 by Arkex-ai · 0 comments

Arkex-ai commented

2026-11-12 02:53:19 +00:00

(Migrated from github.com)

Pipeline-cache persistence still allocates temporary vectors as large as the cache: pcache_file_read, play_file_write and full base-cache saves on main 55b976c0. That private allocation overlaps the driver's retained cache. #2206 provides a useful opt-out, but keeping the warm application cache should not require another complete heap buffer.

Local reproduction:

Authored 128 MiB input, Windows: the extracted reader added 128.25 MiB of private commit; a read-only file view added 0.25 MiB, with equal complete-input checksum. Rejecting allocations above 32 MiB rejected only the vector reader.
RTX 5070 Ti, Windows, warm reverse-order preparation pair: both paths prepared 17,955 graphics pipelines and saved identical cache bytes. The file-backed candidate had a 607 MiB lower sampled process-private peak (100 ms sampling). This is one local preparation comparison, not a cold-driver or FPS claim. Both runs retain one known failed shader from the input program set.
Two candidate processes loaded the same Kokiri checkpoint and exited normally using the persisted cache. These runs use an ef81aee7 host with byte-identical renderer API18 and the candidate renderer based on 55b976c0; they are not a full current-host build.

The scoped implementation keeps the cache format, generation retention, thresholds and existing switches. Reads validate the header first; Linux uses an owned private file snapshot. Writes use an owned sibling temporary and atomic publication of the actual returned length. Incomplete driver copies and I/O failures retain the old cache; a failed read must not delete a valid generation.

Linux snapshots need temporary disk space and extra I/O. Mapped pages still use RAM, and mapped media faults remain native faults. Driver-owned cache memory is not removed. Linux file tests pass; actual Linux Vulkan, AMD and Intel runtime validation are still missing.

Coordination: investigation claim, #2456 scope question. Preserve #2456's Intel synchronization, #2430's hardened opens/limits and #2317's path migration. #882's sandbox will need an intentional temporary-file policy; no in-place overwrite fallback is proposed.

Priority: S3 (1) × provisional reach 0.01 × reproduced confidence 1.0 = 0.01. Reach is an unmeasured planning assumption for large-cache installations, not an installed-base statistic or a claim to fix reported OOM crashes.

Pipeline-cache persistence still allocates temporary vectors as large as the cache: `pcache_file_read`, `play_file_write` and full base-cache saves on main `55b976c0`. That private allocation overlaps the driver's retained cache. #2206 provides a useful opt-out, but keeping the warm application cache should not require another complete heap buffer. Local reproduction: - Authored 128 MiB input, Windows: the extracted reader added 128.25 MiB of private commit; a read-only file view added 0.25 MiB, with equal complete-input checksum. Rejecting allocations above 32 MiB rejected only the vector reader. - RTX 5070 Ti, Windows, warm reverse-order preparation pair: both paths prepared 17,955 graphics pipelines and saved identical cache bytes. The file-backed candidate had a 607 MiB lower sampled process-private peak (100 ms sampling). This is one local preparation comparison, not a cold-driver or FPS claim. Both runs retain one known failed shader from the input program set. - Two candidate processes loaded the same Kokiri checkpoint and exited normally using the persisted cache. These runs use an `ef81aee7` host with byte-identical renderer API18 and the candidate renderer based on `55b976c0`; they are not a full current-host build. The scoped implementation keeps the cache format, generation retention, thresholds and existing switches. Reads validate the header first; Linux uses an owned private file snapshot. Writes use an owned sibling temporary and atomic publication of the actual returned length. Incomplete driver copies and I/O failures retain the old cache; a failed read must not delete a valid generation. Linux snapshots need temporary disk space and extra I/O. Mapped pages still use RAM, and mapped media faults remain native faults. Driver-owned cache memory is not removed. Linux file tests pass; actual Linux Vulkan, AMD and Intel runtime validation are still missing. Coordination: [investigation claim](https://discord.com/channels/1568810484818378804/1570103281856806962/1570255465806565520), [#2456 scope question](https://git.magicbean.dev/MagicBeanDev/Hookshot/pulls/2456). Preserve #2456's Intel synchronization, #2430's hardened opens/limits and #2317's path migration. #882's sandbox will need an intentional temporary-file policy; no in-place overwrite fallback is proposed. Priority: S3 (1) × provisional reach 0.01 × reproduced confidence 1.0 = 0.01. Reach is an unmeasured planning assumption for large-cache installations, not an installed-base statistic or a claim to fix reported OOM crashes.