mirror of
https://github.com/boykopovar/AnyPS5.git
synced 2026-11-03 12:57:27 +00:00
ntamas94andClaude Opus 5.5
baa0ca3dfc
The first commit drained the device for every label recorded while any copied writer was listed, including labels nowhere near it. CPU-copied writers also occur in default runs, so that could cost a drain per label in normal play. While a copied writer is listed, a label of 8 bytes or less (every RELEASE_MEM, the fence WRITE_DATAs) is now not recorded on the GPU at all. A completion action of the open (or newest in-flight) batch stores it when that batch is reaped (AfterCompletions with storedOnGpu false, reason 5). Completions run in push order and batches finish front-first, so it lands after every earlier write-back without a drain. The existing GPU store plus completion re-store path would not keep this order: its GPU store is visible to a polling title before the write-back runs, and its re-store runs only over a write-back that overlapped the label itself. Recorder::PendingCompletionStores counts the guest stores completions still have to make (not end-of-pipe deliveries or write-backs). While it is not zero, later small labels follow the same route even after every writer was written back, so a later GPU label cannot become visible before an earlier completion store or land under it at one address. Stores over 8 bytes keep the GPU store and go behind the completions only over a copied writer's range or a pending completion store of their own range. Queue 0 reaps while it spins on REWIND, so a title polling such a label there cannot wait forever. The debt entry lists what this still leaves out of order (larger stores and GPU-made DMA_DATA copies away from those ranges, and a nested sync from inside a write-back). APS5_DRAIN_COMPLETION_LABELS=1 brings back the drain (reason 2); a second ctest entry runs the test with it. CopiedWriterLabelOrder now skips with a line naming minStorageBufferOffsetAlignment when the device binds the record at +0x1004 in place (alignment 4, AMD's driver and RADV) and nothing is written back. Above 4 the write-back count still fails the test. A new pass checks the order across batches: a later label to the address of a pending completion store is not visible before it and is not overwritten by it at the reap. Co-Authored-By: Claude Opus 5.5 <[email protected]>