Commit aea8856277e7a77152d68bdbb54882569403341c
2023-05-19T16:52:43
Reland "Metal: Optimized BufferSubData per device" This reverts commit ee64836f702332adaca58d9f452063a04b2da955 , relanding the patch stack described there. Between patchsets 1 and 5: - The shadow buffer allocation has been replaced with a multimap of precisely-sized buffers, rather than rounding up buffer sizes. - Garbage collection of shadow buffers is triggered in three situations: - A certain number of context switches have occurred; this number was hand-tuned to avoid GC every frame. - A certain number of command buffer submissions has occurred; this number was hand-tuned to GC no more often than every few seconds on representative workloads. - The total size of the allocated shadow buffers is more than 1 MB, and either more than twice the size at the last garbage collection, or 64 MB more than at the last garbage collection. In this case, aggressive GC is performed in order to reclaim shadow buffers more quickly. Performance before and after these changes appears identical on microbenchmarks. On one Figma test case, comparing GPU memory allocated inside the BufferManager, peak consumption is decreased by over 75%, and steady-state consumption decreases by over 88%. Patchset 6 adds a needed workaround for a bug in the AMDMTLBronzeDriver affecting uploads of client-side data, and therefore some dEQP tests. It also streamlines the aggressive GC. Bug: angleproject:7544 Change-Id: I81b061f0b33c27fa403527fa12d626f4e9c88ebe Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/4497413 Reviewed-by: Shahbaz Youssefi <syoussefi@chromium.org> Reviewed-by: Geoff Lang <geofflang@chromium.org> Commit-Queue: Geoff Lang <geofflang@chromium.org> Commit-Queue: Shahbaz Youssefi <syoussefi@chromium.org>