AnyPS5

0 Fork 0

mirror of https://github.com/boykopovar/AnyPS5.git synced 2026-11-03 12:57:27 +00:00

Files

T

Celegans12 5804207c50 perf(agc): stage the written elements of address-based compute builds in device memory

An address-based compute build bound its written V# elements in place in
the host imports of the cached address space, and its BDA table sent the
same addresses there, so every access went over PCIe. A written element
inside one import region now becomes a region of its own, staged like any
written element (copy-in, device-local shadow, copy-back), and
AddressRanges cuts the import's table entry around it so pointers into
the element reach the shadow as its descriptor does. resolveOverlays
keeps a cluster only when it lies inside one direct import region, is
stagingEligible and is crossed by no in-place range or own region.
Dispatches under APS5_BDA_STAGING_MIN_THREADS (512) threads keep the
imports; APS5_NO_BDA_STAGING=1 turns it off. Test: agc_bda_staging.
The copy-back trade-off of staged elements is listed as technical debt.

AI-assisted: yes

2026-10-21 03:53:21 +02:00

debt

perf(agc): stage the written elements of address-based compute builds in device memory

2026-10-21 03:53:21 +02:00

ARCHITECTURE.md

fix(agc): serve sampled views of arrays past maxImageArrayLayers from their viewed slices

2026-10-19 21:37:26 +02:00

BUILD.md

Merge pull request #8075 from johandavidcm/feat/agc-sample-committer

2026-10-21 14:12:36 +13:00

CONVENTIONS.md

…