[alfalfa-star]

What

The buffer pool's device tier (the detiler's scratch buffers and the staging shadows of written guest buffers) kept at most 512 MiB by default. When a title's working set of these buffers is larger, the driver allocates and frees device memory continually. PPSA14632 had 70761 misses and 69828 evictions per 10 s with 485 of 512 MiB retained, and each one is a vkAllocateMemory or vkFreeMemory.

• The default budget is now an eighth of the largest device-local heap, and at least the old 512 MiB: up to 2 GiB on a 16 GiB card, 1 GiB on an 8 GiB card, 512 MiB below that.
• APS5_STAGING_POOL_MIB still sets it explicitly, and 0 still disables the tier.
• BufferPool::DeviceBudget takes the memory properties and is public, so the rule can be tested.

Tested

Linux 7.2.9 (Arch), Mesa 26.2.4 RADV, RX 9070 XT (16 GiB).

• PPSA14632 (LEGO Horizon Adventures) with a 2 GiB budget: 1312 misses and 99 evictions per 10 s (1628 MiB retained), and presents rose from about 34 to about 39 per 10 s (with feat(agc): import shared direct memory through udmabuf on Linux #943 and the draw snapshot budget PR in place).
• agc_driver_graphics: new stagingPoolBudgetTests (an eighth of a 16 GiB device-local heap is 2 GiB, a 2 GiB heap gives the 512 MiB floor, a host-only heap does not count).
• ctest: 463 of 463 pass on current main.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

🤖 Generated with Claude Code

[alfalfa-star]

Rebased onto current main (be04d4a). The new conventions check flagged the edited BufferPool.hpp comment about the APS5_STAGING_POOL_MIB default. I deleted that comment instead of updating it; the default is described above. ctest: 463/463.

[alfalfa-star]

The linux failure on be04d4a is a timeout in agc_driver_flat_store. The previous head passed linux, and the only change since then removes a comment, so it looks flaky. Could a maintainer re-run the failed job? I don't have the rights to.

[alfalfa-star]

Follow-up on the agc_driver_flat_store timeout, with numbers instead of "looks flaky". On linux every PR runs it three times (--repeat until-fail:3), and the first run is always much slower than the other two, presumably lavapipe compiling its shaders with a cold cache:

run	ctest flags	1st	2nd	3rd
#1147 2625e61	serial	78.6 s	7.7 s	7.9 s
#1152 be2b6d3	serial	62.2 s	8.4 s	8.5 s
#1077 f402eab	--parallel 4	85.0 s	5.2 s	5.0 s
#1147 be04d4a	--parallel 4	timeout (120 s)

main's workflow added --parallel 4 between the two #1147 runs, and this test already used half to two thirds of its 120 s timeout when run serially. So whether it fails depends on what else runs beside it, not on this PR. A re-run should pass most of the time. The root cause, the size of this test's shaders, is already being addressed in #1265, which reports the cold run going from 84.4 s to 27.0 s (measured on Intel ANV, not lavapipe). #1280 and #1282 change the same code. Edit: #1715 (merged) already runs this test with RUN_SERIAL, so a rebase onto current main picks up that fix.

[alfalfa-star]

Rebased onto main 6b56175 (6d1cf27), which includes #1715. agc_driver_flat_store now passes when run alone, but agc_driver_buffer_format timed out instead. Since #639 that test takes 93-108 s on lavapipe for every PR, not only this one (details in #1762). Could a maintainer re-run the linux job? Locally, 479/479 pass.

[adriadam10]

This conflicts with main since commit 28345b0 (device size classes), and it isn't a mechanical rebase:

• BufferPool::DeviceTiered and Capacity(bytes, properties) are static and call DeviceBudget() to know whether the device tier is enabled. With DeviceBudget(memory) there is no Context there, so they can't compute the heap-derived value. Keep the environment check (APS5_STAGING_POOL_MIB=0 disables the tier) in a static helper that doesn't need the heap, and use the heap-derived value only for deviceTier.budget in the constructor, so Capacity and tierFor stay consistent.
• After the rebase, keep stagingPoolBudgetTests and make sure it still covers the 16 GiB / 2 GiB / host-only-heap cases.
• The linux failure on the current head is the agc_driver_buffer_format timeout on lavapipe you already described (#1762), not this change; it should clear on a re-run after the rebase.

[alfalfa-star]

Rebased onto current main (after #1004's size classes) in 3bb4494, as you suggested:

• DeviceTierEnabled() (static, read once) holds the environment check: the tier is in use unless APS5_STAGING_POOL_MIB is set to 0. DeviceTiered, and through it Capacity, Usage and tierFor, use only that, so the size classes and the tier choice can't disagree.
• DeviceBudget(memory) is used only in the constructor, for deviceTier.budget: APS5_STAGING_POOL_MIB when set, else max(512 MiB, largest device-local heap / 8).
• stagingPoolBudgetTests is unchanged and still covers 16 GiB → 2 GiB, 2 GiB → 512 MiB and a host-only heap → 512 MiB.

ctest: 638/638 on RADV (RX 9070 XT).

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#374	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp L8
#425	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp L8
#454	core/libs/prx/libSceAgcDriver/tests/Graphics.cpp L8

[adriadam10]

Checked the rework against the size classes: DeviceTiered/Capacity only depend on the static DeviceTierEnabled(), the heap-derived budget is used only in the constructor. Built and ran agc_driver_graphics on RADV; the new stagingPoolBudgetTests don't compile without the change. Merging.