fix(vk): past the allowance the textures' floors give way, so a 6 GB card's VRAM comes back under it #2082

Merged
HumanBean merged 2 commits from HumanBean/Longshot:fix/vram-floor-past-allowance into main 2026-11-21 17:33:21 +00:00
Owner

Summary

On cards the driver allows under 8 GB, the texture cache has a floor: the 90% rule, the compaction's hold and reserve,
and the 97% rule never cut its budget below 1800 MB (vram_compact::texture_min_floor), and the dynamic floor can be
as high as half the allowance (#637). That stops a nearly full card from evicting the textures in use one pass after
another, which is right while the card is within its allowance. Past the allowance it means no rule can take a texture
once the textures are under the floor, whatever the rest of the renderer holds. The driver then pages VRAM out to
system memory, and every frame reads part of it over PCIe. In #939 a GTX 1060 6 GB sat at 6128 MB of the 5231 MB it
allowed with 1606 MB of textures, already under today's 1800 MB floor, and Hyrule Field went from 31 to 14 fps until a
restart.

With this change the floors apply only while the driver's count is within the allowance
(vram_compact::floor_within, and count_line_floor for the 97% rule). Past it, the eviction's own rule still keeps
every texture used in the last 60 presents, and the textures that have gone unused longer go until the count is back
under the line. Within the allowance nothing changes.

How to see it on main: a 6 GB card (or RX_VK_TEST_VRAM_MB=6144 with an RX_VK_TEST_ALLOWED_MB low enough that the
mirrors, shadows, render targets and blocks' free space take more than the allowance less 1800 MB). Tour a few areas.
The log's vk: VRAM: line shows the count past the allowance and textures ... (budget 1800 MB) while the textures sit
at or under 1800 MB.

Why textures run over their budget in #939 and #1008: in those logs (nightly-189, rc.4, before #637 added today's
floor) the textures were over a budget the rest of VRAM had squeezed to 136-448 MB. Most of them had been used in the
last frames ("used in the last 8 presents 1668 MB" of 2358), so the eviction, which never takes a texture used in the
last 60 presents, could not meet it. The budget is enforced, but it can't go below the working set. On main today
the same squeeze stops at the floor instead. This PR lets it go on past the allowance, down to the working set.

Read first: floor_within and count_line_floor in gpu/vk/vram_compact.h, then the two call sites in
texture_frame (gpu/vk/vk_texture.cpp: active_floor, and the 97% rule's floor).

See #939, #1008, #861, #1443. It does not fix the render-target growth (#1462, merged), the shadow map's size (#1367)
or the target pool's churn (#1996). Those are other parts of the same VRAM.

Testing

Tested: Linux, RTX 3090, NVIDIA 620.70.06, main @ 9dd41326 only. I ran a memory tour (scripts/memtour.py,
repository save kokiri-start, Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death
Mountain Trail, Goron City, two laps, 60 s each) at RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=5200, a 6 GB
card's whole policy. 106 vk: VRAM: samples: median 4815 MB, peak 5227 MB, 2 over 5200. Textures sat at about the
1800 MB floor (budget 1800 MB in most samples) beside mirrors 550-1250 MB, shadows up to 364 MB, 340-1355 MB free in
the blocks, and 350-610 MB unaccounted. So on this PC the rest of the renderer still fits beside the floor. The GTX 1060
of #939 runs at 2560x1440, with Windows' own overhead on top, and there it does not fit.
This PR, in the game: this PR @ f42ab627 as a pack over the same base, through the job queue: a memory tour of 8 areas (Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death Mountain Trail, Goron City), 40 s each, at RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600, where the floor binds on this PC. 149 vk: budget: samples: the driver's count median 4379 MB, peak 4765 MB, 11 samples over 4600; textures median 1818 MB, down to 142 MB at the lowest when it gave way. No crash, no device loss, no texture errors seen, clean quit. main at 4600 MB is queued for the comparison (Measurements).

Still to test:

  • This PC, by hand: the tour above on main and on this PR at RX_VK_TEST_ALLOWED_MB=4600 RX_VK_TEST_VRAM_MB=6144,
    expecting main past 4600 MB with budget 1800 MB and this PR back under it, with textures below 1800 MB.
  • Windows, NVIDIA, a 6 GB card (GTX 1060 / 1660 / RTX 2060 / 3050 6 GB / 4050 Laptop): @Beanstalk below.
  • Windows, AMD, a 6-8 GB card (RX 5600 XT / 6600 / 6600 XT): @Beanstalk below.

@Beanstalk This changes the texture cache's budget on cards with less than 8 GB of video memory: once the driver says
the game is past the video memory it allows, the textures are no longer held at a floor of 1800 MB, so unused ones are
let go until it is back under the allowance. Please run main and this PR on Windows with NVIDIA on a 6 GB card and on
Windows with AMD on a 6 to 8 GB card, at the driver's default settings, 1440p output where the card allows it. Play
Hyrule Field for two minutes, then Kokiri Forest, Kakariko, the Market and Death Mountain Trail for two minutes each
turning the camera, then Hyrule Field again for two minutes. Expect the log's vk: VRAM: lines to stay at or under
the allowance on this PR where main goes past it, and the second Hyrule Field visit to keep its fps and 1% low near the
first one where main's falls. Expect no textures flashing, black, garbled or missing, and no crash or device lost.
Please report the peak vk: VRAM: used and allowed, the textures ... (budget ...) values, fps and 1% low on both
Hyrule Field visits, and the Windows "shared GPU memory" of the game if you can read it.

Measurements

Verdict: at a 6 GB card's 4600 MB allowance the driver's count peaks 574 MB lower (4765 against 5339 MB) and goes over the allowance less often (11 against 15 of 149 samples); the median moves 78 MB down. One run each way, in turn, on the same PC and tour: a single pair, so the peak is the clear part, the counts are close.

Setup: Linux, RTX 3090, NVIDIA 620.70.06; repository save kokiri-start, headless memory tour of 8 areas (Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death Mountain Trail, Goron City), 40 s each, default options; RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600 (where the floor binds on this PC); the vk: budget: lines, 149 samples each.

Before: main @ 9dd41326 After: this PR @ f42ab627
driver's VRAM count, median / peak (allowed 4600) 4457 / 5339 MB 4379 / 4765 MB
samples past the allowance 15 of 149 11 of 149
textures, median / lowest 1848 / 142 MB 1818 / 142 MB
crash, device loss, texture errors none none

An earlier main tour at 5200 MB (106 samples: median 4815, peak 5227, 2 over) showed the floor doesn't bind on this PC at that allowance.

vk: policy: sizes by the allowance at the start (the sites read it later): textures 3584 MB, mirrors 1700, target pool 433, compaction from 325, shadows 650, ...
vk: VRAM: 5187 MB used of 4669 MB allowed (742 MB of it free in the blocks); textures 2408 MB (budget 1800 MB), 1710 textures
vk: VRAM: 4751 MB used of 5200 MB allowed (552 MB of it free in the blocks); textures 1766 MB (budget 1800 MB), 1257 textures

Risks

Past the allowance more textures may be let go than before, down to those used in the last 60 presents (about 2 s at
30 fps). A texture brought back soon after costs an upload and, for ASTC, a decode: a short hitch when turning back to
something, where main paged instead. Within the allowance nothing changes. Cards with 8 GB allowed or more are
affected only past their allowance. No switch; a revert is enough. Nothing changes for players' files or caches.

Checklist

Ready for review

  • One topic, on current main; the title is a Conventional Commit.
  • Tested on my own PC, in the game for a change the game runs, as Testing says.
  • mise run test passes (any failure that also fails on main is named). (Not run on this PC; CI runs it.)
  • Measured against main, or Measurements says why there is nothing to measure.
  • No game files, build output or hand-edited generated files; third_party/ untouched; nothing GPL-incompatible.

When it applies

  • A fix: a game-free test that fails on main and passes here. (tests/test_vram_compact.py: the held_budget checks past the allowance fail on main; not run here, CI runs it.)
  • A check under tests/ that reads changed source is updated, not weakened.
  • Docs updated.
  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.

Ready to merge

  • Everything in Still to test has run, its results in Tested.
  • Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers").
  • Approved in review, every comment answered.
  • CI passes on the head; on current main, measured again or a comment says why the numbers still hold.

🤖 Generated with Claude Code

https://claude.ai/code/session_017vw8e2uxGKQxzDeQ15jjNK

## Summary On cards the driver allows under 8 GB, the texture cache has a floor: the 90% rule, the compaction's hold and reserve, and the 97% rule never cut its budget below 1800 MB (`vram_compact::texture_min_floor`), and the dynamic floor can be as high as half the allowance (#637). That stops a nearly full card from evicting the textures in use one pass after another, which is right while the card is within its allowance. Past the allowance it means no rule can take a texture once the textures are under the floor, whatever the rest of the renderer holds. The driver then pages VRAM out to system memory, and every frame reads part of it over PCIe. In #939 a GTX 1060 6 GB sat at 6128 MB of the 5231 MB it allowed with 1606 MB of textures, already under today's 1800 MB floor, and Hyrule Field went from 31 to 14 fps until a restart. With this change the floors apply only while the driver's count is within the allowance (`vram_compact::floor_within`, and `count_line_floor` for the 97% rule). Past it, the eviction's own rule still keeps every texture used in the last 60 presents, and the textures that have gone unused longer go until the count is back under the line. Within the allowance nothing changes. How to see it on `main`: a 6 GB card (or `RX_VK_TEST_VRAM_MB=6144` with an `RX_VK_TEST_ALLOWED_MB` low enough that the mirrors, shadows, render targets and blocks' free space take more than the allowance less 1800 MB). Tour a few areas. The log's `vk: VRAM:` line shows the count past the allowance and `textures ... (budget 1800 MB)` while the textures sit at or under 1800 MB. Why textures run over their budget in #939 and #1008: in those logs (nightly-189, rc.4, before #637 added today's floor) the textures were over a budget the rest of VRAM had squeezed to 136-448 MB. Most of them had been used in the last frames ("used in the last 8 presents 1668 MB" of 2358), so the eviction, which never takes a texture used in the last 60 presents, could not meet it. The budget is enforced, but it can't go below the working set. On `main` today the same squeeze stops at the floor instead. This PR lets it go on past the allowance, down to the working set. Read first: `floor_within` and `count_line_floor` in `gpu/vk/vram_compact.h`, then the two call sites in `texture_frame` (`gpu/vk/vk_texture.cpp`: `active_floor`, and the 97% rule's `floor`). See #939, #1008, #861, #1443. It does not fix the render-target growth (#1462, merged), the shadow map's size (#1367) or the target pool's churn (#1996). Those are other parts of the same VRAM. ## Testing **Tested:** Linux, RTX 3090, NVIDIA 620.70.06, `main` @ 9dd41326 only. I ran a memory tour (`scripts/memtour.py`, repository save `kokiri-start`, Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death Mountain Trail, Goron City, two laps, 60 s each) at `RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=5200`, a 6 GB card's whole policy. 106 `vk: VRAM:` samples: median 4815 MB, peak 5227 MB, 2 over 5200. Textures sat at about the 1800 MB floor (`budget 1800 MB` in most samples) beside mirrors 550-1250 MB, shadows up to 364 MB, 340-1355 MB free in the blocks, and 350-610 MB unaccounted. So on this PC the rest of the renderer still fits beside the floor. The GTX 1060 of #939 runs at 2560x1440, with Windows' own overhead on top, and there it does not fit. **This PR, in the game:** this PR @ f42ab627 as a pack over the same base, through the job queue: a memory tour of 8 areas (Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death Mountain Trail, Goron City), 40 s each, at `RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600`, where the floor binds on this PC. 149 `vk: budget:` samples: the driver's count median 4379 MB, peak 4765 MB, 11 samples over 4600; textures median 1818 MB, down to 142 MB at the lowest when it gave way. No crash, no device loss, no texture errors seen, clean quit. `main` at 4600 MB is queued for the comparison (Measurements). **Still to test:** - This PC, by hand: the tour above on `main` and on this PR at `RX_VK_TEST_ALLOWED_MB=4600 RX_VK_TEST_VRAM_MB=6144`, expecting `main` past 4600 MB with `budget 1800 MB` and this PR back under it, with textures below 1800 MB. - Windows, NVIDIA, a 6 GB card (GTX 1060 / 1660 / RTX 2060 / 3050 6 GB / 4050 Laptop): @Beanstalk below. - Windows, AMD, a 6-8 GB card (RX 5600 XT / 6600 / 6600 XT): @Beanstalk below. @Beanstalk This changes the texture cache's budget on cards with less than 8 GB of video memory: once the driver says the game is past the video memory it allows, the textures are no longer held at a floor of 1800 MB, so unused ones are let go until it is back under the allowance. Please run main and this PR on Windows with NVIDIA on a 6 GB card and on Windows with AMD on a 6 to 8 GB card, at the driver's default settings, 1440p output where the card allows it. Play Hyrule Field for two minutes, then Kokiri Forest, Kakariko, the Market and Death Mountain Trail for two minutes each turning the camera, then Hyrule Field again for two minutes. Expect the log's `vk: VRAM:` lines to stay at or under the allowance on this PR where main goes past it, and the second Hyrule Field visit to keep its fps and 1% low near the first one where main's falls. Expect no textures flashing, black, garbled or missing, and no crash or device lost. Please report the peak `vk: VRAM:` used and allowed, the `textures ... (budget ...)` values, fps and 1% low on both Hyrule Field visits, and the Windows "shared GPU memory" of the game if you can read it. ## Measurements **Verdict:** at a 6 GB card's 4600 MB allowance the driver's count peaks 574 MB lower (4765 against 5339 MB) and goes over the allowance less often (11 against 15 of 149 samples); the median moves 78 MB down. One run each way, in turn, on the same PC and tour: a single pair, so the peak is the clear part, the counts are close. Setup: Linux, RTX 3090, NVIDIA 620.70.06; repository save `kokiri-start`, headless memory tour of 8 areas (Zora's Domain, Hyrule Field, Kakariko, Market, Kokiri Forest, Lost Woods, Death Mountain Trail, Goron City), 40 s each, default options; `RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600` (where the floor binds on this PC); the `vk: budget:` lines, 149 samples each. | | Before: main @ 9dd41326 | After: this PR @ f42ab627 | |---|---|---| | driver's VRAM count, median / peak (allowed 4600) | 4457 / 5339 MB | 4379 / 4765 MB | | samples past the allowance | 15 of 149 | 11 of 149 | | textures, median / lowest | 1848 / 142 MB | 1818 / 142 MB | | crash, device loss, texture errors | none | none | An earlier `main` tour at 5200 MB (106 samples: median 4815, peak 5227, 2 over) showed the floor doesn't bind on this PC at that allowance. ``` vk: policy: sizes by the allowance at the start (the sites read it later): textures 3584 MB, mirrors 1700, target pool 433, compaction from 325, shadows 650, ... vk: VRAM: 5187 MB used of 4669 MB allowed (742 MB of it free in the blocks); textures 2408 MB (budget 1800 MB), 1710 textures vk: VRAM: 4751 MB used of 5200 MB allowed (552 MB of it free in the blocks); textures 1766 MB (budget 1800 MB), 1257 textures ``` ## Risks Past the allowance more textures may be let go than before, down to those used in the last 60 presents (about 2 s at 30 fps). A texture brought back soon after costs an upload and, for ASTC, a decode: a short hitch when turning back to something, where `main` paged instead. Within the allowance nothing changes. Cards with 8 GB allowed or more are affected only past their allowance. No switch; a revert is enough. Nothing changes for players' files or caches. ## Checklist ### Ready for review - [x] One topic, on current `main`; the title is a Conventional Commit. - [x] Tested on my own PC, in the game for a change the game runs, as Testing says. - [ ] `mise run test` passes (any failure that also fails on `main` is named). (Not run on this PC; CI runs it.) - [x] Measured against `main`, or Measurements says why there is nothing to measure. - [x] No game files, build output or hand-edited generated files; `third_party/` untouched; nothing GPL-incompatible. ### When it applies - [x] A fix: a game-free test that fails on `main` and passes here. (`tests/test_vram_compact.py`: the `held_budget` checks past the allowance fail on `main`; not run here, CI runs it.) - [x] A check under `tests/` that reads changed source is updated, not weakened. - [x] Docs updated. - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. ### Ready to merge - [ ] Everything in Still to test has run, its results in Tested. - [ ] Tier 1 hardware it can affect is tested (docs/dev/standards.md, "Hardware tiers"). - [ ] Approved in review, every comment answered. - [ ] CI passes on the head; on current `main`, measured again or a comment says why the numbers still hold. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_017vw8e2uxGKQxzDeQ15jjNK
A GTX 1060 6 GB in #939 sat at 6128 MB of the 5231 MB its driver allowed with
1606 MB of textures. On main the 90% rule, the compaction's hold and the 97%
rule all clamp the textures' budget at a floor (1800 MB on a card under 8 GB,
the dynamic one up to half the allowance), so with the textures under it no
rule can take one while the driver pages. held_budget's checks here fail on
main; the source checks name the sites in texture_frame.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
fix(vk): past the allowance the textures' floors give way, so a 6 GB card's VRAM comes back under it (#939)
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 8s
CI / Workflow lint (actionlint) (pull_request) Successful in 14s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 14s
Third-party parts / As their notes say (pull_request) Successful in 16s
Website / Website on Cloudflare (pull_request_target) Has been skipped
CI / Website build and checks (pull_request) Successful in 3m34s
PR description / Description follows the template (pull_request_target) Successful in 4s
beanstalk/approval can't be tested as asked: see the farm's comment
f42ab6276a
Within the allowance the textures' budget keeps its floor (1800 MB under
8 GB, the dynamic floor up to half the allowance) so a full card does not
evict the working set pass after pass. Past it the driver pages VRAM out to
system memory, which cost the GTX 1060 of #939 more than half its frame rate
until a restart, and the floor kept every rule from taking a texture there.
vram_compact::floor_within drops the floor once the driver's count is past the
allowance, in the 90% rule, the compaction's hold and reserve (held_budget)
and the 97% rule (count_line_floor). The textures used in the eviction's last
60 presents still stay. No switch: within the allowance nothing changes.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Author
Owner

Tested in game: Linux, RTX 3090, NVIDIA 620.70.06, this PR @ f42ab627 over main 9dd41326's base, through the job queue: memory tour of 8 areas at RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600 (a 6 GB card's policy). Driver count median 4379 MB, peak 4765 MB against 4600 allowed; textures gave way down to 142 MB when needed; no crash or texture errors, clean quit.

Tested in game: Linux, RTX 3090, NVIDIA 620.70.06, this PR @ f42ab627 over main 9dd41326's base, through the job queue: memory tour of 8 areas at RX_VK_TEST_VRAM_MB=6144 RX_VK_TEST_ALLOWED_MB=4600 (a 6 GB card's policy). Driver count median 4379 MB, peak 4765 MB against 4600 allowed; textures gave way down to 142 MB when needed; no crash or texture errors, clean quit.
Author
Owner

@Beanstalk This changes the texture cache's budget on cards with less than 8 GB of video memory: once the driver says
the game is past the video memory it allows, the textures are no longer held at a floor of 1800 MB, so unused ones are
let go until it is back under the allowance. Please run main and this PR on Windows with NVIDIA on a 6 GB card and on
Windows with AMD on a 6 to 8 GB card, at the driver's default settings, 1440p output where the card allows it. Play
Hyrule Field for two minutes, then Kokiri Forest, Kakariko, the Market and Death Mountain Trail for two minutes each
turning the camera, then Hyrule Field again for two minutes. Expect the log's vk: VRAM: lines to stay at or under
the allowance on this PR where main goes past it, and the second Hyrule Field visit to keep its fps and 1% low near the
first one where main's falls. Expect no textures flashing, black, garbled or missing, and no crash or device lost.
Please report the peak vk: VRAM: used and allowed, the textures ... (budget ...) values, fps and 1% low on both
Hyrule Field visits, and the Windows "shared GPU memory" of the game if you can read it.

@Beanstalk This changes the texture cache's budget on cards with less than 8 GB of video memory: once the driver says the game is past the video memory it allows, the textures are no longer held at a floor of 1800 MB, so unused ones are let go until it is back under the allowance. Please run main and this PR on Windows with NVIDIA on a 6 GB card and on Windows with AMD on a 6 to 8 GB card, at the driver's default settings, 1440p output where the card allows it. Play Hyrule Field for two minutes, then Kokiri Forest, Kakariko, the Market and Death Mountain Trail for two minutes each turning the camera, then Hyrule Field again for two minutes. Expect the log's `vk: VRAM:` lines to stay at or under the allowance on this PR where main goes past it, and the second Hyrule Field visit to keep its fps and 1% low near the first one where main's falls. Expect no textures flashing, black, garbled or missing, and no crash or device lost. Please report the peak `vk: VRAM:` used and allowed, the `textures ... (budget ...)` values, fps and 1% low on both Hyrule Field visits, and the Windows "shared GPU memory" of the game if you can read it.
Member

Beanstalk test farm

Commit 6781926 · approved by nobody yet · dashboard

A member of BigBeans can review this commit and comment @Beanstalk to test it.

<!-- beanstalk:sticky --> ### Beanstalk test farm Commit `6781926` · approved by nobody yet · [dashboard](https://beanstalk.magicbean.dev/prs/2082) A member of `BigBeans` can review this commit and comment `@Beanstalk` to test it.
Author
Owner

Measured main against this PR at a 6 GB card's 4600 MB allowance, same tour, one run each in turn: the driver's VRAM peak 5339 → 4765 MB, samples over the allowance 15 → 11 of 149. Measurements updated.

Measured main against this PR at a 6 GB card's 4600 MB allowance, same tour, one run each in turn: the driver's VRAM peak 5339 → 4765 MB, samples over the allowance 15 → 11 of 149. Measurements updated.
Author
Owner

@Beanstalk

@Beanstalk
HumanBean force-pushed fix/vram-floor-past-allowance from f42ab6276a
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 8s
CI / Workflow lint (actionlint) (pull_request) Successful in 14s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 14s
Third-party parts / As their notes say (pull_request) Successful in 16s
Website / Website on Cloudflare (pull_request_target) Has been skipped
CI / Website build and checks (pull_request) Successful in 3m34s
PR description / Description follows the template (pull_request_target) Successful in 4s
beanstalk/approval can't be tested as asked: see the farm's comment
to 6781926dd7
Some checks failed
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped
Third-party parts / The same as upstream (pull_request) Has been skipped
Website / Website on Cloudflare (pull_request) Has been skipped
CI / Conventional Commit title (pull_request) Successful in 15s
CI / Workflow lint (actionlint) (pull_request) Successful in 26s
Guide screenshots / Guide images and local UI source review (pull_request) Failing after 30s
Third-party parts / As their notes say (pull_request) Successful in 38s
PR description / Description follows the template (pull_request_target) Failing after 21s
beanstalk/approval not approved yet
CI / Website build and checks (pull_request) Successful in 7m15s
2026-11-21 17:26:15 +00:00
Compare
Author
Owner

Rebased onto main 4c8210e7. One conflict, text only (a code comment / the architecture doc's section beside main's new 'Each heap's budget'): kept main's text and this PR's. No code change from the tested head. No tests run on this PC (CI runs them).

Rebased onto main 4c8210e7. One conflict, text only (a code comment / the architecture doc's section beside main's new 'Each heap's budget'): kept main's text and this PR's. No code change from the tested head. No tests run on this PC (CI runs them).
HumanBean changed title from WIP: fix(vk): past the allowance the textures' floors give way, so a 6 GB card's VRAM comes back under it to fix(vk): past the allowance the textures' floors give way, so a 6 GB card's VRAM comes back under it 2026-11-21 17:32:31 +00:00
HumanBean deleted branch fix/vram-floor-past-allowance 2026-11-21 17:33:30 +00:00
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!2082
No description provided.