WIP: perf(vk): keep opaque ASTC textures as BC1, half of BC3's VRAM (RX_VK_ASTC_BC1, off by default) #521

Draft

Amateurwizrd wants to merge 4 commits from Amateurwizrd/Longshot:perf/astc-bc1 into main

pull from: Amateurwizrd/Longshot:perf/astc-bc1

merge into: MagicBeanDev:main

MagicBeanDev:main

Conversation 6 Commits 4 Files changed 16 +1589 -21

Amateurwizrd commented 2026-11-14 17:40:01 +00:00

Contributor

• Copy link
• Copy Markdown

Why

On a GPU without ASTC (NVIDIA and most desktop AMD cards) every ASTC texture is kept as BC3, 8 bits a texel, though
the opaque ones need only BC1's 4: a card short of VRAM pays for that alpha nobody uses with evictions and uploads.

What changed

• vk: gpu/vk/astc_opaque.h (new) reads each ASTC block's mode and colour endpoint modes without decoding; void-extent
  blocks with alpha at its maximum and blocks whose every partition's mode carries no alpha are opaque; modes that carry
  alpha, reserved and illegal blocks are not.
• vk: with RX_VK_ASTC_BC1=1, create in gpu/vk/vk_texture.cpp makes a texture whose every block is opaque as BC1
  (its RGB format, so no block decodes as transparent), before the image is made and counted, so the budget and the
  eviction see BC1's true size.
• vk: upload_astc_gpu reads each layer of guest memory once into a private copy, judges that copy, and stages the same
  copy; a layer with alpha makes the image again as BC3 (bc1_to_bc3, the texels it holds kept) before anything is
  reserved or recorded, and a refused remake reserves no staging and is tried again.
• vk: texture_view_of in gpu/vk/vk_draw.cpp resolves the view again when the touch on a cached hit made the image
  again, so the draw samples the new image; a copy between a BC1 and a BC3 texture goes through the encoder
  (gpu/vk/vk_xfer.cpp). The textures' log line counts those kept as BC1 and the MB saved.
• tests: tests/test_astc_opaque.py (hand-built blocks of every mode and 420,000 random blocks, each one judged opaque
  decoding to alpha 255 in the tree's decoder), tests/test_astc_bc1_remake.py (the real upload, remake and view code
  with a modelled device: alpha appearing, a game write between the judging and the staging, a refused remake, a cached
  view); tests/test_native_astc_texture_count.py and tests/texture_staging_publication_test.cpp follow the new code.
• docs: docs/development.md documents RX_VK_ASTC_BC1.

Flags and settings

Flag	What it does	Default now	Recommendation	Helps on	Doesn't help on	Untested on
RX_VK_ASTC_BC1=1	Opaque ASTC textures kept as BC1 instead of BC3 where the GPU has no ASTC	off	off by default; on after testing on 8 GB cards	cards short of VRAM: 158-240 MB less in the scenes measured	a 16 GB card: nothing gained, frame time within 0.05 ms of off here	8 GB cards, AMD, Intel, GPUs with ASTC (unchanged there)
RX_VK_COUNT_BY_ESTIMATE=1 (existing)	The BC3 remake counts its image as create does under this switch; the switch itself is unchanged	off	unchanged	-	-	-

Measurements

Baseline: this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0
The switch is off by default, so main and this PR at their defaults run the same path; the comparison is the switch, on
the same build. The branch's head, 3484acbf, adds only the docs/tested/ files of these runs.
main: 9c9d69ed (2026-11-14)
this PR: perf/astc-bc1 @ 1f9c5eb3
Hardware: RTX 5070 Ti 16 GB, driver 621.01, Ryzen 7 7800X3D, 32 GB, Windows 11 Pro 26200
Scene: Kokiri Forest (child) standing, and a scripted Hyrule Field (child) walk; output 2560x1440, render resolution 75% (1920x1080), frame generation off, uncapped, VSync off, sound off, a 1707x960 window
Method: 60 s of PresentMon (MsBetweenPresents) after 15 s of settling, 2 runs each, alternated (off, on, off, on) after a warm-up run, the median of the runs; VRAM from the last vk: VRAM: line and the textures' line of longshot_log.txt
How to reproduce: a save at each spot, RX_VK_ASTC_BC1=0 and =1 in turn on the same build, the numbers from PresentMon and longshot_log.txt

	this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0	this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1
Kokiri Forest: fps / 1% low	78.6 / 60.9	78.4 / 62.2
Kokiri Forest: median / p99 frame time	12.64 / 14.97 ms	12.69 / 14.82 ms
Kokiri Forest: VRAM in use / textures	6588 / 3768 MB	6562 / 3481 MB
Hyrule Field walk: fps / 1% low	80.7 / 43.5	80.7 / 41.7
Hyrule Field walk: median / p99 frame time	12.09 / 18.77 ms	12.10 / 18.85 ms
Hyrule Field walk: VRAM in use / textures	10742 / 7774 MB	10461 / 7419 MB

RX_VK_ASTC_BC1=0, Kokiri Forest:     median 12.73 / 12.55 ms, p99 15.18 / 14.75 ms; vk: VRAM: 6526 / 6650 MB used, textures 3672 / 3863 MB
RX_VK_ASTC_BC1=1, Kokiri Forest:     median 12.68 / 12.70 ms, p99 14.74 / 14.90 ms; vk: VRAM: 6596 / 6528 MB used, textures 3510 / 3452 MB
                                     (83 kept as BC1, 159 MB less than as BC3, 0 made BC3 again) / (80 kept as BC1, 157 MB less than as BC3, 0 made BC3 again)
RX_VK_ASTC_BC1=0, Hyrule Field walk: median 12.09 / 12.09 ms, p99 18.11 / 19.42 ms; vk: VRAM: 10677 / 10807 MB used, textures 7780 / 7768 MB
RX_VK_ASTC_BC1=1, Hyrule Field walk: median 12.16 / 12.04 ms, p99 19.11 / 18.59 ms; vk: VRAM: 10613 / 10308 MB used, textures 7435 / 7402 MB
                                     (294 kept as BC1, 238 MB less than as BC3, 0 made BC3 again) / (296 kept as BC1, 239 MB less than as BC3, 0 made BC3 again)

An earlier build that carried this change before the review's fixes (3 rounds against the switch off, with Kokiri
Forest at 4K too) agreed: 157-240 MB less VRAM, none made BC3 again, frame time 0.03-0.19 ms longer.

Testing

Tested: python -I -B tests/run_all.py on Windows 11 (llvm-mingw): main @ 9c9d69ed 531 passed, 5 failed; this PR
532 passed and the same 5 failing on main (test_oot_state_tap, test_portable_builder_home,
test_portable_builder_startup, test_tested_configurations, test_walk_key), plus test_gpu_hang_watch, which
fails and passes on reruns of main @ 9c9d69ed itself (a 4 s timing bound) and passed twice alone on this branch. The
renderer builds. In the game, the runs above: no device loss, no texture made BC3 again.

Not tested: frame dumps compared texel by texel (the screenshots are compared by eye); a card short of VRAM,
where the saving matters; GPUs with ASTC, where nothing changes; a Linux build (no Linux machine here).

Needs other testers:

• 8 GB NVIDIA or AMD card, Windows, RX_VK_ASTC_BC1=1: Hyrule Field and Castle Town against the switch off, expecting less VRAM in the textures' line, fewer evictions and no texture that turns transparent or wrong (request to open)
• Linux (x86_64, gcc or clang): the renderer builds with this PR, expecting no error in gpu/vk/vk_texture.cpp or gpu/vk/astc_opaque.h (request to open)

Screenshots

Kokiri Forest from the same save after 15 s of settling, the game's own picture (1280x720), attached in the comments:
"this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0" and "this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1". The same picture: the differences
are the fairies, the mist and the light moving between the two captures; no texture turned transparent or wrong.

Risk and rollback

With the switch on, a texture judged opaque wrongly would lose its alpha (the judging is checked against the decoder on
420,000 random blocks), and a mixed BC1/BC3 copy goes through the encoder, which costs CPU time. Off by default: nothing
changes unless RX_VK_ASTC_BC1=1 is set; unset it, or revert the commit.

Reviewer notes

The idea is Hookshot-old's e9868252; here the choice is made before the image is made and counted, BC1 is used in its
RGB format, the staged bytes are the judged bytes, and a texture that gains alpha is made again as BC3. Written for
Longshot.

Issues

None

Checklist

For everyone

• One topic.
• The title is a Conventional Commit (type(scope): description).
• Rebased on current main.
• python -I -B tests/run_all.py run and passing (any failure that also fails on main is named above).
• Tested in the game, or what wasn't is named under Testing.
• Testing filled in: what was tested, what wasn't, and what needs other testers (with test requests).
• Measured: main and this PR at named commits, on the same PC, scene and settings, at least 2 interleaved runs each, the median reported.
• No game files, dumps, build output or generated files.
• third_party/ untouched (a new folder with a PROVENANCE.md instead).
• The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled.

If applicable

• AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.
• Flags noted, and the switch reference updated (docs/development.md, the README).
• Screenshots for a change that shows on screen.
• A file in docs/tested/ for each new measured run (format in docs/tested-configurations.md; no appended row or assembled view).
• A check under tests/ that reads changed source updated in this PR, not weakened.
• Host code builds on both Windows (llvm-mingw) and Linux, with UTF-8 paths. (Windows only so far: Linux is a test request to open)
• Docs updated.

## Why On a GPU without ASTC (NVIDIA and most desktop AMD cards) every ASTC texture is kept as BC3, 8 bits a texel, though the opaque ones need only BC1's 4: a card short of VRAM pays for that alpha nobody uses with evictions and uploads. ## What changed - vk: `gpu/vk/astc_opaque.h` (new) reads each ASTC block's mode and colour endpoint modes without decoding; void-extent blocks with alpha at its maximum and blocks whose every partition's mode carries no alpha are opaque; modes that carry alpha, reserved and illegal blocks are not. - vk: with `RX_VK_ASTC_BC1=1`, `create` in `gpu/vk/vk_texture.cpp` makes a texture whose every block is opaque as BC1 (its RGB format, so no block decodes as transparent), before the image is made and counted, so the budget and the eviction see BC1's true size. - vk: `upload_astc_gpu` reads each layer of guest memory once into a private copy, judges that copy, and stages the same copy; a layer with alpha makes the image again as BC3 (`bc1_to_bc3`, the texels it holds kept) before anything is reserved or recorded, and a refused remake reserves no staging and is tried again. - vk: `texture_view_of` in `gpu/vk/vk_draw.cpp` resolves the view again when the touch on a cached hit made the image again, so the draw samples the new image; a copy between a BC1 and a BC3 texture goes through the encoder (`gpu/vk/vk_xfer.cpp`). The textures' log line counts those kept as BC1 and the MB saved. - tests: `tests/test_astc_opaque.py` (hand-built blocks of every mode and 420,000 random blocks, each one judged opaque decoding to alpha 255 in the tree's decoder), `tests/test_astc_bc1_remake.py` (the real upload, remake and view code with a modelled device: alpha appearing, a game write between the judging and the staging, a refused remake, a cached view); `tests/test_native_astc_texture_count.py` and `tests/texture_staging_publication_test.cpp` follow the new code. - docs: `docs/development.md` documents `RX_VK_ASTC_BC1`. ## Flags and settings | Flag | What it does | Default now | Recommendation | Helps on | Doesn't help on | Untested on | |---|---|---|---|---|---|---| | `RX_VK_ASTC_BC1=1` | Opaque ASTC textures kept as BC1 instead of BC3 where the GPU has no ASTC | off | off by default; on after testing on 8 GB cards | cards short of VRAM: 158-240 MB less in the scenes measured | a 16 GB card: nothing gained, frame time within 0.05 ms of off here | 8 GB cards, AMD, Intel, GPUs with ASTC (unchanged there) | | `RX_VK_COUNT_BY_ESTIMATE=1` (existing) | The BC3 remake counts its image as `create` does under this switch; the switch itself is unchanged | off | unchanged | - | - | - | ## Measurements Baseline: this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0 The switch is off by default, so main and this PR at their defaults run the same path; the comparison is the switch, on the same build. The branch's head, 3484acbf, adds only the `docs/tested/` files of these runs. main: 9c9d69ed (2026-11-14) this PR: perf/astc-bc1 @ 1f9c5eb3 Hardware: RTX 5070 Ti 16 GB, driver 621.01, Ryzen 7 7800X3D, 32 GB, Windows 11 Pro 26200 Scene: Kokiri Forest (child) standing, and a scripted Hyrule Field (child) walk; output 2560x1440, render resolution 75% (1920x1080), frame generation off, uncapped, VSync off, sound off, a 1707x960 window Method: 60 s of PresentMon (MsBetweenPresents) after 15 s of settling, 2 runs each, alternated (off, on, off, on) after a warm-up run, the median of the runs; VRAM from the last `vk: VRAM:` line and the textures' line of `longshot_log.txt` How to reproduce: a save at each spot, `RX_VK_ASTC_BC1=0` and `=1` in turn on the same build, the numbers from PresentMon and `longshot_log.txt` | | this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0 | this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1 | |---|---|---| | Kokiri Forest: fps / 1% low | 78.6 / 60.9 | 78.4 / 62.2 | | Kokiri Forest: median / p99 frame time | 12.64 / 14.97 ms | 12.69 / 14.82 ms | | Kokiri Forest: VRAM in use / textures | 6588 / 3768 MB | 6562 / 3481 MB | | Hyrule Field walk: fps / 1% low | 80.7 / 43.5 | 80.7 / 41.7 | | Hyrule Field walk: median / p99 frame time | 12.09 / 18.77 ms | 12.10 / 18.85 ms | | Hyrule Field walk: VRAM in use / textures | 10742 / 7774 MB | 10461 / 7419 MB | ``` RX_VK_ASTC_BC1=0, Kokiri Forest: median 12.73 / 12.55 ms, p99 15.18 / 14.75 ms; vk: VRAM: 6526 / 6650 MB used, textures 3672 / 3863 MB RX_VK_ASTC_BC1=1, Kokiri Forest: median 12.68 / 12.70 ms, p99 14.74 / 14.90 ms; vk: VRAM: 6596 / 6528 MB used, textures 3510 / 3452 MB (83 kept as BC1, 159 MB less than as BC3, 0 made BC3 again) / (80 kept as BC1, 157 MB less than as BC3, 0 made BC3 again) RX_VK_ASTC_BC1=0, Hyrule Field walk: median 12.09 / 12.09 ms, p99 18.11 / 19.42 ms; vk: VRAM: 10677 / 10807 MB used, textures 7780 / 7768 MB RX_VK_ASTC_BC1=1, Hyrule Field walk: median 12.16 / 12.04 ms, p99 19.11 / 18.59 ms; vk: VRAM: 10613 / 10308 MB used, textures 7435 / 7402 MB (294 kept as BC1, 238 MB less than as BC3, 0 made BC3 again) / (296 kept as BC1, 239 MB less than as BC3, 0 made BC3 again) ``` An earlier build that carried this change before the review's fixes (3 rounds against the switch off, with Kokiri Forest at 4K too) agreed: 157-240 MB less VRAM, none made BC3 again, frame time 0.03-0.19 ms longer. ## Testing **Tested:** `python -I -B tests/run_all.py` on Windows 11 (llvm-mingw): main @ 9c9d69ed 531 passed, 5 failed; this PR 532 passed and the same 5 failing on main (`test_oot_state_tap`, `test_portable_builder_home`, `test_portable_builder_startup`, `test_tested_configurations`, `test_walk_key`), plus `test_gpu_hang_watch`, which fails and passes on reruns of main @ 9c9d69ed itself (a 4 s timing bound) and passed twice alone on this branch. The renderer builds. In the game, the runs above: no device loss, no texture made BC3 again. **Not tested:** frame dumps compared texel by texel (the screenshots are compared by eye); a card short of VRAM, where the saving matters; GPUs with ASTC, where nothing changes; a Linux build (no Linux machine here). **Needs other testers:** - 8 GB NVIDIA or AMD card, Windows, `RX_VK_ASTC_BC1=1`: Hyrule Field and Castle Town against the switch off, expecting less VRAM in the textures' line, fewer evictions and no texture that turns transparent or wrong (request to open) - Linux (x86_64, gcc or clang): the renderer builds with this PR, expecting no error in `gpu/vk/vk_texture.cpp` or `gpu/vk/astc_opaque.h` (request to open) ## Screenshots Kokiri Forest from the same save after 15 s of settling, the game's own picture (1280x720), attached in the comments: "this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0" and "this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1". The same picture: the differences are the fairies, the mist and the light moving between the two captures; no texture turned transparent or wrong. ## Risk and rollback With the switch on, a texture judged opaque wrongly would lose its alpha (the judging is checked against the decoder on 420,000 random blocks), and a mixed BC1/BC3 copy goes through the encoder, which costs CPU time. Off by default: nothing changes unless `RX_VK_ASTC_BC1=1` is set; unset it, or revert the commit. ## Reviewer notes The idea is Hookshot-old's e9868252; here the choice is made before the image is made and counted, BC1 is used in its RGB format, the staged bytes are the judged bytes, and a texture that gains alpha is made again as BC3. Written for Longshot. ## Issues None ## Checklist ### For everyone - [x] One topic. - [x] The title is a Conventional Commit (`type(scope): description`). - [x] Rebased on current `main`. - [x] `python -I -B tests/run_all.py` run and passing (any failure that also fails on `main` is named above). - [x] Tested in the game, or what wasn't is named under Testing. - [x] Testing filled in: what was tested, what wasn't, and what needs other testers (with test requests). - [x] Measured: main and this PR at named commits, on the same PC, scene and settings, at least 2 interleaved runs each, the median reported. - [x] No game files, dumps, build output or generated files. - [x] `third_party/` untouched (a new folder with a `PROVENANCE.md` instead). - [x] The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled. ### If applicable - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. - [x] Flags noted, and the switch reference updated (docs/development.md, the README). - [x] Screenshots for a change that shows on screen. - [x] A file in `docs/tested/` for each new measured run (format in `docs/tested-configurations.md`; no appended row or assembled view). - [x] A check under `tests/` that reads changed source updated in this PR, not weakened. - [ ] Host code builds on both Windows (llvm-mingw) and Linux, with UTF-8 paths. (Windows only so far: Linux is a test request to open) - [x] Docs updated.

Amateurwizrd added 1 commit 2026-11-14 17:40:01 +00:00

perf(vk): opaque ASTC textures kept as BC1, half the video memory of BC3 (RX_VK_ASTC_BC1) ... 1f9c5eb31d

Where the GPU has no ASTC, ASTC textures are kept as BC3, 8 bits a texel; one
with no alpha needs only BC1's 4. With `RX_VK_ASTC_BC1=1` (off by default until
measured in the game) a texture whose every block is provably opaque is made
as BC1. `gpu/vk/astc_opaque.h` reads each block's mode and colour endpoint
modes without decoding: void-extent blocks with alpha at its maximum, and
blocks whose every partition's mode carries no alpha, are opaque; modes that
carry alpha, reserved and illegal blocks are not.

Unlike Hookshot-old's version, the choice is made before the image is made and
counted, so the budget and the eviction see BC1's true size; BC1 is used in its
RGB format, so no block can decode as transparent; each upload judges the very
bytes it stages, and a texture whose memory later holds alpha is made again as
BC3 before anything is encoded (a draw holding the BC1 image's cached view
resolves the new one); a copy between a BC1 and a BC3 texture goes through the
encoder instead of a block copy. The textures' log line counts those kept as
BC1 and the MB saved.

The idea is Hookshot-old's e9868252 (PR 2554); written anew for Longshot.

Measured on an RTX 5070 Ti, 3 rotated rounds against the same build with it
off: 81 textures kept as BC1 at Kokiri Forest (158 MB less than as BC3), about
295 on a Hyrule Field walk (238-240 MB less), none made BC3 again. Frame time
0.19 ms longer at Kokiri Forest (in each round), 0.06 ms on the walk and 0.03 ms
at 4K (noise). Off by default: the saving is for cards short of VRAM.

Tested: tests/test_astc_opaque.py (hand-built blocks of every mode, and
420,000 random blocks: each one judged opaque decodes to alpha 255 in the
tree's decoder), tests/test_astc_bc1_remake.py (new),
tests/test_native_astc_texture_count.py, the renderer builds.
Not tested: the picture against BC3, a card short of VRAM.

Assisted-by: Claude

mooki added the

area/renderer

gpu/amd

gpu/nvidia

kind

performance

priority

high

platform/linux

platform/windows

labels 2026-11-14 18:15:19 +00:00

Amateurwizrd added 1 commit 2026-11-14 19:25:36 +00:00

docs(tested): RX_VK_ASTC_BC1 off and on at Kokiri Forest and a Hyrule Field walk ... 3484acbff1

Four runs on an RTX 5070 Ti (621.01), Ryzen 7 7800X3D, Windows 11: the switch
off and on, two alternated runs each, at Kokiri Forest standing and on a
Hyrule Field walk, on perf/astc-bc1 1f9c5eb3.

Assisted-by: Claude

Amateurwizrd commented 2026-11-14 19:25:38 +00:00

Author

Contributor

• Copy link
• Copy Markdown

This PR's own run is in: perf/astc-bc1 @ 1f9c5eb3, RX_VK_ASTC_BC1 off and on alternated, 2 runs each after a warm-up, at Kokiri Forest standing and on a Hyrule Field walk (RTX 5070 Ti, 621.01). With it on: 287 MB fewer texture MB at Kokiri Forest (3481 against 3768) and 355 MB on the walk (7419 against 7774), 157-239 MB of it from BC1 in the textures' line, none made BC3 again; frame time within 0.05 ms either way. The Measurements now show this run instead of the earlier build's; 3484acbf adds its four docs/tested files. Screenshots in the next comment.

This PR's own run is in: perf/astc-bc1 @ 1f9c5eb3, RX_VK_ASTC_BC1 off and on alternated, 2 runs each after a warm-up, at Kokiri Forest standing and on a Hyrule Field walk (RTX 5070 Ti, 621.01). With it on: 287 MB fewer texture MB at Kokiri Forest (3481 against 3768) and 355 MB on the walk (7419 against 7774), 157-239 MB of it from BC1 in the textures' line, none made BC3 again; frame time within 0.05 ms either way. The Measurements now show this run instead of the earlier build's; 3484acbf adds its four docs/tested files. Screenshots in the next comment.

Amateurwizrd commented 2026-11-14 19:25:46 +00:00

Author

Contributor

• Copy link
• Copy Markdown

Kokiri Forest from the same save after 15 s of settling, the game's own picture. The same picture both ways: the fairies, the mist and the light move between the captures; no texture turned transparent or wrong.

this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0

this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1

Kokiri Forest from the same save after 15 s of settling, the game's own picture. The same picture both ways: the fairies, the mist and the light move between the captures; no texture turned transparent or wrong. this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=0  this PR @ 1f9c5eb3, RX_VK_ASTC_BC1=1 

astc_bc1_off.png

2.6 MiB

astc_bc1_on.png

2.6 MiB

waryxxful referenced this pull request 2026-11-14 21:07:50 +00:00

Idea: sparse mirrors cost 1-4 ms of GPU a frame on an RX 5600 XT; whole mirrors are faster until their budget fills (runs wanted on other 6 GB cards) #585

HuzZz referenced this pull request 2026-11-14 22:59:21 +00:00

Sun shadows on Link flicker/break dependant on camera view #601

kingo commented 2026-11-15 00:54:10 +00:00

First-time contributor

• Copy link
• Copy Markdown

Hey, I've got an 8 GB RTX 5060 Ti and this PR says it's untested on 8 GB cards. Happy to run it, same route and A/B setup I used for #421, main vs your branch, VRAM and fps per area. Want me to?

Hey, I've got an 8 GB RTX 5060 Ti and this PR says it's untested on 8 GB cards. Happy to run it, same route and A/B setup I used for #421, main vs your branch, VRAM and fps per area. Want me to?

zeldaootr commented 2026-11-15 15:31:03 +00:00

Contributor

• Copy link
• Copy Markdown

We built the same switch on a local branch before seeing this PR. Yours is closer to landing, so we'd rather help it than compete. What we checked at 3484acbf:

• The opacity rule is sound. We fuzzed astc_opaque::block against a port of host_shaders/astc_decoder.comp (the decoder the renderer runs): 92M random blocks over all 14 footprints, 0 judged opaque that decode with alpha below 1.
• Game-free checks pass on Linux with g++ and clang (the 14 that touch these files), and the branch merges into main aa95d35f with no conflicts.

Two things we found:

1. Low: a BC1 → BC1 copy can be lost if an upload later in the same recording makes that texture BC3 (vk_xfer.cpp:274; the conversion's blit runs before the recording, vk_texture.cpp:2583). Rare. We have a small patch with a test.
2. Cost on a slow CPU: each opaque texture is scanned twice, at create and at its first upload. About 100 ms for the terrain array on a Zen 2 CPU. A desktop won't show it; a Deck might.

What we can offer, if you want any of it:

• The patch for 1, and a test that pins the decoder and encoder functions the rule relies on: two commits on your head, on our fork as help-521-bc1-copy-and-pins (308391e0, 2383da57). Cherry-pick them or redo them your way; we won't push to your branch.
• The fuzzer (one C++ file).
• A Steam Deck OLED run of your head at 1280x800, switch off and on, across Hyrule Field, the Lost Woods and the Market. That also gives you a Linux build.

Checked by an AI agent (Claude Code) for zeldaootr.

We built the same switch on a local branch before seeing this PR. Yours is closer to landing, so we'd rather help it than compete. What we checked at 3484acbf: - **The opacity rule is sound.** We fuzzed `astc_opaque::block` against a port of `host_shaders/astc_decoder.comp` (the decoder the renderer runs): 92M random blocks over all 14 footprints, 0 judged opaque that decode with alpha below 1. - **Game-free checks pass on Linux** with g++ and clang (the 14 that touch these files), and the branch merges into `main` aa95d35f with no conflicts. Two things we found: 1. **Low: a BC1 → BC1 copy can be lost** if an upload later in the same recording makes that texture BC3 (`vk_xfer.cpp:274`; the conversion's blit runs before the recording, `vk_texture.cpp:2583`). Rare. We have a small patch with a test. 2. **Cost on a slow CPU: each opaque texture is scanned twice,** at `create` and at its first upload. About 100 ms for the terrain array on a Zen 2 CPU. A desktop won't show it; a Deck might. What we can offer, if you want any of it: - The patch for 1, and a test that pins the decoder and encoder functions the rule relies on: two commits on your head, on our fork as [`help-521-bc1-copy-and-pins`](https://git.magicbean.dev/zeldaootr/Longshot/src/branch/help-521-bc1-copy-and-pins) (308391e0, 2383da57). Cherry-pick them or redo them your way; we won't push to your branch. - The fuzzer (one C++ file). - A Steam Deck OLED run of your head at 1280x800, switch off and on, across Hyrule Field, the Lost Woods and the Market. That also gives you a Linux build. *Checked by an AI agent (Claude Code) for zeldaootr.*

Rogasse referenced this pull request 2026-11-15 17:28:44 +00:00

Game: 13-25 real FPS in Castle Town Market on Nightly 189 with RTX 4050 6 GB #813

HumanBean force-pushed perf/astc-bc1 from 3484acbff1 to 8871fd4f23 2026-11-15 18:43:46 +00:00 Compare

HumanBean commented 2026-11-15 18:46:12 +00:00

Owner

• Copy link
• Copy Markdown

Rebased onto main @ bbe6fdf2. No tests were run here; the test boxes stay yours. No conflicts.

Rebased onto `main` @ bbe6fdf2. No tests were run here; the test boxes stay yours. No conflicts.

zeldaootr referenced this pull request 2026-11-15 19:30:26 +00:00

chore(agents): /prior-work, a skill that looks for the work already done before a change is built #888

HumanBean referenced this pull request from a commit 2026-11-15 19:38:23 +00:00

chore(agents): /prior-work, a skill that looks for the work already done before a change is built (#888)

Amateurwizrd added 2 commits 2026-11-15 19:53:20 +00:00

fix(vk): a copy into a BC1-kept texture makes it BC3 in the recording's order (RX_VK_ASTC_BC1) ... ac2cc16b30

A block-for-block BC1 -> BC1 copy wrote the BC1 image; a later upload in the same recording that found alpha made the
image BC3 from the upload command buffer, which runs before the recording, so the copy landed in the retired image and
was lost. `copy_t2t` now sends every copy into a BC1 texture to `texture_copy_into_bc`, which makes it BC3 first.

Assisted-by: Claude

test(vk): pin the decoder and encoder functions astc_opaque.h and BC1 rest on ... d36ba1f2b1

Assisted-by: Claude

Amateurwizrd commented 2026-11-15 19:53:27 +00:00

Author

Contributor

• Copy link
• Copy Markdown

@kingo yes please, thank you. An 8 GB card is exactly the gap. Same route and setup as #421: main vs this PR with RX_VK_ASTC_BC1=1, plus this PR with =0, VRAM and fps per area. Head is d36ba1f2.

@zeldaootr thanks, that's a thorough check. I took both commits as they are (now ac2cc16b, d36ba1f2), on top of the rebase onto bbe6fdf2. Checks that read this diff, one at a time with the pinned llvm-mingw: 372 run, 291 passed; every failure also fails on main on this PC, and test_astc_opaque and test_astc_bc1_remake pass. Yes to the Steam Deck OLED run at 1280x800, off and on, which also covers the Linux build. On the double scan: invisible on a desktop. If the Deck run shows it, carrying the opacity verdict from create to the first upload is the fix; I'd rather measure first. One question: every copy into a BC1 texture now makes it BC3. Do you know whether OoT copies into them often? If it does, part of the saving goes; kingo's run will show it in the VRAM numbers. I'd keep the fuzzer out of the PR unless the maintainers want it in tests/.

@kingo yes please, thank you. An 8 GB card is exactly the gap. Same route and setup as #421: main vs this PR with `RX_VK_ASTC_BC1=1`, plus this PR with `=0`, VRAM and fps per area. Head is `d36ba1f2`. @zeldaootr thanks, that's a thorough check. I took both commits as they are (now `ac2cc16b`, `d36ba1f2`), on top of the rebase onto `bbe6fdf2`. Checks that read this diff, one at a time with the pinned llvm-mingw: 372 run, 291 passed; every failure also fails on main on this PC, and `test_astc_opaque` and `test_astc_bc1_remake` pass. Yes to the Steam Deck OLED run at 1280x800, off and on, which also covers the Linux build. On the double scan: invisible on a desktop. If the Deck run shows it, carrying the opacity verdict from `create` to the first upload is the fix; I'd rather measure first. One question: every copy into a BC1 texture now makes it BC3. Do you know whether OoT copies into them often? If it does, part of the saving goes; kingo's run will show it in the VRAM numbers. I'd keep the fuzzer out of the PR unless the maintainers want it in `tests/`.

Some checks are pending

Hide checks

CI / Conventional Commit title (pull_request) Blocked by required conditions

Details

CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions

Details

CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions

Details

CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions

Details

CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions

Details

Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions

Details

Third-party parts / As their notes say (pull_request) Blocked by required conditions

Details

Third-party parts / The same as upstream (pull_request) Blocked by required conditions

Details

Website / Website on Cloudflare (pull_request) Blocked by required conditions

Details

PR description / Description follows the template (pull_request_target) Successful in 4s

Details

This pull request has changes conflicting with the target branch.

• gpu/vk/vk_texture.cpp

Some workflows are waiting to be reviewed.

View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Merge commit title WIP: perf(vk): keep opaque ASTC textures as BC1, half of BC3's VRAM (RX_VK_ASTC_BC1, off by default) (#521)

Merge commit body Reviewed-on: https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/521

WIP: perf(vk): keep opaque ASTC textures as BC1, half of BC3's VRAM (RX_VK_ASTC_BC1, off by default) (#521) Reviewed-on: https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/521 Copy merge commit message

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u 'perf/astc-bc1:Amateurwizrd-perf/astc-bc1'

git switch 'Amateurwizrd-perf/astc-bc1'

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

area/accessibility

Barriers for people with disabilities

  

area/audio

Sound and music

  

area/input

Controllers, keyboard, mouse

  

area/install

Installing, updating, building from your game

  

area/menu

In-game menu and settings

  

area/renderer

Vulkan renderer, shaders, effects

  

area/saves

Save files

  

area/tooling

Developer tools: scripts, CI, build and release tooling

  

area/upscaling

DLSS, FSR, frame generation

  

beanstalk: approved @ 1697603

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 2c5a62e

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 4fd82b7

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 64d6d09

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 6f91fae

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 70a78f9

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 8015c9a

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 8d361dc

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 9d8356e

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ abf2227

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ d82f202

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ e5f1e40

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ ee8ab5f

Set by Beanstalk for show; the farm's database is what counts

  

ci: approved

A maintainer approved CircleCI for this fork PR (sticky; removed after 14 quiet days)

  

good first issue

A good start for a new contributor

  

gpu/amd

AMD GPUs

  

gpu/igpu

Integrated GPUs

  

gpu/intel

Intel GPUs

  

gpu/nvidia

NVIDIA GPUs

  

help wanted

Extra hands or knowledge wanted

  

kind

bug

Something doesn't work as it should

  

kind

chore

Build, CI, dependencies, tidying

  

kind

docs

Documentation

  

kind

feature

A new feature or an improvement

  

kind

performance

Frame rate, frame time, loading, memory

  

kind

question

A question rather than a change

  

needs measurements

Waiting for before/after measurements

  

platform/linux

Affects Linux (a change for every system carries both Windows and Linux)

  

platform/macos

macOS (Apple silicon or Intel Macs)

  

platform/steam-deck

Steam Deck

  

platform/windows

Affects Windows (a change for every system carries both Windows and Linux)

  

preview

Maintainers: deploy a preview of this PR's website (read its changes first). Removed after the run.

  

priority

critical

Crashes, data loss, or blocks most players

  

priority

high

Important: next in line

  

priority

low

Nice to have

  

regression

Worked in an earlier build, broken in a newer one: highest priority, above other bugs

  

status

blocked

Waits on something else first

  

status

confirmed

Reproduced or agreed on

  

status

duplicate

Already reported or proposed

  

status

invalid

Not a problem in this project

  

status

needs-info

Waiting for more information from the reporter

  

status

wontfix

Will not be worked on

  

testing wanted

Wanted: a test on other hardware (posts a test request in #test-requests)

  

ci: approved

No labels

area/accessibility

area/audio

area/input

area/install

area/menu

area/renderer

area/saves

area/tooling

area/upscaling

beanstalk: approved @ 1697603

beanstalk: approved @ 2c5a62e

beanstalk: approved @ 4fd82b7

beanstalk: approved @ 64d6d09

beanstalk: approved @ 6f91fae

beanstalk: approved @ 70a78f9

beanstalk: approved @ 8015c9a

beanstalk: approved @ 8d361dc

beanstalk: approved @ 9d8356e

beanstalk: approved @ abf2227

beanstalk: approved @ d82f202

beanstalk: approved @ e5f1e40

beanstalk: approved @ ee8ab5f

ci: approved

good first issue

gpu/amd

gpu/igpu

gpu/intel

gpu/nvidia

help wanted

kind

bug

kind

chore

kind

docs

kind

feature

kind

performance

kind

question

needs measurements

platform/linux

platform/macos

platform/steam-deck

platform/windows

preview

priority

critical

priority

high

priority

low

regression

status

blocked

status

confirmed

status

duplicate

status

invalid

status

needs-info

status

wontfix

testing wanted

ci: approved

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

4 participants

Notifications

Subscribe

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!521

No description provided.