fix(vk): check the cooperative matrix probe's mapping, fence and timeout #2742

Open

github-sync wants to merge 1 commit from github/pr/2742 into main

github-sync commented

2026-11-12 10:34:20 +00:00

Owner

@juliotec opened this pull request on GitHub · 2026-11-12 10:31 UTC · original

run_shader, the startup probe for cooperative matrices, wrote through the
mapped pointer without checking vkMapMemory, so a failed mapping crashed at
startup. And when the probe's dispatch was submitted but did not finish within
the 5 s wait (a slow or hung driver), it destroyed the pipeline, descriptor pool,
command pool, buffer and memory the GPU could still be using.

What changed (gpu/vk/vk_coopmat.cpp, run_shader):

a failed vkMapMemory frees the buffer and memory and fails the probe
a submitted dispatch that did not complete leaves its resources alive (a
one-time leak at startup) instead of destroying them under the GPU
a fence that could not be created skips the submit and fails the probe, with
everything freed
The probe's result and every successful path are unchanged.

How tested: release build; Zelda OoT on Radeon RX 9070 XT (Windows,
Adrenalin).

Assisted-by: GitHub Copilot

> [@juliotec](https://github.com/juliotec) opened this pull request on GitHub · 2026-11-12 10:31 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2742) run_shader, the startup probe for cooperative matrices, wrote through the mapped pointer without checking vkMapMemory, so a failed mapping crashed at startup. And when the probe's dispatch was submitted but did not finish within the 5 s wait (a slow or hung driver), it destroyed the pipeline, descriptor pool, command pool, buffer and memory the GPU could still be using. What changed (gpu/vk/vk_coopmat.cpp, run_shader): - a failed vkMapMemory frees the buffer and memory and fails the probe - a submitted dispatch that did not complete leaves its resources alive (a one-time leak at startup) instead of destroying them under the GPU - a fence that could not be created skips the submit and fails the probe, with everything freed The probe's result and every successful path are unchanged. How tested: release build; Zelda OoT on Radeon RX 9070 XT (Windows, Adrenalin). Assisted-by: GitHub Copilot

github-sync added 1 commit

2026-11-12 10:34:20 +00:00

fix(vk): check the cooperative matrix probe's mapping, fence and timeout de34c991e0

run_shader, the startup probe for cooperative matrices, wrote through the
mapped pointer without checking vkMapMemory, so a failed mapping crashed at
startup. And when the probe's dispatch was submitted but did not finish within
the 5 s wait (a slow or hung driver), it destroyed the pipeline, descriptor pool,
command pool, buffer and memory the GPU could still be using.

What changed (gpu/vk/vk_coopmat.cpp, run_shader):
- a failed vkMapMemory frees the buffer and memory and fails the probe
- a submitted dispatch that did not complete leaves its resources alive (a
  one-time leak at startup) instead of destroying them under the GPU
- a fence that could not be created skips the submit and fails the probe, with
  everything freed
The probe's result and every successful path are unchanged.

How tested: release build; Zelda OoT on Radeon RX 9070 XT (Windows,
Adrenalin).

Assisted-by: GitHub Copilot

github-sync referenced this pull request

2026-11-12 14:41:43 +00:00

feat(vk): the game's anti-aliasing networks on AMD RDNA3/4 matrix units (16x16x16 cooperative matrices) #1975

github-sync referenced this pull request

2026-11-12 14:44:37 +00:00

fix(vk): skip clears of textures kept in ASTC blocks #2718

github-sync commented

2026-11-12 14:45:11 +00:00

Author

Owner

@Joey-Magnus approved on GitHub · 2026-11-12 14:34 UTC · original

CI never ran here. I merged this with the other four of your Vulkan fixes onto current main, compiled vk_coopmat.cpp with clang++ using the back end's flags, and ran tests/run_all.py with the Vulkan headers. No failure comes from this file.

The changes are right. A failed vkMapMemory now frees the buffer and memory instead of memcpy into null. A missing fence skips the submit, and everything is freed because nothing was queued. A submitted dispatch that times out leaves its pool, pipeline, buffer and memory alive rather than destroying them while the GPU may still use them (gpu/vk/vk_coopmat.cpp:215). Destroying the shader module there is fine after pipeline creation. On a device that really hung, later submits queue behind it anyway, so the one-time leak is the right trade.

#1975 (cooperative matrices for the AA networks) changes run_shader's signature. It merges cleanly with this one, but whoever lands second should check that the new early return still frees what #1975 adds.

> [@Joey-Magnus](https://github.com/Joey-Magnus) approved on GitHub · 2026-11-12 14:34 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2742#pullrequestreview-5876244151) CI never ran here. I merged this with the other four of your Vulkan fixes onto current main, compiled vk_coopmat.cpp with clang++ using the back end's flags, and ran tests/run_all.py with the Vulkan headers. No failure comes from this file. The changes are right. A failed vkMapMemory now frees the buffer and memory instead of memcpy into null. A missing fence skips the submit, and everything is freed because nothing was queued. A submitted dispatch that times out leaves its pool, pipeline, buffer and memory alive rather than destroying them while the GPU may still use them (gpu/vk/vk_coopmat.cpp:215). Destroying the shader module there is fine after pipeline creation. On a device that really hung, later submits queue behind it anyway, so the one-time leak is the right trade. #1975 (cooperative matrices for the AA networks) changes run_shader's signature. It merges cleanly with this one, but whoever lands second should check that the new early return still frees what #1975 adds.

github-sync referenced this pull request

2026-11-12 14:45:14 +00:00

fix(vk): retry sparse blocks whose bind failed #2743

github-sync force-pushed github/pr/2742 from de34c991e0 to 5713e26850

2026-11-12 19:00:53 +00:00

Compare

Commenting is not possible because the repository is archived.

No reviewers