Skip to content

mesa: Backport tu/autotune and reg_size_vec4 fixes for a702 - #3119

Open
kandru42 wants to merge 2 commits into
qualcomm-linux:wrynosefrom
kandru42:a702_wrynose_fix
Open

mesa: Backport tu/autotune and reg_size_vec4 fixes for a702#3119
kandru42 wants to merge 2 commits into
qualcomm-linux:wrynosefrom
kandru42:a702_wrynose_fix

Conversation

@kandru42

Copy link
Copy Markdown
Contributor

In the current freedreno driver of mesa currently shipped in wrynose branch does not have these changes which causes GPU hangs with few vulkan samples. master branch's mesa version has been updated to a newer version where these fixes are already present.

Pull the mesa commits mentioned as backport to solve the GPU hangs in current meta-qcom wrynose build.

QLI-2.0 P0 critical fix.

CR fixed: 4634531

lkondred and others added 2 commits September 10, 2026 19:15
The autotuner in Mesa 26.0.5 does not disable tuning for small
renderpasses (fewer than 5 draw calls). These RPs are difficult
to hash stably across command buffers, causing the autotuner to
make inconsistent GMEM/SYSMEM decisions that result in a GPU hang
on A704 (Shikra).

This hang reproduces with the following Vulkan sample applications:
  - computecullandlod
  - deferredmultisampling
  - indirectdraw

Pull in upstream Mesa commit bf2777c013 (tu/autotune: Disable
autotuning for small renderpasses by default) as an adapted
backport until oe-core advances to a mesa version that already
contains it.

Signed-off-by: Lakshman Chandu Kondreddy <lkondred@qti.qualcomm.com>
The current mesa(26.0.5) in wrynose has reg_size_vec4 for a702 as 48,
which is incorrect. Backport this commit(6a282e8f18) until oe-core
advances to a newer mesa version which contains this fix.

This hang reproduces with the following Vulkan sample applications:
  - computenbody
  - timelinesemaphore

The upstream commit also updates
src/freedreno/ci/freedreno-a702-fails.txt.
This file does not exist in mesa 26.0.5 used by the wrynose branch,
therefore the CI-specific hunk was omitted while preserving the
functional GPU hang fix.

Signed-off-by: Sahitya Kandru <kandru@qti.qualcomm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants