Skip to content

asahi: give the 10-bit X2 formats the RGB10A2 end-of-tile selector - #282

Merged
DjDeveloperr merged 1 commit into
aurora-wipfrom
mcgl/egl-g17p-x2
Nov 13, 2026
Merged

DjDeveloperr merged 1 commit into
aurora-wipfrom
mcgl/egl-g17p-x2

Conversation

@DjDeveloperr

Copy link
Copy Markdown

Summary

On the Neo (A18 Pro, G17P), 14 dEQP-EGL cases abort: color_clears.{single_context,multi_context,multi_thread}.{gles2,gles3}.other and render.{…}.{gles2,gles3,gles2_gles3}.other. Each one aborts at the first multisampled 10-bit XRGB config, with:

MESA: error: asahi: the end-of-tile shader did not compile for this chip
MESA: error: asahi: G17P background shader is unproven: op 3, … 2 samples, …
MESA: error: asahi: background shader has no proven native descriptor encoding on G17P

A multisampled EGL config resolves into its window buffer at the end of the render. On G17P that store-and-resolve end-of-tile program takes its per-target selector from agx_g17p_mrt_eot_selector(). The table listed R10G10B10A2_UNORM and B10G10R10A2_UNORM (0x75) but not their X2 spellings. So a multisampled B10G10R10X2 config got selector 0, the packer refused it, the end-of-tile program failed to compile, and the driver aborted.

X2 and A2 use the same tile format, and the single-sample end-of-tile path already treats X2 as A2. That is why the single-sample XRGB2101010 configs pass. This change maps both X2 formats to 0x75 and adds the two checks to test-agx-g17p-mrt-mixed.

Evidence

  • Root cause, measured on the Neo, aurora-wip 8f013bf: each of the 14 cases, captured alone, completes EGL configs 1-20 and aborts at config 21. eglinfo lists config 21 as XRGB2101010, 30-bit, 2 samples; config 20 is the same format at 1 sample.
  • The fix: being tested on hardware in the Neo batch gate (composite batch1), which includes the 14 cases and the dEQP-EGL sample. I'll add the result here before merge.
  • Expected: after the fix, the multisampled X2 configs (21-30) and the RGBA16 configs (106-120) run for the first time.
  • Other users of the table:
    • the M5's multi-target end-of-tile path now accepts X2 instead of failing;
    • G18G's mixed-format fallback gets 0x75 instead of 0 for X2;
    • older chips don't use the table.

🤖 Generated with Claude Code

A multisampled EGL config resolves into its window buffer at the end of
the render, and on G17P that store takes its per-target selector from
agx_g17p_mrt_eot_selector(). The table listed the RGB10A2 formats but
not their X2 spellings, so a multisampled B10G10R10X2 config got
selector 0, its end-of-tile program failed to compile, and the driver
aborted. dEQP-EGL's color_clears and render .other cases hit this at
the first 2-sample X2 config.

X2 and A2 share the same tile format; map X2 to 0x75 like A2.

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr

DjDeveloperr commented Nov 13, 2026 •

Copy link
Copy Markdown
Author

Gate evidence for merging #282 (batch1, step 7 of a tree-checked sequence):

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.

1 participant