[jargnar]

Fixes #44.

One commit on integration/aurora at b7b03c5b3b8a. It runs the existing vertex and fragment nir_opt_large_constants branch on the AGX2 chips too, so a large constant table is read from the shader's constant blob instead of being stored to scratch by every invocation. Skyrim Special Edition through DXVK 3 goes from 2.8 fps to 43 fps.

The change

• hk_preprocess_nir_internal() in hk_shader.c: the chip condition now guards only the compute arm. The vertex and fragment arm keeps its hk_has_scratch_sized_array() check and its 256-byte threshold, and now applies to every chip.
• Nothing after it needed a change. lower_load_constant, the blob that hk_compile_nir() keeps and the upload in hk_flush_shaders() already serve vertex and fragment shaders on every chip; Honeykrisp on AGX2 fills a large constant table in scratch for every pixel: Skyrim SE at 2.8 fps through DXVK 3 #44 has the links.

What stays as it was

• The native AGX3 and M5 paths: the same passes with the same thresholds for the same stages.
• AGX2 compute. Only hk_dispatch_m5_native() and hk_dispatch_g17p_native() publish cs.constant_data_addr, so the pass is not run for it.
• An array the shader writes to, and any shader without an array above the scratch threshold.
• The upload reuse in hk_cmd_upload_constants(), which is still native-only. On AGX2 a bound blob is uploaded again at each shader flush; for the 16 KB table here that is one upload a frame. It is left out to keep this to the gate, and can go in here if wanted.

MEASURED

On one AGX2 machine, kernel linux-aurora 7.2.11.aurora14-1. Two builds: b7b03c5b3b8a as it is, and this head, aecd98f83e8f. Vulkan driver only, debugoptimized with assertions, tools stage with LLVM 23, loaded with VK_DRIVER_FILES. Nothing was installed and no boot stack was touched.

• The standalone reproducer in Honeykrisp on AGX2 fills a large constant table in scratch for every pixel: Skyrim SE at 2.8 fps through DXVK 3 #44, every case PASS with no wrong pixel in both builds:

  Case	base	this head
  table in the fragment shader, 1920x1080	421.09 ms	0.30 ms
  the same, declared as DXVK declares it	421.24 ms	0.35 ms
  a table the fragment shader writes to, 256x256	14.99 ms	15.05 ms
  table in the vertex shader, 1005 points	0.95 ms	0.27 ms
  the same through a pass-through geometry shader	0.80 ms	0.34 ms

• What the driver compiled: the fragment shader goes from 3680 instructions, 1893 spills and 19696 bytes of scratch to 15 instructions; the vertex shader from 3701 instructions to 27; the shader that writes to its table keeps the same hash.

• Skyrim Special Edition (Proton Experimental ARM64, DXVK v3.1.1-188-gcf0a0d90f896514, 1920x1080, Ultra preset): 2.8 fps in two runs of the base, 43.4, 42.5 and 42.5 fps in three runs with the change. The shadow mask pass goes from 339 ms to 0.49 ms, its shader from 5987 instructions with 16080 bytes of scratch to 578 with none. These runs used a build of the same tree made before the commit existed.

• No GPU timeout or fault in the kernel log in any run. The kernel does log an occasional Job canceled on DRM scheduler teardown when one of the reproducer's short cases exits. Over 96 runs per driver it logged 18 such lines with the system package, 15 with the unpatched build and 32 with this head, whose short cases finish sooner; every run passed. It is there without this change, and which job it cancels was not looked into.

• python3 tools/siliconforest/check.py passes on this head.

• The CI probe .github/mesa-ci/pipeline.c passes on the drm-shim profiles g13g, g13x, g14g, g15s, g15c, g15g and g17p, with its own three shaders and with the table shaders in place of the interpolation pair. Run without the validation layer, which is not installed here.

NOT measured

• ci / result: on a first pull request from a fork it waits for a maintainer to start it.
• Vulkan CTS.
• Any other AGX2 chip on hardware, and the native AGX3 and M5 paths on hardware.
• Tessellation.
• Other games.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Extends large constant-table lowering to AGX2 vertex and fragment shaders.

• AGX2 vertex and fragment shaders read large tables from constant blobs.

_{Reviews (2) · Last reviewed commit: "hk: read large vertex and fragment const..." · Reviewed by Greptile}