Respect unsigned storage-image formats #848

Closed

klimovich008 wants to merge 2 commits from codex/unsigned-image-sample-types into main

klimovich008 commented

2026-11-08 23:09:36 +00:00

(Migrated from github.com)

What this fixes

Compute preparation can emit a float-sampled storage image with an explicit unsigned format when the texture-format answer is false. Vulkan rejects the module with VUID-StandaloneSpirv-Image-04965, including the R32ui case reproduced on RX 9070 XT.

Why merge

Explicit unsigned formats determine the sampled type before consulting the environment.
Typeless and signed formats retain their existing queries and behavior.
Keep current main's depth guard and unrelated changes; preserve read, write and atomic work.

Change

gpu/vk/eden/shader_recompiler/ir_opt/texture_pass.cpp: recognize the five explicit unsigned image formats.
tests/unsigned_image_types_ir.cpp: reuse the real-pass/emitter read, write, buffer-write, atomic and Typeless fixture.
nx2/test_unsigned_image_types.py: validate fourteen modules, answer invariance and query counts.
tests/run_all.py: register the existing compiler/header/validator bindings.

Tested

python -I -B tests/run_all.py --jobs 1 --strict unsigned_image_types.

Fails on the old code, passes now: earlier CPU checks found five old float-answer failures; all fourteen fixed modules validated and nine controls stayed byte-identical. Earlier RX 9070 XT/API4 checks built four fresh compute variants with no VUIDs on #772 plus #834 and this fix. All four rebased CI jobs pass, including the Linux fourteen-module regression. Windows passes with the expected validator-dependent skip.

Not covered

Current-main GPU and full-game validation remain pending. Earlier hardware evidence used the combined #772/#834 build.
Signed formats and separate extended-format/buffer-read capability gaps are unchanged.

## What this fixes Compute preparation can emit a float-sampled storage image with an explicit unsigned format when the texture-format answer is false. Vulkan rejects the module with `VUID-StandaloneSpirv-Image-04965`, including the R32ui case reproduced on RX 9070 XT. ## Why merge - Explicit unsigned formats determine the sampled type before consulting the environment. - Typeless and signed formats retain their existing queries and behavior. - Keep current main's depth guard and unrelated changes; preserve read, write and atomic work. ## Change - `gpu/vk/eden/shader_recompiler/ir_opt/texture_pass.cpp`: recognize the five explicit unsigned image formats. - `tests/unsigned_image_types_ir.cpp`: reuse the real-pass/emitter read, write, buffer-write, atomic and Typeless fixture. - `nx2/test_unsigned_image_types.py`: validate fourteen modules, answer invariance and query counts. - `tests/run_all.py`: register the existing compiler/header/validator bindings. ## Tested `python -I -B tests/run_all.py --jobs 1 --strict unsigned_image_types`. Fails on the old code, passes now: earlier CPU checks found five old float-answer failures; all fourteen fixed modules validated and nine controls stayed byte-identical. Earlier RX 9070 XT/API4 checks built four fresh compute variants with no VUIDs on #772 plus #834 and this fix. All four [rebased CI jobs](https://github.com/klimovich008/Hookshot/actions/runs/42578982395) pass, including the Linux fourteen-module regression. Windows passes with the expected validator-dependent skip. ## Not covered - **Current-main GPU and full-game validation remain pending.** Earlier hardware evidence used the combined #772/#834 build. - Signed formats and separate extended-format/buffer-read capability gaps are unchanged.

NotRexed commented

2026-11-09 05:30:35 +00:00

(Migrated from github.com)