[SteelCrab]

What

Add a WGSL backend and anyps5-wgsl tools for WebGPU while keeping native shader cache keys, SPIR-V and encoded artifacts compatible with current main. WGSL requests use distinct cache keys. The production compiler source-version stamp still changes when compiler sources change.

This is the WGSL prerequisite of #6879 and replaces closed #9094. The relinker fix #10017 and Emscripten library portability #11341 have merged. Unsupported WGSL cases and host limitations are recorded in the shader debt file.

Tested

Ubuntu 26.04.1 x86-64, GCC 15.2, Node 26.12.0, Chromium 151.0.7922.34; main 399e7fb57e2e202e009227286212fc19a69d2eab, tested head 2125187a957e614bc4f4ba85e0f02fa71b7cf6b0.

• Native shader-tools Release builds on main and the change passed. shader_native_snapshot: 1/1 passed on each; wgsl_native_compat: 1/1 passed on the change.
• Three public compute fixtures produced 9/9 byte-identical native files between main and this head: cache keys, SPIR-V and encoded artifacts. Native-to-WGSL-to-native cache reuse and distinct target keys passed.
• Browser WebGPU checked all three generated WGSL compute fixtures and created 3/3 compute pipelines on Chromium's SwiftShader Vulkan backend. Runtime format placeholders were set to representative values for this check. Five repeated browser checks passed; a malformed WGSL negative control was rejected.
• Both verbose CTest runs used ANYPS5_REQUIRE_VULKAN=1, ANYPS5_NO_SHADER_CACHE=1 and VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation on the same host; the native snapshot tests themselves are CPU-only. The browser pipeline check used the same variables with Chromium's SwiftShader Vulkan backend. Neither CTest log nor the pipeline log contained a missing-layer message, VUID- error or skipped case.
• The full native library libs target built successfully on the change, including the production shader recompiler and NID patching.
• Existing shader_cache_file_read and shader_cache_native_paths passed 2/2 on each tree, compiled from the respective sources through the same standalone CMake harness with the three Vulkan environment variables above. The change's full project build also ran those tests 2/2. In shader_cache_native_paths, two Windows-only Python cases were skipped on each Linux run.
• Independent Windows report: contributor tested this head merged onto their main snapshot; the standalone shader tools built and the two CTests passed three times. This is not a full Windows build.
• Local code and complete PR-template Conventions and git diff --check passed. Full product runtime suites, title execution and physical GPU tests have not run.

CI coverage: WGSL backend / native-compatibility-and-wgsl builds native main and the change, compares the nine native files, checks cache reuse, creates three Chromium WebGPU pipelines and rejects an invalid shader; it checks validation logs for missing layers, skipped cases and VUID- errors. Build / linux covers the full native library build and existing cache tests when a maintainer dispatches it; the automatic Linux job was skipped. Conventions / conventions checks repository and PR rules. Remote results for this updated head are pending. The earlier Naga/wgpu job passed on the previous head and is historical.

Draft work remaining: full product runtime tests, current-head hosted checks, and author coordination for the overlapping PRs below. The Chromium checker uses only Node built-ins and adds no third-party source or Cargo packages to the project.

Overlap review:

• At scan base 6f1d2c70ae68b6326dbbf5965d90571f71752129, tools/pr_overlap.py reported direct line conflicts with feat(agc): test the depth bounds in the pixel program when the device lacks the depthBounds feature #11551 in serialized artifact size assertions and fix(shader): synchronize GLC and DLC buffer accesses on MoltenVK with device atomics and the guest's release and acquire #14217, perf(shader): keep the device's f16 NaN results with IEEE_MODE clear when a startup probe matches RDNA #15719 and fix(shader): inline mesh helper functions on NVIDIA #15917 in the target options/cache-key append. Proposed handling is to retain the native target fields and serialization changes alongside the WGSL-only key suffix, then re-run native byte comparisons. No landing order has been agreed with those authors.
• The scan exited 2 because 49 open PRs already conflicted with main and could not be automatically compared. Branch-only diffs were extracted and compared against this branch. In particular, feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642, perf(agc): GPU descriptor heap on VK_EXT_descriptor_heap (APS5_GPU_HEAP) #9508, feat(agc): serve lane-keyed, pointer and runtime image-table columns #10193, feat(agc): resolve table samplers through device image keys #13928 and feat(agc): resolve runtime sampler addresses through device image keys #13929 touch variant/materialization or artifact size paths; feat(agc): run mesh-path draws through a compute pass on devices without mesh shaders #9356, feat(agc): emulate mesh-fallback barycentrics through tessellation #9387 and perf(agc): share one render pass between mesh-path draws run as compute #15548 touch shader target/emission and artifact sizes; perf(shader): run wave64 compute at one lane where no lane can tell #12324 touches the native emitter call. Their integrated results are untested. Other reported PRs share build or debt files. The overlap report lists the affected branches; their merge order remains unresolved.
• This draft makes no claim that those changes are compatible or that the overlap check passed. Coordinate the overlapping PRs before marking the change ready.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: Google SwiftShader software Vulkan via Chromium 151.0.7922.34, Ubuntu 26.04.1; no physical GPU tested
• AI-assisted: yes

[JJenkx]

Windows check of head 2125187a95, merged onto main bfbb25cb72 (fde1deb30e, clean merge).

Setup: Windows 11 Enterprise 22621, WinLibs GCC 15.2 UCRT, Release. Configured as your workflow does: -DANYPS5_BUILD_SHADER_TOOLS=ON -DBUILD_TESTING=ON, in a separate build directory.

• The standalone shader-tools configuration configures and builds cleanly with MinGW.
• wgsl_native_compat and shader_native_snapshot pass in 3 of 3 runs.
• Without -DBUILD_TESTING=ON that configuration registers no tests, because it returns before the usual test setup. Your workflow passes it explicitly, so that's only a note for anyone building it by hand.

Not covered here: the native-main snapshot comparison, the Rust wgsl-check tool, and the full emulator build with the changed Recompiler.cpp and ShaderDiskCache.cpp.

AI-assisted: yes.