Repository navigation
ci: compiled shaders are built in CI and checked against their HLSL - #405
Merged
Merged
Conversation
foxnne
changed the title
`float3(x)` from one scalar is Metal and GLSL, not HLSL: `(float3)x`. The workflow goes on past a shader that does not compile, uploads what did, and says where a stale one differs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Compiled by this workflow (shadercross 1ff05bec, DXC v1.9.2602), the toolchain the backend's default shaders reproduce with byte for byte. The DXIL's interface (signatures, pipeline state, flags) is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
foxnne
marked this pull request as ready for review
October 16, 2026 15:41
foxnne
added a commit
that referenced
this pull request
Oct 16, 2026
**Stacked on #405** (its commits show here until it merges). The glass written once, in HLSL; its Metal and WebGL GLSL generated from it and checked by #405's job, instead of three programs kept in step by hand. Keeping them in step is where the shader bugs came from: #373's `float3(x)` was valid Metal and GLSL and broke the HLSL, and the compiled forms went stale twice. ## What changes - **`scripts/compile-shaders.sh`**: every `X.hlsl` with compiled forms committed beside it, to each of them. SPIR-V, DXIL and Metal by shadercross (the Metal SDL_GPU takes, as the backend's own default shaders already shipped). GLSL ES 3.00 for WebGL2 by spirv-cross from the SPIR-V, with NaN checks relaxed (HLSL's `min`/`max` otherwise come out wrapped in `isnan` tests at every pixel), float precision high, and the names the web binds (`vColor`, `vTextureCoord`, `uSampler`, `uTex1`, `uData`). The workflow runs it and compares every form. - **The glass loads the generated Metal and GLSL** (`LiquidField`); the hand-written `liquid_glass.metal` and `liquid_glass.glsl` are gone. The HLSL is the one source and carries the explanations the GLSL had. - **The web compiles a complete generated source as it is** (`web/web.js`, a `#version` line), on WebGL2 only: on WebGL1 the program fails and the glass falls back to its meshes (generated loops are not WebGL1's constant-bounded ones). - `program_example`'s committed Metal, hand-written to show the interface, is now generated like the rest (same interface). ## Verified - **CI's compile check**: the backend's default shaders reproduce byte for byte in all three of SPIR-V, DXIL and Metal with this toolchain. - **macOS**: the same in-app About dialog captured through the agent with the hand-written Metal (#405's build) and the generated one (this build): plain glass differs by a mean of 0.18 of 255, at most 2 (the dither); elsewhere only edges, where the dialog's contents sat a fraction of a pixel apart between the two captures. - **Web (Chromium, WebGL2)**: the generated GLSL compiles and links against the web's own vertex shader, and `uData`, `uSampler`, `uTex1` and `uMatrix` all resolve; this build's menus draw their glass through it, no program failed, and it looks as main's did. - `zig build`, `zig build test` (91/91 steps), `zig build test-integration` (53/53 steps), `zig build check-web`. - Not seen: Windows (D3D12) and Linux (Vulkan), as for #405; Firefox and Safari on the web. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Compiles the SPIR-V and DXIL that fizzy's native backend embeds (Vulkan, D3D12) from their HLSL in CI, and fails when what is committed is not what the HLSL compiles to, uploading the fresh files as an artifact to commit.
Why. Nothing noticed when an HLSL changed and its compiled files did not: Windows and Linux quietly drew the previous glass after #227 (fixed by hand in #242), and again since #373, whose shader changes are in the HLSL but not in the DXIL or SPIR-V. Its first run here found worse: main's glass HLSL has not compiled since #373 (
float3(x)from one scalar is Metal and GLSL, not HLSL). Fixed here ((float3)x).The toolchain is SDL_shadercross at
1ff05bec, with the DirectXShaderCompiler release it pins itself (Microsoft's prebuilt v1.9.2602, by SHA-256), built as shadercross's own CI builds it: only SPIRV-Cross and shadercross, a few minutes, cached by the pin. With it the backend's default shaders, unchanged since they were compiled, reproduce byte for byte, SPIR-V and DXIL: the committed files were made with this compiler. (An earlier run built shadercross's vendored DXC from source instead, half an hour: its SPIR-V matched, its DXIL differed in DXC's version string.)The glass SPIR-V and DXIL are compiled again from #373's HLSL (with the fix above). The DXIL's interface (input and output signatures, pipeline state, feature flags) is identical to the old one's, so the backend's pipeline for it is unchanged.
First step of making the glass one source. Next, in its own PR: the glass written once, in HLSL, its Metal and WebGL GLSL generated from it by SPIRV-Cross and checked by this same job, so the forms cannot drift apart as they did here.
Not seen on Windows or Linux yet: the new glass on D3D12 and Vulkan wants a look on those machines.
zig buildandzig build check-webpass on macOS.🤖 Generated with Claude Code