Shader packs: Iris translation spike results and implementation order

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ludovic Marie
2026-10-15 15:15:06 +02:00
parent 304ea50a5f
commit eb466fe2d2
+65 -2
View File
@@ -15,7 +15,7 @@ Both are off by default. With no pack selected the renderer draws exactly what i
| Renderer layout (below) | today's code |
| Horizon shader packs: `composite` stage, uniforms, hot reload, Shader Packs screen | first slice planned |
| Horizon shader packs: `deferred`, `gbuffer`, `shadow` stages, custom textures | planned |
| Java shader packs: preprocessing + GLSL parse coverage | spike (results below once run) |
| Java shader packs: preprocessing + GLSL parse coverage | spike done: 450 of 470 programs reach valid WGSL (below) |
| Java shader packs: Iris pipeline emulation | planned |
## How the renderer draws a frame today
@@ -265,4 +265,67 @@ players bring their own, as with Iris.
## Spike results
Not run yet: the counts of programs that parse, the failures and their causes go here.
Run on 2026-10-15 with naga 30.0.1 over Complementary Reimagined r5.9.3, BSL v10.1.8 and Sildur's Vibrant v2.02
Extreme (the 26.3 builds from Modrinth, kept outside the repository). Every program of the overworld, Nether and End
folders and the shared root was counted; the `dh_*`, `clrwl_*` and `voxy*` programs, which Iris only loads with
Distant Horizons, Colorwheel or Voxy installed, were left out.
| Pack | Programs | naga as is | Preprocessed + legalised: parse | validate | WGSL written |
| --- | ---: | ---: | ---: | ---: | ---: |
| Complementary Reimagined | 183 | 0 | 180 | 180 | 180 |
| BSL | 173 | 0 | 170 | 165 | 165 |
| Sildur's Vibrant | 114 | 0 | 114 | 105 | 105 |
| **Total** | **470** | **0** | **464** | **450** | **450 (96%)** |
"As is" is naga's GLSL front end on the files with `#include` inlined and Iris's macros defined: it accepts only
`#version 440`–`460` with Vulkan-style bindings, and every pack is written for `#version 120`–`330` compatibility
profiles. "Preprocessed + legalised" is the translation of steps 1–3 above, done by a throwaway token-level tool:
Iris's `#include`, its macros (`MC_VERSION`, `MC_RENDER_STAGE_*`, ...) and the packs' `#if`s expanded with naga's own
preprocessor (`pp-rs`), then:
- `#version 450 core`; `attribute`/`varying` to `in`/`out` with locations, one per declarator;
- loose uniforms into one `std140` block (`bool`s stored as `int`);
- `sampler2D` & co. split into texture + sampler bindings, uses rebuilt as `sampler2D(t, s)`, also through function
parameters, and not where a local of the same name hides the uniform;
- compatibility built-ins (`gl_Vertex`, `gl_MultiTexCoord*`, `gl_Color`, `gl_ModelViewMatrix`, `gl_Fog`, `ftransform()`,
`gl_FragColor`, `gl_FragData[n]`) and Iris's core-profile names (`vaPosition`, `modelViewMatrix`, ...) declared;
`texture2D`, `shadow2D`, `texture2DLod`, ... renamed;
- global variables with non-constant initialisers (accepted by NVIDIA and AMD drivers, used throughout
Complementary) declared at global scope and assigned at the start of `main`; unsized global arrays sized from
their initialiser;
- `texture(shadowSampler, ..).x`: the float widened to `vec4` first (naga has no scalar swizzle);
- `isnan`, `isinf` and `modf` replaced by helper functions (WGSL has no `isnan`/`isinf`).
What is left, by cause:
| Programs | Cause | Fix |
| ---: | --- | --- |
| 9 | Sildur's passes a `mat3` between stages (`tbnMatrix`); WGSL can't | split matrix varyings into columns |
| 6 | `writeonly image2D` without a format (`shadowcomp.csh` in Complementary and BSL) | take the format from the pack's image declaration (`shaders.properties` / `const int` format) |
| 5 | naga GLSL front-end bug: a call result inside a constructor (`vec3(TAAJitter(..), z)`) leaves an unvisited expression (BSL) | spill call results to temporaries; report to naga |
So the GLSL side is feasible: naga's front end plus a legaliser reaches 96% on the three packs, and every remaining
failure has a mechanical fix. Valid WGSL isn't a working pack yet: the output still has to be bound to Horizon's
pipeline (render targets, the uniforms' values, vertex attributes for `gbuffers_*`), which is the emulation work.
The spike also showed what the real translator needs that a token rewriter does badly: scopes (a local named like a
uniform), types (which `texture` call returns a float) and declarations split across `#if`s, so it should rewrite a
GLSL syntax tree after preprocessing, not tokens.
### Order to implement
1. **Translator crate** (`horizon-shaderpack`): the preprocessor (includes, option `#define`s with the player's
values, Iris macros) and a syntax-tree legaliser covering every rewrite above plus the three fixes in the table;
a test running it over the user's own copies of the three packs when they're present (never committed), and
small hand-written GLSL fixtures in the repo.
2. **`shaders.properties`** and option screens: the option list, sliders, profiles, `program.*.enabled`, blend modes.
3. **Full-screen programs**: `deferred*`, `composite*`, `final` with `colortex0..15` (formats from the pack),
`depthtex0..2`, `noisetex` and the uniforms (most are already computed for the world shaders). This alone runs
Sildur's and BSL post-processing over Horizon's own world.
4. **`gbuffers_terrain`, `gbuffers_water`, `gbuffers_sky*`**: Horizon's terrain vertices extended with Iris's
attributes (`mc_Entity` from `block.properties`, `mc_midTexCoord`, `at_tangent`, lightmap), a second set of
terrain pipelines writing the pack's render targets.
5. **Entities, block entities, hand, weather, clouds, particles** `gbuffers_*`.
6. **Shadow pass**: terrain and entities from the sun into `shadowtex0/1` and `shadowcolor0/1`, culled from the
light's view, with the pack's distortion.
7. **The rest**: custom uniforms (`uniform.*` expressions), custom textures, compute programs and images, per-dimension
folders, `entity.properties`/`item.properties`.