Skip to main content

concinnity_render/render_graph/
passes.rs

1// src/render_graph/passes.rs
2//
3// Stable identity for every render-graph pass. Used by:
4//
5//   - The graph itself, as the dispatch key the executor matches on.
6//   - The per-pass GPU timer (the Metal backend's `pass_timing`), which keys
7//     its sample-buffer slots off the same integer. Adding a new pass = a
8//     new variant here + a new entry in [`PASS_NAMES`] + a bumped
9//     `PASS_COUNT`; nothing else needs to change for timing to flow. The
10//     `every_pass_id_round_trips_to_its_name` test forces all three edits
11//     at compile time (a missed registration would otherwise report zero
12//     GPU time for the pass), and `pass_timing::slot_pair` debug_asserts
13//     the index at runtime.
14//
15// Variants are intentionally `#[repr(u32)]` so a `PassId` round-trips
16// through `as usize` into the [`PASS_NAMES`] / counter-sample-buffer
17// slot index. Adding a variant in the middle of the list will renumber
18// later variants and silently shift every timing slot; append-only.
19
20/// Stable display name for each pass. Index = `PassId as usize`. Used by
21/// the WS `profile.passes` reply and the per-pass timing readback.
22pub const PASS_NAMES: [&str; PASS_COUNT] = [
23    "cull",
24    "shadow",
25    "ssr_prepass",
26    "ssao_prepass",
27    "ssao_kernel",
28    "ssao_blur",
29    "main",
30    "auto_exposure",
31    "decals",
32    "fog",
33    "particles_sim",
34    "particles_draw",
35    "ssr_resolve",
36    "velocity",
37    "taa_resolve",
38    "bloom",
39    "composite",
40    "fog_froxel",
41    "upscale",
42    "transparent",
43    "raymarch",
44    "hiz_build",
45    "cull2",
46    "main2",
47    "ssgi",
48    "rt_reflections",
49    "gbuffer_prepass",
50    "reflection_composite",
51    "light_cull",
52    "spot_shadow",
53    "lines",
54    "hiz_final",
55];
56
57/// Number of distinct passes the engine times. Sized to match
58/// [`PASS_NAMES`]; the per-pass timing array in
59/// [`crate::profile::RenderStats`] is sized to at least this many
60/// slots.
61pub const PASS_COUNT: usize = 32;
62
63/// One per-pass identity. Cast to `usize` to index [`PASS_NAMES`] or any
64/// `[T; PASS_COUNT]` companion array.
65#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
66#[repr(u32)]
67pub enum PassId {
68    /// GPU visibility cull; produces the indirect draw arguments.
69    Cull = 0,
70    /// Directional shadow cascades and spot shadow slices.
71    Shadow = 1,
72    /// Depth / normal prepass feeding screen-space reflections.
73    SsrPrepass = 2,
74    /// Depth / normal prepass feeding ambient occlusion.
75    SsaoPrepass = 3,
76    /// The ambient-occlusion gather.
77    SsaoKernel = 4,
78    /// Bilateral blur over the occlusion buffer.
79    SsaoBlur = 5,
80    /// The lit forward pass.
81    Main = 6,
82    /// Luminance reduction driving auto-exposure.
83    AutoExposure = 7,
84    /// Projected decals.
85    Decals = 8,
86    /// Volumetric fog composite.
87    Fog = 9,
88    /// Particle simulation compute.
89    ParticlesSim = 10,
90    /// Particle draw.
91    ParticlesDraw = 11,
92    /// Reflection trace and composite.
93    SsrResolve = 12,
94    /// Screen-space velocity, for TAA and motion blur.
95    Velocity = 13,
96    /// Temporal anti-aliasing resolve.
97    TaaResolve = 14,
98    /// Bloom down/upsample chain.
99    Bloom = 15,
100    /// Tonemap, grade, and present.
101    Composite = 16,
102    /// Volumetric-fog froxel-volume compute pass. Populates a 3D
103    /// `(scattered, transmittance)` texture once per frame, sampled by the
104    /// fullscreen `Fog` render pass instead of an inline ray-march. Metal
105    /// only; DX/Vulkan keep the ray-march and never insert this pass.
106    FogFroxel = 17,
107    /// Temporal upscaling pass. When the world's `PostProcessConfig`
108    /// enables `temporal_upscaling`, the renderer draws the 3D scene at a
109    /// fraction of drawable size and inserts this pass between the post-SSR
110    /// scene and the Bloom + Composite stack. The backend runs its
111    /// platform-native temporal upscaler (MetalFX on macOS; FSR / DLSS /
112    /// XeSS slots on the Windows backends are placeholders today) to
113    /// reconstruct a drawable-resolution image. Replaces `TaaResolve`:
114    /// the upscaler does temporal accumulation itself, so adding both
115    /// would double-temporal the scene.
116    Upscale = 18,
117    /// Transparent / translucent geometry pass. Runs after `SsrResolve`
118    /// (so water + glass see opaque reflections) and before
119    /// `TaaResolve` / `Upscale` (so translucents pick up temporal
120    /// accumulation). Reads the latest scene-pre-taa colour + main
121    /// depth as sampled textures; writes scene-pre-taa blended
122    /// (SRC_ALPHA / ONE_MINUS_SRC_ALPHA). Each transparent draw owns
123    /// its own pipeline + descriptor set; the pass aggregates them as
124    /// a back-to-front sorted list at encode time. Gated on
125    /// `FrameGraphInputs::transparent_enabled`; when no consumer is
126    /// in the world, the slot is omitted entirely.
127    Transparent = 19,
128    /// Raymarched SDF volume pass. Rasterises the back faces of each
129    /// `SdfVolume`'s world-space bounding box and runs the user-authored
130    /// fragment shader, which sphere-traces a signed distance field
131    /// inside the box. Hit fragments write opaque colour into
132    /// `hdr_resolve` (RMW between `AutoExposure` and `Decals`) and
133    /// update the main depth attachment so the raymarched surface
134    /// composites with rasterised geometry naturally: decals, fog,
135    /// SSR-resolve, and TAA all consume the post-Raymarch depth and
136    /// colour. Gated on `FrameGraphInputs::raymarch_enabled`; when no
137    /// `SdfVolume` is in the world the slot is omitted entirely.
138    Raymarch = 20,
139    /// Mid-frame Hi-Z (depth-mip pyramid) rebuild for two-pass occlusion
140    /// culling. Inserted only when `FrameGraphInputs::two_pass_occlusion_enabled`
141    /// is on: after `Main` (phase 1) has written this frame's depth, this
142    /// compute pass reduces it into the Hi-Z pyramid so `Cull2` can re-test
143    /// the objects phase 1 occluded against up-to-date depth. Distinct from
144    /// the end-of-frame Hi-Z build (which feeds the *next* frame's phase-1
145    /// cull and stays an inline action, not a graph node). Every backend
146    /// implements the node; whether it appears is `two_pass_occlusion_enabled`,
147    /// which each seeds from its own two-pass state.
148    HizBuild = 21,
149    /// Phase-2 GPU cull for two-pass occlusion. Re-tests the objects `Cull`
150    /// (phase 1) marked Hi-Z-occluded against the freshly rebuilt pyramid
151    /// (`HizBuild`) and encodes a draw for any that turn out visible into a
152    /// second indirect command buffer `Main2` consumes. Reads the per-object
153    /// status buffer phase-1 cull wrote + the `draw_args2` buffer it writes.
154    /// Gated on `FrameGraphInputs::two_pass_occlusion_enabled`.
155    Cull2 = 22,
156    /// Phase-2 main pass for two-pass occlusion. Loads (does not clear) the
157    /// HDR colour + depth `Main` wrote and re-runs only the bindless-static
158    /// indirect draw through `Cull2`'s command buffer, depth-compositing the
159    /// disoccluded geometry with phase 1. Instanced + skinned geometry is not
160    /// Hi-Z-culled, so it is fully drawn in phase 1 and not repeated here.
161    /// Becomes the new head of the hdr_resolve post-decoration chain (so
162    /// AutoExposure / Decals / Fog / SSR see the combined result). Gated on
163    /// `FrameGraphInputs::two_pass_occlusion_enabled`.
164    Main2 = 23,
165    /// Screen-space global illumination. A refinement of SSR: it reuses the
166    /// SSR depth + normal pre-pass G-buffer and screen-space ray-march, but
167    /// integrates bounced radiance over a cosine-weighted hemisphere instead
168    /// of along one reflection vector. Sits on the hdr_resolve RMW chain (after
169    /// `Raymarch`, before `Decals`): it reads the lit scene as the bounce
170    /// radiance source and additively composites the gathered + denoised
171    /// indirect term back into it, so the near-field colour bleed layers on top
172    /// of the IBL ambient. Gated on `FrameGraphInputs::ssgi_enabled`; when
173    /// `indirect_lighting` is IBL-only the slot is omitted entirely.
174    Ssgi = 24,
175    /// Hardware ray-traced reflections. Occupies the same scene-pre-taa slot as
176    /// `SsrResolve` (reads the post-decoration `hdr_resolve`, writes
177    /// `scene_pre_taa`) and takes precedence over it: when this pass is live the
178    /// builder inserts it and omits `SsrResolve` (a world may author both; RT
179    /// runs where available, SSR is the fallback). It still relies on the SSR
180    /// depth + normal + roughness
181    /// pre-pass (so `SsrPrepass` is forced on), but instead of a screen-space
182    /// march it traces a world-space reflection ray against an acceleration
183    /// structure built over the static scene geometry, so off-screen reflected
184    /// geometry appears. Gated on `FrameGraphInputs::rt_reflections_enabled`,
185    /// and so only on GPUs that report ray-tracing support.
186    RtReflections = 25,
187    /// Unified geometry G-buffer pre-pass. One jittered traversal of the visible
188    /// set writes view-space normal + linear depth, perceptual roughness, and
189    /// screen-space motion into a single MRT (plus a sampleable depth), replacing
190    /// the separate `SsrPrepass` + `Velocity` (and the SSAO-owned prepass): every
191    /// consumer (SSR, SSAO, SSGI, RT, TAA, upscaler) reads this one output. Gated
192    /// on `FrameGraphInputs::unified_gbuffer_prepass`.
193    GBufferPrepass = 26,
194    /// Roughness-aware reflection composite. Not a standalone graph node: it is
195    /// encoded inline at the tail of the `SsrResolve` / `RtReflections` pass
196    /// (both write a reflection target, then blur it by roughness and composite
197    /// it over the scene). It carries its own timing slot so its cost is visible
198    /// separately from the trace/march that precedes it. Inline on every
199    /// backend, so no backend's graph executor ever dispatches this id: each
200    /// treats it as a programming error the way it treats the bundled SSAO
201    /// sub-passes.
202    ReflectionComposite = 27,
203    /// Clustered light-binning compute pass. Once per frame, before Main: bins the
204    /// scene's local lights (the GpuLight buffer) into a per-cluster index list
205    /// over a screen-tiled, exponential-depth froxel grid, which the forward pass
206    /// reads to shade each fragment from only its cluster's lights instead of
207    /// iterating every light. Runs when the world has local lights. Writes a
208    /// storage buffer Main reads (RAW edge).
209    LightCull = 28,
210    /// Depth-only render of each shadowed spot light's cone into one slice of the
211    /// spot shadow map array. Local lights are static, so the projections are
212    /// built once; only the depth contents refresh, one slice per frame under
213    /// `ShadowUpdate::Hybrid`. Runs when the world has a shadow-casting spot.
214    SpotShadow = 29,
215    /// World-space line geometry (trajectories, tethers, path previews, the
216    /// editor's origin axes). Blend-writes the resolved scene colour after the
217    /// world decorations, sampling the resolved scene depth so a line behind
218    /// geometry is occluded by it. Gated on `FrameGraphInputs::lines_enabled`:
219    /// a frame that submits no lines omits the node entirely, so a frame that
220    /// draws none never pays for it.
221    Lines = 30,
222    /// Terminal Hi-Z (depth-mip pyramid) build. Reduces the frame's final main
223    /// depth into the pyramid the *next* frame's phase-1 `Cull` tests against,
224    /// so it is declared last and reads the depth every decoration pass has
225    /// finished with. Distinct from `HizBuild`, which rebuilds the same pyramid
226    /// mid-frame from phase-1 depth for `Cull2`; when two-pass occlusion is on
227    /// both run and this one supersedes it for the next frame. Present whenever
228    /// the GPU-cull path built a pyramid (`FrameGraphInputs::hiz_build_enabled`).
229    HizFinal = 31,
230}
231
232impl PassId {
233    /// Stable display name, looked up in [`PASS_NAMES`]. `'static` since
234    /// the table is `const`.
235    pub fn name(self) -> &'static str {
236        PASS_NAMES[self as usize]
237    }
238}
239
240#[cfg(test)]
241mod tests {
242    use super::*;
243
244    // Every `PassId` variant, in declaration (index) order. This is a sized
245    // `[PassId; PASS_COUNT]` array, so adding a variant forces three edits to
246    // keep this compiling: bump `PASS_COUNT`, add the `PASS_NAMES` entry, and
247    // list the variant here. Combined with `expected_name`'s wildcard-free
248    // match below, a new graph pass cannot ship without a timing name (which
249    // would otherwise read as zero GPU time).
250    const ALL: [PassId; PASS_COUNT] = [
251        PassId::Cull,
252        PassId::Shadow,
253        PassId::SsrPrepass,
254        PassId::SsaoPrepass,
255        PassId::SsaoKernel,
256        PassId::SsaoBlur,
257        PassId::Main,
258        PassId::AutoExposure,
259        PassId::Decals,
260        PassId::Fog,
261        PassId::ParticlesSim,
262        PassId::ParticlesDraw,
263        PassId::SsrResolve,
264        PassId::Velocity,
265        PassId::TaaResolve,
266        PassId::Bloom,
267        PassId::Composite,
268        PassId::FogFroxel,
269        PassId::Upscale,
270        PassId::Transparent,
271        PassId::Raymarch,
272        PassId::HizBuild,
273        PassId::Cull2,
274        PassId::Main2,
275        PassId::Ssgi,
276        PassId::RtReflections,
277        PassId::GBufferPrepass,
278        PassId::ReflectionComposite,
279        PassId::LightCull,
280        PassId::SpotShadow,
281        PassId::Lines,
282        PassId::HizFinal,
283    ];
284
285    // Expected timing name per variant. The match has no wildcard arm, so
286    // adding a `PassId` variant fails to compile here until it is named. This
287    // is the forcing function for the timing-name registration gotcha.
288    fn expected_name(pass: PassId) -> &'static str {
289        match pass {
290            PassId::Cull => "cull",
291            PassId::Shadow => "shadow",
292            PassId::SsrPrepass => "ssr_prepass",
293            PassId::SsaoPrepass => "ssao_prepass",
294            PassId::SsaoKernel => "ssao_kernel",
295            PassId::SsaoBlur => "ssao_blur",
296            PassId::Main => "main",
297            PassId::AutoExposure => "auto_exposure",
298            PassId::Decals => "decals",
299            PassId::Fog => "fog",
300            PassId::ParticlesSim => "particles_sim",
301            PassId::ParticlesDraw => "particles_draw",
302            PassId::SsrResolve => "ssr_resolve",
303            PassId::Velocity => "velocity",
304            PassId::TaaResolve => "taa_resolve",
305            PassId::Bloom => "bloom",
306            PassId::Composite => "composite",
307            PassId::FogFroxel => "fog_froxel",
308            PassId::Upscale => "upscale",
309            PassId::Transparent => "transparent",
310            PassId::Raymarch => "raymarch",
311            PassId::HizBuild => "hiz_build",
312            PassId::Cull2 => "cull2",
313            PassId::Main2 => "main2",
314            PassId::Ssgi => "ssgi",
315            PassId::RtReflections => "rt_reflections",
316            PassId::GBufferPrepass => "gbuffer_prepass",
317            PassId::ReflectionComposite => "reflection_composite",
318            PassId::LightCull => "light_cull",
319            PassId::SpotShadow => "spot_shadow",
320            PassId::Lines => "lines",
321            PassId::HizFinal => "hiz_final",
322        }
323    }
324
325    #[test]
326    fn pass_names_match_pass_count() {
327        assert_eq!(PASS_NAMES.len(), PASS_COUNT);
328    }
329
330    #[test]
331    fn every_pass_id_round_trips_to_its_name() {
332        // Each variant's integer value must equal its position in both `ALL`
333        // and `PASS_NAMES` so the per-pass timing arrays stay aligned, and
334        // every pass must carry a non-empty, expected name.
335        for (i, &pass) in ALL.iter().enumerate() {
336            assert_eq!(pass as usize, i, "{pass:?} index out of order");
337            assert_eq!(pass.name(), PASS_NAMES[i], "{pass:?} name table mismatch");
338            assert_eq!(pass.name(), expected_name(pass), "{pass:?} name drifted");
339            assert!(!pass.name().is_empty(), "{pass:?} has an empty name");
340        }
341        // The last listed variant pins the high end of the index range.
342        assert_eq!(ALL[PASS_COUNT - 1] as usize, PASS_COUNT - 1);
343    }
344}