Render graph completion + TAA sky fix + froxel light clustering (audit tasks #15, #23)#66
Merged
Merged
Conversation
Cluster 9 prep: the GTAO dispatch (temporal EMA ping-pong, Halton-5 rotation) leaves end_frame_with_scene; p22/p32 derive inside the method. Goldens pixel-identical.
Cluster 10 prep: the Phase-4b translucent/refractive/additive dispatch (back-to-front sort, scene-color snapshot for reads_scene materials, impulse-field update) becomes record_translucent_pass. Goldens pixel-identical.
Cluster 11 prep: the measure+adapt pass becomes record_auto_exposure; its luminance source derives from composite_source_view() internally (it measures whatever the composite will read), and the caller keeps the src/dst ping-pong indices because the composite binds the same dst view. Goldens pixel-identical.
…jection With zero sky velocity, TAA reprojected sky pixels POSITIONALLY: the far-plane world reconstruction divides by a near-zero w, landing prev_uv on arbitrary scene geometry. The luma-only YCoCg history clamp (correct for sparkle) then locks the wrong CHROMA in forever — skies tinted toward whatever bright object the broken UV hit (uniform green with a green sphere in frame; reddish with a red cube). Every TAA-on game shipped with this. Sky is at infinity, so far-depth pixels now reproject the view DIRECTION (w=0 through prev_vp): exact under camera rotation, translation-invariant by definition, with an in.uv fallback for degenerate w. Found by the new TAA-on golden (lit_primitives_taa), which now pins both the fix and the TAA branch of the post-FX cascade. TAA-off goldens unaffected (the change is inside the far-depth else-branch).
…b complete Every render pass between geometry upload and the terminal composite now executes as a PassNode in one frame Graph: shadow → hdr_scene → translucent → hiz_build → occlusion_capture → gtao → ssao_blur → ssr_march → ssr_temporal → ssgi → bloom → compose → postfx_tail → auto_exposure. Reads/writes declare the real data dependencies (Shadow/HdrColor/MRT/Depth named outputs; Transient tokens for the intermediates the enum doesn't name); each node additionally carries a with_after pin to its predecessor so the schedule reproduces the hand-tuned order exactly. Loosening those pins so the scheduler can interleave independent passes is the documented next refinement — do it dependency-by-dependency with the goldens watching. The context owns &mut Renderer (the cluster-1 pattern at full scale): closures borrow nothing at build time and call the record_* methods. Feature toggles are checked inside closures/methods, never by omitting nodes (with_after on a missing node is a schedule error by design). Immediate-mode uploads moved ahead of the graph build — queue write-buffer ordering is submission-scoped, so within this one encoder the move is semantically identical. All 6 goldens (incl. the new TAA-on scene) pixel-identical.
Point lights were O(live lights) per fragment since the 256-light cap raise. A 16x9x24 froxel grid now restores O(cluster lights) on every backend with fragment-stage storage buffers: - froxel.rs: compute pass assigns lights to clusters (sphere vs froxel AABB, log-distributed depth slices); cluster counts + index list in storage buffers; light data stays in the existing UBO. - The clustered point-light loop is spliced into SCENE_SHADER between the BEGIN/END-POINT-LIGHT-LOOP markers at pipeline build; the plain loop remains the WebGL2 fallback and semantic reference. - lighting_layout grows bindings 10-12 when clustered; pipeline_3d shares the layout unaffected (immediate-mode keeps the plain loop). - froxel_assign runs as a render-graph node before hdr_scene. - BLOOM_DISABLE_FROXEL=1 kill-switch forces the reference loop. Parity: new golden many_point_lights_clustered_scene drives the retained scene path (scene_pipeline) under 40 lights; its golden was generated with the reference loop and the clustered path reproduces it bit-exactly (mean/max diff 0.0). The assignment is strictly conservative, so the two paths are mathematically identical. Also extracts lighting.rs: the group-1 bind-group entry list existed in four hand-synced copies (a binding-drift hazard); all lighting bind groups now go through one builder. mod.rs ratchets 11871->11775.
proggeramlug
changed the title
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.
Completes the final two open items from the architecture audit: the render-graph migration (#15) and froxel light clustering (#23). Also fixes a real TAA bug the new coverage caught.
Render graph (#15)
record_*methods.Froxel light clustering (#23)
The 256-light cap raise (task #14) removed the capability ceiling but left the scene shader paying O(live lights) per fragment. Now:
BEGIN/END-POINT-LIGHT-LOOPmarkers; the plain loop remains the WebGL2 fallback and the semantic reference. Capability-gated on fragment-stage storage buffers (FroxelPass::supported),BLOOM_DISABLE_FROXEL=1kill-switch for field bisection.many_point_lights_clustered_scenedrives the retained scene path under 40 lights; its golden was generated with the reference loop and the clustered path reproduces it with mean/max diff 0.0 (the assignment is strictly conservative, so the paths are mathematically identical). The test asserts the clustered path is actually active so it can't silently regress to testing the fallback against itself.All goldens pixel-identical (except the two new ones, by definition), 97 unit tests green, validators clean.