Ticket 016 V1-V4 landed first-order importance-sampling approximations (temporal + spatial jitter decorrelation, history-luma-scaled jitter magnitude, variance-adaptive EMA alpha). The ticket spec's fully-general approach — per-probe 2D PDF from prev-frame radiance × BRDF, precomputed inverse CDF, inverse-CDF ray direction sampling with per-ray PDF weighting for unbiased accumulation — was deferred as not load-bearing at our current 64-rays/probe density.
When to do this
- If a case appears where a probe's 64 rays waste most of their budget on low-contribution directions (e.g., tight interior rooms where only the doorway is lit).
- If ray budget ever drops to 16 or 32 rays/probe for perf reasons.
Scope
PROBE_CDF_PRECOMPUTE_WGSL — one compute pass per probe that reads the 8×8 octahedral history and writes row-sum + row-CDF textures (2× the atlas size).- Trace shader update (SW / SDF / HW): replace the fixed octel-to-direction mapping with inverse-CDF lookup (two
textureLoad calls per ray) and divide accumulation by the sampled PDF.
- Bind CDF textures on all 3 trace pipelines.
Why not yet
See docs/perf/016-lumen-importance-sampling.md §V4 closure — V1-V3 already redirect sample budget toward productive directions without the CDF machinery, and at 64 rays/probe the marginal gain over V3 is expected to be small.
Ticket 016 V1-V4 landed first-order importance-sampling approximations (temporal + spatial jitter decorrelation, history-luma-scaled jitter magnitude, variance-adaptive EMA alpha). The ticket spec's fully-general approach — per-probe 2D PDF from prev-frame radiance × BRDF, precomputed inverse CDF, inverse-CDF ray direction sampling with per-ray PDF weighting for unbiased accumulation — was deferred as not load-bearing at our current 64-rays/probe density.
When to do this
Scope
PROBE_CDF_PRECOMPUTE_WGSL— one compute pass per probe that reads the 8×8 octahedral history and writes row-sum + row-CDF textures (2× the atlas size).textureLoadcalls per ray) and divide accumulation by the sampled PDF.Why not yet
See docs/perf/016-lumen-importance-sampling.md §V4 closure — V1-V3 already redirect sample budget toward productive directions without the CDF machinery, and at 64 rays/probe the marginal gain over V3 is expected to be small.