Builtin Library
- ..
- README.md2.2 KB
- module.wgsl8.0 KB
README.md
temporal_resolve
The temporal resolve: how a frame's sample is folded into the history a
temporal reconstruction carries. Temporal antialiasing (taa_resolve, the
resolve at render scale 1) and temporal upscaling
(@builtin::systems.temporalUpscale.taau_resolve, the same resolve onto a
display grid larger than the raster) both accumulate through it.
#include "@builtin::systems.antiAliasing.temporal_resolve"
| function | what it is |
|---|---|
temporal_compress / temporal_expand | Light into the reversible tonemapped space every sample is weighed in (each colour over one plus its luminance), and back. |
temporal_ycocg / temporal_rgb | The YCoCg space the clip runs in, and back. |
temporal_tap_weight(d) | The weight a sample landing d pixels from a pixel's centre carries: a Gaussian fitted to a Blackman-Harris window. |
temporal_tap_weight_axis(d) | That weight along one axis; the Gaussian separates, so temporal_tap_weight(d) is the product of its two axes' weights and a gather over a grid of taps takes one per row and one per column. |
temporal_history_catmull_rom(history, p, dims) | The history read at p through a Catmull-Rom spline over the 4x4 footprint less its corners, which keeps the contrast a bilinear read averages away on every reprojection. |
temporal_neighbourhood_new / temporal_neighbourhood_add | The current frame's samples around the pixel, as a running sum and sum of squares in the compressed YCoCg space. |
temporal_clip(lo, hi, center, history) | The history moved toward the centre only as far as it takes to enter the box. |
temporal_accumulate(nb, current, history, gamma, alpha, reset) | One frame folded in: variance clipping of the history to the neighbourhood's mean plus or minus gamma standard deviations, the frame's share halved where the history passed the clip untouched (anti-flicker), and taken toward the current sample alone as reset rises (a disoccluded pixel, or one reprojected off the frame). Returns light. |
Each pass supplies what is particular to its grids: where its current sample comes from, where its history is read, and how far it trusts the reprojection.