adminandClaude Opus 5.5
df61b08bd1
A blur already worked with Sample texture, but only by wiring every sample by hand: 25 nodes for a 5x5. Averaging their colours also gave a dark seam wherever a neighbour was transparent. The Blur node writes the loop itself and averages the right way. - Blur takes radius (pixels of the picture, 0 to 8, fractions allowed) and u/v like Sample texture, and gives color and alpha. - Kind is Box (default) or Gaussian, Direction is Both, Horizontal or Vertical. Both are arguments of one prelude helper, blur(). - The samples are averaged premultiplied and only then made straight, so a transparent neighbour lowers the opacity and adds no black. - The outermost ring counts by the part of the radius that reaches it, so a radius from a Property changes smoothly and compiles nothing again. - A node may now declare prepare, a vec4 the compiler works out once into a local. Color and alpha read that one sum instead of blurring twice. - A reached Blur counts as a shifted sample, so the picture comes with the Output's Edge. It cannot feed the Vertex output. - Both perf tools take GM_PERF_BLUR=<radius>. Wider blurs and bloom need several passes. That is #301. Measured on stein2 (HD 630, one run per value, fps), Box blur: - 48x48 sprites at 100/400/800: radius 2 59.8/58.4/55.6, radius 4 53.1/51.4/36.3, radius 8 46.6/24.4/14.4. - 256x256 field shaders at 1/10/50: radius 2 60.2/56.0/49.0, radius 4 60.1/49.0/27.1, radius 8 52.4/34.9/11.4. Tests: - Gate: 224 files, 3467 tests, CHECK EXIT 0 (+7). - e2e 28: ALL PASSED, 94 checks (+5 in 28l-vi). Pixels match the arithmetic to 1: the last red texel at radius 0.5/1/4 reads 187/173/154 (expected 188/173/154). - Counter-checks: Blur not counted as shifted fails 1 unit test. Blur output left premultiplied fails the seam and spot checks. - e2e total 1712 is computed (1707 + 5). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>