mirror of
https://github.com/storytold/photocraft.git
synced 2026-11-21 01:18:43 +00:00
Every compositor built all ~30 pipelines when created. On the Windows CI runners (WARP, FXC) that took 3 to 7.5 s in every GPU test process, and a test that draws one layer needs two or three of them. The pipelines now sit in one table per compositor (shared with its siblings), one entry per kernel, each built the first time a draw needs it. try_new_with_format_threads(_classified), which the app uses for its background build, still builds every pipeline up front in parallel through the same table, so the app keeps its start-up error and its cached deterministic-failure path. A pipeline that fails on first use makes that render Unsupported (CPU fallback); a failed effect-map pipeline turns layer effects to the CPU as before. The #9173 rules hold: each kernel has its own vertex entry point, a kernel's formats are built on one thread, and every build in the process takes the same lock. every_compositor_pipeline_builds builds all pipelines through the threaded build, so FXC still compiles every shader on Windows CI. New count test: one layer builds a small share of the pipelines, and a second render or a sibling builds none.