Files
photocraft/crates/gpu
05246233b6 Kept composites follow the text gamma and free their memory; GPU timing survives an OOM retry (#8520, #8521) (#8522)
* Background redraw: key kept composites on the text gamma, trim them every frame (#8520)

Two gaps in the kept background redraws:

- Color Settings > "Blend Text Colors Using Gamma" is an application
  setting that type layers composite with, but it wasn't part of the key.
  After toggling it, going back to a kept state (an Alt-click solo and
  again, undo/redo) uploaded type edges drawn with the old gamma. The key
  now holds the gamma, and a redraw that ran while the gamma changed
  (even if changed back, and also while another document was shown)
  isn't kept. Every running redraw is checked every frame; two changes
  within one frame aren't seen.
- Kept composites were trimmed only when the next deferred redraw
  started. A state History dropped (or purged) kept its texels, and
  lowering Memory Usage or a small edit that grew a kept document didn't
  release them. Each frame now drops dead entries, and the cap is taken
  again when Memory Usage or the revision of any kept document changes.

Tests (each fails without its part of the fix):
a_state_drawn_before_the_text_gamma_changed_is_composited_again,
a_redraw_across_a_text_gamma_change_is_not_kept,
a_redraw_across_a_text_gamma_change_and_back_is_not_kept,
a_hidden_documents_redraw_across_a_text_gamma_change_and_back_is_not_kept,
kept_composites_go_when_memory_usage_drops_between_redraws,
kept_composites_go_when_a_small_edit_grows_the_document_past_the_cap,
an_older_kept_composite_goes_when_its_document_grows_past_the_cap,
a_kept_state_history_purged_goes_at_the_next_frame.

Co-authored-by: Michael Higgins <55243596+Higgs32584@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* GPU compositor: test the layer page out-of-memory paths, keep GPU timing after one (#8521)

The PR's tests passed with its fix reverted: main's after-cell evict()
already meets every assert of the zero-budget test, so the out-of-memory
fallback had no test. A test hook (Compositor::fail_layer_allocs) now makes
the next n page textures fail as out of memory, and a parity test covers
the in-place retry, the submit-and-retry, the CPU fallback with a healthy
device, and the next refresh back on the GPU. Each of those fails when its
part of the fix is removed.

A retry that failed after its submit also switched GPU timing off for the
rest of the session: the submit marks the measurement flushed, and
cancelling a flushed measurement stops timing for good. The retry waits
for the GPU, so it now clears that mark (GpuTimer::drained);
an_out_of_memory_refresh_does_not_stop_gpu_timing fails without it.

Co-authored-by: Michael Higgins <55243596+Higgs32584@users.noreply.github.com>
Co-authored-by: FredApps <153173854+FredApps@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Address CodeRabbit review: serialise the OOM timing test with GPU_LOCK

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Michael Higgins <55243596+Higgs32584@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: FredApps <153173854+FredApps@users.noreply.github.com>
2026-10-25 10:19:25 -04:00
..