perf(vk): DLSS keeps its feature through loading screens (a 44-78 ms stutter at each area entry) #1520

Merged
NotRexed merged 1 commit from fix/dlss-keep-feature into main 2026-11-10 15:25:19 +00:00
NotRexed commented 2026-11-10 06:04:52 +00:00 (Migrated from github.com)

With DLSS on, the feature was released after 120 frames with no gameplay upscale. Every area change's loading screen lasts longer than that, so the feature was made again on the next area's first frame, costing 44-78 ms of the replay thread. In a 15-minute test walk that happened 12 times. The release was meant for a player who turns DLSS off in the menu, so the feature and its images are now kept while DLSS / DLAA is still the Upscaler's choice. RX_VK_DLSS_IDLE_RELEASE=1 restores the old behaviour.

Measured on an RTX 4060 Ti at 2560x1440 (tonight's DLSS A/B): the feature was made 12 times per walk, feature N made (again after it was let go; 44.0-78.4 ms of CPU). The re-test with this change is queued.

🤖 Generated with Claude Code

With DLSS on, the feature was released after 120 frames with no gameplay upscale. Every area change's loading screen lasts longer than that, so the feature was made again on the next area's first frame, costing 44-78 ms of the replay thread. In a 15-minute test walk that happened 12 times. The release was meant for a player who turns DLSS off in the menu, so the feature and its images are now kept while DLSS / DLAA is still the Upscaler's choice. `RX_VK_DLSS_IDLE_RELEASE=1` restores the old behaviour. Measured on an RTX 4060 Ti at 2560x1440 (tonight's DLSS A/B): the feature was made 12 times per walk, `feature N made (again after it was let go; 44.0-78.4 ms of CPU)`. The re-test with this change is queued. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Commenting is not possible because the repository is archived.
No description provided.