Screen capture cannot keep 60 fps at ultrawide and 4K sizes (measured in debug builds) #880

Open
opened 2026-11-18 07:39:18 +00:00 by charles · 0 comments
Owner

Found while measuring #876 / #589 (PR #879), debug builds only so far: at a 3440×1440 capture source and above, the capture thread cannot keep up with 60 fps — 21–35 converted frames per second out of 60 at 3440×1440, 7 fps at 5120×2160 "Source". At the same sizes libwebrtc hands over 3–4× more torn copies than at 1920×1080 (17–27 per 20 s), all of which the #763 check dropped.

To do. Measure the same in a release build (the nightly's profile) with tools/capture-consistency/run.sh --chain --setting at 3440×1440, 3840×1600 and 5120×2160 for 1080p, 1440p and Source, on DMA-BUF and MemFd: converted fps, time per stage (pull, verify second look, convert/scale, hand to encoder), dropped-as-torn count. If the release build also falls short of the target frame rate, find the stage that costs the time and fix it (the verification pass comparing whole frames, a full-size conversion before scaling instead of scaling first, copies). A sharer with a 3440×1440 monitor (RTX 4080, Wayland, nightly 373) still sees the #589 top strip; whether that is related is unknown — a capture path that is saturated behaves differently from the one the #763 fix was measured on.

Found while measuring #876 / #589 (PR #879), **debug builds only so far**: at a 3440×1440 capture source and above, the capture thread cannot keep up with 60 fps — 21–35 converted frames per second out of 60 at 3440×1440, 7 fps at 5120×2160 "Source". At the same sizes libwebrtc hands over 3–4× more torn copies than at 1920×1080 (17–27 per 20 s), all of which the #763 check dropped. **To do.** Measure the same in a release build (the nightly's profile) with `tools/capture-consistency/run.sh --chain --setting` at 3440×1440, 3840×1600 and 5120×2160 for 1080p, 1440p and Source, on DMA-BUF and MemFd: converted fps, time per stage (pull, verify second look, convert/scale, hand to encoder), dropped-as-torn count. If the release build also falls short of the target frame rate, find the stage that costs the time and fix it (the verification pass comparing whole frames, a full-size conversion before scaling instead of scaling first, copies). A sharer with a 3440×1440 monitor (RTX 4080, Wayland, nightly 373) still sees the #589 top strip; whether that is related is unknown — a capture path that is saturated behaves differently from the one the #763 fix was measured on.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
charles/fluxer-native#880
No description provided.