Details
-
Task
-
Resolution: Unresolved
-
P2: Important
-
None
-
6.5, 6.6, 6.7, 6.8
-
None
-
Linux/Yocto
-
342ae435a (dev), d33437c4f (6.7), e544c665d (6.6), 56f9a5248 (tqtc/lts-6.5), 828b4ab8d (dev), c31fcb1e6 (6.7), 0b41f67f3 (tqtc/lts-6.5)
-
Multimedia Wk 13-14
Description
The current implementation of the screen capture takes too much CPU and slow-down the UI. Taking a single frame takes around 65ms
The factors causing the performance problems.
- The capturing goes in the main thread because of OpenGL restrictions (the opengl compositor cannot write the content to FBO in another thread).
- Currently, we use CPU implementation, calling fbo.toImage(), which consists of 2 steps: downloading from the texture, and flipping.
Possible optimizations
1. retain opengl texture (fbo) in the video buffer and lazily (by the 1st request) map to memory (convert with no flipping). The approach needs to implement thread-dependent OpenGL context, and make the recorder able to handle transformed frames. This approach depends on https://bugreports.qt.io/browse/QTBUG-121832.
2. find a way to flip the texture by OpenGL.