Slow performance causing dropped frame for main streaming/recording

[norihiro]

Issue description

When enable this filter, sometimes it causes dropped frame for main streaming and recording.

Current status of investigation

Profile

Checking rendering time with adding profile.

ajaof_render: min=2.579 ms, median=4.488 ms, max=10.32 ms, 99th percentile=9.094 ms

• Trial 1 - Separate stage-surface
  With separating staging (until gs_stage_texture) and mapping (gs_stagesurface_map and later) into different iteration, rendering time is reduced and the number of dropped frame looks reduced. (I had 3 runs and the numbers are slightly different.)
  The code is 2ac4699.

ajaof_render: min=0.017 ms, median=0.602 ms, max=19.639 ms, 99th percentile=17.333 ms
ajaof_render_render: min=0.017 ms, median=0.272 ms, max=5.12 ms, 99th percentile=0.861 ms
ajaof_render_output: min=0.565 ms, median=0.92 ms, max=19.637 ms, 99th percentile=17.748 ms
ajaof_render: min=0.055 ms, median=0.584 ms, max=15.644 ms, 99th percentile=10.016 ms
ajaof_render_render: min=0.054 ms, median=0.259 ms, max=4.331 ms, 99th percentile=0.64 ms
ajaof_render_output: min=0.565 ms, median=0.693 ms, max=15.644 ms, 99th percentile=12.179 ms
ajaof_render: min=0.029 ms, median=0.669 ms, max=19.953 ms, 99th percentile=16.839 ms
ajaof_render_render: min=0.028 ms, median=0.285 ms, max=3.744 ms, 99th percentile=0.788 ms
ajaof_render_output: min=0.61 ms, median=1.297 ms, max=19.952 ms, 99th percentile=17.677 ms

• Trial 2 - Breakdown profile more
  It is not clear which is actually slow. Maybe gs_stagesurface_map or video_output_lock_frame. If later, prepare a new thread to call video_output_lock_frame.
• Trial 3 - Run stage-surface as the graphics task
  The graphics tasks are executed after all other rendering while the rendering routine ajaof_render is called at very beginning of rendering the main texture. Assuming gs_stage_texture is actually executed in pipeline, this approach might improve performance without slowing the output.
  If Trial 2 reveals the necessity to have another thread, another idea is call gs_stagesurface_map from the new thread. The call will be automatically delayed by obs_enter_graphics().
• Misc - memcpy
  If linesize == video_linesize, call just one memcpy.
  On i7-7700, This improve time of copying 1920x1080 frame data from 1.122 ms to 0.607 ms