You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[BUG] Remove HiFi code duplication in HiFi buffer management #5273
Turns out github links don't point to resolved comments properly. This bug is track the deduplication of *_setup_circular(source) functions in HiFi sections.
I've noticed that audio_stream_bytes_without_wrap() performs better in many cases than using same circular start-end registers for several circular buffers in the hot loop. I will test the most important components with such change and change code if my test devices show processing time improvement.
@kv2019i@cujomalainey Please keep 2.6 as target even if this would be a small change. We have other bugs fixing going on, better to not risk those for 2.5.
I have test several method to remove that duplication in TDFB HiFi part, but unfortunately, I found there is only about 0.2MCPS improvement but it needs about 19+ MCPS. So I did not create a PR for TDFB since such improvement seems meaningless. So I think we can close this issue now.
Activity
cujomalainey commentedon Jan 26, 2022
Turns out github links don't point to resolved comments properly. This bug is track the deduplication of *_setup_circular(source) functions in HiFi sections.
lgirdwood commentedon Jan 27, 2022
@cujomalainey fyi @singalsu is creating a spreadsheet for all the SIMD optimizations/cleanups. Will be able to share it soon.
singalsu commentedon Feb 1, 2022
I've noticed that audio_stream_bytes_without_wrap() performs better in many cases than using same circular start-end registers for several circular buffers in the hot loop. I will test the most important components with such change and change code if my test devices show processing time improvement.
lgirdwood commentedon Feb 16, 2022
50% complete, should be fully complete for v2.2
kv2019i commentedon Feb 17, 2023
@singalsu @andrula-song Can you update on status? I put milestone tentatively to v2.6, but we can pull in to v2.5 if we are ready.
cujomalainey commentedon Feb 21, 2023
@singalsu any chance you can share that doc?
1 remaining item
singalsu commentedon Mar 3, 2023
@kv2019i @cujomalainey Please keep 2.6 as target even if this would be a small change. We have other bugs fixing going on, better to not risk those for 2.5.
lgirdwood commentedon May 10, 2023
@singalsu @andrula-song now fixed ?
andrula-song commentedon May 11, 2023
PR #7499 is under review.
alex-cri commentedon May 25, 2023
@andrula-song, @singalsu, with PR #7499 merged can we close this?
andrula-song commentedon May 26, 2023
I have test several method to remove that duplication in TDFB HiFi part, but unfortunately, I found there is only about 0.2MCPS improvement but it needs about 19+ MCPS. So I did not create a PR for TDFB since such improvement seems meaningless. So I think we can close this issue now.
cujomalainey commentedon May 26, 2023
Can you share the data in a draft PR for us to see the diff, MCPS isn't the only metric of importance here
andrula-song commentedon May 29, 2023
here is the draft #7700
alex-cri commentedon May 30, 2023
@cujomalainey, is there further feedback that you feel is relevant for this issue?
lgirdwood commentedon Jun 1, 2023
Move to v2.7 since #7700 wont be ready for v2.6 stable branch fork.
Audio: TDFB: Optimize the circular buffer usage