Audio: Volume: Optimize read frag usage from zero crossing search - #5328
Conversation
|
Here's a plot of execution times while ramping volume. To see better the impact this topology was changed to use 800 ms long ramp with zero-crossings detect. The last 200 ms in 1000 ms long capture is lower load due to completed ramp. Upper row is original and lower row is optimized.
|
|
Something wrong -- I'm getting different zero crossings locations (debugged with testbench printtf()) with optimized 32 bit, finding out why ... |
This patch optimizes the buffer seek for zero crossing. The audio_stream_read_frag_s16/24/32() is replaced with block based search. The search block is defined by new stream function audio_stream_rewind_bytes_without_wrap(). The other added stream function audio_stream_rewind_wrap() handles the pointer wrap when reading backwards. The saving during volume ramping is from 12.4 MCPS to 10.8 MCPS (1.6 MCPS) for 16 bit play, and 16.4 MCPS to 14.4 MCPS (2 MCPS) for 24 bit and 32 bit playback. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>
9a35400 to
4ea90b7
Compare
|
I think there's a mistake in zero-crossing search in the original code. This patch is not changing the functionality, just speeds it. The prev_sum is not updated when crossing is found so I think it misses every other. Also I think in start the prev_sum should be set to sum of last frame, then seek backwards and update sum and compare. If issue is confirmed I would do other PR. |
|
@singalsu any update - ready to merge now ? |
This is OK for me. I didn't change zero crossing search algorithm, this performs identically to current version but is faster. |
2 participants
This patch optimizes the buffer seek for zero crossing. The
audio_stream_read_frag_s16/24/32() is replaced with block
based search. The search block is defined by new stream function
audio_stream_rewind_bytes_without_wrap(). The other added stream
function audio_stream_rewind_wrap() handles the pointer wrap
when reading backwards.
The saving during volume ramping is from 12.4 MCPS to 10.8 MCPS
(1.6 MCPS) for 16 bit play, and 16.4 MCPS to 14.4 MCPS (2 MCPS)
for 24 bit and 32 bit playback.
Signed-off-by: Seppo Ingalsuo seppo.ingalsuo@linux.intel.com