Audio: Volume: Replace read/write frag buffers in generic C version #5300
Conversation
This patch replaces the non efficient per sample fragment functions with audio_stream_bytes_without_wrap() aided block processing. The volume component MCPS is optimized from 24 MCPS to 8 MCPS. The saving is similar in all s16, s24, s32 volume copy() functions. The test was done in TGL-H platform with HiFi3 code version disabled and with xt-xcc compiler. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>
|
Here's plot of copy() execution time vs. playback time for a 30s long clip. Surprisingly all the volume handlers have quite similar load. (I checked with additional traces that in sof-hda-generic-2ch.tplg the speaker/headset PCM uses in volume component the aplay input format, that's good for this kind of formats impact testing. The DAI converts the pipeline format to DAI)
|
|
@singalsu which plots are before and after ? What's the speedup ? |
singalsu
requested review from
dbaluta,
lbetlej,
mmaka1 and
plbossart
as code owners
The upper row is optimized s16/s24/s32, the bottom row is same with git main. Optimized is average 19 us and original 59-60 us per copy, so the speed up is 3.1x. or 32% of previous load. |
|
rerun CI as CML DUT was missing. |
|
SOFCI TEST |
|
CI showing known "sof-logger was already dead" on some Zephyr tests. Unrelated. |
2 participants
This patch replaces the non efficient per sample fragment functions
with audio_stream_bytes_without_wrap() aided block processing.
The volume component MCPS is optimized from 24 MCPS to 8 MCPS. The
saving is similar in all s16, s24, s32 volume copy() functions.
The test was done in TGL-H platform with HiFi3 code version disabled
and with xt-xcc compiler.
Signed-off-by: Seppo Ingalsuo seppo.ingalsuo@linux.intel.com