Audio: pcm_converter: Optimize buffers access in generic version #5019
Conversation
lgirdwood
left a comment
lgirdwood
left a comment
There was a problem hiding this comment.
Whats the performance improvement ?
|
@lgirdwood This version seems to fail CI in every platform before cAVS. I need to test and fix. I'm now setting up a BYT device to measure the execute speed difference. |
singalsu
force-pushed
the
pcm_converter_generic_optimize
branch
from
7a51057 to
2ac1d2a
Compare
|
Here's execution time for s16 to s24 conversion measured in CML platform with gcc build. I could not yet get up my BYT device where this code would be default. The execution time average dropped from 15 us to 13 us that corresponds to 6 MCPS to 5 MCPS. But since we can have potentially several instances of these running the saving can be larger.
|
singalsu
requested review from
dbaluta,
lbetlej,
mmaka1 and
plbossart
as code owners
singalsu
force-pushed
the
pcm_converter_generic_optimize
branch
from
2ac1d2a to
c108e96
Compare
singalsu
changed the title
|
This push contains temporary disable of hifi3 code version to increase CI test coverage for this patch. |
This code is used in non-HiFi3 platforms and in all gcc builds. The inefficient audio_stream_read/write_frag_s16/s32() usage is replaced by block processing with help of function audio_stream_bytes_without_wrap(). The circular buffers pointers checks and wraps are done only when necessary. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>
singalsu
force-pushed
the
pcm_converter_generic_optimize
branch
from
c108e96 to
5f26007
Compare
singalsu
changed the title
|
Can one of the admins verify this patch? |
|
CI shows logger and FW presence error unrelated to this PR. |
4 participants
This code is used in non-HiFi3 platforms and in all gcc builds.
The inefficient audio_stream_read/write_frag_s16/s32() usage is
replaced by block processing with help of function
audio_stream_bytes_without_wrap(). The circular buffers pointers
are checked only in the end or when one of the pointers are at end.
Signed-off-by: Seppo Ingalsuo seppo.ingalsuo@linux.intel.com