[alsa-devel] [PATCH] speaker-test: Fix dropped samples at the end of test

[Jarkko Nikula]

On 09/15/2014 06:23 PM, Takashi Iwai wrote:
> At Fri, 12 Sep 2014 15:14:28 +0300,
> Jarkko Nikula wrote:
>> Commit 6d1673526b0f ("Avoid unnecessary drain/restart in speaker-test")
>> drains only when buffer is bigger than audio sample. This has a drawback
>> that up to buffer size amount of data may not be heard at the end of audio
>> sample.
>>
>> This was noted with "speaker-test -c 2 -t wav -s 2" test on a
>> hardware that has a buffer size of 24000 samples and 48 kHz sample rate.
>> Instead of playing "front right" it played something like "front ra".
>>
>> Reverse buffer size vs sample size test wouldn't work either since then
>> samples smaller than buffer are dropped.
>>
>> Fix this by removing buffer_size tests from write_loop() and do
>> drain/restart always when not aborting.
>>
>> Signed-off-by: Jarkko Nikula <jarkko.nikula at linux.intel.com>
>> Reported-by: Vidal, Guillaume-florianX <guillaume-florianx.vidal at intel.com>
>> ---
>> This was originally noted on Baytrail ADSP hw (default buffer size 24000)
>> but can be heard also on Intel HDA (default buffer size 8192) when audio
>> sample is small enough but bigger than buffer. For instance 100 ms sample
>> finishes too shortly (buffer size 8192, sample size 9600) but 50 ms plays
>> ok (buffer size 8192, sample size 4800).
>>
>> I guess some optimization can be done for snd_pcm_prepare() when not looping
>> but that's not necessary for this fix.
> Won't it suffice by just putting snd_pcm_drain() at the end of the
> whole operation like below?  Doing snd_pcm_drain() and
> snd_pcm_prepare() at each time causes often undesired pop noises or
> such.
Yeah, that works too and pop avoidance is a good argument.
> OTOH, doing drain there is good for showing the text at the right
> time.  So, we'll likely want to have both options managed by a command
> line option.
>
I see slight difference between text print and sample starts time on 
Baytrail ADSP when doing loop test with your patch. Perhaps about half 
second since buffer size is 24000 and SR 48 kHZ but I don't think that 
matters much since it's hardly noticeable.

Tested-by: Jarkko Nikula <jarkko.nikula at linux.intel.com>