Re: [alsa-devel] mx31 snd and mc13783 codec status
On Thu, Apr 01, 2010 at 06:01:38PM +0200, Valentin Longchamp wrote:
Hi Sascha,
Sascha Hauer wrote:
Find the latest version of my code here:
The following changes since commit 01e77706cdde7c0b47e5ca1f4284a795504c7c40: Linus Torvalds (1): Merge branch 'for-linus' of git://gitorious.org/linux-omap-dss2/linux
are available in the git repository at:
git://git.pengutronix.de/git/sha/linux-2.6.git mc13783
Sascha Hauer (3): add a mc13783 codec driver add phycore-mc13783 sound support pcm038: add sound support
arch/arm/mach-mx2/mach-pcm038.c | 23 ++- sound/soc/codecs/Kconfig | 4 + sound/soc/codecs/Makefile | 2 + sound/soc/codecs/mc13783.c | 727 +++++++++++++++++++++++++++++++++++++++ sound/soc/codecs/mc13783.h | 32 ++ sound/soc/imx/Kconfig | 9 + sound/soc/imx/Makefile | 3 + sound/soc/imx/phycore-mc13783.c | 160 +++++++++ 8 files changed, 959 insertions(+), 1 deletions(-) create mode 100644 sound/soc/codecs/mc13783.c create mode 100644 sound/soc/codecs/mc13783.h create mode 100644 sound/soc/imx/phycore-mc13783.c
And do you know if your initial mc13783 codec support coupled with mx31 had some limitations ? Our setup is quite straightforward, we have direct connection from the mx31 to the mc13783 on a single SSI.
Our board uses both SSI channels of the MC13783 which we then mux into one channel in the DAM unit. I don't know how this affects you.
I have struggled with the DAM unit (this thing is an awful bulk of wires) and now I get some sound on the loudspeaker.
The thing is that I get a less than a second sound loop (I use aplay to test, so userspace app should be ok), as if the buffer that the fiq asm interrupt (from ssi_fiq.S) copies to the SSI hardware never was updated.
If I have understood the fiq behaviour correctly, you have a asm fiq interrupt that does copy a larger tx buffer into the SSI hardware. Besides it, you have the imx_ssi_timer_callback that checks when the tx buffer was completely copied. If it is the case, then a new buffer tx buffer is "issued" with the snd_pcm_period_elapsed call (and then snd_pcm_update_hw_ptr0). Is this behaviour correct ?
Yes.
If then it looks like on my system, I have a problem with the snd_pcm_update_hw_ptr0 call.
You could try reverting b4e82b5b785670b68136765059d1afc65c0ae023. Though I have tested it on my boards it may have some implications.
Sascha
On Fri, Apr 02, 2010 at 11:53:58AM +0200, Sascha Hauer wrote:
On Thu, Apr 01, 2010 at 06:01:38PM +0200, Valentin Longchamp wrote:
The thing is that I get a less than a second sound loop (I use aplay to test, so userspace app should be ok), as if the buffer that the fiq asm interrupt (from ssi_fiq.S) copies to the SSI hardware never was updated.
If I have understood the fiq behaviour correctly, you have a asm fiq interrupt that does copy a larger tx buffer into the SSI hardware. Besides it, you have the imx_ssi_timer_callback that checks when the tx buffer was completely copied. If it is the case, then a new buffer tx buffer is "issued" with the snd_pcm_period_elapsed call (and then snd_pcm_update_hw_ptr0). Is this behaviour correct ?
Yes.
What sample rate are you trying to play and what buffer size? In my testing the FIQ was really struggling with most applications at sample rates over ~16kHz since you need each audio period to be long enough to at least fill the interval between timer polls but applications wanted to select buffer sizes that were consumed faster than the timer tick.
Using an application like speaker-test which has minimal overhead and no I/O to worry about helps too.
If then it looks like on my system, I have a problem with the snd_pcm_update_hw_ptr0 call.
You could try reverting b4e82b5b785670b68136765059d1afc65c0ae023. Though I have tested it on my boards it may have some implications.
I was seeing similar behaviour before and after that patch too, the main effect there was to improve the error handling when the FIQ and the timer get out of sync with each other.
Mark Brown wrote:
On Fri, Apr 02, 2010 at 11:53:58AM +0200, Sascha Hauer wrote:
On Thu, Apr 01, 2010 at 06:01:38PM +0200, Valentin Longchamp wrote:
What sample rate are you trying to play and what buffer size? In my testing the FIQ was really struggling with most applications at sample rates over ~16kHz since you need each audio period to be long enough to at least fill the interval between timer polls but applications wanted to select buffer sizes that were consumed faster than the timer tick.
Using an application like speaker-test which has minimal overhead and no I/O to worry about helps too.
OK, I have done some tests with speaker-test and I experienced the same as you did: with rates over 16kHz, the FIQ starts to have some problems.
If then it looks like on my system, I have a problem with the snd_pcm_update_hw_ptr0 call.
You could try reverting b4e82b5b785670b68136765059d1afc65c0ae023. Though I have tested it on my boards it may have some implications.
I was seeing similar behaviour before and after that patch too, the main effect there was to improve the error handling when the FIQ and the timer get out of sync with each other.
Mark is right, the behavior is similar with or without this patch, it does not improve the FIQ problems with rates over 16 kHz.
Val
On Fri, Apr 02, 2010 at 11:42:42AM +0100, Mark Brown wrote:
On Fri, Apr 02, 2010 at 11:53:58AM +0200, Sascha Hauer wrote:
On Thu, Apr 01, 2010 at 06:01:38PM +0200, Valentin Longchamp wrote:
The thing is that I get a less than a second sound loop (I use aplay to test, so userspace app should be ok), as if the buffer that the fiq asm interrupt (from ssi_fiq.S) copies to the SSI hardware never was updated.
If I have understood the fiq behaviour correctly, you have a asm fiq interrupt that does copy a larger tx buffer into the SSI hardware. Besides it, you have the imx_ssi_timer_callback that checks when the tx buffer was completely copied. If it is the case, then a new buffer tx buffer is "issued" with the snd_pcm_period_elapsed call (and then snd_pcm_update_hw_ptr0). Is this behaviour correct ?
Yes.
What sample rate are you trying to play and what buffer size? In my testing the FIQ was really struggling with most applications at sample rates over ~16kHz since you need each audio period to be long enough to at least fill the interval between timer polls but applications wanted to select buffer sizes that were consumed faster than the timer tick.
I just stumbled upon a board which had something in /etc/asound.conf which decreased the buffer sizes. The result was choppy sound and a cpu utilisation of ~40%.
After deleting the file I could play sounds (and record simultaniously) with rates up to 44100kHz without visible cpu utilisation.
Looking at it I realised that poll_time is 0, so the timer gets reloaded with the actual jiffies value which of course is a bad idea. We should probably use a hrtimer here.
Sascha
On Wed, Apr 07, 2010 at 01:03:50PM +0200, Sascha Hauer wrote:
On Fri, Apr 02, 2010 at 11:42:42AM +0100, Mark Brown wrote:
What sample rate are you trying to play and what buffer size? In my testing the FIQ was really struggling with most applications at sample rates over ~16kHz since you need each audio period to be long enough to at least fill the interval between timer polls but applications wanted to select buffer sizes that were consumed faster than the timer tick.
After deleting the file I could play sounds (and record simultaniously) with rates up to 44100kHz without visible cpu utilisation.
Yes, if you tweak the buffer sizes (most of the ALSA test utilities support this) then you can use higher sample rates - pulseaudio's high latency modes should work great, for example.
Looking at it I realised that poll_time is 0, so the timer gets reloaded with the actual jiffies value which of course is a bad idea. We should probably use a hrtimer here.
Yes, by default you're only going to be able to poll at 1ms intervals so the poll time ends up going down to zero. I did look at switching over to a hrtimer when I was fixing up the sample rate slips but I wasn't able to convince myself that there was a hrtimer offering sufficient resolution available on i.MX.
participants (3)
-
Mark Brown -
Sascha Hauer -
Valentin Longchamp