I think I know what this is. But I spent the best part of my breakfast wondering about it. I think using Pulse Audio might work around this problem, but I think it's a driver, configuration issue on the ALSA side or a misunderstanding of the API on my part, which is why I'm sending this to the ALSA list.
So here's the problem: Swfdec uses ALSA in non-blocking mode and polls its file descriptor(s). Whenever the audio queue becomes empty enough that it's useful to render more audio, the file descriptor wakes up from the poll.
This setting is apparently configurable using various hw params, however, Swfdec does not configure it and has been relying on the fact that the sound device picks proper defaults, thinking that a sound card with a small buffer will likely want to wake up more often than something like Pulse Audio's new design, which keeps a 2+ second buffer for the audio.
It seems though that your sound card wakes up roughly whenever a new sample can be rendered, which is 44100 times per second (or more, as from your profile it looks like ALSA resamples the audio). This causes the whole rendering code inside Swfdec to run for every sample, which is so much overhead that it kills all your performance.
My question to you ALSA people now is: Who's responsible for ensuring that the number of wakeups is sane? Should I do some configuring of the audio device and if so, what kind of configuration, or should ALSA be able to pick reasonable defaults for me?
Cheers, Benjamin
On Sun, Dec 7, 2008 at 2:46 AM, Jesse Zbikowski embeddedlinuxguy@gmail.com wrote:
Hi Riccardo,
On Fri, Dec 5, 2008 at 2:32 PM, Riccardo Magliocchetti riccardo.magliocchetti@gmail.com wrote:
Give it a run with oprofile, this is my opreport
Thanks for this data. It looks like I may have an audio bottleneck. opreport -t3 shows
31.5811 no-vmlinux 14.4701 libswfdec-0.9.so.3.0.0 8.3210 libasound_module_rate_speexrate.so 6.0088 libc-2.8.90.so 4.7060 libpixman-1.so.0.12.0 3.6615 python2.5 3.2831 libcairo.so.2.10800.0
If I just look at the libraries I see the top libswfdec routines are all audio related:
14.1497 libswfdec-0.9.so.3.0.0 swfdec_sound_buffer_get_n_samples 2.1666 libswfdec-0.9.so.3.0.0 swfdec_audio_format_get_channels 1.9369 libswfdec-0.9.so.3.0.0 swfdec_audio_format_new 1.7904 libswfdec-0.9.so.3.0.0 swfdec_audio_flv_render 1.3205 libswfdec-0.9.so.3.0.0 __i686.get_pc_thunk.bx 1.2073 libswfdec-0.9.so.3.0.0 swfdec_audio_format_get_granularity
I could try using the PulseAudio backend instead of ALSA.
You can try to improve your performance by using latest stable version of cairo and pixman
OK, I am currently using the latest Ubuntu versions which are a bit behind latest stable. I will see if this makes a difference.
Package Ibex Latest xorg-xserver 1.5.2 1.5.3 cairo 1.8.0 1.8.4 pixman 0.12 0.13.2
play a bit with your X video driver config ie. exa vs xaa.
I did not see libexa.so in my oprofile output although it is installed, so I will make sure this is enabled.
Thanks again, Jesse Z. _______________________________________________ Swfdec mailing list Swfdec@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/swfdec