Improving the ALSA backend

RedDwarf · Sat May 31, 2008 3:36 pm

I have an openSUSE 10.3 x86-64, Core2Duo E6420, ASUS P5B system with two sound cards: a Sound Blaster Live! and a AD1988 integrated codec. ALSA drivers are "1.0.16.20080601" and alsa-lib is "1.0.16.hg20080510" (your learn to love the openSUSE Build Service ;-)

).
With the ALSA backend in bsnes 0.032a the SBLive! crashes and AD1988 has distorsioned sound.
I think the crashes are because


	Quote: snd_pcm_sframes_t written = snd_pcm_writei(device.handle, buffer.data, buffer.length);

should be


	Quote: snd_pcm_sframes_t written = snd_pcm_writei(device.handle, buffer.data, buffer.length/2);

since length is in frames and no in samples.

Also adding a


	Code: snd_output_t *setup; int err = snd_output_stdio_attach(&setup, stdout, 0); if (err < 0) { printf("Ups!\n"); } snd_pcm_dump_setup(device.handle, setup);

to the end of the init() function to see exactly how the sound card has been configured.

The SB Live! is configured this way:


	Quote: stream : PLAYBACK access : RW_INTERLEAVED format : S16_LE subformat : STD channels : 2 rate : 32000 exact rate : 32000 (32000/1) msbits : 16 buffer_size : 64 period_size : 16 period_time : 500 tstamp_mode : NONE period_step : 1 avail_min : 16 start_threshold : 64 stop_threshold : 64 silence_threshold: 0 silence_size : 0 boundary : 4611686018427387904

The AD1988 is configured this way:


	Quote: stream : PLAYBACK access : RW_INTERLEAVED format : S16_LE subformat : STD channels : 2 rate : 32000 exact rate : 32000 (32000/1) msbits : 16 buffer_size : 1365 period_size : 682 period_time : 21333 tstamp_mode : NONE period_step : 1 avail_min : 682 start_threshold : 1364 stop_threshold : 1365 silence_threshold: 0 silence_size : 0 boundary : 3073706745680363520

So even if the code asked for a 90us latency, what we have is a 2ms latency for the SB Live! and a 43ms latency for AD1988.
Now, was 90us really intended? It's very low, isn't? Since I had problems to make the SB Live! work with a 2ms latency without buffer underruns I have changed "device.latency = 90;" for "device.latency = 90000;". If 90us was really intented (what OSS archieves with policy = 4?) say me.

New configuration is:
- SB Live!:


	Quote: stream : PLAYBACK access : RW_INTERLEAVED format : S16_LE subformat : STD channels : 2 rate : 32000 exact rate : 32000 (32000/1) msbits : 16 buffer_size : 2880 period_size : 720 period_time : 22500 tstamp_mode : NONE period_step : 1 avail_min : 720 start_threshold : 2880 stop_threshold : 2880 silence_threshold: 0 silence_size : 0 boundary : 6485183463413514240

Exactly 90ms latency, with 4 periods.

- AD1988


	Quote: stream : PLAYBACK access : RW_INTERLEAVED format : S16_LE subformat : STD channels : 2 rate : 32000 exact rate : 32000 (32000/1) msbits : 16 buffer_size : 2730 period_size : 682 period_time : 21333 tstamp_mode : NONE period_step : 1 avail_min : 682 start_threshold : 2728 stop_threshold : 2730 silence_threshold: 0 silence_size : 0 boundary : 3073706745680363520

~85ms latency, with ~4 periods.

At the end, I have found that this alternative code works fine for both SB Live! and AD1988, speed regulation included.


	Code: #include <alsa/asoundlib.h> #include <ruby/ruby.h> namespace ruby { #include "alsa.h" class pAudioALSA { public: AudioALSA &self; struct { snd_pcm_t handle; snd_pcm_format_t format; int channels; const char name; unsigned latency; } device; struct { uint32_t data; unsigned length; } buffer; struct { unsigned frequency; snd_pcm_uframes_t buffer_size; snd_pcm_uframes_t period_size; } settings; bool cap(Audio::Setting setting) { if(setting == Audio::Frequency) return true; return false; } uintptr_t get(Audio::Setting setting) { if(setting == Audio::Frequency) return settings.frequency; return false; } bool set(Audio::Setting setting, uintptr_t param) { if(setting == Audio::Frequency) { settings.frequency = param; if(device.handle) { term(); init(); } return true; } return false; } void sample(uint16_t left, uint16_t right) { if(!device.handle) return; buffer.data[buffer.length++] = left + (right << 16); if(buffer.length < settings.period_size) return; uint32_t buffer_ptr = buffer.data; do { snd_pcm_sframes_t written = snd_pcm_writei(device.handle, buffer_ptr, buffer.length); if(written < 0) { snd_pcm_recover(device.handle, written, 1); } else if(written < buffer.length) { //only some samples written buffer.length -= written; buffer_ptr += written; } else { //all samples written buffer.length = 0; } } while(buffer.length != 0); } bool init() { if(snd_pcm_open(&device.handle, device.name, SND_PCM_STREAM_PLAYBACK, 0) < 0) { //failed to initialize term(); return false; } if(snd_pcm_set_params(device.handle, device.format, SND_PCM_ACCESS_RW_INTERLEAVED, device.channels, settings.frequency, 1, device.latency) < 0) { //failed to set device parameters term(); return false; } if(snd_pcm_get_params(device.handle, &settings.buffer_size, &settings.period_size) < 0) { //I suppose ALSA always uses 4 periods/buffer, and I expect buffer_size to give me the exact latency requested... better than nothing settings.period_size = device.latency1e-6settings.frequency/4; } buffer.data = new uint32_t[settings.period_size]; return true; } void term() { if(device.handle) { snd_pcm_drain(device.handle); snd_pcm_close(device.handle); device.handle = 0; } if(buffer.data) { delete[] buffer.data; buffer.data = 0; } } pAudioALSA(AudioALSA &self_) : self(self_) { device.handle = 0; device.format = SND_PCM_FORMAT_S16_LE; device.channels = 2; device.name = "default"; device.latency = 100000; buffer.data = 0; buffer.length = 0; settings.frequency = 32000; } ~pAudioALSA() { term(); } }; bool AudioALSA::cap(Setting setting) { return p.cap(setting); } uintptr_t AudioALSA::get(Setting setting) { return p.get(setting); } bool AudioALSA::set(Setting setting, uintptr_t param) { return p.set(setting, param); } void AudioALSA::sample(uint16_t left, uint16_t right) { p.sample(left, right); } bool AudioALSA::init() { return p.init(); } void AudioALSA::term() { p.term(); } AudioALSA::AudioALSA() : p(new pAudioALSA(this)) {} AudioALSA::~AudioALSA() { delete &p; } } //namespace ruby

Patch:


	Code: --- src/lib/ruby/audio/alsa.cpp +++ src/lib/ruby/audio/alsa.cpp @@ -19,13 +19,14 @@ } device; struct { - uint16_t data; + uint32_t data; unsigned length; - unsigned size; } buffer; struct { unsigned frequency; + snd_pcm_uframes_t buffer_size; + snd_pcm_uframes_t period_size; } settings; bool cap(Audio::Setting setting) { @@ -53,30 +54,26 @@ void sample(uint16_t left, uint16_t right) { if(!device.handle) return; - buffer.data[buffer.length++] = left; - buffer.data[buffer.length++] = right; - if(buffer.length + 2 < buffer.size) return; //will crash in some cases if not stopped two before - - snd_pcm_sframes_t written = snd_pcm_writei(device.handle, buffer.data, buffer.length); - if(written < 0) { - snd_pcm_recover(device.handle, written, 1); - //no samples written, drop one sample to prevent possible emulation stall - buffer.length -= 2; - memmove(buffer.data, buffer.data + 2, buffer.length * sizeof(uint16_t)); - } else if(written < buffer.length) { - //only some samples written - buffer.length -= written; - memmove(buffer.data, buffer.data + written, buffer.length * sizeof(uint16_t)); - } else { - //all samples written - buffer.length = 0; - } + buffer.data[buffer.length++] = left + (right << 16); + if(buffer.length < settings.period_size) return; + uint32_t buffer_ptr = buffer.data; + do { + snd_pcm_sframes_t written = snd_pcm_writei(device.handle, buffer_ptr, buffer.length); + if(written < 0) { + snd_pcm_recover(device.handle, written, 1); + } else if(written < buffer.length) { + //only some samples written + buffer.length -= written; + buffer_ptr += written; + } else { + //all samples written + buffer.length = 0; + } + } while(buffer.length != 0); } bool init() { - buffer.data = new uint16_t[buffer.size]; - - if(snd_pcm_open(&device.handle, device.name, SND_PCM_STREAM_PLAYBACK, SND_PCM_NONBLOCK) < 0) { + if(snd_pcm_open(&device.handle, device.name, SND_PCM_STREAM_PLAYBACK, 0) < 0) { //failed to initialize term(); return false; @@ -89,6 +86,13 @@ return false; } + if(snd_pcm_get_params(device.handle, &settings.buffer_size, &settings.period_size) < 0) { + //I suppose ALSA always uses 4 periods/buffer, and I expect buffer_size to give me the exact latency requested... better than nothing + settings.period_size = device.latency1e-6settings.frequency/4; + } + + buffer.data = new uint32_t[settings.period_size]; + return true; } @@ -110,13 +114,12 @@ device.format = SND_PCM_FORMAT_S16_LE; device.channels = 2; device.name = "default"; - device.latency = 90; + device.latency = 100000; buffer.data = 0; buffer.length = 0; - buffer.size = device.latency 32; - settings.frequency = 22050; + settings.frequency = 32000; } ~pAudioALSA() {

Also would be good to add an option to select the device. With SB Live! I have no problems using HQ2x and so, but since the AD1988 doesn't makes hardware mixing the CPU usage is too high when using it with HQ2x. Using "plughw" insted of "default" the CPU usage with AD1988 is lower (even if I loss sound in the rest of the system).

If someone else also has problems try this alternative and say me...

Edit: Sorry for the error in the code. Corrected.
Edit2: Forget the blocking thing.
Edit3: Changed latency value to 100ms and readded application audio buffer to reduce CPU usage.
At this point I think the code is so good like it can be. Application buffer size and latency are the only two numbers that could be tweaked. So I open the poll.

Improving the ALSA backend

Who is online