Notes on status of CoreAudio Implementation of PortAudio Document Last Updated December 9, 2005 There are currently two implementations of PortAudio for Mac Core Audio. The original is in pa_mac_core_old.c, and the newer, default implementation is in pa_mac_core.c. Only pa_mac_core.c is currently developed and supported as it uses apple's current core audio technology. To select use the old implementation, replace pa_mac_core.c with pa_mac_core_old.c (eg. "cp pa_mac_core_auhal.c pa_mac_core.c"), then run configure and make as usual. ------------------------------------------- Notes on Newer/Default AUHAL implementation: by Bjorn Roche Last Updated December 9, 2005 Principle of Operation: This implementation uses AUHAL for audio I/O. To some extent, it also operates at the "HAL" Layer, though this behavior can be limited by platform specific flags (see pa_mac_core.h for details). The default settings should be reasonable: they don't change the SR of the device and don't cause interruptions if other devices are using the device. Major Software Elements Used: Apple's HAL AUs provide output SR conversion transparently, however, only on output, so this implementation uses AudioConverters to convert the sample rate on input. A PortAudio ring buffer is used to buffer input when sample rate conversion is required or when separate audio units are used for duplex IO. Finally, a PortAudio buffer processor is used to convert formats and provide additional buffers if needed. Internally, interleaved floating point data streams are used exclusively - the audio unit converts from the audio hardware's native format to interleaved float PCM and PortAudio's Buffer processor is used for conversion to user formats. Simplex Input: Simplex input uses a single callback. If sample rate conversion is required, a ring buffer and AudioConverter are used as well. Simplex output: Simplex output uses a single callback. No ring buffer or audio converter is used because AUHAL does its own output SR conversion. Duplex, one device (no SR conversion): When one device is used, a single callback is used. This achieves very low latency. Duplex, separate devices or SR conversion: When SR conversion is required, data must be buffered before it is converted and data is not always available at the same times on input and output, so SR conversion requires the same treatment as separate devices. The input callback reads data and puts it in the ring buffer. The output callback reads the data off the ring buffer, into an audio converter and finally to the buffer processor. Platform Specific Options: By using the flags in pa_mac_core.h, the user may specify several options. For example, the user can specify the sample-rate conversion quality, and the extent to which PA will attempt to "play nice" and to what extent it will interrupt other apps to improve performance. For example, if 44100 Hz sample rate is requested but the device is set at 48000 Hz, PA can either change the device for optimal playback ("Pro" mode), which may interrupt other programs playing back audio, or simple use a sample-rate conversion, which allows for friendlier sharing of the device ("Play Nice" mode). Additionally, the user may define a "channel mapping" by calling paSetupMacCoreChannelMap() on their stream info structure before opening the stream with it. See below for creating a channel map. Known issues: - Buffering: No buffering beyond that provided by core audio is provided except where absolutely needed for the implementation to work. This may cause issues with large framesPerBuffer settings and it also means that no additional latency will be provided even if a large latency setting is selected. - Latency: Latency settings are generally ignored. They may be used as a hint for buffer size in paHostFramesPerBufferUnspecified, or the value may be used in cases where additional buffering is needed, such as doing input and output on separate devices. Latency settings are always automatically bound to "safe" values, however, so setting extreme values here should not be an issue. - Buffer Size: paHostFramesPerBufferUnspecified and specific host buffer sizes are supported. paHostFramesPerBufferUnspecified works best in "pro" mode, where the buffer size and sample rate of the audio device is most likely to match the expected values. In the case of paHostFramesPerBuffer, an appropriate framesPerBuffer value will be used that guarantees minimum requested latency if that's possible. - Timing info. It reports on stream time, but I'm probably doing something wrong since patest_sine_time often reports negative latency numbers. Also, there are currently issues with some devices whehn plugging/unplugging devices. - xrun detection: The only xrun detection performed is when reading and writing the ring buffer. There is probably more that can be done. - abort/stop issues: stopping a stream is always a complete operation, but latency should be low enough to make the lack of a separate abort unnecessary. Apple clarifies its AudioOutputUnitStop() call here: http://lists.apple.com/archives/coreaudio-api/2005/Dec/msg00055.html - blocking interface: should work fine. - multichannel: It has been tested successfully on multichannel hardware from MOTU: traveler and 896HD. Also Presonus firepod and others. It is believed to work with all Core Audio devices, including virtual devices such as soundflower. - sample rate conversion quality: By default, SR conversion is the maximum available. This can be tweaked using flags pa_mac_core.h. Note that the AU render quyality property is used to set the sample rate conversion quality as "documented" here: http://lists.apple.com/archives/coreaudio-api/2004/Jan/msg00141.html - x86/Universal Binary: Universal binaries can be build. Creating a channel map: How to create the map array - Text taken From AUHAL.rtfd : [3] Channel Maps Clients can tell the AUHAL units which channels of the device they are interested in. For example, the client may be processing stereo data, but outputting to a six-channel device. This is done by using the kAudioOutputUnitProperty_ChannelMap property. To use this property: For Output: Create an array of SInt32 that is the size of the number of channels of the device (Get the Format of the AUHAL's output Element == 0) Initialize each of the array's values to -1 (-1 indicates that that channel is NOT to be presented in the conversion.) Next, for each channel of your app's output, set: channelMapArray[deviceOutputChannel] = desiredAppOutputChannel. For example: we have a 6 channel output device and our application has a stereo source it wants to provide to the device. Suppose we want that stereo source to go to the 3rd and 4th channels of the device. The channel map would look like this: { -1, -1, 0, 1, -1, -1 } Where the formats are: Input Element == 0: 2 channels (- client format - settable) Output Element == 0: 6 channels (- device format - NOT settable) So channel 2 (zero-based) of the device will take the first channel of output and channel 3 will take the second channel of output. (This translates to the 3rd and 4th plugs of the 6 output plugs of the device of course!) For Input: Create an array of SInt32 that is the size of the number of channels of the format you require for input. Get (or Set in this case as needed) the AUHAL's output Element == 1. Next, for each channel of input you require, set: channelMapArray[desiredAppInputChannel] = deviceOutputChannel; For example: we have a 6 channel input device from which we wish to receive stereo input from the 3rd and 4th channels. The channel map looks like this: { 2, 3 } Where the formats are: Input Element == 0: 2 channels (- device format - NOT settable) Output Element == 0: 6 channels (- client format - settable) ---------------------------------------- Notes on Original implementation: by Phil Burk and Darren Gibbs Last updated March 20, 2002 WHAT WORKS Output with very low latency, <10 msec. Half duplex input or output. Full duplex on the same CoreAudio device. The paFLoat32, paInt16, paInt8, paUInt8 sample formats. Pa_GetCPULoad() Pa_StreamTime() KNOWN BUGS OR LIMITATIONS We do not yet support simultaneous input and output on different devices. Note that some CoreAudio devices like the Roland UH30 look like one device but are actually two different CoreAudio devices. The Built-In audio is typically one CoreAudio device. Mono doesn't work. DEVICE MAPPING CoreAudio devices can support both input and output. But the sample rates supported may be different. So we have map one or two PortAudio device to each CoreAudio device depending on whether it supports input, output or both. When we query devices, we first get a list of CoreAudio devices. Then we scan the list and add a PortAudio device for each CoreAudio device that supports input. Then we make a scan for output devices.