Simultaneous write and read

[kolia]

Hi @ssfrr

How would you recommend playing a sound while simultaneously recording? (Not a duplex stream, I’m not trying to play what is being recorded, I’m trying to play a predefined sound)

Thanks

[ssfrr]

It seems like this is still a duplex stream. Duplex just means input and output are happening simultaneously.

Something like this should work:

using PortAudio

N = 48000*5
output = randn(N) * 0.1
input = PortAudioStream() do str
    @async write(str, output)
    read(str, N)
end

Note that if deterministic latency is important (e.g. if you're measuring it, or measuring impulse responses) then you'll want to pass the synced=true keyword argument to PortAudioStream. The trade-off is that you need to make sure to read the same number of sample as you're writing, or the buffers get hung up.

[mchitre]

@ssfrr is the synced keyword working?

julia> PortAudioStream(1, 1; synced=true) do stream
         write(stream, 0.01*randn(4800))
         read(stream, 0.1s)
       end
ERROR: MethodError: no method matching PortAudioStream(::PortAudio.PortAudioDevice, ::PortAudio.PortAudioDevice, ::Int64, ::Int64; synced=true)
Closest candidates are:
  PortAudioStream(::PortAudio.PortAudioDevice, ::PortAudio.PortAudioDevice, ::Any, ::Any; eltype, samplerate, latency, warn_xruns, recover_xruns)
at /Users/mandar/.julia/packages/PortAudio/JRDxf/src/PortAudio.jl:154 got unsupported keyword argument "synced"

[bramtayl]

Well, the synced = true argument isn't present anymore, but opening duplex streams seems to work fine. Correct me if I'm wrong.

[fhchl]

It seems like the synced kwarg has been removed in code, but not from README.md.

[rob-luke]

Thanks for noticing that detail @fhchl. If it's in the readme then it should work. Or a note should be added to the readme on how to achieve the same functionality now.

@bramtayl do you know why was the synced keyword removed? And how should users ensure that the input and output ringbuffers are kept synchronized now?

Going forward, any time functionality is removed it should involve a code review (where the change is specifically highlighted), and we should be sure to add depreciation warnings and note it in the documentation and release note.

[rob-luke]

Ok I’m still digging in to this, but synced seems to have been removed in 9eb565e January 2020.

[rob-luke]

@ssfrr is there an alternative suggestion on how to achieve deterministic latency with synced removed? Or should we just remove it from the README?

[bramtayl]

I don't know why synced was removed, sorry

[ssfrr]

We lost synced when I switched from the callback implemented in C to using libportaudio's streaming interface. It made the code a lot simpler but we lost some low-level control and are relying on portaudio to manage the io buffers.

Previously the C callback would stop pushing/pulling from the write/read ringbuffers when either one of them was full/empty (respectively). That ensured that the ringbuffer state was always synchronized, but made things trickier to manage because the system would hang up if you tried to read from a stream without simultaneously writing to it, or vice-versa.

In general it was a headache and it seemed like the vast majority of users don't really care about deterministic round-trip latency, so I gave it up. This isn't great if you're using it to try to do impulse response measurements or something, but changes in the round-trip latency should only happen on under/overflow, so your measurement would be hosed anyways.

[rob-luke]

Thanks @ssfrr. I'll clean it from the readme then and close this issue after. Cheers