On Fri 30 Apr 2021 at 16:55, Pavel Hofman <pavel.hofman@xxxxxxxxxxx> wrote: > Dne 30. 04. 21 v 16:26 Jerome Brunet napsal(a): >> From: Ruslan Bilovol <ruslan.bilovol@xxxxxxxxx> >> >> As per USB and UAC2 specs, asynchronous audio sink endpoint >> requires explicit synchronization mechanism (Isochronous >> Feedback Endpoint) >> >> Implement feedback companion endpoint for ISO OUT endpoint >> >> This patch adds all required infrastructure and USB requests >> handling for feedback endpoint. Syncrhonization itself is >> still dummy (feedback ep always reports 'nomimal frequency' >> e.g. no adjustement is needed). This satisfies hosts that >> require feedback endpoint (like Win10) and poll it periodically >> >> Actual synchronization mechanism should be implemented >> separately >> >> Signed-off-by: Ruslan Bilovol <ruslan.bilovol@xxxxxxxxx> >> Signed-off-by: Jerome Brunet <jbrunet@xxxxxxxxxxxx> > > Hi, > > The HS calculation of Q16.16 feedback value > overflows at some 524kHz, disallowing use of larger samplerates (e.g. > 768kHz or higher). > > I tested the formula used in alsa USB driver > https://github.com/torvalds/linux/blob/d99676af540c2dc829999928fb81c58c80a1dce4/sound/usb/endpoint.c#L80 > which uses only 10bit shift. The feedback control in UAC2 gadget now > works up to 4M samplerate with 1Hz precision (tested on RPi4 with > bInterval = 1, checked in stream0 proc file on linux host). > > --- a/drivers/usb/gadget/function/u_audio.c > +++ b/drivers/usb/gadget/function/u_audio.c > @@ -118,7 +119,8 @@ static void u_audio_set_fback_frequency(enum > usb_device_speed speed, > * Prevent integer overflow by calculating in Q12.13 > format and > * then shifting to Q16.16 > */ > - ff = DIV_ROUND_UP((freq << 13), (8*1000)) << 3; > + ff = ((freq << 10) + 62) / 125; Pavel, The code posted is a little different from snip here. While I understand the "<< 10" and "/ 125", the "+ 62" would welcome a comment. Also in the final patch, the calculation is a bit different and moved to "long long" ... but I'm sure the same type of improvement could be done. > } > *(__le32 *)buf = cpu_to_le32(ff); > } > > > Best regards, > > Pavel.
