Hi Pierre,
On 1/4/2023 3:19 PM, Pierre-Louis Bossart wrote:
On 12/23/22 17:31, Wesley Cheng wrote:
Several Qualcomm based chipsets can support USB audio offloading to a
dedicated audio DSP, which can take over issuing transfers to the USB
host controller. The intention is to reduce the load on the main
processors in the SoC, and allow them to be placed into lower power modes.
It would be nice to clarify what you want to offload
a) audio data transfers for isoc ports
b) control for e.g. volume settings (those go to endpoint 0 IIRC)
c) Both?
Thanks for sharing your experience, and inputs!
It would be the audio related endpoints only, so ISOC and potentially
feedback ep.
This has a lot of implications on the design. ASoC/DPCM is mainly
intended for audio data transfers, control is a separate problem with
configurations handled with register settings or bus-specific commands.
Control would still be handled by the main processor.
There are several parts to this design:
1. Adding ASoC binding layer
2. Create a USB backend for Q6DSP
3. Introduce XHCI interrupter support
4. Create vendor ops for the USB SND driver
Adding ASoC binding layer:
soc-usb: Intention is to treat a USB port similar to a headphone jack.
The port is always present on the device, but cable/pin status can be
enabled/disabled. Expose mechanisms for USB backend ASoC drivers to
communicate with USB SND.
Create a USB backend for Q6DSP:
q6usb: Basic backend driver that will be responsible for maintaining the
resources needed to initiate a playback stream using the Q6DSP. Will
be the entity that checks to make sure the connected USB audio device
supports the requested PCM format. If it does not, the PCM open call will
fail, and userpsace ALSA can take action accordingly.
Introduce XHCI interrupter support:
XHCI HCD supports multiple interrupters, which allows for events to be routed
to different event rings. This is determined by "Interrupter Target" field
specified in Section "6.4.1.1 Normal TRB" of the XHCI specification.
Events in the offloading case will be routed to an event ring that is assigned
to the audio DSP.
Create vendor ops for the USB SND driver:
qc_audio_offload: This particular driver has several components associated
with it:
- QMI stream request handler
- XHCI interrupter and resource management
- audio DSP memory management
When the audio DSP wants to enable a playback stream, the request is first
received by the ASoC platform sound card. Depending on the selected route,
ASoC will bring up the individual DAIs in the path. The Q6USB backend DAI
will send an AFE port start command (with enabling the USB playback path), and
the audio DSP will handle the request accordingly.
Part of the AFE USB port start handling will have an exchange of control
messages using the QMI protocol. The qc_audio_offload driver will populate the
buffer information:
- Event ring base address
- EP transfer ring base address
and pass it along to the audio DSP. All endpoint management will now be handed
over to the DSP, and the main processor is not involved in transfers.
Overall, implementing this feature will still expose separate sound card and PCM
devices for both the platorm card and USB audio device:
0 [SM8250MTPWCD938]: sm8250 - SM8250-MTP-WCD9380-WSA8810-VA-D
SM8250-MTP-WCD9380-WSA8810-VA-DMIC
1 [Audio ]: USB-Audio - USB Audio
Generic USB Audio at usb-xhci-hcd.1.auto-1.4, high speed
This is to ensure that userspace ALSA entities can decide which route to take
when executing the audio playback. In the above, if card#1 is selected, then
USB audio data will take the legacy path over the USB PCM drivers, etc...
You would still need some sort of mutual exclusion to make sure the isoc
endpoints are not used concurrently by the two cards. Relying on
userspace intelligence to enforce that exclusion is not safe IMHO.
Sure, I think we can make the USB card as being used if the offloading
path is currently being enabled. Kernel could return an error to
userspace when this situation happens.
Intel looked at this sort of offload support a while ago and our
directions were very different - for a variety of reasons USB offload is
enabled on Windows platforms but remains a TODO for Linux. Rather than
having two cards, you could have a single card and addition subdevices
that expose the paths through the DSP. The benefits were that there was
a single set of controls that userspace needed to know about, and volume
settings were the same no matter which path you used (legacy or
DSP-optimized paths). That's consistent with the directions to use 'Deep
Buffer' PCM paths for local playback, it's the same idea of reducing
power consumption with optimized routing.
Volume control would still be done through the legacy path as mentioned
above. For example, if a USB headset w/ a HID interface exposed (for
volume control) was connected, those HID events would be routed to
userspace to adjust volume accordingly on the main processor. (although
you're right about having separate controls still present - one for the
ASoC card and another for USB card)
Another point is that there may be cases where the DSP paths are not
available if the DSP memory and MCPS budget is exceeded. In those cases,
the DSP parts needs the ability to notify userspace that the legacy path
should be used.
If we ran into this scenario, the audio DSP AFE port start command would
fail, and this would be propagated to the userspace entity. It could
then potentially re-route to the legacy/non-offload path.
Another case to handle is that some USB devices can handle way more data
than DSPs can chew, for example Pro audio boxes that can deal with 8ch
192kHz will typically use the legacy paths. Some also handle specific
formats such as DSD over PCM. So it's quite likely that PCM devices for
card0 and card1 above do NOT expose support for the same formats, or put
differently that only a subset of the USB device capabilities are
handled through the DSP.
Same as the above. We have programmed the USB backend to support the
profiles that the audio DSP can handle. I assume if there was any other
request, the userspace entity would fail the PCM open for that requested
profile.
And last, power optimizations with DSPs typically come from additional
latency helping put the SoC in low-power modes. That's not necessarily
ideal for all usages, e.g. for music recording and mixing I am not
convinced the DSP path would help at all.
That's true. At the same time, this feature is more for power related
benefits, not specifically for performance. (although we haven't seen
any performance related issues w/ this approach on the audio profiles
the DSP supports) I think if its an audio profile that supports a high
sample rate and large number of channels, then the DSP wouldn't be able
to support it anyway, and userspace could still use the legacy path.
This would allow for those high-performance audio devices to not be
affected.
Thanks
Wesley Cheng
