Nouveau - Nov 2013 - H.264 engine differences between fermi and tesla cards

On 11/19/2013 08:16 PM, Ilia Mirkin wrote:
> Hello,
> 
> I hope this is an appropriate style of request for this forum. I added
> code to support video decoding on the tesla cards that have a
> similar-style video decoding engine to fermi cards (i.e. G98, GT21x,
> the IGP's -- the falcon-controlled decoding engines, rather than the
> xtensa-controlled ones), by using pretty much the same logic that we
> had for the fermi cards. This worked great for MPEG-2 and VC-1.
> However for H.264 videos, it appears to decode a few frames, and then
> the engine hangs.
> 
> In traces, I noticed that the nvidia driver reloads the BSP/VP/PPP
> engines every second or so. Is this done as a powersaving technique,
> or is it done as a workaround for some issue? Does nouveau need to do
> the same thing? If so, any specifics on the reload condition?
> 
> Any other ideas as to what might be going wrong? Are there some subtle
> differences between the fermi and pre-fermi engines? Or a difference
> when decoding H.264 files vs MPEG2/VC1 files? Perhaps there's other
> information I can provide. BTW, this is with using the firmware blobs
> from the NVIDIA proprietary driver.
> 
> Thanks,
> 
>   -ilia
As you observed, the nvidia driver unloads the video engines on certain GPUs
when they go idle to save power.  You can disable this behavior by loading the
nvidia kernel module with: modprobe nvidia
NVreg_RegistryDwords="RMPowerFeature=64"

Regarding your H.264 hangs, the most likely cause is mis-programming the video
engine.  I suggest double-checking that the nouveau driver sends the exact same
parameters for each decode operation as the nvidia driver does.  In particular,
check that buffer alignments match up, as those may vary between GPU
generations.

Thanks,
Ben

On Thu, Nov 21, 2013 at 5:07 PM, Benjamin Morris <bmorris at nvidia.com>
wrote:
> On 11/19/2013 08:16 PM, Ilia Mirkin wrote:
>> Hello,
>>
>> I hope this is an appropriate style of request for this forum. I added
>> code to support video decoding on the tesla cards that have a
>> similar-style video decoding engine to fermi cards (i.e. G98, GT21x,
>> the IGP's -- the falcon-controlled decoding engines, rather than
the
>> xtensa-controlled ones), by using pretty much the same logic that we
>> had for the fermi cards. This worked great for MPEG-2 and VC-1.
>> However for H.264 videos, it appears to decode a few frames, and then
>> the engine hangs.
>>
>> In traces, I noticed that the nvidia driver reloads the BSP/VP/PPP
>> engines every second or so. Is this done as a powersaving technique,
>> or is it done as a workaround for some issue? Does nouveau need to do
>> the same thing? If so, any specifics on the reload condition?
>>
>> Any other ideas as to what might be going wrong? Are there some subtle
>> differences between the fermi and pre-fermi engines? Or a difference
>> when decoding H.264 files vs MPEG2/VC1 files? Perhaps there's other
>> information I can provide. BTW, this is with using the firmware blobs
>> from the NVIDIA proprietary driver.
>>
>> Thanks,
>>
>>   -ilia
>
> As you observed, the nvidia driver unloads the video engines on certain
GPUs when they go idle to save power.  You can disable this behavior by loading
the nvidia kernel module with: modprobe nvidia
NVreg_RegistryDwords="RMPowerFeature=64"
>
> Regarding your H.264 hangs, the most likely cause is mis-programming the
video engine.  I suggest double-checking that the nouveau driver sends the exact
same parameters for each decode operation as the nvidia driver does.  In
particular, check that buffer alignments match up, as those may vary between GPU
generations.
Thanks a lot for the response! I've set aside some time this weekend
to debug this some more, I'll be sure to pay special attention to how
we're computing the various buffer sizes and their alignments.

  -ilia

On Thu, Nov 21, 2013 at 5:22 PM, Ilia Mirkin <imirkin at alum.mit.edu>
wrote:
> On Thu, Nov 21, 2013 at 5:07 PM, Benjamin Morris <bmorris at
nvidia.com> wrote:
>> On 11/19/2013 08:16 PM, Ilia Mirkin wrote:
>>> Hello,
>>>
>>> I hope this is an appropriate style of request for this forum. I
added
>>> code to support video decoding on the tesla cards that have a
>>> similar-style video decoding engine to fermi cards (i.e. G98,
GT21x,
>>> the IGP's -- the falcon-controlled decoding engines, rather
than the
>>> xtensa-controlled ones), by using pretty much the same logic that
we
>>> had for the fermi cards. This worked great for MPEG-2 and VC-1.
>>> However for H.264 videos, it appears to decode a few frames, and
then
>>> the engine hangs.
>>>
>>> In traces, I noticed that the nvidia driver reloads the BSP/VP/PPP
>>> engines every second or so. Is this done as a powersaving
technique,
>>> or is it done as a workaround for some issue? Does nouveau need to
do
>>> the same thing? If so, any specifics on the reload condition?
>>>
>>> Any other ideas as to what might be going wrong? Are there some
subtle
>>> differences between the fermi and pre-fermi engines? Or a
difference
>>> when decoding H.264 files vs MPEG2/VC1 files? Perhaps there's
other
>>> information I can provide. BTW, this is with using the firmware
blobs
>>> from the NVIDIA proprietary driver.
>>>
>>> Thanks,
>>>
>>>   -ilia
>>
>> As you observed, the nvidia driver unloads the video engines on certain
GPUs when they go idle to save power.  You can disable this behavior by loading
the nvidia kernel module with: modprobe nvidia
NVreg_RegistryDwords="RMPowerFeature=64"
>>
>> Regarding your H.264 hangs, the most likely cause is mis-programming
the video engine.  I suggest double-checking that the nouveau driver sends the
exact same parameters for each decode operation as the nvidia driver does.  In
particular, check that buffer alignments match up, as those may vary between GPU
generations.
>
> Thanks a lot for the response! I've set aside some time this weekend
> to debug this some more, I'll be sure to pay special attention to how
> we're computing the various buffer sizes and their alignments.
So... I just did some experimenting, and it's not looking good for the
buffer alignment theory. For the same video, with the same nouveau
code, it plays back a variable amount. It normally gets through
150-160 frames before the VP engine hangs. (This isn't a hard hang, it
just never finishes processing the frame.) I've looked at what we send
on the different runs into the FIFO, and it appears to be completely
identical between runs, down to the exact addresses of all the
buffers. So there's some non-determinism in there somewhere.

I analyzed the data being pushed fairly carefully, both by the nvidia
driver and nouveau. I did note some differences, but making
adjustments to the nouveau code just made things worse, it would only
get through 1-50 frames before hanging in the same way. I probably
didn't quite understand something.

I should have asked this directly in my original request, but is there
any chance that NVIDIA could release the ABI docs for its video
playback firmware? I wouldn't need a full-on spec, just enough bits to
get H.264 going (since the rest work just fine already). Specifically
buffer sizing/alignment, and what any "non-obvious" values are in the
parameters passed to BSP/VP/PPP engines. No need to talk about
reference frame management or the crypto stuff.

Thanks,

  -ilia