Increased GPU consumption after migrating to oneVPL dispatcher on Skylake

[AlexTheGreat]

Hi all,

I am fairly stuck on this, so I hope someone's expert advice will help me here. I've migrated a decoder app from Media SDK to oneVPL. Pretty straightforward, I replaced the headers and libs (I am on Windows) and could compile right away. My test computer is Skylake (HD Graphics 530 with latest drivers). If I decode a 30 FPS H264 stream, GPU-Z shows about 15 percent of GPU consumption (exact same GPU consumption as before), which is expected for my high resolution stream.

Now the mystery begins. As a next step I replaced the legacy code (MFXInit) with the recommended oneVPL dispatcher (MFXLoad and MFXCreateSession). Replacing MFXInit with the new way of creating the session, however, increases the GPU consumption to 45 - 50 percent for the very same video stream, no other code changes. Reverting back to MFXInit restores the old significantly lower GPU consumption.

I've not tested other GPUs but since my users still run older architectures like Skylake, it seems I cannot use the new dispatcher code at the moment.

Does anybody have an idea what I might be overlooking? I can provide more info if needed.

Thanks.

Alex

[AlexTheGreat]

For what it's worth, this huge GPU consumption increase happens for H264 only. Testing with H265, the GPU consumption is still the same regardless of how the session is created (MFXInit or MFXLoad and MFXCreateSession).

A.

[AthiraM_Intel]

Hi,

Thank you for posting in Intel Communities.

Could you please share the following details:

1. Did you see any time difference in running the code after migrating to oneVPL? If yes, please share the screenshots.
2. The versions of MediaSDK and oneVPL
3. OS and Hardware details
4. Sample reproducer code and the exact steps.

Thanks

[AlexTheGreat · Accepted solution]

Hi,

yes, I was already working on a small repro for this and realized that my problem was not even reproducible with all H264 video streams. I finally isolated the cause which turned out to be a strange (not a multiple of 16) height of my original stream. I was enforcing these dimensions when allocating the surfaces. Aligning to a multiple of 16 and cropping afterwards restores the original GPU consumption level.

I still don't fully understand the background and how it was affected by the way I create the session, but I can finalize the migration to oneVPL now.

Thanks.

Alex

[AthiraM_Intel]

Hi,

Glad to know that your issue is resolved. Could you please let us know whether we can go ahead and close this thread?

Thanks

[AthiraM_Intel]

Hi,

We have not heard back from you. Can we go ahead and close this thread?

Thanks

[AthiraM_Intel]

Hi,

We have not heard back from you. This thread will no longer be monitored by Intel. If you need further assistance, please post a new question.

Thanks