Composable Overlay vis-a-vis Gstreamer and VVAS

Support
1

What are the trade-off’s of using PYNQ Composable overlay versus Gstreamer or Vitis Video Analytics SDK (VVAS) for preprocessing a video pipeline feeding to a model in PYNQ DPU.

Rahul

2

Hi @Rahuldee,

Can you provide more background? What do you want to do? Are you familiar with any of these two?

Mario

3

Hi Mario,

I understand Gstreamer is used to compose a CPU based software pipeline (using gst-launch-1.0) where one can define source, sink and intermediate nodes ( Example:gst-launch-1.0 filesrc location=music.mp3 ! mad ! audioconvert !``audioresample ! osssink)

VVAS, I understand, achieves functionality similar to Gstreamer using FPGA fabric (PL section).

Composable pipeline (though not exactly gstreamer) is helpful in creating a data pipeline using static and dynamic IP. using the PL section. There are certain IP’s available now and I think more or custom blocks will be added in the future.

Now, if I wish to tap off from the Composable pipeline to feed data to DPU I will have to switch between two different overlays. I am not sure if this is feasible or a good idea. Gstreamer allows me to do this at a lower speed (since it uses PS).

Please let me know your thoughts.

Regards,
Rahul

4

Hi @Rahuldee,

You are correct, I believe that VVAS could also allow you to accelerate certain function on the PL.

Now, if I wish to tap off from the Composable pipeline to feed data to DPU I will have to switch between two different overlays

This is feasible, but the overhead of switching overlays will erode any speed up you get from accelerating functions on the PL.

If you want to combine composable + DPU, the best alternative is to put the two things in the same overlay. This is something that would require some work on the hardware side.

What board are you targeting?

Mario

5

Hi Mario,

I am working on the Kria Starter kit with Ubuntu 20.04 and Kria-PYNQ installation.

6

Hi,

I am not sure if you will have enough resources to combine DPU + composable on that device. You could reduce the size of the composable.
We do not have any design like that at the moment, so I cannot comment how easy/hard will be to accomplish it.

Mario