Re: Frame rate performance (was: firewire video on RTEMS-4.6.x-MVME5500)

[Steve Shoaf <shoaf@aps.anl.gov>]

Emmanuel,

On Oct 10, 2007, at 1:42 PM, Emmanuel Mayssat wrote:

> Hello all,
>
> I tried frame grabber, etc. recently decided to move to an AXIS video
> server. The asyn/epics driver is in development. We are reading the
> mjpeg video stream (using the GPL code from axmjpeg). The  
> throughput of
> the video stream on the network can be throttled on the server  
> side. We
> use interrupt to trigger frame processing.
>
> We completed performance testing. We can achieve the 30 frames per
> second benchmark. I was shocked with the achieved performance. This
> solution rocks! I gave myself a pat on the back just for thinking  
> of it!
> Load on cpu < 5%, load on local area network... negligible.
>
> What takes time is the processing of each frame (background  
> subtraction,
> ROI, intensity monitor, etc.), so this part needs to be optimized.  
> This
> is actively being developed/reviewed.

We had a similar situation at the APS when we were upgrading our  
analog based frame grabber system.  The amount of time to perform  
simple operations was adding up, and we wanted the system to perform  
all of the analysis operations in real time.  I started using vector  
processing on the image data and saw substantial speed improvements.   
For instance, the background subtraction on an 8-bit grayscale vga  
size image took ~4ms using a simple for-next loop.  The vector based  
subtraction completes in under 1 ms.  I presented a poster on this at  
PAC over the summer and can email you a copy of the paper if you want.

Steve

> --
> Emmanuel
>
> > > Here at the CLS we're using linux-based soft IOCs to
> > > communicate with Flea and Flea2 cameras.
> > > We're having difficulty in achieving video frame rates
> > > greater than about 10 Hz at 640x480 resolution. I'm
> > > interested to know how you are getting such fine performance
> > > (aside from using RTEMS :-)