memcopy from image buffer is very slow DM8168 vc3 demo

BenM

Hello,

For background I am working with the Netra dm8168 EZSDK 5.00 I currently have a project set up based out of the vc3 demo with a custom daughtercard that reads in 12 bit monochrome images from a CMOS imager. In my current implementation I am using OMX to grab pointers to image buffers filled by the M3. I am currently capturing 1920 x 1280 images at 45 fps.

The problem arises when I try to process these images. When I make a copy of the the image buffer it takes 50 - 70ms using memcpy. While if I declare an arbitrary buffer of the same size and copy it, it copies in 3ms. Any idea what would slow the copying of a video port image buffer compared to a non-video port buffer?

Thanks,

Ben

over 14 years ago

0 Iain Hunter over 14 years ago

TI__Genius 16130 points

Ben,

The capture buffers are all allocated from non-cached memory which is why you see poor performance on a memcpy(). This is to avoid cache coherency problems between cores and also reflects the fact that all the components actually use DMA to empry/fill buffers.

When you do a memcpy() on your arbitrary buffer it will have been allocated from the heap and so will be cached.

Iain

0 BenM over 14 years ago in reply to Iain Hunter

Expert 2220 points

Thanks for the info. So is it correct to assume that the only way to do real time processing on images is to work on the image buffers with the DSP and not the ARM?

0 Iain Hunter over 14 years ago in reply to BenM

TI__Genius 16130 points

Ben,

Yes. Optimal way would be to use DMA to bring buffers in in slices to internal RAM in DSP and then process them and use DMA to write back. This of course assumes your algorithm can be sliced.

Iain

Processors

Processors forum

memcopy from image buffer is very slow DM8168 vc3 demo