OpenCV OCL with Epiphany/Parallella

[Kjetil]

I know the development with Epiphany/Parallella is at an early stage, but I’d like to throw a question out here so I can plan accordingly; Can we expect to see the Epiphany/Parallella hardware supported by in the future

[Kjetil]

Noone knows?

[aolofsson]

Sorry for the slow reply. The straight vector usage model of OpenCL is not well suited for the Epiphany architecture due to our memory architecture. Epiphany runs more effectively on moderately sized kernels on chunks of data. We looked into OpenCV OCL briefly and it did not seem like a great fit Would appreciate inputs from others..
Andreas

[dar]

I see two issues that could be problematic. First, image support was not prioritized in the OpenCL baseline implementation available now. Second, the OpenCV OpenCL code will be tuned to GPUs and so I would expect that there might be issues running the code without modification, without tuning for Epiphany processor which is different from a GPU. Both issues could be addressed in the future, but OpenCV is a large project and so it would require some effort to port the code to Epiphany.

[stealthpaladin]

Just to note, I would expect feature detection features similar to OpenCV to be available, as I will be providing an open source package for them around the first of next year. This is one of the primary focuses of my signal-based projects and will likely be one of the first libraries a small team of us develop.

I can't fully ensure what role OpenCL will play right now. It may be implemented largely on FPGA logic with Epiphany crunching the data.
However I'm practically in love with OpenCL so I'd expect some hooks to make it work =)

[Gravis]

i think the better way of going about this is to modify to use Epiphany. though, i'm certain this will take some creative programming if you really want to harness the entire power of the Epiphany's ability to scale to lots of cores.

[harking]

[Gravis]