Intel Begins Working On A Vulkan Compute Back-End For OpenCV Library

[phoronix]

Phoronix: Intel Begins Working On A Vulkan Compute Back-End For OpenCV Library

As perhaps a sign of where Intel is heading for their GPU computing strategy with their in-development discrete GPUs, they are developing a Vulkan compute back-end for the widely-used OpenCV library. This Vulkan back-end is for handling GPU-based compute for neural networks with this Open Computer Vision library as an alternative to the CUDA and OpenCL GPU compute support...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-Vulkan-OpenCV-Backend

[schmidtbag]

This is one of the most interesting uses of Vulkan compute we've seen thus far.

I completely agree. OpenCV is already very CPU intensive even with T-API. The reduced CPU overhead of Vulkan ought to have a distinctly large performance increase.
This is a very welcome addition.

[mannerov]

I'm quite surprised. To me the biggest problem with OpenCV is not the cpu overhead, but a lot of inneficient kernels, and the loss of bandwith caused by not merging small kernels together.

[Guest]

Awesome!

[schmidtbag]

I'm quite surprised. To me the biggest problem with OpenCV is not the cpu overhead, but a lot of inneficient kernels, and the loss of bandwith caused by not merging small kernels together.

I suppose that in and of itself could be a cause for higher CPU overhead, but depending on your workflow, keeping the kernels separate might offer a performance advantage. I'm not sure there's an easy way to address this without re-writing a lot of the library.
That being said, if we take the library for what it is, a common use application for it is robots and IoT devices, which have very limited processing power. OpenCV gets exponentially slower as you increase the resolution, so if such a device needs to handle an image larger than 640x480 (which most devices can handle just fine), anything to help reduce CPU overhead will be very beneficial to such devices.