Working on GPU in a generic way

asked 2019-09-12 07:39:17 -0600

alon3
1

Are there any tools or libraries that belong to opencv that enable GPU work in a generic way?

edit retag flag offensive close merge delete

Comments

No. For that you'd need something like OpenGL 4.3, which has compute shaders. Here is a link to a sample compute shader app:

https://github.com/sjhalayka/qjs_comp...

Note that the code requires the GLUT and GLEW libraries.

sjhalayka ( 2019-09-12 12:00:01 -0600 )edit

Compute shaders were made official in 2012, with OpenGL 4.3. They're not new, and they are widely used.

sjhalayka ( 2019-09-13 09:55:32 -0600 )edit

add a comment

1 answer

Sort by » oldest newest most voted

answered 2019-09-13 04:27:40 -0600

kbarni

4134 ●1 ●13 ●49

There's a way to execute custom OpenCL code on the GPU from OpenCV using the cv::ocl classes and UMat type images.

Here you can find an example for doing this.

A word of warning: OpenCL code can be difficult to develop. It's massively parallel, has no concurrency checks (so it will crash for any bugs) and it's difficult to debug. OTOH OpenCL is very portable, it runs on most GPUs and on CPUs too (CUDA runs only on nVidia hardware).

I personally prefer using CPU parallelisation using TBB, it's much easier to implement.

Concerning CUDA, there is a cv::cuda namespace for CUDA-based operations. It has several image processing algorithms implemented, but I don't know if it's possible to run generic CUDA code.