I'm porting an image post-processing library to OpenCV, and testing the UMat OpenCL acceleration features.
First tests work quite well, with complex filters using lots of arithmetic and math operations, I have a consistent speed increase (such as x2 to x4) with setUseOpenCL(true) vs setUseOpenCL(false). So far so good.
However, cv::gaussianBlur suspiciously offers the exact same performance, such as a 1500ms time for a 4096x4096 image and a 321x321 pixels kernel, with and without OpenCL.
So I suspect it's not using OpenCL in my case, but why?