Own server for Java programms to use the NVidia GPU

Android/Java
1

I use OpenCV under Java, I wanted to ask whether I can also use performance gains there if I compile OpenCV like this.

And the next question: Are all operations then shifted to the GPU, for example

SURF detector = SURF.create(…

and

detector.detectAndCompute(…

?

Are “normal” matrix operations also outsourced to the GPU?

And since the Tesla C2050 is no longer available, what do you buy for it today?

2

unfortunately, no (or, not so easy)
you need to use special api’s / classes to use CUDA/OpenCL code, which are not exposed to the java api

IF you’re good with c++, you can still wrap your gpu pipeline into JNI code, and call that from java

this is cpu only, to use the CUDA optimization, you need a different api – but again, no java bindings ;(

3

I use the official API:

// C++: void cv::Feature2D::detectAndCompute(Mat image, Mat mask, vector_KeyPoint& keypoints, Mat& descriptors, bool useProvidedKeypoints = false)

So I have to write my own JNI so that I can access it via Java and then use these functions instead of the “official” ones?

4

that’s still the CPU API (you need GPUMat instead, look e.g. here:

5

What do I have to consider specifically when I go NonFree?

Switch from
cv::Feature2D Class Reference
[2D Features Framework] » [Feature Detection and Description]

To:

cv::cuda::SURF_CUDA Class Reference
[Extra 2D Features Framework] » [Non-free 2D Features Algorithms]

6

you have to enable the nonfree code at build time of the libs:

cmake -DOPENCV_ENABLE_NONFREE=ON

(and ofc. need to build with the contrib repo, where all CUDA code was moved)

7

Thanks for the tip, I will try it out in the next few weeks.

8

Ok, technically I know, I mean legally, if I don’t publish parts of my software but I want to use these non-public routines myself?