Unstable result of opencv_core.useOpenCL()

[Daniel-Alievsky]

I detected very strange bug. In my test, I intensively call opencv functions from different threads (I repeat all calculations 150 times in a simple loop). Every time I call opencv_core.useOpenCL() method to decide, whether I should use Mat or UMat version of my function. And... sometimes opencv_core.useOpenCL() returns false!

Of course, usually it returns true on my computer. However, in the same program, moreover, in a single Java loop (!) this method sometimes return false.

How is it possible? Maybe, it is also a problem of insufficient synchronization, like in #650 ? Can you advise me some workaround (probably, I need to call this function only once and store the results in some Java constant)?

I work with JavaCPP 1.4 version (OpenCV 3.4.0-1.4)

[Daniel-Alievsky]

In the latest JavaCPP 1.4.4-SNAPSHOT (OpenCV 4.0.0) there is the same behavior: sometimes useOpenCL() returns false, but usually true,
What does it mean? What exactly means the result of this method?

[saudet]

Sounds like a bug in your video driver. Upgrade to the latest version, but as usual if that doesn't fix it, please report this issue upstream!

[Daniel-Alievsky]

I've updated my driver to the latest version from https://www.nvidia.com/Download/index.aspx My version is GeForce GTX 770M, 417.35 (latest for today). It does not help.

I just added the following line to the start of every my OpenCV function:

        System.out.println("haveOpenCL: " 
            + opencv_core.haveOpenCL() + "; useOpenCL: " + opencv_core.useOpenCL()
            + " - " + getClass());

When I call my functions from single thread, results are always correct:
haveOpenCL: true; useOpenCL: true - class com.siams.stare....

But while calling from different parallel threads, the results are another: sometimes I see
haveOpenCL: true; useOpenCL: false - class com.siams.stare...
And the same behavior in the latest 4.0.0-1.4.4-SNAPSHOT

C++ implementation of useOpenCL is almost the same in 3.3.0 and 4.0.0. For example, in 4.0.0 it returns
data->useOpenCL = (int)(haveOpenCL() && Device::getDefault().ptr() && Device::getDefault().available()) ? 1 : 0;

It seems that it checks whether the device in "available" now, and return false if not. What does it mean?

Can you advise me to use always opencv_core.haveOpenCL() function instead of useOpenCL? I don't need to know the current device status, if it can change so quickly; I just need to select between 2 versions of the algorithms: UMat-oriented or Mat-oriented.

[saudet]

So, sounds like another bug in OpenCV. Please report upstream!

[Daniel-Alievsky]

Hello Samuel, it seems I found the problem: see opencv/opencv#13472 (comment)
Waiting for their reply...

[saudet]

I don't think that's issue, unless your OpenCL implementation has a bug and isn't thread-safe, which may very well be the case...

[saudet]

Let's close this issue as it is now being tracked upstream.
Make sure to ping them once in a while for progress!