Support for CUDA parallel programming

[gsaponaro]

Out of curiosity, which iCub modules take advantage of parallel computing frameworks (e.g. CUDA, OpenCL, OpenMP, TBB)?

It would be helpful to have a brief list, in order to know which modules (from the icub-main applications list) to deploy on the Nvidia machine of our cluster and which ones on the non-Nvidia machines. For example, I know that the motionCUT module takes advantage of CUDA.

[pattacini]

motionCUT was used to take advantage of OpenMP while now it benefits from TBB, not CUDA at all.

icub-main does not contain CUDA dependent code, if I'm not mistaken.
By contrast, repositories such as iol, stereo-vision and himrep require the SiftGPU package and thus CUDA.

Regarding the TBB layer, I think that only motionCUT documentation does explicitly mention this requirement. However, it is always a good rule to enable TBB binding while compiling OpenCV.

[gsaponaro]

Thanks for the clarification.

Regarding motionCUT, part of the reason I got confused among the different frameworks, was that it sets an internal macro with OpenMP in its name:

    // check if OpenCV supports OpenMP multi-threading
    #if (CV_MAJOR_VERSION>0) && (CV_MAJOR_VERSION<3)
        #define _MOTIONCUT_MULTITHREADING_OPENMP
    #endif

but I guess motionCUT is agnostic to whatever optimization OpenCV is configured/compiled to use.

[pattacini]

Well, motionCUT is not completely agnostic, in a sense that it knows that some old versions of OpenCV were designed to work with OpenMP, while the most recent versions deal with the new TBB layer.

Of course, if OpenCV is not configured/compiled to use those multi-threading layers, then this condition is resolved internally to OpenCV.