Add support for Nvidia NVCaffe

[HBadertscher]

Hi,

This pull request fixes issue #1002 by adding a new CMake flag NV_CAFFE and #ifdef where needed to distinguish between the standard Caffe and NVCaffe.

This has been tested with the official Nvidia Docker image nvcr.io/nvidia/caffe:18.12-py2.

Do you think it would be possible to test this automatically using Travis too?

I'm happy about any feedback.

cc @chrigima

[soulslicer]

I just ran it with nv_caffe, and it runs about 25% slower

[gineshidalgo99]

Thanks a lot! I will check it in a couple weeks, I find myself in a super busy weeks but I will be back to this eventually!

[gineshidalgo99]

EDITED1:
Do you think it would be possible to test this automatically using Travis too? Yes and no. I.e., Travis only allow CPU-based testing (it does not provide GPU support), so I would not be able to test its runtime behavior. But we could test whether it compiles (which is what we do with the GPU version).

If you are able to provide a script on how to download NVcaffe, compile it, and then link OP to it, I should be able to adapt it to Travis.

EDITED2:
Merged! I have also pushed an additional commit to combine OPTION(NV_CAFFE) inside DL_FRAMEWORK, so I believe that would make more sense. Thanks again for the important PR!

And feel free to keep updating it with new pull requests! :)

[peteruhrig]

I have written a step-by-step guide to install a portable container with NVCaffe and support for multiple NVidia cards as well as CPU. Maybe it helps someone here. http://peter-uhrig.de/openpose-with-nvcaffe-in-a-singularity-container-with-support-for-multiple-architectures/

[gineshidalgo99]

@peteruhrig Thanks! Your link will be pushed soon into the Custom NVIDIA NVCaffe section in the doc/installation.md!

[peteruhrig]

I'm still updating the page. It turns out NVCaffe does not support older cards. Officially they support only Pascal, Volta and Turing architectures. But Maxwell (GTX 980) works for me, too. Kepler cards (K40m) do not, however. Thus I'm building a third version for legacy GPU, i.e. is just the standard one with custom Caffe and GPU support.

[gineshidalgo99]

I am simply linking your URL, so as long as you do those edits on that URL, there will be no issue :)

Otherwise, simply post your final URL later and I'll update it.