No OpenCL support

[fnordahl]

Why do you lock this technology to a single hardware vendor by writing the accelerated bits in nvidia proprietary CUDA and not industry standard OpenCL?

[Grabber]

@fnordahl, Darknet first codes were done in OpenCL but at this time it was much easier to use CUDA.

If you really want to see it working, you may start helping the Darknet community by proposing the architecture and pushing code.

[fnordahl]

Cool, I would love to do that. Is there a dev branch lying around somewhere with the initial OpenCL code that can be published as a WIP to work from?

[TaihuLight]

You can fork one and modify the code.

[mrdeveloperdude]

I will gladly help. Why was it easier in CUDA?

[spedagadi]

Here is my OpenCL version of tiny yolo inference engine. Produces runtime speeds of ~208fps.

https://github.com/sat8/YoloOCLInference

Note that this is inference engine only.

[santo4ul]

Hi Sat8, The link (https://github.com/sat8/YoloOCLInference) you provided above is broken. Could you post the correct link?

Thanks,
San

[pramenku]

Is anyone have OpenCL based source code to build the test? If yes, can someone please point the location..
As a Tester, I want to try and run the tests.

[ganyc717]

I have port the Darknet to OpenCL. You may check this https://github.com/ganyc717/Darknet-On-OpenCL

[pramenku]

thanks ganyc717 for the info.

[sowson]

I have ported the DarkNet to OpenCL as well. You may check this https://github.com/sowson/darknet It works on AMD & Intel & NVidia on macOS & GNU/Linux. It supports Yolo3-spp, Yolo3, Yolo2, Yolo1, CIFAR-10, both execution, and training. I used other developers code, but to enable the training process and make it perform well I had to write some kernels from scratches. In Yolo2 (608x608) on external WebCam on CentOS, I have 30 FPS, that is the limitation of WebCam I have :-). I even tested this on RPi3 with VC4CL if you are interested in the results please look into http://codingbytodesign.net/2018/08/01/darknet-in-opencl/ Thanks!

[r0l1]

@sowson great work! Could you please open the issue section at your repo?

[sowson]

@r0l1 done! ✅

[wkoszek]

@pjreddie What's your take on @sowson project? Do you think it could be merged to the main darknet?

[sowson]

@pjreddie thank you so much for your work... It was a difficult journey to went... today (27 Jan 2019) Yolo1, Yolo2, Yolo3 training start to decrease less as expected :D. OMG, it was a more than a year to make it... :D. Is there anyone with strong AMD setup PC to try it? Thanks!

[sowson]

My project had some issues but now seams to the most of them have been solved at https://github.com/sowson/darknet it is Darknet in OpenCL... I am looking for contributor who help with pull request that add Windows setup possibility. Thanks! P.S. @pjreddie any comments on this?