OpenCL

[sowson]

Hello, recently I made the DarkNet on OpenCL that is technology that really passionate me and I started recently PhD studies on AI field. I am using few different GPUs, recently 2 x AMD Radeon VII that works very well on my research. I have few issues... well minor issues but still.

First is that I can use only my trivial GEMM implementation for MultiGPU.

Do you know why clBLAS and CLBlast not support multi-threading environment?

Second I was wonder how fast comparing to Radeon VII is WX Pro series.

I was wonder can someone can test it on the AMD Pro GPUs that I do not have?

You may see my comparison at: DarkNet CUDA vs OpenCL and CPU vs NVIDIA vs AMD – coding by to design

My project is at: GitHub - sowson/darknet: Convolutional Neural Networks on OpenCL on Intel & NVidia & AMD & ARM for m...

Enjoy! .

[sowson]

There is one more update that may be nice to read here. I compared Nvidia Titan RTX ($3000) on GNU/Linux with AMD Radeon VII ($700) on macOS... gues what GPU won? Here is the description and results in a table: Is OpenCL beats CUDA? Thanks!

[colesdav]

Norton is telling me this link: https://community.amd.com/external-link.jspa?url=https%3A%2F%2Fgithub.com%2Fsowson%2Fdarknet 
Is a dangerous website...

[sowson]

Hi, what do you mean by that? I was thinking it is a phishing first but looks safe

[sowson]

Hi, what do you mean by that? I was thinking it is a phishing first but looks safe

[colesdav]

I am interested in what you are working on.

I clicked on your link and I see this:

[sowson]

This may happen because I have a new domain on the Internet and old one only has permanent redirect enabled... new blog address is iblog.isowa.io sorry I do not know how to edit this entry... is there any moderator that can change the entry from codingbytofrdign.net => iblog.isowa.io Again sorry by that...

NEW LINKS MAY INTEREST YOU ALL

https://iblog.isowa.io/2019/02/02/darknet-in-opencl-on-asus-thinker-board-s/

https://iblog.isowa.io/2019/08/18/the-fastest-darknet-in-opencl-on-the-planet/

https://iblog.isowa.io/2019/11/05/gpu-computing-on-opencl/

https://iblog.isowa.io/2019/12/29/darknet-cuda-vs-opencl-and-cpu-vs-nvidia-vs-amd/

https://iblog.isowa.io/2020/01/04/gpu-opencl-fine-tuning-problem/

https://iblog.isowa.io/2020/02/08/pc-for-phd-studies/

https://iblog.isowa.io/2020/03/02/hania-pc-well-it-needs-macos/

https://iblog.isowa.io/2020/03/03/is-opencl-beats-cuda/

Thanks!

[sowson]

I am happy to say, about the first, pre-publication. At https://www.preprints.org/manuscript/202007.0506/v1 ;-).

[sowson]

Hello... I am not sure if that DNN / AI / ML interested you, but here it is YOLO4 on OpenCL from Today... and CUDA version has 2 very heavy to find trick that does not work for training, I squash those in my port on OpenCL... Look for it at https://github.com/sowson/darknet Thanks! Edit: Screenshots added.

p .

[elad]

sowson‌ any hope to be able to run this on Windows?

[sowson]

Yes, take a look on the https://iblog.isowa.io/2021/11/20/darknet-on-opencl-on-windows-11-x64 ;-).

[sowson]

Once it uses CMake it is possible, only one dependency is clBLAS, and that DLL needs to be built as well. But to be honest I do not have Windows on any PC I could experiment with this. But GNU/Linux (Ubuntu or CentOS) are fully supported)...