All OpenCL versions form all vendors have this issue. It is a wrong computation. Please take a look at my blog describing it in detail. Can this be fixed on AMD OpenCL anyhow?

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

it is part of GitHub - sowson/darknet: Convolutional Neural Networks on OpenCL on Intel & NVidia & AMD & Mali GPUs for macOS & GNU/Lin… project please help!

Thanks!

There is one more thing if tuning is equal 1 then the computation is more accurate and when tuning is equal 16 there is faster but no as much accurate, thanks!