Open set-soft opened 1 year ago
@set-soft
I think this is an interesting question which I didn't awared before. The kdb should be a pre-compiled database that told miopen to find the right strategy to improve speeding for related net layer.
I didn't dig on kdb before, because I think it should be compiled on the runtime. But you are right, it may need to find out the real effects.
I cannot make a certain deadline, if I have any results, I will feed back here.
Thanks you @xuhuisheng for even considering it! I appreciate your hard work!
Environment
What is the expected behavior
This is not a bug report, just a question.
As you correctly state "Please test override hsa version first. Then ROCm will use gfx1030 fatbin to run on gfx1012, which needn't re-compile". This is my case, I can fake my board to be gfx1030, things work. But Stable Diffusion WebUI takes around 3 minutes to compile the kernels after complaining:
MIOpen(HIP): Warning [SQLiteBase] Missing system database file: gfx1030_11.kdb Performance may degrade. Please follow instructions to install: https://github.com/ROCmSoftwarePlatform/MIOpen#installing-miopen-kernels-package
Of course I can't install gfx1030_11.kdb because it doesn't exist. I know almost nothing about the ROCm internals so perhaps I'm asking something ridiculous. Do you have such a .kdb? Is mandatory to have the actual GPU to compile them? Can I compile them using the actual GPU, but without needing to compile the whole ROCm (which could take hours in my system, assuming I can free enough disk)
Thanks in advance