uvos

Looks good, i will pull this into mce and force push our upower branch to this, some one else can then do the status applet, i dont have commit access...

Unfortunately i investigated in the kernel for arcturus (mi100) and it looks like this is simply not implemented at all at the kernel level, with some evidence that this is...

I know, however i dont believe its any different. you can try runtime-pm and report back if it works. switching the uclk clock domain is defiantly not supported on the...

You need to set amdgpu.runpm=1 and set power/control to auto in device sysfs. Note that you won't see the device suspended in SMI as reading any of the values that...

> Hi [@65a](https://github.com/65a), unfortunately, both runtime PM and `mclk` configuration w/rocm-smi are not supported at the driver level for MI200, see [ROCm/rocm_smi_lib#117](https://github.com/ROCm/rocm_smi_lib/issues/117) for a similar issue. > > > Idle...

@AlexVlx I am the hip maintainer of https://github.com/ggml-org/llama.cpp. I would like to make the case that the deprecation of `__AMDGCN_WAVEFRONT_SIZE__` inside device code has bad implications for us. We maintain...

Your example already shows why this is a bad idea. We also compile for and run on gfx803 which you just broke. There is also no reason to believe that...

> As for "the `mwavefrontsize64` option isn't intended for production use, and taking a dependence on it is unsupported." Compiler documentation just states that the option is available, and RDNA...

> Oh! I didn't know that, thanks! I apologize for my ignorance, but what is "the compiler" then? Do you mean the clang in question, but without hip/hip_runtime.h? Either way,...

rocm uses gpu isa directly. To get good performance you have to have kernels for various problem sizes. This means that for etach function in rocm libaries like rocblas there...