AnsweredAssumed Answered

amdgpu: Understanding disable_cu and USER_SHADER_ARRAY_CONFIG

Question asked by jungle on May 23, 2019

I am trying to understand how to use disable_cu in amdgpu. So the docs say to apply it like   dsiable_cu=se.sh.cu

I can successfully switch of cus as Open CL reports a different count. However I am not sure how they are addressed, other people online seem to start from 0.0.0 and go to 3.0.15 for a vega 64, I am assuming that and RX580 would go from 0.0.0 to 3.0.9 or something ? and the middle number is always 0 ??? I couldnt find any concrete docs on this.

 

Also how can it be set on a per card basis ? the code in the linux kernel seems to take a device:

static void gfx_v8_0_set_user_cu_inactive_bitmap(struct amdgpu_device *adev,                                u32 bitmap)

If I disable all of the CUs in a shader engine Linux doesn't boot properly, is there a way to get it to continue ?

Or is there a way in OpenCL to target certain shader engines ?

 

Does this actually do anything on a hardware level ? Can a shader engine be switched of or can the clock be turned off ?

 

 

Thanks

Outcomes