GPU hang-up occurs due to low GPU clock.

[hsaito]

Hello,

Operating environment is as follows

i.MX8QM

Linux: L5.4.47_2.2.0

TensorflowLite: v2.2.0

When using NNAPI with tensorflowLite, the GPU clock is reduced (1/64 clock) due to CPU temperature rise.
At this time, the processing speed is reduced, but the system is working properly.
After a while, the CPU temperature drops and the GPU clock returns to normal.
At this time, if the Invoke() function is still running, the GPU hangs.

Is it normal for the GPU to hang up due to fluctuations in the GPU clock?
Also, do you know why this phenomenon occurs?

[Bio_TICFSL]

Hello hsaito,

It's normal that gpu drop as CPU temperature is high, but it does not hangs, please send your code in order to check it.

Regards

[hsaito]

Hello Bio_TICFSL,

Attached is the source code where the hang-up occurs.
The hangup occurs in the Invoke() process within inference() with NNAPI enabled.

[Bio_TICFSL]

Hello,

GPU should not recover the clock when temperature is back to normal. It shouldn't hang.

It is possible that sometimes gpu clock isn't recovering within the timeout, which makes it think gpu is stuck.

Could you change the gdcGPU_TIMEOUT to be a little longer and rebuilt your kernel to see if this helps?  By default, it is 30 seconds.  We know this helps some customers.

It is in gpu driver code  gpu-viv6/driver/hal/inc/gc_hal_options.h.

5.4.47 is too old.  I am afraid we are not supporting it anymore.

Regards