Hi,

Pocl's CPU driver does not touch the affinity, by default. IOW it leaves it to 
the kernel. By default it creates as much threads as there are logical CPU 
cores; this can be limited by
POCL_MAX_PTHREAD_COUNT env variable.

If you want to force Pocl to set the affinity of threads, set POCL_AFFINITY env 
var to 1.

As for:

"When subdevices are created using CL_DEVICE_PARTITION_EQUALLY, the number of 
utilised cores seems to be one less than actually specified"

This has been discussed already in https://github.com/pocl/pocl/issues/716.
The short version is that while kernel commands are split and processed on all 
driver threads, the other commands (clReadBuffer) are never split, and are 
processed in a single thread.

Regards,
-- mb

________________________________
From: Srijeeta Maity <[email protected]>
Sent: Wednesday, April 1, 2020 5:49 PM
To: [email protected] <[email protected]>
Subject: [pocl-devel] Subdevices andCPU affinity mismatch


Hello,

I found some observation while checking the CPU core affinity after creating 
OpenCL subdevices using PoCL.

1. I created 8 subdevices in a platform having 8CPU cores (Odroid XU4) and 
assigned some tasks on different subdevices . I checked  with htop  that the 
device affinity of those subdevices is not constant for the whole runtime (i.e. 
it is reset from time to time).
2. Some of those subdevices seem to be set to the same affinity. This can also 
be a side effect of the refresh rate in htop, so that I am not able to see when 
the affinity has changed.
3. When subdevices are created using CL_DEVICE_PARTITION_EQUALLY, the number of 
utilised cores seems to be one less than actually specified (i.e. partition 
equally to subdevices each having 4 compute units, will actually only use 3 cpu 
cores).

While checking for solution I found there was a similar bug reported for OpenCL 
in 2013.  https://software.intel.com/en-us/forums/opencl/topic/377784

Can you please share some pointer to solve this issue?

Thanks and regards,
Srijeeta

_______________________________________________
pocl-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/pocl-devel

Reply via email to