Archives Discussions

[dragonxi4amd]

Services from AMD to restart OpenCL resources

Q1. How to restart an OpenCL compatible device provided by AMD ?

a. GPU ?

- warm start instead of power switch on/off ?

b. CPU

- just OpenCL runtime and OpenCL tasks i.e not other apps ?

Q2. How to reset context ?

a. just setting reference count to zero ?

b. killing resources (tasks/processes) using the context ?

[LeeHowes]

What do you want to achieve?

[dragonxi4amd]

Hi Lee,

1. to kill OpenCL processes/tasks/threads not responding

Achievment: system and other apps continue normally

2. to release resources (RAM, queues, stack...) reserved by OpenCL   dead / not responding/ faulty system 

Achievment: necessary resources usable to OS and other apps

3. to continue by restarting just OpenCL system - not OS and other apps

Achievment: no side effects & decreased quality of service to other apps 

4. to avoid power off/shutdown

Achievment: power on and system stays online   

5. in case of several OpenCL devices to restart just a dead device

Achievment:  other devices keep on offering their services

[LeeHowes]

I would imagine that you wouldn't be able to do most of that programmatically. The device is owned by the operating system, after all. I don't know if the driver has any calls to reset it but I doubt there would ever be a wish to put such a thing into the OpenCL standard.

[dragonxi4amd]

Thanks for your answer Lee !

Referring to Topic Title: Detecting device resets
http://forums.amd.com/forum/messageview.cfm?catid=390&threadid=147247&highlight_key=y&keyword1=Detec...

where TDR = Timeout Detection and Recovery of GPU

Otterz:
* was using windows 7, with TDR enabled
* tried registering a callback when creating the context, but that callback did not seem to be called !
 
Raistmer: "There is no correct way for program to be informed of driver failure/restart. And this is really bad, especially when app should run unattended".

My feedback:

~ callback is defined in OpenCL specs and should be called in exceptions

~  I agree (so do our clients) with Raistmer - apps should run unattended

My Question: is the user defined callback function called in the latest AMD SDK version in case GPU makes reset ?