barrier and driver crash.

LLess — Mon, 01 Apr 2013 04:39:31 GMT

Hi,

Today I used the barrier call for the fist time and it was causing a crash of the graphic driver (HD4000, Win7 x64, 9.17.10.2932) (Lucky for me Win7 was able to restart it nicely every time).

The issue was that I had a barrier inside an IF statement and that not all work items were taking the branch.

Now, it would be good in the offline compiler was able to at least raise a warning about that if not an error actually.

After moving the barrier my kernel worked fixing the code, so no big deal here but just to let you know.

Ideally the kernel compilation should detect that before crashing the driver :)

Laurent.

Laurent,

Raghupathi_M_Intel — Wed, 03 Apr 2013 20:11:58 GMT

Laurent,

Thanks for the suggestion. The optimization guide clearly mentions the pitfalls of using barriers in branchy code. Though it is a nice idea for the tools to WARN about potential problems, I am not sure if such a capability is beyond what Kernel Builder can offer. If you have more suggestion please do post here. I will gather all the feedback and talk to the product team.

Thanks,
Raghu

topic Laurent, in OpenCL* for CPU

barrier and driver crash.

Laurent,