Re: Issues testing HSA driver

I solved the first issue myself:

Due to the weak consistency memory model, accesses to counter caused an unexpected behavior.

It was fixed by using a mem_fence:

counter = (counter + 1) % 10 ;

mem_fence(CLK_LOCAL_MEM_FENCE);

Anyway, avoiding the mem_fence caused certain instabilities in the driver, that maybe could be reviewed by the development team.

I got no way to solve the second issue. Could be posssible to force the use of c11 atomic in a OpenCL kernel launched in the CPU?

Thanks.