topic Thank you! Let me know what in Software Archive

OpenCL error

peastman — Tue, 02 Jul 2013 22:24:01 GMT

I'm porting an OpenCL application to run on Xeon Phi. I've encountered what I'm pretty sure is a bug in the compiler or runtime. With some work, I've narrowed it down to the following minimal test case. To observe the problem, just launch testKernel() using a single workgroup of 64 work units. It repeatedly prints out

1 <= 0????

void testFunction(__local short* a, __local int* b) {
   if (get_local_id(0) == 0)
       *a = 50;
   barrier(CLK_LOCAL_MEM_FENCE);
   int c = *a/32;
   if (c > 0) {
       if (get_local_id(0) == 0)
           *b = 0;
   }
   else
       printf("%d <= 0????\n", c);
}

__kernel void testKernel() {
   __local short a;
   __local int b;
   testFunction(&a, &b);
}

Any idea what the problem is, or how I can work around it?

Peter

Hi,

Yuri_K_Intel — Wed, 10 Jul 2013 07:25:41 GMT

Hi, I will look into this issue. A couple of comments. It would be better (faster to reproduce) if you could provide full source code, including host part. There is a separate forum for OpenCL - http://software.intel.com/en-us/forums/intel-opencl-sdk, please use it for OpenCL related issues/questions. Thanks, Yuri

Hi Yuri,

peastman — Wed, 10 Jul 2013 17:53:27 GMT

Hi Yuri,

Thanks for looking into it. My contact at Intel specifically told me to post on this forum. Do you want me to repost it on the other one?

I don't have a simple, self contained host program for launching it. I'm working from a very large code base. But all it really needs to do is the generic boilerplate found in any OpenCL program: select the platform, figure out which device is the Xeon Phi, create a context, compile the kernel, and launch it.

Peter

Hi Peter,

Yuri_K_Intel — Fri, 12 Jul 2013 13:47:48 GMT

Hi Peter, No need to repost. I was able to reproduce the issue and I have submitted a ticket for it. I will inform you of the further investigation results. Thanks, Yuri

Thank you! Let me know what

peastman — Fri, 12 Jul 2013 17:31:08 GMT

Thank you! Let me know what you find out, especially if you can suggest a workaround.

Peter