topic Hello Konrad, in Software Archive

OpenCL; MIC specific compilation errors

Konrad_L_ — Tue, 20 May 2014 17:56:52 GMT

Dear all,

I'm trying to compile one OpenCL Kernel which is actually quite long for the MIC, however the compilation for the MIC doesn't work. I was able to compile the same OpenCL code with the AMD SDK for a AMD GPU and also for my Intel CPU (using the AMD SDK); however when I try to compile the code for the MIC I get following error: "LLVM ERROR: Buffer too small for code", when I try to compile my entire code with multiple kernels [There is already a forum entry for this: https://software.intel.com/en-us/forums/topic/499450]. When I try to reduce the number of kernels to the one I'm mainly interested in I get a segmentation fault, or the external kernel builder crashes.

Are there any workaround regarding this? Is there already a release plan for the next OpenCL SDK for the XeonPhi?

Thank you,

Konrad

Dear Konrad,Thanks for

Arik_N_Intel — Wed, 21 May 2014 10:41:15 GMT

Dear Konrad,

Thanks for reporting the issue.

This problem is known and will be fixed in our opcoming release of the OpenCL runtime. (No final date yet)

The crash that you reports is new to me. Can you please share a reproducer for the crash?

Thanks,

Arik

Hello Konrad,

Yuri_K_Intel — Tue, 25 Nov 2014 12:37:36 GMT

Hello Konrad, Sorry for the delay. Current solution/workaround is to prevent automatic vectorization either by CL_CONFIG_USE_VECTORIZER or by supplying a vector type in vec_type_hint. See https://software.intel.com/en-us/node/531302, https://software.intel.com/en-us/node/531303 Thanks, Yuri