OpenCL CL_BUILD_ERROR in CUDA 7 but not in CUDA 6

FangQ · August 17, 2015, 5:19pm

can anyone take a look at my question posted on the OpenCL forum?

https://www.khronos.org/message_boards/showthread.php/11664-OpenCL-support-difference-in-cuda-6-vs-7

I am having trouble running my opencl code compiled with CUDA 7, but it works fine for previous versions of CUDAs.

little_jimmy · August 18, 2015, 4:58am

is it just for this particular opencl application, or multiple said applications (with cuda 7) - i.e. with cuda 7, is it only this opencl application, or others too?

and is it for specific architectures, or all architectures (you have multiple devices, so you could verify this)

FangQ · August 18, 2015, 6:19pm

I tried a hello-world sample kernel, it compiles and runs fine.

so, I supposed it is about my kernel. Something was changed from 6 to 7 to prevent it from building, and the building error was not helpful at all.

can anyone give it a try and let me know what might be wrong? my cl unit can be downloaded from

github.com

fangq/mcxcl/blob/master/src/mcx_core.cl

/***************************************************************************//**
**  \mainpage Monte Carlo eXtreme - GPU accelerated Monte Carlo Photon Migration \
**      -- OpenCL edition
**  \author Qianqian Fang <q.fang at neu.edu>
**  \copyright Qianqian Fang, 2009-2018
**
**  \section sref Reference:
**  \li \c (\b Yu2018) Leiming Yu, Fanny Nina-Paravecino, David Kaeli, and Qianqian Fang,
**         "Scalable and massively parallel Monte Carlo photon transport simulations
**         for heterogeneous computing platforms," J. Biomed. Optics, 23(1), 010504 (2018)
**
**  \section slicense License
**          GPL v3, see LICENSE.txt for details
*******************************************************************************/

#ifdef MCX_SAVE_DETECTORS
    #pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
#endif

#ifdef USE_HALF

This file has been truncated. show original

the building code is

github.com

fangq/mcxcl/blob/master/src/mcx_host.cpp#L371-L376


      
                  activeplatform = platform;
              }
          } else {
              cuid++;
              memcpy((*info) + ((*activedev)++), &cuinfo, sizeof(GPUInfo));
          }

FangQ · August 18, 2015, 6:23pm

as mentioned in my original post, the return status of clBuildProgram was different for 980Ti vs 590/730. For 980Ti, the status was CL_BUILD_ERROR, and the msg returned from clGetProgramBuildInfo was simply “GeForce GTX 980 Ti”. For 590/730, the return status of clBuildProgram was CL_BUILD_NONE.

I just upgraded cuda from 7 to 7.5. errors stay the same, except 980Ti loses its name and show as a device named “Graphics Device” :(