search

CNugteren / CLBlast

Tuned OpenCL BLAS
Apache License 2.0
1.06k stars 202 forks source link

Segmentation fault for "_routine_" tuners #556

Open diverger opened 4 months ago

diverger commented 4 months ago

Hi,

On a Mali-G610 device, OpenCL 3.0 platform, the below tuners failed.

./clblast_tuner_routine_xgemm -precision 32
./clblast_tuner_routine_xgemm -precision 3232
./clblast_tuner_routine_xgemm -precision 16
./clblast_tuner_routine_xtrsv -precision 32
./clblast_tuner_routine_xtrsv -precision 3232

All of them show at the end:

* Writing a total of x results to 'clblast_routine_xxxxx.json'
* Completed tuning process

make[3]: *** [CMakeFiles/alltuners.dir/build.make:109: CMakeFiles/alltuners] Segmentation fault (core dumped)
make[2]: *** [CMakeFiles/Makefile2:545: CMakeFiles/alltuners.dir/all] Error 2
make[1]: *** [CMakeFiles/Makefile2:552: CMakeFiles/alltuners.dir/rule] Error 2
make: *** [Makefile:377: alltuners] Error 2

Are the tuning results usable in this situation?

BTW: I commented out the '64 6464' precisions in the top CMakefileList.txt.

CNugteren commented 3 months ago

The tuning results are valid in this case, given that it wrote:

  • Writing a total of x results to 'clblast_routine_xxxxx.json'
  • Completed tuning process

So most likely the issue is with some OpenCL cleaning-up code that doesn't work on your platform. If you can provide me with a backtrace (compile in debug mode) then I can have a look, but also feel free to ignore this issue.