Open Quuxplusone opened 4 years ago
Bugzilla Link | PR45401 |
Status | REOPENED |
Importance | P normal |
Reported by | Ye Luo (xw111luoye@gmail.com) |
Reported on | 2020-04-02 06:54:26 -0700 |
Last modified on | 2021-03-04 06:26:05 -0800 |
Version | unspecified |
Hardware | PC Linux |
CC | bruno.turcksin@gmail.com, llvm-bugs@lists.llvm.org |
Fixed by commit(s) | |
Attachments | |
Blocks | |
Blocked by | |
See also |
When I built from the master today, the problem remains with cuda 10.1/10.2 but goes away with CUDA 11.
After a bit more investigation, I found it related with the flag -march
intended for the host compilation.
I invoke the compiler with
-march=native -O3 -ffast-math -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda
some functions failed to be loaded.
If I put -march=skylake-avx512, it still fails.
Once I changed the flag to -march=skylake, my application runs.
So my hypothesis is, some avx512 codes contaminated my device functions.
The issue was caused by the difference between host and device pass during
compilation.
See "background" description in https://github.com/QMCPACK/qmcpack/pull/2981