On 18/11/18 19:49, Rebecca N. Palmer wrote: > This appears to be a crash in a loop unrolling optimization pass > (enabled only when -cl-fast-relaxed-math is set, hence only crashing then). > > Disabling this pass unconditionally (see attached) avoids the crash, but > may reduce performance. >
Hi Rebecca, Just tested the patch (on top of the debian-beignet-git-repo) with llvm 7. The workaround works for me, in a sense everything is better than a segfault! So the opencv_perf_dnn test now completes (although there are still some other opencl related errors in part with build-options, but those could very well be opencv bugs). Thanks for the help and patch! -- Sander