From: Timo B. <tim...@gm...> - 2021-07-18 20:23:16
|
Hi, thanks for the tips. I have opened a bug report at https://github.com/conda-forge/pocl-feedstock/issues/67. The following is copied over from there as response to Michal's tips: --------------------- Tried out the following recommendation: llc --version in conda correctly identifies my CPU as skylake Running the kernel in debug mode with gdb gives a segfault in the following function in the conda version of pocl: Thread 1 "python" received signal SIGSEGV, Segmentation fault. 0x00007fff651618cc in llvm::TargetPassConfig::addPass(llvm::Pass*, bool, bool) () from /home/betcke/miniconda3/envs/bempp_test/lib/./libLLVM-11.so My distribution version of pocl is built on top of llvm 12. The conda version is built on top of LLVM 11. If I manually try to upgrade the conda version to libllvm 12 the conda dependency resolution wants to remove pocl. -------------------------- Hopefully the gdb message may give some clues. Best wishes Timo On Fri, 16 Jul 2021 at 14:33, Michal Babej (TAU) <mic...@tu...> wrote: > > Hello, > > __private double8' (vector of 8 'double' values) without 'avx512f' enabled changes the ABI > > Can you make sure that LLVM inside the Conda environment doesn't misdetect the CPU ? > Can be checked with llc command from LLVM package. Running llc --version should print something like: > > Optimized build. > Default target: x86_64-pc-linux-gnu > Host CPU: znver1 > > Finally, after all these warnings the code segfaults, which I did not see in the distribution provided pocl version. > > This definitely shouldn't happen even with the warnings. > > code can be generated that includes the type double16, but is never actually called as function in the kernel. > > Using wider vectors than CPU has should work properly, so i think that's fine. Can you try looking into doc/sphinx/source/debug.rst document, the "Debugging with GDB" section has a guide on compiling & running non-optimized -O0 version of your kernel. If it doesn't crash with -O0 code, it could be a bug somewhere in pocl's LLVM code. > > Regards, > Michal > > _______________________________________________ > pocl-devel mailing list > poc...@li... > https://lists.sourceforge.net/lists/listinfo/pocl-devel -- Timo Betcke Professor of Computational Mathematics University College London Department of Mathematics E-Mail: t.b...@uc... Tel.: +44 (0) 20-3108-4068 |