Cray: hugepage_alloc test fails due to compiler bug #1557
Labels
Comments
havogt
added a commit
that referenced
this issue
Sep 25, 2020
Environment changes on daint: - CUDA is now loaded as `cudatoolkit` instead of a specific version - Cray clang is now used without `-fno-cray`. `-fno-cray` doesn't work currently and is tracked here https://webrt.cscs.ch/Ticket/Display.html?id=40241. A Cray ticket is filed. However, the `test_hugepage_alloc` fails, which looks like an invalid optimization, see #1557. Performance changes: We see quite some chagnes both on GPU and CPU after the upgrade. Most significant: - CUDA: vertical advection on small domain with float is about 40% slower (in all variants: clang-cuda, nvcc-clang, nvcc-gcc) - cpu_kfirst: factor 10x slowdown for layout_transformation, maybe a vectorization issue
|
The wrong optimization is actually in the testing code. Allocation works properly. |
|
The open issue after #1560 will be to re-enable the test for Cray once we have a compiler where this is fixed. |
fthaler
changed the title
|
The problem is still in Cray clang 11.0.0 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
test_hugepage_allocfails on Cray clang version 10.0.2.It's the first time we are able to use Cray without
-fno-crayand it's the only failing test.The text was updated successfully, but these errors were encountered: