Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent) #1919
Conversation
… restrictive and allow more tests to pass. For Intel 17, instead of -fp-model source, use -fp-model consistent. For Intel 18, instead of -fp-model source, use -fp-model consistent -fimf-use-svml. Fixes #1477
Shorter paths for case2 using system_tests_compare_two The recently-changed version of system_tests_compare_two nested case2's bld and run directories under $CIME_OUTPUT_ROOT/TESTNAME/case2/TESTNAME. Since the TESTNAME is often long, this was leading to very long paths to some things in the bld directory. This was causing some compilation errors. With this PR, case2's bld and run directories appear in $CIME_OUTPUT_ROOT/TESTNAME/case2/bld and $CIME_OUTPUT_ROOT/TESTNAME/case2/run. This shortens the bld directory paths enough to get some formerly-failing builds to pass. The change in run directory is just for consistency. I was hoping to do this just via setting EXEROOT and RUNDIR in system_tests_compare_two, to avoid cluttering up create_clone with arguments that are just needed for tests. But it turns out that we need to do this in create_clone, because create_clone calls case_setup, which creates the EXEROOT and RUNDIR directories. Test suite: scripts_regression_tests on cheyenne Also confirmed that the following tests build successfully: ERP.f10_f10_musgs.I1850Clm45BgcCru.yellowstone_pgi ERP_P60x2_Lm36.f10_f10_musgs.I2000Clm50BgcCrop.yellowstone_intel.clm-clm50cropIrrigMonth_interp Test baseline: N/A Test namelist changes: none Test status: bit for bit Fixes ESMCI/cime#1914 User interface changes?: none Update gh-pages html (Y/N)?: Code review: jedwards
|
I think the issue comes up in |
|
I have my own set of tests that include acme_developer, acme_integration, the hires test, as well as others that have given us issues along the way. Results are here: |
|
Can I merge this to next? I'd like to then make Intel 18 the default for Cori in another PR. |
|
What's the previous pass/fail status: can't find v18 original? |
|
Sorry, I typically delete the cs.status files. Is there something here that can answer your question? |
|
It looks like this change introduces a new fail with the HOMME test. Machine-specific flags can be added to |
|
Yea, you're right. The HOMME test was not affected in the same way by the e3sm flags and seems so sensitive that I was treating it specially. I'm not sure Mark wants to add these flags to HOMME though? |
|
I'd like to move homme to -fp-model=fast. I think the standalone HOMME tests are probably less sensitive and shouldn't we be able to get away with fast? |
|
Can I merge this to next? |
|
I'd really like to get this thru first before making other changes. |
…ml (PR #1919) Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent)
|
This should have fixed github 1850 as well |
…ml (PR #1919) Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent)
…ml (PR #1919) Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent)
…ml (PR #1919) Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent)
…ml (PR #1919) Adjust fp-model flags for Intel compiler on cori-knl (use fp-model consistent)
For cori-knl and Intel compiler, adjust the compiler flags to be more restrictive and allow more tests to pass.
For Intel 17, instead of -fp-model source, use -fp-model consistent.
For Intel 18, instead of -fp-model source, use -fp-model consistent -fimf-use-svml.
Fixes #1477