Initialize arrays and scalars to signaling NaNs for Intel in debug mode #5801
Conversation
amametjanov
added
enhancement
Machine Files
BFB
|
This passes |
| @@ -1757,6 +1757,8 @@ subroutine diag_phys_writeout(state, psl) | |||
|
|
|||
| !! Boundary layer atmospheric stability, temperature, water vapor diagnostics | |||
|
|
|||
| p_surf_t1 = 0._r8 | |||
| p_surf_t2 = 0._r8 | |||
There was a problem hiding this comment.
Nice catch, @amametjanov . In the same spirit, we probably should also initialize p_surf_q1 and p_surf_q2. vertinterp calculations only loop over ncol. Just like p_surf_t1 and p_surf_t2, p_surf_q1 and p_surf_q2 could also be invoked in full array operation when certain output variable is specified.
| @@ -1833,6 +1833,7 @@ subroutine clubb_tend_cam( & | |||
| ! At each CLUBB call, initialize mean momentum and thermo CLUBB state | |||
| ! from the CAM state | |||
|
|
|||
| rvm = 0._r8 | |||
There was a problem hiding this comment.
Certainly an initialization of it has no harm. I do wonder if an error was detected with the new debug compilation options. Only 1 to ncol in rvm are used in calculation, and they always have proper values to begin with (from state1%q). If an invalid is reported without the rvm initialization, it seems an over-kill.
There was a problem hiding this comment.
Array accesses are over pverp also in rvm(1:ncol,1:pverp):
1944 do i=1,ncol ! loop over columns
...
2047 do k=1,pverp
...
2060 rvm_in(k) = real(rvm(i,pverp-k+1), kind = core_rknd)
This was needed for SMS_D_Ln5.ne4_oQU240.F2010.chrysalis_intel.eam-clubb_sp. Without this initialization, error is
16: forrtl: error (182): floating invalid - possible uninitialized real/complex variable.
16: Image PC Routine Line Source
16: libpnetcdf.so.3.0 000015555101868C for__signal_handl Unknown Unknown
16: libpthread-2.28.s 000015554E03DB20 Unknown Unknown Unknown
16: e3sm.exe 0000000002EECC7C clubb_intr_mp_clu 2060 clubb_intr.F90
16: e3sm.exe 0000000003546686 physpkg_mp_tphysb 2712 physpkg.F90
16: e3sm.exe 000000000350D8CF physpkg_mp_phys_r 1150 physpkg.F90
16: libiomp5.so 00001555515583F3 __kmp_invoke_micr Unknown Unknown
16: libiomp5.so 00001555514DC273 Unknown Unknown Unknown
and rvm = 0._r8 was the minimal initialization to get past this issue.
| sh_cldice(i,k)=sh_icwmr(i,k)* fice(i,k) *sh_frac(i,k) | ||
| endif | ||
| end do | ||
| end do |
There was a problem hiding this comment.
Need to monitor tests with COSP on, esp. with MMF, as icmwr has 0s for configurations that use CLUBB or SHOC.
There was a problem hiding this comment.
Please clarify.
- If
ficehas a NaN, thensh_cldliqwas also get assigned a NaN (and now getting a 0 instead). - If no NaNs in
fice, thensh_cldliqwas and will be getting a 0 wheneversh_icwmrhas a 0.
There was a problem hiding this comment.
discussion on telecon: integration passed and that includes COSP and MMF so this is ready.
amametjanov
force-pushed
the
azamat/machines/add-init-snan-intel-debug
branch
from
8c657b4
to
ba16f5b
Compare
(PR #5801) Initialize arrays and scalars to signaling NaNs for Intel in debug mode Add `-init=snan,arrays` to Intel's debug-compile flags. Also, initialize EAM's and ELM's uninitialized arrays, which were flagged by compiler's run-time checks. [BFB]
|
Rebased onto today's master. |
(PR #5801) Initialize arrays and scalars to signaling NaNs for Intel in debug mode Re-merge to bring the new commit since last merge.
|
Noticed something odd while using these new debugging options with the MOAB coupler on chrysalis. The code runs fine without debugging but with it on, we got this stack trace: line 639 in seq_map_mod.F90 is an executable statement but the other 2 files/lines are comments. Removing -init=snan,arrays let the code run to completion. |
|
Likely that some scalars or arrays at that line are uninitialized. When I added |
Add
-init=snan,arraysto Intel's debug-compile flags.Also, initialize EAM's and ELM's uninitialized arrays, which were flagged by compiler's run-time checks.
[BFB]