nrn ExternalProject SUNDIALS 3.2.1

include work from cvode3 branch

closes #1328

[x] Cvode version 3 test data must be visually compared to the legacy data for acceptable similarity.
[ ] Must use nrn-modeldb-ci to visually compare the differences.
[x] test coverage substantially complete
[x] merge rxd testdata and update submodule commit

Aug 20 '22 12:08 alexsavulescu

Logfiles from GitLab pipeline #89893 (:no_entry:) have been uploaded here!

Status and direct links:

Dec 11 '22 20:12 bbpbuildbot

Logfiles from GitLab pipeline #119379 (:no_entry:) have been uploaded here!

Status and direct links:

May 01 '23 11:05 bbpbuildbot

Logfiles from GitLab pipeline #120887 (:no_entry:) have been uploaded here!

Status and direct links:

May 04 '23 13:05 bbpbuildbot

@alkino Just to let you know, I have a fix for ctest -R netpar but need to extend to NRNMPI_DYNAMICLOAD.

May 05 '23 07:05 nrnhines

@alkino Just to let you know, I have a fix for ctest -R netpar but need to extend to NRNMPI_DYNAMICLOAD.

NRNMPI_DYNAMICLOAD should be cleaned to be the same than in coreneuron (double implentation today): https://github.com/neuronsimulator/nrn/issues/2285, so feel free to rework everything

May 05 '23 08:05 alkino

Logfiles from GitLab pipeline #121401 (:no_entry:) have been uploaded here!

Status and direct links:

May 05 '23 13:05 bbpbuildbot

Logfiles from GitLab pipeline #121436 (:no_entry:) have been uploaded here!

Status and direct links:

May 05 '23 14:05 bbpbuildbot

Logfiles from GitLab pipeline #121578 (:no_entry:) have been uploaded here!

Status and direct links:

May 06 '23 15:05 bbpbuildbot

Logfiles from GitLab pipeline #122857 (:no_entry:) have been uploaded here!

Status and direct links:

May 11 '23 07:05 bbpbuildbot

Logfiles from GitLab pipeline #124658 (:no_entry:) have been uploaded here!

Status and direct links:

May 17 '23 09:05 bbpbuildbot

src/nrnmpi/nrndynam.cpp has the comment

// Only works if caller is compiled with same MPI version that is
// dynamically loaded.

Which was intended to imply that because sundials uses mpi, NRNMPI_DYNAMIC_LOAD was incomplete because libnrniv.so (which contains all the sundials functions) still referred to some MPI methods which made the entire distribution specific to the build machine version of MPI.

However this is not the case. NRNMPI_DYNAMIC_LOAD can be built with several versions of MPI and will work on any user machine that has one of those MPI versions installed. It turns out that the only sundials library archive that refers to MPI is libsundials_nvecparallel.a and during link time, nothing from that archive is added to the shared library because all the relevant functions are added by earlier linking against nrn/src/nrniv/nvector_parallel.c (which is basically a modified copy of external/sundials/src/sundials-external/src/nvec_par/nvector_parallel.c)

To make this clear, I recommend modifying src/nrniv/CMakeLists.txt to remove the fragment

if(NRN_ENABLE_MPI)
  target_link_libraries(nrniv_lib ${SUNDIALS_LIB_DIR}/libsundials_nvecparallel.a)
endif()

I'll make and test this change.

May 17 '23 12:05 nrnhines

Logfiles from GitLab pipeline #124872 (:no_entry:) have been uploaded here!

Status and direct links:

May 17 '23 14:05 bbpbuildbot

Logfiles from GitLab pipeline #124926 (:no_entry:) have been uploaded here!

Status and direct links:

May 17 '23 16:05 bbpbuildbot

Logfiles from GitLab pipeline #126459 (:no_entry:) have been uploaded here!

Status and direct links:

May 25 '23 10:05 bbpbuildbot

Logfiles from GitLab pipeline #126966 (:no_entry:) have been uploaded here!

Status and direct links:

May 28 '23 12:05 bbpbuildbot

Logfiles from GitLab pipeline #128011 (:no_entry:) have been uploaded here!

Status and direct links:

Jun 01 '23 10:06 bbpbuildbot

It is difficult getting tests to pass in a principled manner with respect to quantitative identity of results. Differences are sometimes seen with respect to architecture, operating system, floating point differences between transcendental functions (pow(x,y) for example), compiler version, order of summation (eg. threads, MPI), and, now, cvode version. I'm guessing that it might be worthwhile to use a test strategy that does a series of cvode runs with a series of a absolute tolerances at fixed exact time points and event times and verify that the "results" approach an asymptotic standard result.

I've gotten cvodeobj.cpp to compile with Sundials v 6 and when the whole thing compiles will see if I can get some tests such as test_nrntest_fast.py modified to pass for cv2, 3, 6 on windows/mac/linux

Jun 01 '23 13:06 nrnhines

It is difficult getting tests to pass in a principled manner with respect to quantitative identity of results. Differences are sometimes seen with respect to architecture, operating system, floating point differences between transcendental functions (pow(x,y) for example), compiler version, order of summation (eg. threads, MPI), and, now, cvode version. I'm guessing that it might be worthwhile to use a test strategy that does a series of cvode runs with a series of a absolute tolerances at fixed exact time points and event times and verify that the "results" approach an asymptotic standard result.

I've gotten cvodeobj.cpp to compile with Sundials v 6 and when the whole thing compiles will see if I can get some tests such as test_nrntest_fast.py modified to pass for cv2, 3, 6 on windows/mac/linux

Should we pass to sundials 6 directly?

Jun 01 '23 13:06 alkino

Should we pass to sundials 6 directly?

I don't think so. Undoubtedly there are bugs in my modifications. At this point, it is just another example see if we can eliminate version specific data for CI tests.

Jun 01 '23 14:06 nrnhines

Logfiles from GitLab pipeline #128303 (:no_entry:) have been uploaded here!

Status and direct links: