[CI] Improve the tests for the linear solvers

[amontoison]

No description provided.

[amontoison]

@nimgould
I added CI tests with MUMPS v5.7.2 in #318 and all tests passed with Int32 and Int64. 👍
However I added CI tests with PASTIX v6.3.0 in this PR (only for Int32) and the tests failed. 👎

[nimgould]

OK, I'll have a look. I think my pastix is out of date, so I'll install the latest first

[nimgould]

The makefile double precision tests (both standalone via pastixt.F90 and as part of sls/sbls)
all work fine with pastix 6.3.2. However, the single precision ones don't get started; the spmGetArray function doesn't allocate the values array to the correct length. I will investigate further, but I have no idea why the meson double tests are failing

[amontoison]

I will upgrade the version installed with CI (6.3.0 -> 6.3.2) to check if the tests passed with this version.

[amontoison]

@nimgould I updated PaStiX to use the version 6.3.2 but single and double precision tests of PaStiX are still failing:

=================================== 74/460 ===================================
test:         GALAHAD:pastix+single+fortran / pastixt_single
start time:   17:16:18
duration:     0.15s
result:       killed by signal 11 SIGSEGV
command:      ASAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1 MALLOC_PERTURB_=88 MSAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1:print_stacktrace=1 UBSAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1:print_stacktrace=1 LD_LIBRARY_PATH=/home/runner/work/GALAHAD/GALAHAD/builddir/:/opt/hostedtoolcache/Python/3.12.4/x64/lib:/home/runner/work/GALAHAD/GALAHAD/galahad/lib:/home/runner/work/GALAHAD/GALAHAD/../deps/lib:/home/runner/work/GALAHAD/GALAHAD/../CUTEst/lib /home/runner/work/GALAHAD/GALAHAD/builddir/pastixt_single
----------------------------------- stdout -----------------------------------
  eps    1.0000000474974513E-003
single
  size of row, col, val =            7           7           7
----------------------------------- stderr -----------------------------------
ischedInit: The thread number has been automatically set to 2

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7fca35823970 in ???
#1  0x7fca35822ad5 in ???
#2  0x7fca3544251f in ???
#3  0x7fca354a53fe in ???
#4  0x7fca352f8d5f in ???
#5  0x7fca353e2fba in ???
#6  0x7fca35384b0d in ???
#7  0x7fca35394c6b in ???
#8  0x7fca35b68018 in ???
#9  0x55d624ceec3c in test_pastix
	at ../src/external/pastix/pastixt.F90:226
#10  0x55d624cef47b in main
	at ../src/external/pastix/pastixt.F90:7
==============================================================================

=================================== 75/460 ===================================
test:         GALAHAD:pastix+double+fortran / pastixt_double
start time:   17:16:18
duration:     0.15s
result:       killed by signal 11 SIGSEGV
command:      ASAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1 MSAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1:print_stacktrace=1 UBSAN_OPTIONS=halt_on_error=1:abort_on_error=1:print_summary=1:print_stacktrace=1 MALLOC_PERTURB_=104 LD_LIBRARY_PATH=/home/runner/work/GALAHAD/GALAHAD/builddir/:/opt/hostedtoolcache/Python/3.12.4/x64/lib:/home/runner/work/GALAHAD/GALAHAD/galahad/lib:/home/runner/work/GALAHAD/GALAHAD/../deps/lib:/home/runner/work/GALAHAD/GALAHAD/../CUTEst/lib /home/runner/work/GALAHAD/GALAHAD/builddir/pastixt_double
----------------------------------- stdout -----------------------------------
  eps    9.9999999999999998E-013
double
  size of row, col, val =            7           7           7
----------------------------------- stderr -----------------------------------
ischedInit: The thread number has been automatically set to 2

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

Backtrace for this error:
#0  0x7f4f19423970 in ???
#1  0x7f4f19422ad5 in ???
#2  0x7f4f1904251f in ???
#3  0x7f4f190a53fe in ???
#4  0x7f4f18ef8d5f in ???
#5  0x7f4f18fe2fba in ???
#6  0x7f4f18f84b0d in ???
#7  0x7f4f18f94c6b in ???
#8  0x7f4f1976a018 in ???
#9  0x561ca02f7c36 in test_pastix
	at ../src/external/pastix/pastixt.F90:226
#10  0x561ca02f8475 in main
	at ../src/external/pastix/pastixt.F90:7
==============================================================================

[amontoison]

If I compile PASTIX without SCOTCH, the test pastixt in double precision works. 🤔

[nimgould]

Oh, I should say that I took the advice from the Pastix page to set scotch on and metis off. This works for me. I suppose that it might well be a metis version clash (again!)

[amontoison]

@nimgould
I have this compilation error, I don't understand how a .mod can be broken?!

 [2842/4669] Compiling Fortran object pastixt_quadruple.p/src_external_pastix_pastixt.F90.o
FAILED: pastixt_quadruple.p/src_external_pastix_pastixt.F90.o 
gfortran -Ipastixt_quadruple.p -I. -I.. -Iinclude -I../include -I../src/dum/include -I../src/metis/include -Isrc/ampl -I../src/ampl -I../../deps/modules -I../../CUTEst/modules -I../../deps/include/spm -I../../deps/include/pastix -Ilibgalahad_quadruple.so.p -fdiagnostics-color=always -D_FILE_OFFSET_BITS=64 -O0 -g -fopenmp -DLANCELOT_USE_MA57 -cpp -DREAL_128 -DGALAHAD_BLAS -DGALAHAD_LAPACK -DDUMMY_QMUMPS -DDUMMY_MKL_PARDISO -DDUMMY_PARDISO -DDUMMY_WSMP -DDUMMY_MPI -Jpastixt_quadruple.p -o pastixt_quadruple.p/src_external_pastix_pastixt.F90.o -c ../src/external/pastix/pastixt.F90
f951: Fatal Error: Reading module ‘libgalahad_quadruple.so.p/spmf_interfaces_quadruple.mod’ at line 304 column 51: Expected left parenthesis

I would like to test GALAHAD with PaStiX v6.4.0.
I was able to cross-compile this version, which means that all Julia users can use it in the next release of GALAHAD_jll.jl / GALAHAD.jl if we want.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 66.66667% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 26.92%. Comparing base (0103718) to head (dcc0966).

Files with missing lines	Patch %	Lines
include/galahad_sls_defaults_ls.h	66.66%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #319   +/-   ##
=======================================
  Coverage   26.92%   26.92%           
=======================================
  Files         164      165    +1     
  Lines       29202    29208    +6     
  Branches     8804     8806    +2     
=======================================
+ Hits         7863     7865    +2     
- Misses      18767    18769    +2     
- Partials     2572     2574    +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[nimgould]

I have no idea. As you say, a bug in a mod file is unusual, I have never seen one before. Certainly there is no issue here with the quad test using the dummy pastix (but then I don't know if the real one supports 128 bit reals)

[amontoison]

We need to support PASTIX 6.4.0 if we want to use / test it. The interface in GALAHAD is for an older version and they changed the API...

[nimgould]

Sigh ... do these people not understand how much it annoys their users if every minor version is different. This will be the third change in a year.

I'll have a look when I get a chance, do you know what changed?

[amontoison]

Nick, I have an issue when we deallocate the vectors in pastixt.
Do you have any idea why?
https://github.com/ralna/GALAHAD/blob/master/src/external/pastix/pastixt.F90#L234

[nimgould]

It is not obvious to me. I'll try using a STAT = status for each allocate/deallocate

[nimgould]

I am happy with this, both single and double precision stand-alone and sls tests pass. Thank you, Alexis!

[nimgould]

Alas, it doesn't work with in64, but this is actually a pastix error as compiling with the -DPASTIX_INT64=ON flag fails as it requires 64-bit Scotch ... which is available as part of the scotch package. Seemingly Cmake just doesn't find it

[nimgould]

Update: adding
-DSCOTCH_INCLUDE_DIRS="/usr/include/scotch-int64"
-DSCOTCH_DIR="/usr/lib/x86_64-linux-gnu/scotch-int64"
-DCMAKE_PREFIX_PATH="/usr/lib/x86_64-linux-gnu/scotch-int64"

to the pastix Cmake built the 64-bit version. But now,

at line 4443 of file sls.F90
Fortran runtime error: Index '6' of dimension 1 of array 'data' outside of expected range (1:1)

which seems to be a result of some data misalignment following
CALL spmAlloc( data%spm )
...
CALL spmGetArray( ...

Inserting

write(6,*) SIZE( data%PTR ), SIZE( data%matrix%PTR )

before the failing line, I see

1 6

while for the 32-bit integer it is

6 6

Thus it looks like spmGetArray is not allocating the correct size

[amontoison]

Alas, it doesn't work with in64, but this is actually a pastix error as compiling with the -DPASTIX_INT64=ON flag fails as it requires 64-bit Scotch ... which is available as part of the scotch package. Seemingly Cmake just doesn't find it

I compile PaStiX with only METIS for the CI and it works (-DPASTIX_ORDERING_SCOTCH=OFF and -DPASTIX_ORDERING_METIS=ON).
I also got errors locally with SCOTCH but I am 99% sure that the issue is not in GALAHAD.

The build that failed for the CI was on Mac and I compile PaStiX only on linux (unrelated to this PR).
The failure was related to rqst_c_quadruple and it is just because it reached the time-out.

We should send the bug to Mathieu for SCOTCH + PaStiX.

[nimgould]

This works with the makefile build. I am not sure I understand the "conflicts" message above, the version I get from a pull on pastix_ci seems fine