Normalisation fixes

[dschwoerer]

Both commits are required to have the 2D-production test for Hermes-3. I tested locally only the 2nd patch, and in CI the first patch only. Both failed on there own.

[github-actions]

clang-tidy made some suggestions

[github-actions]

warning: braces around scalar initializer [clang-diagnostic-braced-scalar-init]

Suggested change

	.g11_mul{true},
	.g11_mul true,

[dschwoerer]

Not sure what clang-tidy is trying to do here, but g++ 15.2.0 is not happy with the suggested change.

[ZedThree]

I don't understand this at all -- why is this not equivalent? Is it just loss of precision?

[dschwoerer]

Yes, and that seems to cause the failures in Hermes-3. If you think we should not do this, and just update the expected results, then we can do that, too. But I thought it might be good if we do not change results without a somewhat good reason, and the _mul additions do not seem like a huge addition of code.

[dschwoerer]

For the record, without these changes, Hermes-3 fails with:

      Start 14: 2D-production
14/20 Test #14: 2D-production ..............................***Failed   12.69 sec

libgomp: Invalid value for environment variable OMP_NUM_THREADS: 
2D-Production test: download took 0.00 seconds
2D-Production test: zip extraction took 1.77 seconds
2D-Production test: simulation took 2.08 seconds
2D-Production test: loading results took 6.95 seconds
2D-production test: Failed in 11.24 seconds
2D-production test failures:

inner_lower_target [ddt(Pe)]:

Not equal to tolerance rtol=1e-05, atol=1e-08

Mismatched elements: 1 / 16 (6.25%)
Mismatch at index:
 [8]: 0.03679871893512578 (ACTUAL), 0.036799153418756995 (DESIRED)
Max absolute difference among violations: 4.34483631e-07
Max relative difference among violations: 1.1806892e-05
 ACTUAL: array([-0.050165, -0.067434, -0.095253, -0.140382, -0.18905 , -0.238565,
       -0.344202, -0.562915,  0.036799, -0.091465, -0.586319, -0.396325,
       -0.224284, -0.064263,  0.009454,  0.049401])
 DESIRED: array([-0.050165, -0.067434, -0.095253, -0.140382, -0.18905 , -0.238565,
       -0.344202, -0.562915,  0.036799, -0.091465, -0.586319, -0.396325,
       -0.224284, -0.064263,  0.009454,  0.049401])

sol_ring [ddt(Pe)]:

Not equal to tolerance rtol=1e-05, atol=1e-08

Mismatched elements: 1 / 50 (2%)
Mismatch at index:
 [0]: 0.03679871893512578 (ACTUAL), 0.036799153418756995 (DESIRED)
Max absolute difference among violations: 4.34483631e-07
Max relative difference among violations: 1.1806892e-05
 ACTUAL: array([ 0.036799, -1.645953, -1.76924 , -1.600019, -1.298552, -0.991483,
       -0.854056, -0.782628, -0.775318, -0.861349, -0.864964, -0.775591,
       -0.790239, -0.855372, -0.999383, -1.299618, -1.591131, -1.75628 ,...
 DESIRED: array([ 0.036799, -1.645953, -1.76924 , -1.600019, -1.298552, -0.991483,
       -0.854056, -0.782628, -0.775318, -0.861349, -0.864964, -0.775591,
       -0.790239, -0.855372, -0.999383, -1.299618, -1.591131, -1.75628 ,...

So it seems the rounding errors in doubles, cause differences in 1e-5 range.

[dschwoerer]

There are rounding differences between B/Bnorm and B * (1/Bnorm) - they are numerically not the same. It seems such differences cause enough of a differences to cause larger differences in the end, i.e. the 2D-production test is very sensitive to such grid details

The PR address this by normalising just the same way that Hermes-3 used to normalise the metrics