Bad performance of Spice3BenchmarkFourBitBinaryAdder dominating MSL 3.2.1 tests

[Adrian Pop]

The one model Modelica.Electrical.Spice3.Examples.Spice3BenchmarkFourBitBinaryAdder dominates the time needed for testing MSL3.2.1 without producing good results.

The C runtime takes 6m40s to simulate that model. Afterwards the verification of 46/47 signals fails. Looking at the results, it appears possible that the C runtime doesn't appropriately adapt to the extreme short experiment(StopTime=1e-006, Interval=1e-009).

The C++ code generation takes 4m3s in the Templates phase (opposed to 20s for C code generation). C++ generates code for 171 algebraic loops (two code files for each of them). Many of the loops have a dimension of 1. See:

$ grep _dimAEq OMCpp*FourBit*Algloop*.cpp

Many of those are linear (checked Algloop 825, 920, 938, ...).

Shouldn't linear algebraic loops of dimension 1 be treated more efficiently?

[Niklas Worschech]

Maybe the problem is, that after the linear tearing the dimension of the new linear system is not checked. If the new system has the dimension one, it should not be generated a linear system for it.

[Rüdiger Franke]

Also linear systems of dim 2 should possibly be solved directly using Cramer's rule.

[Martin Sjölund]

The performance of the C++ code generation got even worse:

Notification: Performance of Templates: time 165.7/210.4, allocations: 85.97 GB / 93.65 GB, free: 164.1 MB / 2.633 GB

For C, it's 3.5 seconds.

[Niklas Worschech]

Is it possible to disable this test, until we fixed the problem for the code generation of 171 algebraic loops?

[Martin Sjölund]

Not really for the library coverage. We run the C runtime version for pull requests but since a long time ago disabled cppruntime.