Windows fixes about emdna

[nicocvn]

Trace of the only test that fail on Windows (in BpGeometryFunctionsTest.dXiMatrix):

C:\Users\velin\Documents\emDNA.git\emDNA\tests\utest_BpGeometryFunctions.cpp(67): error: The difference between mm_dXi2(i,j) and dXi_mats[1](i,j) is 1.4554329963445412e-15, which exceeds Real(1E-15), where
mm_dXi2(i,j) evaluates to -0.011379443932765363,
dXi_mats[1](i,j) evaluates to -0.011379443932763907, and
Real(1E-15) evaluates to 1.0000000000000001e-15.

So missing the target by ~0.455e-15 ... might be some issues with inputs and rounding; need to take a look.

[nicocvn]

Instructions:

• need to run build+install from the "Developer prompt" installed with Visual Studio
• config: cmake -G"Visual Studio 16 2019" -A x64 -S .\emDNA.git\ -B .\emDNA-build\ -DCMAKE_BUILD_TYPE=Release (for latest visual studio otherwise generator needs to be adjusted)
• then same commands

[nicocvn]

Link about fp contract in MSVC: https://docs.microsoft.com/en-us/cpp/build/reference/fp-specify-floating-point-behavior?view=msvc-160

Similarly Clang and GCC have the -ffp-contract option

[nicocvn]

So after further investigation I do not think the failing test on Windows is a big issue; here are the numbers for the value that fails with MSVC:

• GCC on Linux and Apple Clang on macOS both compute: -0.011379443932764877
• on Windows using Visual Studio 16 2019 toolchain the value is computed as: -0.011379443932763907

The difference is 9.7e-16 (the machine epsilon on the test setup is 2.22045e-16).

Looking at the computation of that value it is pretty heavy on trigonometric functions and expressions that are likely to be sensitivity to FMA type of optimization (fuse multiply add). It is likely that GCC/Clang and MSVC perform different types of optimizations.

Godbolt link: https://godbolt.org/z/W5nWTd MSVC seems to call some optimized sse2 library for cos and sin while GCC relies on a single call to get sin and cos.

Note that the reference value in the test was computed using Mathematica on macOS so it is also possible (likely) that this introduces some bias.

[nicocvn]

For now the test is disabled on MSVC.