RAMBO is unstable in single precision

[valassi]

RAMBO is unstable in single precision

I just realised that for gg_tt in single precision the compareTest test gives deviations larger than the 3% tolerance for some specific builds (debug mode with gcc)

[avalassi@itscrd90 gcc11/usr] /data/avalassi/GPU2023/madgraph4gpuX/epochX/cudacpp/gg_tt.mad/SubProcesses/P1_gg_ttx> make cleanall; make -f cudacpp.mk -j debug FPTYPE=f 
...
[avalassi@itscrd90 gcc11/usr] /data/avalassi/GPU2023/madgraph4gpuX/epochX/cudacpp/gg_tt.mad/SubProcesses/P1_gg_ttx> ./runTest.exe 
...
[----------] 1 test from SIGMA_SM_GG_TTX_GPU/MadgraphTest
[ RUN      ] SIGMA_SM_GG_TTX_GPU/MadgraphTest.CompareMomentaAndME/0
INFO: Opening reference file ../../test/ref/dump_CPUTest.Sigma_sm_gg_ttx.txt
MadgraphTest.h:298: Failure
Value of: momentumErrors.str().empty()
  Actual: false
Expected: true

particle 2      component 3
         madGraph:   1.079908162355423e-01
         reference:  1.041966527700424e-01
         rel delta:  3.641348704695702e-02 exceeds tolerance of 2.999999932944775e-02
Google Test trace:
MadgraphTest.h:280: In comparing event 46 from iteration 0
   0  7.500000000000000e+02  0.000000000000000e+00  0.000000000000000e+00  7.500000000000000e+02
ref0  7.500000000000000e+02  0.000000000000000e+00  0.000000000000000e+00  7.500000000000000e+02

   1  7.500000000000000e+02  0.000000000000000e+00  0.000000000000000e+00 -7.500000000000000e+02
ref1  7.500000000000000e+02  0.000000000000000e+00  0.000000000000000e+00 -7.500000000000000e+02

   2  7.500137329101562e+02 -7.050960083007812e+02  2.555634918212891e+02  1.079908162355423e-01
ref2  7.500000000000000e+02 -7.051112670898438e+02  2.555740966796875e+02  1.041966527700424e-01

   3  7.499924926757812e+02  7.051048583984375e+02 -2.555704498291016e+02 -1.049910709261894e-01
ref3  7.500000000000000e+02  7.051112670898438e+02 -2.555740966796875e+02 -1.041966527700424e-01

  ME  3.608205318450928e-01
r.ME  3.608229160308838e-01

[Thread 0x7fffeef02000 (LWP 4069063) exited]
[  FAILED  ] SIGMA_SM_GG_TTX_GPU/MadgraphTest.CompareMomentaAndME/0, where GetParam() = 0x675e60 (133 ms)

Note, this is only for debug builds..

[valassi]

I will probably fix this specifc case by increasing the tolerance from 3E-2 to 4E-2. However it may be a good idea to always have rambo in double precision, so as to test only the precision of the ME calculation?