Open valassi opened 1 year ago
This is getting really weird... I find a difference in the second significant digit???
*** (2-sse4) Compare CMADEVENT_CUDACPP x10 events.lhe to MADEVENT events.lhe reference (with dummy colors and helicities) ***
882c882
< 6 1 7.2E-07 0.6713289E+03 0.7546771E-02 0.9767560E-01
---
> 6 1 7.1E-07 0.6713289E+03 0.7546771E-02 0.9767560E-01
ERROR! events.lhe.cpp.10 and events.lhe.ref.10 differ!
I will put it to one significant digit, but I need to understand this with @oliviermattelaer and/or @zeniheisser
Note, I mentioned this in aaf58d9 and 64594ce, which were in valassi's branches. But these were eventually included in master during MR #499.
I have just implemented the tmad comparison of lhe files when cudacpp is run in float rather than double
I had to add an lheFloat.sh script that massages the lhe file to reduce the precision of only one field, the XWGTUP... which I guess comes itself from the estimation of a cross section somewhere?...
Anyway, for a first test on ggtt I had used five significant digits that was ok, but now I have run ggttg and here I get one 2.04 differing from 2.05, so not even three digits are enough, ie a difference at the percent level! This is clearly too much... to be understood better with @oliviermattelaer
Anyway: our plan is to move from "d" not to "f" (full float) but to a new situation (I will call it "m" for mixed?) where onlythe color algebra is in float, while the FFVs are in double. One of the things to do there will be to increase the precision of these comparisons to sensible levels... a onepercent difference is clearly no good for physics I think...
Todod will be increase the precision for lhe comparisons (now two significant digits only), but also for xsec comparisons (now relative diff up to 2E-4 which is also too high)