Reduce machine epsilon and add a distance epsilon #560
Conversation
|
This requires a review of uses of |
There was a problem hiding this comment.
@ftessier I have tested this and I get the warning:
warning: left shift count >= width of type [-Wshift-count-overflow]
This happens in bit-wise shift expressions when the number being shifted (1) is recognized as an int which is only 32 bits. If I change 1<<50 to 1ULL<<50 the warning disappears. It is probably compiler related since in principle there shouldn't be a warning. I am using here at home gcc 7.5.0.
|
Oh wow, good catch, I don't think there is any harm in using |
|
Rebased on develop and added fixed |
Reduce the value of the egs++ floating-point epsilon from 10e-10 to a value that is 4 times (2 bits away) from the true machine epsilon, and appropriately for single and double precision versions of the code. Add a distanceEpsilon constant that is independent of the machine epsilon, and much larger, but still entirely sufficient for physical dimensions (well below atomic dimensions in double precision). By default, the base geometry sets its boundaryTolerance to this constant. Its value is close to the previous value of epsilon (1e-10), which should curb any change in floating point error triggers in the geometry library at this point. Earlier in #177, the epsilon constant was introduced to remove hard-coded floating point comparison guards. The 1e-10 value was chosen because it matched most of the hard-coded values in the geometry library. However, this epsilon is now starting to be used as a general float guard (see #525 for example), where 1e-10 is not appropriate, as it effectively reduces the floating-point precision of the computation. Hence the need for independent epsilon values for computations and for geometrical purposes.
|
Rebased on develop. |
Reduce the value of the egs++ floating-point epsilon from 1e-10 to
a value that is 4 times (2 bits away) from the true machine epsilon, and
appropriately for single and double precision versions of the code.
Add a distanceEpsilon constant that is independent of the machine
epsilon, and much larger, but still entirely sufficient for physical
dimensions (well below atomic dimensions in double precision). By
default, the base geometry sets its boundaryTolerance to this
constant. Its value is close to the previous value of epsilon (1e-10),
which should curb any change in floating point error triggers in the
geometry library at this point.
Earlier in #177, the epsilon constant was introduced to remove
hard-coded floating point comparison guards. The 1e-10 value was chosen
because it matched most of the hard-coded values in the geometry
library. However, this epsilon is now starting to be used as a general
float guard (see #525 for example), where 1e-10 is not appropriate, as
it effectively reduces the floating-point precision of the computation.
Hence the need for independent epsilon values for computations and for
geometrical purposes.