Marc:
You should probably set a C compiler option enable SSE/SSe2/... instructions on x86 and x86-64 processors when they're available. (That might be the default in gcc for x86-64.) Right how Gambit sets the rounding precision to "double" to remove some problems with double roundings (to "extended" and then to "double"), but there still can be problems because of the differences in the exponent range between the extended precision floats that the registers in x87 use and the double precision floats that are boxed (or put in f64vectors, ...). Using the SSE instruction set would match the model that Gambit uses for floating-point.
Brad
Afficher les réponses par date