nimforum mirror - high precission -> f1024

i've doing some research on nim high precision resources, my goal is working with f1024 numbers,

coded an almost direct c long double wrapper f128 useful for some cases: https://github.com/rforcen/nim/blob/main/ld.nim

and also hpa wrapper (real+complex) for f1024 and beyond,

hpa enables setting fixed precision in Makefile variable,

MANTSIZE_MACRO= -DXMANTISSA_SIZE=62 (default 7) in GNUmakefile

next step is working with GMP/mpfr as a fast and proven solution that lets you set dynamic precision

is there any other alternative?

Nice!; Minifloat AKA half-float would be nice too. :)

found some source for f16 @found some source for f16 https://github.com/artyom-beilis/float16/blob/master/c_src/float16.c i'm curious abort real performance compared w/f32

Cool!, maybe f8 too.

i've used a c source to support f16 as a cshort, directly embedded it w/emit and coded a basic wrapper

in terms of performance it's about 15% slower than nim float (f64)

For these things the new user-defined literals come in handy. See https://github.com/nim-lang/bigints/blob/master/src/bigints.nim#L890 and https://github.com/nim-lang/bigints/blob/master/src/literals.nim for how to introduce them so that older Nim versions are supported too.

Mirror of forum.nim-lang.org