DoubleFloats.jl is available for v0.7.
Arithmetic and elementary functions. Expect 85 bit accuracy using Double64.
| op | speedup |
|---|---|
| + | 11x |
| * | 18x |
| \ | 7x |
| trig | 3x-6x |
Is there a summary somewhere of how this compares to other options besides BigFloat?
GitHub - saschatimme/HigherPrecision.jl: High precision replacements for Float64 Looks like that may only support 0.6 (for now?) and require compiling Julia from source.
https://github.com/JuliaArbTypes/ArbFloats.jl There’s a 0.7 branch but it was failing.
HigherPrecision.jl is v0.6 and its author is using DoubleFloats.jl. The 0.7 branch for ArbFloats.jl was experimental and discontinued. Wait for JuliaCon.
OK got it, thanks!
now v1 ready, some buglets addressed and a few more features added:
d64"0.3" works like big"0.3", maxintfloat, div, cld, fld, rem, mod available.
DoubleFloats v0.7 is here with some docs.
v0.7 exports ComplexD64 with Double64 (and so for 32, 16 bits).
bikeshedding: Do you plan to eventually rename the types to Float128, Complex128,etc? Using “double” to mean “floating-point” has always seemed to me like a misnomer.
The name comes from the software technology, doubling the floating point format by pairing two floats to form an extended precision value. To your point, here “double” does not mean floating-point, it means two-of.
These types are not IEEE 754 compliant so naming them Float128, Complex128 is not an option (this has been suggested, discussed by the core Julia developers and not used for that reason). Julia does expect to have a legit Float128 type sometime (it has an Int128 type already). One difference is that standard Float128s have a larger exponent range than do the Double64s – their exponents are the same as a Float64.
Thanks for the explanation.
DoubleFloats.jl v0.9.0 is released.
The docs are here. Let me know if something in the docs needs more clarification.