Lower accuracy and faster math

[mattfidler]

Because ode optimization may be spending too much time in exp() when it could be reduced to match the accuracy of the ode solving method.

[mattfidler]

https://github.com/simonpf/fastexp

[mattfidler]

Julia uses the fastmath from the compiler

https://github.com/JuliaLang/julia/blob/master/base/fastmath.jl

[mattfidler]

It also means that the protection options should be turned off because NaN and infinite results cannot be used/found.

[mattfidler]

And functions like is.finite is.nan will only return a single type of value

[mattfidler]

https://github.com/cran/RVowpalWabbit/blob/65fd42ff945d7621a6c0b88d910bcef1122ad47e/ChangeLog#L231-L232

[mattfidler]

Upon further investigation, this option is available already, but not well known (since I didn't know it either).

You can use:

options(rxode2.compile.O="fast")

And it will turn on fast math as well as some other non-IEEE compliant options.

Perhaps this should make it into the speed vignette.

[mattfidler]

This works for both clang and gcc compiler.s

[mattfidler]

Probably intel too. Add to Speed vignette.

[mattfidler]

This is what I got on my unbuntu flavored linux laptop:

So it is faster but not hugely different. It could add up with a long running problem though...