Something I find myself doing often is write an algorithm in Float64 (without bothering with types, as I would in Matlab say), and then try to change the default precision to Float32 or BigFloat (to test the sensitivity to numerical errors, to see if I can gain in performance…) Julia makes it very e…

I played around with it, and it is pretty feasible to write a fairly capable macro that not only changes floating-point literals, but also handles functions like rand and ones, while preserving cases like rand(Float64) that explicitly specify a precision. I’ve posted a little (unregistered) package…

[image] antoine-levitt: there’s no particular reason to impose Float64 as default rather than Float32 after all, other than this is what’s commonly done (e.g. in numpy, matlab). You make it sound like a historical accident, but it isn’t. For scientific computation, the trade-off is usually be…

You make it sound like a historical accident, but it isn’t. For scientific computation, the trade-off is usually between economizing on CPU and memory use, and worrying about conditioning & catastrophic cancellation. Float64 turns out to be a sweet spot for a lot of numerical work. That’s why many …

[image] antoine-levitt: a default floating point type Yes and it’s not what I prefer, it is what it really is and it makes a fundamental difference as for what you can do and what you can’t. [image] antoine-levitt: there is no fundamental reason why it should be double precision I don…

[image] yuyichao: As I said, the macro that does local change to your code is perfectly fine. However, because it’s a type and not a precision and so you must not change the return types of other functions. Again, why is this exactly? Isn’t this what defaultRNG() does in https://github.com/Ju…

An illustrative example of why it could be harmful to redefine functions: julia> Base.similar(A::Matrix, ::Type{T}, dims::Dims{2}) where T = similar(A, Int) julia> rand(2,2)*rand(2,2) ERROR: InexactError() Stacktrace: [1] convert(::Type{Int64}, ::Float64) at ./float.jl:679 [2] matmul2x2!(::Array…

This I get : it is not redefining functions in itself which is problematic, it is that the redefined function will be, well, redefined, and cause bad interactions with other functions calling that function (what I call “confuse the expectations” above). My hope is that library code does not call arr…

Sorry for hopping in, but what is wrong with using the functions’ overloads to have ::Type{T} as their first argument? Above, you mentioned about C and some historical reasons of why float corresponds to single precision. Yet, when you write in C/C++ 1.0 it is double. float x = 1.0 uses implicit co…

[image] antoine-levitt: Isn’t this what defaultRNG() does in https://github.com/JuliaLang/julia/pull/23205 ? Or is that PR invalid? That is roughly what the PR does and the PR is invalid. Though that PR doesn’t change the value types returned by rand so it’s not as bad. [image] antoine-le…

For this task, I am using a patched version of julia which redefines the floating point literals. The conversion of all literals is done at the parser level and one can switch the default floating point precision on the fly. The base julia library has never been designed to handle this type con…

Experiments with this (i.e. creating a fork) is fine. As long as it is made clear that (as I mentioend above) this is not at all supported, should not be used in remotely serious work and no attempt will be made to maintain compatibility with it.

Change default precision

Internals & Design

greg_plowman October 25, 2017, 8:42pm 52

Cross-referencing a nice use for this: Performing arithmetic operations in quadruple precision - #14 by greg_plowman

Topic		Replies	Views
Default Real (Float) kind? Internals & Design float	50	1221	June 7, 2025
ANN: ChangePrecision.jl Tooling package , announcement	11	2422	March 27, 2019
Letting users change the default types Internals & Design	13	1136	August 23, 2021
Let julia return default Int32 or Float32-type data? Internals & Design question	1	668	May 12, 2020
Disable promotion from Float64 to BigFloat General Usage	20	456	June 2, 2025

Change default precision

Related topics