Add possibility to pass float values or range to --parameter-scan

[guilhermeleobas]

When using hyperfine, I am missing the possibility of specifying a range or floating point values on --parameter-scan:

hyperfine --parameter-scan i 1 10 2 "sleep {i}"

or

hyperfine --parameter-scan i 0.1 1.0 "sleep {i}"

[sharkdp]

Thank you for the feedback. What CLI would you propose for this? We can't let --parameter-scan take an optional fourth argument.

[guilhermeleobas]

I was thinking in use --parameter-scan with an optional fourth argument but since that is not possible, maybe use --parameter-range-scan with the same semantics of the former but with a parameter indicating the range increment.

edit: Can --parameter-scan take a float parameter?

edit2: Or instead of adding a floating point option, add a way to format the variable generate by --parameter-scan using the printf notation. For instance:

> hyperfine --parameter-scan var 1 5 --format "0.%.f"  "echo {var}"
0.1
0.2
0.3
0.4
0.5

[sharkdp]

Another possibility would be to keep --parameter-scan as it is, and add a new --parameter-step-size <step> (or --step-size <step>) option.

[sharkdp]

edit2: Or instead of adding a floating point option, add a way to format the variable generate by --parameter-scan using the printf notation.

You can already do this - sort of.

With bc:

hyperfine -P var 1 5 'sleep $(echo "{var}/10" | bc -l)'

With zsh:

hyperfine -S zsh -P var 1 5 'sleep $(({var}.0/10))'

[piyushrungta25]

@sharkdp Opened #203 to add --parameter-step-size argument.

[sharkdp]

@piyushrungta25 has added support for --parameter-step-size in #203. I think it would be great if we could also add support for floating point values, such that we can run things like:

hyperfine --parameter-scan var 0.0 1.6 --parameter-step-size 0.1 …

[piyushrungta25]

I think it would be great if we could also add support for floating point values

I see two issues with this.

1. Choosing default value of the step. Something like 0.1 as the default doesn't seem very intuitive. One thing we can do is make the step size argument required for ranges with float bounds.

2. The range iterators in std library(or even num crate) does not take float parameters. So if we are okay with rolling our own iterator, I can come up with an implementation.

[sharkdp]

1. Choosing default value of the step. Something like 0.1 as the default doesn't seem very intuitive. One thing we can do is make the step size argument required for ranges with float bounds.

Good point. I would definitely not choose 0.1 as the default, yes. Requiring the step size to be set for float arguments sounds like a good idea to me!

2. The range iterators in std library(or even num crate) does not take float parameters. So if we are okay with rolling our own iterator,

I'd be okay with this, as long as it comes with proper unit tests. I could imagine that the reason for the absence in the standard library is the following unfortunate property of floating point numbers:

Imagine that someone chooses --parameter-scan var 0.0 1.0 and a step size of 0.1. I think that most users would expect the 1.0 boundary to be inclusive. However, due to rounding errors, a naive implementation could miss the 1.0 boundary due to 0.0 + 10 * 0.1 being too large.

I guess this could probably be circumvented by computing N = (end - start) / stepsize and testing if it is close to an integer number. In this special case, we could round N to that nearby integer and make sure that we perform exactly N incrementations (or N+1 benchmarks).

I can come up with an implementation.

👍

[piyushrungta25]

unfortunate property of floating point numbers

Perhaps a decimal data type would be more appropriate here? This would prevent substitutions like echo 0.30000000000000004 from happening.

[sharkdp]

This has been fully implemented by @piyushrungta25 🎆

[sharkdp]

Released in hyperfine v1.7.0