I posted this and it picked up steam over night, so I thought I'd add how I'm using it:<p>I work on 3D/4D math in F#. As part of the testing strategy for algorithms, I've set up a custom agent with an F# script that instruments Roslyn to find FP and FP-in-loop hotspots across the codebase.<p>The agent then reasons through the implementation and writes core expressions into an FPCore file next to the existing tests, running several passes, refining the pres based on realistic caller input. This logs Herbie's proposed improvements as output FPCore transformations. The agent then reasons through solutions (which is required, Herbie doesn't know algorithm design intent, see e.g. this for a good case study: <a href="https://pavpanchekha.com/blog/herbie-rust.html" rel="nofollow">https://pavpanchekha.com/blog/herbie-rust.html</a>), and once convinced of a gap, creates additional unit tests and property tests (FsCheck/QuickCheck) to prove impact. Then every once in a while I review a batch to see what's next.<p>Generally there are multiple types of issues that can be flagged:<p>a) Expression-level imprecision over realistic input ranges: this is Herbie's core strength. Usually this catches "just copied the textbook formula" instance of naive math. Cancellation, Inf/NaN propagation, etc. The fixes are consistently using fma for accumulation, biggest-factor scaling to prevent Inf, hypot use, etc.<p>b) Ill-conditioned algorithms. Sometimes the text books lie to you, and the algorithms themselves are unfit for purpose, especially in boundary regions. If there are multiple expressions that have a <60% precision and only a 1 to 2% improvement across seeds, it's a good sign the algo is bad - there's no form that adequately performs on target inputs.<p>c) Round-off, accumulation errors. This is more a consequence of agent reasoning, but often happens after an apparent "100% -> 100%" pass. The agent is able to, via failing tests, identify parts of an algorithm that can benefit from upgrading the context to e.g. double-word arithmetic for additional precision.
> I work on 3D/4D math in F#. As part of the testing strategy for algorithms, I've set up a custom agent with an F# script that instruments Roslyn to find FP and FP-in-loop hotspots across the codebase.<p>I don't know if there is an equivalent in Roslyn, but in Julia you can have the agent inspect the LLVM output to surface problems in hot loops.
Really cool, numerical stability can be tricky as errors can accumulate with each operation and suddenly 53 bits of precision is not enough.<p>Also nice to see an article thats not about AI or politics
Some trivial cases produce... interesting results.<p>For x in [−1.79e308, 1.79e308]:<p>Initial Program: 100.0% accurate, 1.0× speedup<p><pre><code> def code(x):
return math.sqrt((x + 1.0))
</code></pre>
Alternative 1: 67.5% accurate, 5.6× speedup<p><pre><code> def code(x):
return 1.0</code></pre>
That <i>does</i> make sense, because a half of all available fp numbers are less than 1 in their magnitude. In particular there should be a plenty of numbers x such that |x| << 1 so x + 1 ~= 1; in fact, the proportion should be just shy of 50%.
But I guess using the density distribution of floating points is rarely useful in a problem. Your actual distribution will almost surely be way different. Imo, the tool presented here should provide a way to manually provide a custom density function (with some common presets like uniform and normal distributions).
Author here! Yes, the float distribution isn't what you want in practice, but distribution selector isn't really the right thing either, because a low probability bad result can still be pretty bad! Hence the range selector; the float distribution is good at picking extreme values that trigger FP error.<p>We usually recommend looking for 90%+ accuracy or carefully examining the accuracy plot
That is indeed one of the problems with IEEE floats. There are only 10^80 atoms in the universe, and a Planck length is 1^-60th of the radius of the universe. But 64-bit floats have an <i>absurd</i> range of over 10^±300! Worse than that, notice that there are as many bit patterns in the never-used range between 10^300 and 10^301 as there are in the super-important range between 1 and 10! Super wasteful. Not to mention the quadrillions of values reserved to represent "NaN"...<p>This is one of the problems that alternative formats such as the Posit aim to solve. It's quite interesting: I've got an implementation in rust here if you want to play with it <a href="https://github.com/andrepd/posit-rust" rel="nofollow">https://github.com/andrepd/posit-rust</a>
Note that the logarithmic distribution of float density is also key to certain kinds of efficient hardware float implementations, because it means you can use the fixed mantissa bits alone as table indices. Unums have proven difficult to build efficient HW implementations for.<p>IEEE floats have a few warts like any other 1980s standard, but they're a fantastic design.
Not really. 1+x/2, however, would be a good approximation to sqrt(1+x) for small (in absolute value) x.
How about:<p>(* Mathematica Notation, Assume x>0 <i>)
If[ x < 10^(-10), 1+ x/2, (</i> order x^2 error <i>)
If[ x> 10^10, Sqrt[x], (</i> order 1/Sqrt[x] error <i>)
(</i>else<i>) Sqrt[x +1] ] ]
(</i> I guess the If statements take too much time. *)
I resurrected the Rust Herbie lint (now using dylint) a while ago: <a href="https://github.com/urschrei/herbie-lint" rel="nofollow">https://github.com/urschrei/herbie-lint</a>
How useful is this when you are using numbers in a reasonable range, like 10^-12 to 10^12?
Generally I try to scale my numbers to be in this range, whether by picking the right units or scaling constraints and objectives when doing nonlinear programming/ optimization.<p>Like looking at this example,<p><a href="https://herbie.uwplse.org/demo/b070b371a661191752fe37ce0321c3e813691aa6.0cc3b6492c83efca5bd11399e0830e7873c749d9/graph.html" rel="nofollow">https://herbie.uwplse.org/demo/b070b371a661191752fe37ce0321c...</a><p>It is claimed that for the function f(x) =sqrt(x+1) -1<p>Accuracy is increased by from 8.5%
accuracy to 98% for alternative 5
Which has f(x) = 0.5x<p>Ok so x=99, the right answer is sqrt(100) -1 = 9<p>But 0.5 * 99 = 49.5 which doesn't seem too accurate to me.
The precondition on the link you shared has -1 <= x && x <= 1, so 99 is way outside of that range. But even so, testing for x=1, which is supposed to be inside that range, 0.5 doesn't seem tolerably close to 0.4142.
I have a suspicion that the accuracy number is the mean of accuracies over all valid floats in the range (or something approximating that), which is going to be weighted towards zero where the accuracy is higher, and perhaps where sqrt near 1 has some artefacts.
The input should be the range and the distribution of probability on this range. Intuitively we have a tendency to assume an uniform probability for range [-1, 1] which is not the case if we check every doubles.
It is, there's a page in the documentation about how errors are defined. Let me also add: Herbie generally gives the most accurate option it found first, and then the other stuff might be useful for speed (0.5x is way faster than two square roots and a divide!) but it's not as accurate
Check the specification at the top. The range for x is [-1, 1]. For the range you provided the accuracy of the 0.5x alternative is reported as only 33%: <a href="https://herbie.uwplse.org/demo/570b973df0f1f4a78fe791858038a84f635de8c6.0cc3b6492c83efca5bd11399e0830e7873c749d9/graph.html" rel="nofollow">https://herbie.uwplse.org/demo/570b973df0f1f4a78fe791858038a...</a>
You're right I misread the graph. That said though I have played around with Herbie before, trying it out on a few of the more gnarly expressions I had in my code (analytical partial derivatives if equations of motion if launch vehicle in rotating spherical frame) and didn't see much appreciable improvement over the expected range of values, but then again I didn't check every single one.<p>What would be cool is if you could some how have this kind of analysis done automatically for your whole program where it finds the needle in the haystack expression that can be improved, assuming you gave expected ranges for your variables
Author here. I've got a few papers about this problem (including one in submission), but it is very very hard to do, especially with acceptable overhead. The state of the art is maybe 100x overhead.
I wonder who decided to use a step function for the speed accuracy plot. They must have thought the convex hull would be wrong because you can't really make linear combinations of algorithms (you could, but you'd have to use time not speed to make it linear). So I get <i>why</i> you would use step functions, but the step is the wrong way around. The current plot suggests accuracy doesn't drop if you need higher speeds
This is an awesome piece of software, one of my favorite little pieces of magic. Finding more precise or more stable floating point formulas is often arduous and requires a lot of familiarity with the behavior of floats. This finds good formulas completely automatically. Super useful for numerical computation.
I wonder, is there a way to only request reformulations that don’t involve branches? The tool already seems quite nice, but that might be a good feature.<p>Also, I’m not sure I understand the speedup. Is it latency or throughput?
Author here. The speed up is modeled throughput, though the model is relatively naive. It's possible to disable branches by turning off the regimes flag, see <a href="https://herbie.uwplse.org/doc/1.0/options.html" rel="nofollow">https://herbie.uwplse.org/doc/1.0/options.html</a>
Thanks for reminding me to try this for audio approximations. :)
I don't quite understand how they define "accuracy".
[flagged]
You didn’t even look at what the tool does, did you?<p>> If the issue is that people write bad floating-point expressions, a code-writing tutorial would be a better solution.<p>Yeah you are just criticizing this without even looking at it. Shame.
Did you read what this does? Because I get the feeling you didn’t…<p>This isn’t a library, you don’t include in your application, and it doesn’t try to replace an understanding of floating point issues on the programmers part.<p>Is this comment written by AI?