Rust f32::atanh

Percentage Accurate: 93.1% → 97.2%
Time: 6.2s
Alternatives: 3
Speedup: 13.9×

Specification

?
\[\begin{array}{l} \\ \tanh^{-1} x \end{array} \]
(FPCore (x) :precision binary32 (atanh x))
float code(float x) {
	return atanhf(x);
}

function code(x)
	return atanh(x)
end

function tmp = code(x)
	tmp = atanh(x);
end

\begin{array}{l}

\\
\tanh^{-1} x
\end{array}

Sampling outcomes in binary32 precision:

Local Percentage Accuracy vs ?

The average percentage accuracy by input value. Horizontal axis shows value of an input variable; the variable is choosen in the title. Vertical axis is accuracy; higher is better. Red represent the original program, while blue represents Herbie's suggestion. These can be toggled with buttons below the plot. The line is an average while dots represent individual samples.

Accuracy vs Speed?

Herbie found 3 alternatives:

AlternativeAccuracySpeedup
The accuracy (vertical axis) and speed (horizontal axis) of each alternatives. Up and to the right is better. The red square shows the initial program, and each blue circle shows an alternative.The line shows the best available speed-accuracy tradeoffs.

Initial Program: 93.1% accurate, 1.0× speedup?

\[\begin{array}{l} \\ 0.5 \cdot \mathsf{log1p}\left(\frac{2 \cdot x}{1 - x}\right) \end{array} \]
(FPCore (x) :precision binary32 (* 0.5 (log1p (/ (* 2.0 x) (- 1.0 x)))))
float code(float x) {
	return 0.5f * log1pf(((2.0f * x) / (1.0f - x)));
}

function code(x)
	return Float32(Float32(0.5) * log1p(Float32(Float32(Float32(2.0) * x) / Float32(Float32(1.0) - x))))
end

\begin{array}{l}

\\
0.5 \cdot \mathsf{log1p}\left(\frac{2 \cdot x}{1 - x}\right)
\end{array}

Alternative 1: 97.2% accurate, 0.9× speedup?

\[\begin{array}{l} \\ 0.5 \cdot \mathsf{log1p}\left(\frac{x}{\left|-1 + x\right|} - \left(x - 1\right) \cdot x\right) \end{array} \]
(FPCore (x)
 :precision binary32
 (* 0.5 (log1p (- (/ x (fabs (+ -1.0 x))) (* (- x 1.0) x)))))
float code(float x) {
	return 0.5f * log1pf(((x / fabsf((-1.0f + x))) - ((x - 1.0f) * x)));
}

function code(x)
	return Float32(Float32(0.5) * log1p(Float32(Float32(x / abs(Float32(Float32(-1.0) + x))) - Float32(Float32(x - Float32(1.0)) * x))))
end

\begin{array}{l}

\\
0.5 \cdot \mathsf{log1p}\left(\frac{x}{\left|-1 + x\right|} - \left(x - 1\right) \cdot x\right)
\end{array}
Derivation

  1. Initial program 88.5%

    \[0.5 \cdot \mathsf{log1p}\left(\frac{2 \cdot x}{1 - x}\right) \]
  2. Add Preprocessing

  4. Applied rewrites95.4%

    \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\color{blue}{\frac{x}{1 - x} - \left(-\frac{x}{1 - x \cdot x}\right) \cdot \left(1 - x\right)}\right) \]

  5. Taylor expanded in x around 0

    \[\leadsto \frac{1}{2} \cdot \mathsf{log1p}\left(\frac{x}{1 - x} - \color{blue}{x \cdot \left(x - 1\right)}\right) \]
  6. Step-by-step derivation

    1. *-commutativeN/A

      \[\leadsto \frac{1}{2} \cdot \mathsf{log1p}\left(\frac{x}{1 - x} - \color{blue}{\left(x - 1\right) \cdot x}\right) \]

    2. lower-*.f32N/A

      \[\leadsto \frac{1}{2} \cdot \mathsf{log1p}\left(\frac{x}{1 - x} - \color{blue}{\left(x - 1\right) \cdot x}\right) \]

    3. lower--.f3294.5

      \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\frac{x}{1 - x} - \color{blue}{\left(x - 1\right)} \cdot x\right) \]

  7. Applied rewrites89.6%

    \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\frac{x}{1 - x} - \color{blue}{\left(x - 1\right) \cdot x}\right) \]

  9. Applied rewrites96.9%

    \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\color{blue}{\frac{x}{\left|1 - x\right|}} - \left(x - 1\right) \cdot x\right) \]

  10. Final simplification96.9%

    \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\frac{x}{\left|-1 + x\right|} - \left(x - 1\right) \cdot x\right) \]
  11. Add Preprocessing

Alternative 2: 96.9% accurate, 13.9× speedup?

\[\begin{array}{l} \\ 0.5 \cdot \left(x + x\right) \end{array} \]
(FPCore (x) :precision binary32 (* 0.5 (+ x x)))
float code(float x) {
	return 0.5f * (x + x);
}

real(4) function code(x)
    real(4), intent (in) :: x
    code = 0.5e0 * (x + x)
end function

function code(x)
	return Float32(Float32(0.5) * Float32(x + x))
end

function tmp = code(x)
	tmp = single(0.5) * (x + x);
end

\begin{array}{l}

\\
0.5 \cdot \left(x + x\right)
\end{array}
Derivation

  1. Initial program 87.7%

    \[0.5 \cdot \mathsf{log1p}\left(\frac{2 \cdot x}{1 - x}\right) \]
  2. Add Preprocessing

  3. Taylor expanded in x around 0

    \[\leadsto \frac{1}{2} \cdot \color{blue}{\left(2 \cdot x\right)} \]
  4. Step-by-step derivation

    1. lower-*.f3296.7

      \[\leadsto 0.5 \cdot \color{blue}{\left(2 \cdot x\right)} \]

  5. Applied rewrites96.7%

    \[\leadsto 0.5 \cdot \color{blue}{\left(2 \cdot x\right)} \]
  6. Step-by-step derivation

    1. Applied rewrites96.7%

      \[\leadsto 0.5 \cdot \left(x + \color{blue}{x}\right) \]
    2. Add Preprocessing

    Alternative 3: 10.8% accurate, 125.0× speedup?

    \[\begin{array}{l} \\ 0 \end{array} \]
    (FPCore (x) :precision binary32 0.0)
    float code(float x) {
	return 0.0f;
}

    real(4) function code(x)
    real(4), intent (in) :: x
    code = 0.0e0
end function

    function code(x)
	return Float32(0.0)
end

    function tmp = code(x)
	tmp = single(0.0);
end

    \begin{array}{l}

\\
0
\end{array}
    Derivation

    1. Initial program 85.3%

      \[0.5 \cdot \mathsf{log1p}\left(\frac{2 \cdot x}{1 - x}\right) \]
    2. Add Preprocessing

    4. Applied rewrites11.0%

      \[\leadsto 0.5 \cdot \mathsf{log1p}\left(\frac{\color{blue}{{\left(\left(-2 \cdot x\right) \cdot x\right)}^{3} \cdot -8}}{1 - x}\right) \]

    5. Taylor expanded in x around 0

      \[\leadsto \color{blue}{0} \]
    6. Step-by-step derivation

      1. Applied rewrites11.0%

        \[\leadsto \color{blue}{0} \]
      2. Add Preprocessing

      Reproduce

      ?
      herbie shell --seed 2024337 
(FPCore (x)
  :name "Rust f32::atanh"
  :precision binary32
  (* 0.5 (log1p (/ (* 2.0 x) (- 1.0 x)))))