bug366, discussion (missed optimization)

Average Error: 31.1 → 0.7

Time: 3.3s

Precision: binary64

Cost: 708

Math

\[\sqrt{a \cdot a - b \cdot b} \]

↓

\[\begin{array}{l} \mathbf{if}\;a \leq -7.920695166384969 \cdot 10^{-250}:\\ \;\;\;\;-a\\ \mathbf{else}:\\ \;\;\;\;a + \frac{b \cdot -0.5}{\frac{a}{b}}\\ \end{array} \]

FPCore

(FPCore (a b) :precision binary64 (sqrt (- (* a a) (* b b))))

↓

(FPCore (a b)
 :precision binary64
 (if (<= a -7.920695166384969e-250) (- a) (+ a (/ (* b -0.5) (/ a b)))))

C

double code(double a, double b) {
	return sqrt(((a * a) - (b * b)));
}

↓

double code(double a, double b) {
	double tmp;
	if (a <= -7.920695166384969e-250) {
		tmp = -a;
	} else {
		tmp = a + ((b * -0.5) / (a / b));
	}
	return tmp;
}

Fortran

real(8) function code(a, b)
    real(8), intent (in) :: a
    real(8), intent (in) :: b
    code = sqrt(((a * a) - (b * b)))
end function

↓

real(8) function code(a, b)
    real(8), intent (in) :: a
    real(8), intent (in) :: b
    real(8) :: tmp
    if (a <= (-7.920695166384969d-250)) then
        tmp = -a
    else
        tmp = a + ((b * (-0.5d0)) / (a / b))
    end if
    code = tmp
end function

Java

public static double code(double a, double b) {
	return Math.sqrt(((a * a) - (b * b)));
}

↓

public static double code(double a, double b) {
	double tmp;
	if (a <= -7.920695166384969e-250) {
		tmp = -a;
	} else {
		tmp = a + ((b * -0.5) / (a / b));
	}
	return tmp;
}

Python

def code(a, b):
	return math.sqrt(((a * a) - (b * b)))

↓

def code(a, b):
	tmp = 0
	if a <= -7.920695166384969e-250:
		tmp = -a
	else:
		tmp = a + ((b * -0.5) / (a / b))
	return tmp

Julia

function code(a, b)
	return sqrt(Float64(Float64(a * a) - Float64(b * b)))
end

↓

function code(a, b)
	tmp = 0.0
	if (a <= -7.920695166384969e-250)
		tmp = Float64(-a);
	else
		tmp = Float64(a + Float64(Float64(b * -0.5) / Float64(a / b)));
	end
	return tmp
end

MATLAB

function tmp = code(a, b)
	tmp = sqrt(((a * a) - (b * b)));
end

↓

function tmp_2 = code(a, b)
	tmp = 0.0;
	if (a <= -7.920695166384969e-250)
		tmp = -a;
	else
		tmp = a + ((b * -0.5) / (a / b));
	end
	tmp_2 = tmp;
end

Wolfram

code[a_, b_] := N[Sqrt[N[(N[(a * a), $MachinePrecision] - N[(b * b), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]

↓

code[a_, b_] := If[LessEqual[a, -7.920695166384969e-250], (-a), N[(a + N[(N[(b * -0.5), $MachinePrecision] / N[(a / b), $MachinePrecision]), $MachinePrecision]), $MachinePrecision]]

Target

Original	31.1
Target	0.5
Herbie	0.7

\[\sqrt{\left|a\right| + \left|b\right|} \cdot \sqrt{\left|a\right| - \left|b\right|} \]

Derivation

1. Split input into 2 regimes

2. if a < -7.92069516638496914e-250

   1. Initial program 31.1

      \[\sqrt{a \cdot a - b \cdot b} \]

   2. Taylor expanded in a around -inf 0.5

      \[\leadsto \color{blue}{-1 \cdot a} \]

   3. Simplified 0.5

      \[\leadsto \color{blue}{-a} \]
      Proof

      [Start] 0.5	\[ -1 \cdot a \]
      mul-1-neg [=>] 0.5	\[ \color{blue}{-a} \]

   if -7.92069516638496914e-250 < a

   1. Initial program 31.2

      \[\sqrt{a \cdot a - b \cdot b} \]

   2. Taylor expanded in a around inf 5.0

      \[\leadsto \color{blue}{a + -0.5 \cdot \frac{{b}^{2}}{a}} \]

   3. Simplified 5.0

      \[\leadsto \color{blue}{a + -0.5 \cdot \frac{b \cdot b}{a}} \]
      Proof

      [Start] 5.0	\[ a + -0.5 \cdot \frac{{b}^{2}}{a} \]
      unpow2 [=>] 5.0	\[ a + -0.5 \cdot \frac{\color{blue}{b \cdot b}}{a} \]

   4. Applied egg-rr 0.8

      \[\leadsto a + \color{blue}{\frac{b \cdot -0.5}{\frac{a}{b}}} \]
3. Recombined 2 regimes into one program.

4. Final simplification 0.7

   \[\leadsto \begin{array}{l} \mathbf{if}\;a \leq -7.920695166384969 \cdot 10^{-250}:\\ \;\;\;\;-a\\ \mathbf{else}:\\ \;\;\;\;a + \frac{b \cdot -0.5}{\frac{a}{b}}\\ \end{array} \]

Alternatives

Alternative 1
Error	0.8
Cost	260

\[\begin{array}{l} \mathbf{if}\;a \leq -7.920695166384969 \cdot 10^{-250}:\\ \;\;\;\;-a\\ \mathbf{else}:\\ \;\;\;\;a\\ \end{array} \]

Alternative 2
Error	31.6
Cost	64

\[a \]

Reproduce

herbie shell --seed 2022356 
(FPCore (a b)
  :name "bug366, discussion (missed optimization)"
  :precision binary64

  :herbie-target
  (* (sqrt (+ (fabs a) (fabs b))) (sqrt (- (fabs a) (fabs b))))

  (sqrt (- (* a a) (* b b))))