ENA, Section 1.4, Exercise 4b, n=2

Percentage Accurate: 74.8% → 100.0%

Time: 6.8s

Alternatives: 4

Speedup: 17.4×

Specification

\[\left(-1000000000 \leq x \land x \leq 1000000000\right) \land \left(-1 \leq \varepsilon \land \varepsilon \leq 1\right)\]

Math

\[\begin{array}{l} \\ {\left(x + \varepsilon\right)}^{2} - {x}^{2} \end{array} \]

FPCore

(FPCore (x eps) :precision binary64 (- (pow (+ x eps) 2.0) (pow x 2.0)))

C

double code(double x, double eps) {
	return pow((x + eps), 2.0) - pow(x, 2.0);
}

Fortran

real(8) function code(x, eps)
    real(8), intent (in) :: x
    real(8), intent (in) :: eps
    code = ((x + eps) ** 2.0d0) - (x ** 2.0d0)
end function

Java

public static double code(double x, double eps) {
	return Math.pow((x + eps), 2.0) - Math.pow(x, 2.0);
}

Python

def code(x, eps):
	return math.pow((x + eps), 2.0) - math.pow(x, 2.0)

Julia

function code(x, eps)
	return Float64((Float64(x + eps) ^ 2.0) - (x ^ 2.0))
end

MATLAB

function tmp = code(x, eps)
	tmp = ((x + eps) ^ 2.0) - (x ^ 2.0);
end

Wolfram

code[x_, eps_] := N[(N[Power[N[(x + eps), $MachinePrecision], 2.0], $MachinePrecision] - N[Power[x, 2.0], $MachinePrecision]), $MachinePrecision]

Initial Program: 74.8% accurate, 1.0× speedup

Math

\[\begin{array}{l} \\ {\left(x + \varepsilon\right)}^{2} - {x}^{2} \end{array} \]

FPCore

(FPCore (x eps) :precision binary64 (- (pow (+ x eps) 2.0) (pow x 2.0)))

C

double code(double x, double eps) {
	return pow((x + eps), 2.0) - pow(x, 2.0);
}

Fortran

real(8) function code(x, eps)
    real(8), intent (in) :: x
    real(8), intent (in) :: eps
    code = ((x + eps) ** 2.0d0) - (x ** 2.0d0)
end function

Java

public static double code(double x, double eps) {
	return Math.pow((x + eps), 2.0) - Math.pow(x, 2.0);
}

Python

def code(x, eps):
	return math.pow((x + eps), 2.0) - math.pow(x, 2.0)

Julia

function code(x, eps)
	return Float64((Float64(x + eps) ^ 2.0) - (x ^ 2.0))
end

MATLAB

function tmp = code(x, eps)
	tmp = ((x + eps) ^ 2.0) - (x ^ 2.0);
end

Wolfram

code[x_, eps_] := N[(N[Power[N[(x + eps), $MachinePrecision], 2.0], $MachinePrecision] - N[Power[x, 2.0], $MachinePrecision]), $MachinePrecision]

Alternative 1: 100.0% accurate, 17.4× speedup

Math

\[\begin{array}{l} \\ \mathsf{fma}\left(2, x, \varepsilon\right) \cdot \varepsilon \end{array} \]

FPCore

(FPCore (x eps) :precision binary64 (* (fma 2.0 x eps) eps))

C

double code(double x, double eps) {
	return fma(2.0, x, eps) * eps;
}

Julia

function code(x, eps)
	return Float64(fma(2.0, x, eps) * eps)
end

Wolfram

code[x_, eps_] := N[(N[(2.0 * x + eps), $MachinePrecision] * eps), $MachinePrecision]

Derivation

1. Initial program 79.6%

   \[{\left(x + \varepsilon\right)}^{2} - {x}^{2} \]
2. Add Preprocessing

3. Taylor expanded in x around 0

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

   1. +-commutative N/A

      \[\leadsto \color{blue}{{\varepsilon}^{2} + 2 \cdot \left(\varepsilon \cdot x\right)} \]

   2. count-2-rev N/A

      \[\leadsto {\varepsilon}^{2} + \color{blue}{\left(\varepsilon \cdot x + \varepsilon \cdot x\right)} \]

   3. unpow2 N/A

      \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} + \left(\varepsilon \cdot x + \varepsilon \cdot x\right) \]

   4. distribute-lft-in N/A

      \[\leadsto \varepsilon \cdot \varepsilon + \color{blue}{\varepsilon \cdot \left(x + x\right)} \]

   5. count-2-rev N/A

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

   6. distribute-lft-in N/A

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

   7. *-commutative N/A

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

   8. lower-*.f64 N/A

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

   9. +-commutative N/A

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

   10. lower-fma.f64 100.0

       \[\leadsto \color{blue}{\mathsf{fma}\left(2, x, \varepsilon\right)} \cdot \varepsilon \]

5. Applied rewrites 100.0%

   \[\leadsto \color{blue}{\mathsf{fma}\left(2, x, \varepsilon\right) \cdot \varepsilon} \]
6. Add Preprocessing

Alternative 2: 97.3% accurate, 0.9× speedup

Math

\[\begin{array}{l} \\ \begin{array}{l} \mathbf{if}\;{\left(x + \varepsilon\right)}^{2} - {x}^{2} \leq 5 \cdot 10^{-324}:\\ \;\;\;\;\left(\varepsilon + \varepsilon\right) \cdot x\\ \mathbf{else}:\\ \;\;\;\;\varepsilon \cdot \varepsilon\\ \end{array} \end{array} \]

FPCore

(FPCore (x eps)
 :precision binary64
 (if (<= (- (pow (+ x eps) 2.0) (pow x 2.0)) 5e-324)
   (* (+ eps eps) x)
   (* eps eps)))

C

double code(double x, double eps) {
	double tmp;
	if ((pow((x + eps), 2.0) - pow(x, 2.0)) <= 5e-324) {
		tmp = (eps + eps) * x;
	} else {
		tmp = eps * eps;
	}
	return tmp;
}

Fortran

real(8) function code(x, eps)
    real(8), intent (in) :: x
    real(8), intent (in) :: eps
    real(8) :: tmp
    if ((((x + eps) ** 2.0d0) - (x ** 2.0d0)) <= 5d-324) then
        tmp = (eps + eps) * x
    else
        tmp = eps * eps
    end if
    code = tmp
end function

Java

public static double code(double x, double eps) {
	double tmp;
	if ((Math.pow((x + eps), 2.0) - Math.pow(x, 2.0)) <= 5e-324) {
		tmp = (eps + eps) * x;
	} else {
		tmp = eps * eps;
	}
	return tmp;
}

Python

def code(x, eps):
	tmp = 0
	if (math.pow((x + eps), 2.0) - math.pow(x, 2.0)) <= 5e-324:
		tmp = (eps + eps) * x
	else:
		tmp = eps * eps
	return tmp

Julia

function code(x, eps)
	tmp = 0.0
	if (Float64((Float64(x + eps) ^ 2.0) - (x ^ 2.0)) <= 5e-324)
		tmp = Float64(Float64(eps + eps) * x);
	else
		tmp = Float64(eps * eps);
	end
	return tmp
end

MATLAB

function tmp_2 = code(x, eps)
	tmp = 0.0;
	if ((((x + eps) ^ 2.0) - (x ^ 2.0)) <= 5e-324)
		tmp = (eps + eps) * x;
	else
		tmp = eps * eps;
	end
	tmp_2 = tmp;
end

Wolfram

code[x_, eps_] := If[LessEqual[N[(N[Power[N[(x + eps), $MachinePrecision], 2.0], $MachinePrecision] - N[Power[x, 2.0], $MachinePrecision]), $MachinePrecision], 5e-324], N[(N[(eps + eps), $MachinePrecision] * x), $MachinePrecision], N[(eps * eps), $MachinePrecision]]

Derivation

1. Split input into 2 regimes

2. if (-.f64 (pow.f64 (+.f64 x eps) #s(literal 2 binary64)) (pow.f64 x #s(literal 2 binary64))) < 4.94066e-324

   1. Initial program 66.5%

      \[{\left(x + \varepsilon\right)}^{2} - {x}^{2} \]
   2. Add Preprocessing

   3. Taylor expanded in x around inf

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

      1. *-commutative N/A

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

      2. lower-*.f64 N/A

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

      3. lower-*.f64 99.4

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

   5. Applied rewrites 99.4%

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

   7. Applied rewrites 99.5%

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

   9. Applied rewrites 99.5%

      \[\leadsto \left(\varepsilon + \varepsilon\right) \cdot x \]

   if 4.94066e-324 < (-.f64 (pow.f64 (+.f64 x eps) #s(literal 2 binary64)) (pow.f64 x #s(literal 2 binary64)))

   1. Initial program 99.3%

      \[{\left(x + \varepsilon\right)}^{2} - {x}^{2} \]
   2. Add Preprocessing

   3. Taylor expanded in x around 0

      \[\leadsto \color{blue}{{\varepsilon}^{2}} \]
   4. Step-by-step derivation

      1. unpow2 N/A

         \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]

      2. lower-*.f64 95.9

         \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]

   5. Applied rewrites 95.9%

      \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 3: 100.0% accurate, 17.4× speedup

Math

\[\begin{array}{l} \\ \left(\left(\varepsilon + x\right) + x\right) \cdot \varepsilon \end{array} \]

FPCore

(FPCore (x eps) :precision binary64 (* (+ (+ eps x) x) eps))

C

double code(double x, double eps) {
	return ((eps + x) + x) * eps;
}

Fortran

real(8) function code(x, eps)
    real(8), intent (in) :: x
    real(8), intent (in) :: eps
    code = ((eps + x) + x) * eps
end function

Java

public static double code(double x, double eps) {
	return ((eps + x) + x) * eps;
}

Python

def code(x, eps):
	return ((eps + x) + x) * eps

Julia

function code(x, eps)
	return Float64(Float64(Float64(eps + x) + x) * eps)
end

MATLAB

function tmp = code(x, eps)
	tmp = ((eps + x) + x) * eps;
end

Wolfram

code[x_, eps_] := N[(N[(N[(eps + x), $MachinePrecision] + x), $MachinePrecision] * eps), $MachinePrecision]

Derivation

1. Initial program 79.6%

   \[{\left(x + \varepsilon\right)}^{2} - {x}^{2} \]
2. Add Preprocessing

3. Taylor expanded in x around 0

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

   1. +-commutative N/A

      \[\leadsto \color{blue}{{\varepsilon}^{2} + 2 \cdot \left(\varepsilon \cdot x\right)} \]

   2. count-2-rev N/A

      \[\leadsto {\varepsilon}^{2} + \color{blue}{\left(\varepsilon \cdot x + \varepsilon \cdot x\right)} \]

   3. unpow2 N/A

      \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} + \left(\varepsilon \cdot x + \varepsilon \cdot x\right) \]

   4. distribute-lft-in N/A

      \[\leadsto \varepsilon \cdot \varepsilon + \color{blue}{\varepsilon \cdot \left(x + x\right)} \]

   5. count-2-rev N/A

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

   6. distribute-lft-in N/A

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

   7. *-commutative N/A

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

   8. lower-*.f64 N/A

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

   9. +-commutative N/A

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

   10. lower-fma.f64 100.0

       \[\leadsto \color{blue}{\mathsf{fma}\left(2, x, \varepsilon\right)} \cdot \varepsilon \]

5. Applied rewrites 100.0%

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

7. Applied rewrites 100.0%

   \[\leadsto \left(\left(\varepsilon + x\right) + x\right) \cdot \varepsilon \]
8. Add Preprocessing

Alternative 4: 72.5% accurate, 34.8× speedup

Math

\[\begin{array}{l} \\ \varepsilon \cdot \varepsilon \end{array} \]

FPCore

(FPCore (x eps) :precision binary64 (* eps eps))

C

double code(double x, double eps) {
	return eps * eps;
}

Fortran

real(8) function code(x, eps)
    real(8), intent (in) :: x
    real(8), intent (in) :: eps
    code = eps * eps
end function

Java

public static double code(double x, double eps) {
	return eps * eps;
}

Python

def code(x, eps):
	return eps * eps

Julia

function code(x, eps)
	return Float64(eps * eps)
end

MATLAB

function tmp = code(x, eps)
	tmp = eps * eps;
end

Wolfram

code[x_, eps_] := N[(eps * eps), $MachinePrecision]

Derivation

1. Initial program 79.6%

   \[{\left(x + \varepsilon\right)}^{2} - {x}^{2} \]
2. Add Preprocessing

3. Taylor expanded in x around 0

   \[\leadsto \color{blue}{{\varepsilon}^{2}} \]
4. Step-by-step derivation

   1. unpow2 N/A

      \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]

   2. lower-*.f64 77.8

      \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]

5. Applied rewrites 77.8%

   \[\leadsto \color{blue}{\varepsilon \cdot \varepsilon} \]
6. Add Preprocessing

Reproduce

herbie shell --seed 2024329 
(FPCore (x eps)
  :name "ENA, Section 1.4, Exercise 4b, n=2"
  :precision binary64
  :pre (and (and (<= -1000000000.0 x) (<= x 1000000000.0)) (and (<= -1.0 eps) (<= eps 1.0)))
  (- (pow (+ x eps) 2.0) (pow x 2.0)))