
(FPCore (g h a) :precision binary64 (let* ((t_0 (/ 1.0 (* 2.0 a))) (t_1 (sqrt (- (* g g) (* h h))))) (+ (cbrt (* t_0 (+ (- g) t_1))) (cbrt (* t_0 (- (- g) t_1))))))
double code(double g, double h, double a) {
double t_0 = 1.0 / (2.0 * a);
double t_1 = sqrt(((g * g) - (h * h)));
return cbrt((t_0 * (-g + t_1))) + cbrt((t_0 * (-g - t_1)));
}
public static double code(double g, double h, double a) {
double t_0 = 1.0 / (2.0 * a);
double t_1 = Math.sqrt(((g * g) - (h * h)));
return Math.cbrt((t_0 * (-g + t_1))) + Math.cbrt((t_0 * (-g - t_1)));
}
function code(g, h, a) t_0 = Float64(1.0 / Float64(2.0 * a)) t_1 = sqrt(Float64(Float64(g * g) - Float64(h * h))) return Float64(cbrt(Float64(t_0 * Float64(Float64(-g) + t_1))) + cbrt(Float64(t_0 * Float64(Float64(-g) - t_1)))) end
code[g_, h_, a_] := Block[{t$95$0 = N[(1.0 / N[(2.0 * a), $MachinePrecision]), $MachinePrecision]}, Block[{t$95$1 = N[Sqrt[N[(N[(g * g), $MachinePrecision] - N[(h * h), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]}, N[(N[Power[N[(t$95$0 * N[((-g) + t$95$1), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision] + N[Power[N[(t$95$0 * N[((-g) - t$95$1), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision]]]
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{2 \cdot a}\\
t_1 := \sqrt{g \cdot g - h \cdot h}\\
\sqrt[3]{t\_0 \cdot \left(\left(-g\right) + t\_1\right)} + \sqrt[3]{t\_0 \cdot \left(\left(-g\right) - t\_1\right)}
\end{array}
\end{array}
Sampling outcomes in binary64 precision:
Herbie found 5 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (g h a) :precision binary64 (let* ((t_0 (/ 1.0 (* 2.0 a))) (t_1 (sqrt (- (* g g) (* h h))))) (+ (cbrt (* t_0 (+ (- g) t_1))) (cbrt (* t_0 (- (- g) t_1))))))
double code(double g, double h, double a) {
double t_0 = 1.0 / (2.0 * a);
double t_1 = sqrt(((g * g) - (h * h)));
return cbrt((t_0 * (-g + t_1))) + cbrt((t_0 * (-g - t_1)));
}
public static double code(double g, double h, double a) {
double t_0 = 1.0 / (2.0 * a);
double t_1 = Math.sqrt(((g * g) - (h * h)));
return Math.cbrt((t_0 * (-g + t_1))) + Math.cbrt((t_0 * (-g - t_1)));
}
function code(g, h, a) t_0 = Float64(1.0 / Float64(2.0 * a)) t_1 = sqrt(Float64(Float64(g * g) - Float64(h * h))) return Float64(cbrt(Float64(t_0 * Float64(Float64(-g) + t_1))) + cbrt(Float64(t_0 * Float64(Float64(-g) - t_1)))) end
code[g_, h_, a_] := Block[{t$95$0 = N[(1.0 / N[(2.0 * a), $MachinePrecision]), $MachinePrecision]}, Block[{t$95$1 = N[Sqrt[N[(N[(g * g), $MachinePrecision] - N[(h * h), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]}, N[(N[Power[N[(t$95$0 * N[((-g) + t$95$1), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision] + N[Power[N[(t$95$0 * N[((-g) - t$95$1), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision]]]
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{2 \cdot a}\\
t_1 := \sqrt{g \cdot g - h \cdot h}\\
\sqrt[3]{t\_0 \cdot \left(\left(-g\right) + t\_1\right)} + \sqrt[3]{t\_0 \cdot \left(\left(-g\right) - t\_1\right)}
\end{array}
\end{array}
(FPCore (g h a) :precision binary64 (/ (* (cbrt g) (* (cbrt -1.0) (cbrt 2.0))) (cbrt (* a 2.0))))
double code(double g, double h, double a) {
return (cbrt(g) * (cbrt(-1.0) * cbrt(2.0))) / cbrt((a * 2.0));
}
public static double code(double g, double h, double a) {
return (Math.cbrt(g) * (Math.cbrt(-1.0) * Math.cbrt(2.0))) / Math.cbrt((a * 2.0));
}
function code(g, h, a) return Float64(Float64(cbrt(g) * Float64(cbrt(-1.0) * cbrt(2.0))) / cbrt(Float64(a * 2.0))) end
code[g_, h_, a_] := N[(N[(N[Power[g, 1/3], $MachinePrecision] * N[(N[Power[-1.0, 1/3], $MachinePrecision] * N[Power[2.0, 1/3], $MachinePrecision]), $MachinePrecision]), $MachinePrecision] / N[Power[N[(a * 2.0), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision]
\begin{array}{l}
\\
\frac{\sqrt[3]{g} \cdot \left(\sqrt[3]{-1} \cdot \sqrt[3]{2}\right)}{\sqrt[3]{a \cdot 2}}
\end{array}
Initial program 45.0%
Applied rewrites14.1%
Taylor expanded in g around inf
lower-*.f64N/A
lower-cbrt.f64N/A
lower-*.f64N/A
lower-cbrt.f64N/A
lower-cbrt.f6495.8
Applied rewrites95.8%
(FPCore (g h a)
:precision binary64
(let* ((t_0 (sqrt (- (* g g) (* h h))))
(t_1 (cbrt (* (pow (* 2.0 a) -1.0) (+ (- g) t_0))))
(t_2 (+ t_1 (cbrt (* (/ -1.0 (* 2.0 a)) (+ g t_0))))))
(if (or (<= t_2 -5e-95) (not (<= t_2 1e-104)))
(fma (cbrt (/ g a)) (cbrt -1.0) (cbrt (* -0.25 (* (/ h a) (/ h g)))))
(+ t_1 (* (cbrt (/ 0.5 a)) (cbrt (- (+ g g))))))))
double code(double g, double h, double a) {
double t_0 = sqrt(((g * g) - (h * h)));
double t_1 = cbrt((pow((2.0 * a), -1.0) * (-g + t_0)));
double t_2 = t_1 + cbrt(((-1.0 / (2.0 * a)) * (g + t_0)));
double tmp;
if ((t_2 <= -5e-95) || !(t_2 <= 1e-104)) {
tmp = fma(cbrt((g / a)), cbrt(-1.0), cbrt((-0.25 * ((h / a) * (h / g)))));
} else {
tmp = t_1 + (cbrt((0.5 / a)) * cbrt(-(g + g)));
}
return tmp;
}
function code(g, h, a) t_0 = sqrt(Float64(Float64(g * g) - Float64(h * h))) t_1 = cbrt(Float64((Float64(2.0 * a) ^ -1.0) * Float64(Float64(-g) + t_0))) t_2 = Float64(t_1 + cbrt(Float64(Float64(-1.0 / Float64(2.0 * a)) * Float64(g + t_0)))) tmp = 0.0 if ((t_2 <= -5e-95) || !(t_2 <= 1e-104)) tmp = fma(cbrt(Float64(g / a)), cbrt(-1.0), cbrt(Float64(-0.25 * Float64(Float64(h / a) * Float64(h / g))))); else tmp = Float64(t_1 + Float64(cbrt(Float64(0.5 / a)) * cbrt(Float64(-Float64(g + g))))); end return tmp end
code[g_, h_, a_] := Block[{t$95$0 = N[Sqrt[N[(N[(g * g), $MachinePrecision] - N[(h * h), $MachinePrecision]), $MachinePrecision]], $MachinePrecision]}, Block[{t$95$1 = N[Power[N[(N[Power[N[(2.0 * a), $MachinePrecision], -1.0], $MachinePrecision] * N[((-g) + t$95$0), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]}, Block[{t$95$2 = N[(t$95$1 + N[Power[N[(N[(-1.0 / N[(2.0 * a), $MachinePrecision]), $MachinePrecision] * N[(g + t$95$0), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision]}, If[Or[LessEqual[t$95$2, -5e-95], N[Not[LessEqual[t$95$2, 1e-104]], $MachinePrecision]], N[(N[Power[N[(g / a), $MachinePrecision], 1/3], $MachinePrecision] * N[Power[-1.0, 1/3], $MachinePrecision] + N[Power[N[(-0.25 * N[(N[(h / a), $MachinePrecision] * N[(h / g), $MachinePrecision]), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision], N[(t$95$1 + N[(N[Power[N[(0.5 / a), $MachinePrecision], 1/3], $MachinePrecision] * N[Power[(-N[(g + g), $MachinePrecision]), 1/3], $MachinePrecision]), $MachinePrecision]), $MachinePrecision]]]]]
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \sqrt{g \cdot g - h \cdot h}\\
t_1 := \sqrt[3]{{\left(2 \cdot a\right)}^{-1} \cdot \left(\left(-g\right) + t\_0\right)}\\
t_2 := t\_1 + \sqrt[3]{\frac{-1}{2 \cdot a} \cdot \left(g + t\_0\right)}\\
\mathbf{if}\;t\_2 \leq -5 \cdot 10^{-95} \lor \neg \left(t\_2 \leq 10^{-104}\right):\\
\;\;\;\;\mathsf{fma}\left(\sqrt[3]{\frac{g}{a}}, \sqrt[3]{-1}, \sqrt[3]{-0.25 \cdot \left(\frac{h}{a} \cdot \frac{h}{g}\right)}\right)\\
\mathbf{else}:\\
\;\;\;\;t\_1 + \sqrt[3]{\frac{0.5}{a}} \cdot \sqrt[3]{-\left(g + g\right)}\\
\end{array}
\end{array}
if (+.f64 (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (+.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h)))))) (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (-.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h))))))) < -4.9999999999999998e-95 or 9.99999999999999927e-105 < (+.f64 (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (+.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h)))))) (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (-.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h))))))) Initial program 47.2%
Taylor expanded in h around 0
associate-*r*N/A
lower-fma.f64N/A
lower-*.f64N/A
lower-cbrt.f64N/A
lower-/.f64N/A
lower-cbrt.f64N/A
lower-cbrt.f64N/A
lower-*.f64N/A
lower-cbrt.f64N/A
unpow2N/A
*-commutativeN/A
times-fracN/A
lower-*.f64N/A
lower-/.f64N/A
lower-/.f64N/A
*-commutativeN/A
lower-*.f64N/A
lower-cbrt.f64N/A
lower-cbrt.f6478.2
Applied rewrites78.2%
Applied rewrites79.1%
if -4.9999999999999998e-95 < (+.f64 (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (+.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h)))))) (cbrt.f64 (*.f64 (/.f64 #s(literal 1 binary64) (*.f64 #s(literal 2 binary64) a)) (-.f64 (neg.f64 g) (sqrt.f64 (-.f64 (*.f64 g g) (*.f64 h h))))))) < 9.99999999999999927e-105Initial program 9.3%
lift-cbrt.f64N/A
lift-*.f64N/A
cbrt-prodN/A
pow1/3N/A
lower-*.f64N/A
pow1/3N/A
lower-cbrt.f64N/A
lift-/.f64N/A
lift-*.f64N/A
associate-/r*N/A
metadata-evalN/A
lower-/.f64N/A
lower-cbrt.f6437.2
lift--.f64N/A
lift-*.f64N/A
lift-*.f64N/A
difference-of-squaresN/A
*-commutativeN/A
lower-*.f64N/A
lower--.f64N/A
Applied rewrites37.2%
Taylor expanded in g around -inf
mul-1-negN/A
*-commutativeN/A
unpow2N/A
rem-square-sqrtN/A
lower-neg.f64N/A
mul-1-negN/A
lower-neg.f6494.5
Applied rewrites94.5%
Final simplification80.0%
(FPCore (g h a) :precision binary64 (fma (cbrt (/ g a)) (cbrt -1.0) (cbrt (* -0.25 (* (/ h a) (/ h g))))))
double code(double g, double h, double a) {
return fma(cbrt((g / a)), cbrt(-1.0), cbrt((-0.25 * ((h / a) * (h / g)))));
}
function code(g, h, a) return fma(cbrt(Float64(g / a)), cbrt(-1.0), cbrt(Float64(-0.25 * Float64(Float64(h / a) * Float64(h / g))))) end
code[g_, h_, a_] := N[(N[Power[N[(g / a), $MachinePrecision], 1/3], $MachinePrecision] * N[Power[-1.0, 1/3], $MachinePrecision] + N[Power[N[(-0.25 * N[(N[(h / a), $MachinePrecision] * N[(h / g), $MachinePrecision]), $MachinePrecision]), $MachinePrecision], 1/3], $MachinePrecision]), $MachinePrecision]
\begin{array}{l}
\\
\mathsf{fma}\left(\sqrt[3]{\frac{g}{a}}, \sqrt[3]{-1}, \sqrt[3]{-0.25 \cdot \left(\frac{h}{a} \cdot \frac{h}{g}\right)}\right)
\end{array}
Initial program 45.0%
Taylor expanded in h around 0
associate-*r*N/A
lower-fma.f64N/A
lower-*.f64N/A
lower-cbrt.f64N/A
lower-/.f64N/A
lower-cbrt.f64N/A
lower-cbrt.f64N/A
lower-*.f64N/A
lower-cbrt.f64N/A
unpow2N/A
*-commutativeN/A
times-fracN/A
lower-*.f64N/A
lower-/.f64N/A
lower-/.f64N/A
*-commutativeN/A
lower-*.f64N/A
lower-cbrt.f64N/A
lower-cbrt.f6474.2
Applied rewrites74.2%
Applied rewrites75.0%
(FPCore (g h a) :precision binary64 (* (cbrt (/ g a)) (cbrt -1.0)))
double code(double g, double h, double a) {
return cbrt((g / a)) * cbrt(-1.0);
}
public static double code(double g, double h, double a) {
return Math.cbrt((g / a)) * Math.cbrt(-1.0);
}
function code(g, h, a) return Float64(cbrt(Float64(g / a)) * cbrt(-1.0)) end
code[g_, h_, a_] := N[(N[Power[N[(g / a), $MachinePrecision], 1/3], $MachinePrecision] * N[Power[-1.0, 1/3], $MachinePrecision]), $MachinePrecision]
\begin{array}{l}
\\
\sqrt[3]{\frac{g}{a}} \cdot \sqrt[3]{-1}
\end{array}
Initial program 45.0%
Taylor expanded in g around inf
associate-*r/N/A
mul-1-negN/A
lower-/.f64N/A
lower-neg.f6430.1
Applied rewrites30.1%
lift-+.f64N/A
+-commutativeN/A
lift-cbrt.f64N/A
pow1/3N/A
sqr-powN/A
lower-fma.f64N/A
Applied rewrites11.7%
Taylor expanded in g around inf
lower-*.f64N/A
lower-cbrt.f64N/A
lower-/.f64N/A
lower-cbrt.f6473.4
Applied rewrites73.4%
(FPCore (g h a) :precision binary64 0.0)
double code(double g, double h, double a) {
return 0.0;
}
module fmin_fmax_functions
implicit none
private
public fmax
public fmin
interface fmax
module procedure fmax88
module procedure fmax44
module procedure fmax84
module procedure fmax48
end interface
interface fmin
module procedure fmin88
module procedure fmin44
module procedure fmin84
module procedure fmin48
end interface
contains
real(8) function fmax88(x, y) result (res)
real(8), intent (in) :: x
real(8), intent (in) :: y
res = merge(y, merge(x, max(x, y), y /= y), x /= x)
end function
real(4) function fmax44(x, y) result (res)
real(4), intent (in) :: x
real(4), intent (in) :: y
res = merge(y, merge(x, max(x, y), y /= y), x /= x)
end function
real(8) function fmax84(x, y) result(res)
real(8), intent (in) :: x
real(4), intent (in) :: y
res = merge(dble(y), merge(x, max(x, dble(y)), y /= y), x /= x)
end function
real(8) function fmax48(x, y) result(res)
real(4), intent (in) :: x
real(8), intent (in) :: y
res = merge(y, merge(dble(x), max(dble(x), y), y /= y), x /= x)
end function
real(8) function fmin88(x, y) result (res)
real(8), intent (in) :: x
real(8), intent (in) :: y
res = merge(y, merge(x, min(x, y), y /= y), x /= x)
end function
real(4) function fmin44(x, y) result (res)
real(4), intent (in) :: x
real(4), intent (in) :: y
res = merge(y, merge(x, min(x, y), y /= y), x /= x)
end function
real(8) function fmin84(x, y) result(res)
real(8), intent (in) :: x
real(4), intent (in) :: y
res = merge(dble(y), merge(x, min(x, dble(y)), y /= y), x /= x)
end function
real(8) function fmin48(x, y) result(res)
real(4), intent (in) :: x
real(8), intent (in) :: y
res = merge(y, merge(dble(x), min(dble(x), y), y /= y), x /= x)
end function
end module
real(8) function code(g, h, a)
use fmin_fmax_functions
real(8), intent (in) :: g
real(8), intent (in) :: h
real(8), intent (in) :: a
code = 0.0d0
end function
public static double code(double g, double h, double a) {
return 0.0;
}
def code(g, h, a): return 0.0
function code(g, h, a) return 0.0 end
function tmp = code(g, h, a) tmp = 0.0; end
code[g_, h_, a_] := 0.0
\begin{array}{l}
\\
0
\end{array}
Initial program 45.0%
lift-sqrt.f64N/A
pow1/2N/A
lift--.f64N/A
lift-*.f64N/A
lift-*.f64N/A
difference-of-squaresN/A
*-commutativeN/A
unpow-prod-downN/A
lower-*.f64N/A
pow1/2N/A
lower-sqrt.f64N/A
lower--.f64N/A
pow1/2N/A
lower-sqrt.f64N/A
+-commutativeN/A
lower-+.f6420.3
Applied rewrites20.3%
Taylor expanded in g around -inf
mul-1-negN/A
lower-neg.f64N/A
lower-*.f64N/A
*-commutativeN/A
+-commutativeN/A
unpow2N/A
rem-square-sqrtN/A
distribute-rgt1-inN/A
lower-cbrt.f64N/A
distribute-rgt1-inN/A
metadata-evalN/A
*-lft-identityN/A
times-fracN/A
metadata-evalN/A
lower-*.f64N/A
lower-/.f64N/A
lower-cbrt.f642.5
Applied rewrites2.5%
Applied rewrites3.0%
herbie shell --seed 2024359
(FPCore (g h a)
:name "2-ancestry mixing, positive discriminant"
:precision binary64
(+ (cbrt (* (/ 1.0 (* 2.0 a)) (+ (- g) (sqrt (- (* g g) (* h h)))))) (cbrt (* (/ 1.0 (* 2.0 a)) (- (- g) (sqrt (- (* g g) (* h h))))))))