Logistic distribution

Percentage Accurate: 99.6% → 99.6%

Time: 8.3s

Alternatives: 8

Speedup: 1.1×

• Could not converge on a ground truth

  1. x = 30616.029296875
  2. s = 0.09156948328018188

Specification

\[0 \leq s \land s \leq 1.0651631\]

Math

\[\begin{array}{l} \\ \begin{array}{l} t_0 := e^{\frac{-\left|x\right|}{s}}\\ t_1 := 1 + t\_0\\ \frac{t\_0}{\left(s \cdot t\_1\right) \cdot t\_1} \end{array} \end{array} \]

FPCore

(FPCore (x s)
 :precision binary32
 (let* ((t_0 (exp (/ (- (fabs x)) s))) (t_1 (+ 1.0 t_0)))
   (/ t_0 (* (* s t_1) t_1))))

C

float code(float x, float s) {
	float t_0 = expf((-fabsf(x) / s));
	float t_1 = 1.0f + t_0;
	return t_0 / ((s * t_1) * t_1);
}

Fortran

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(4) function code(x, s)
use fmin_fmax_functions
    real(4), intent (in) :: x
    real(4), intent (in) :: s
    real(4) :: t_0
    real(4) :: t_1
    t_0 = exp((-abs(x) / s))
    t_1 = 1.0e0 + t_0
    code = t_0 / ((s * t_1) * t_1)
end function

Julia

function code(x, s)
	t_0 = exp(Float32(Float32(-abs(x)) / s))
	t_1 = Float32(Float32(1.0) + t_0)
	return Float32(t_0 / Float32(Float32(s * t_1) * t_1))
end

MATLAB

function tmp = code(x, s)
	t_0 = exp((-abs(x) / s));
	t_1 = single(1.0) + t_0;
	tmp = t_0 / ((s * t_1) * t_1);
end

Initial Program: 99.6% accurate, 1.0× speedup

Math

\[\begin{array}{l} \\ \begin{array}{l} t_0 := e^{\frac{-\left|x\right|}{s}}\\ t_1 := 1 + t\_0\\ \frac{t\_0}{\left(s \cdot t\_1\right) \cdot t\_1} \end{array} \end{array} \]

FPCore

(FPCore (x s)
 :precision binary32
 (let* ((t_0 (exp (/ (- (fabs x)) s))) (t_1 (+ 1.0 t_0)))
   (/ t_0 (* (* s t_1) t_1))))

C

float code(float x, float s) {
	float t_0 = expf((-fabsf(x) / s));
	float t_1 = 1.0f + t_0;
	return t_0 / ((s * t_1) * t_1);
}

Fortran

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(4) function code(x, s)
use fmin_fmax_functions
    real(4), intent (in) :: x
    real(4), intent (in) :: s
    real(4) :: t_0
    real(4) :: t_1
    t_0 = exp((-abs(x) / s))
    t_1 = 1.0e0 + t_0
    code = t_0 / ((s * t_1) * t_1)
end function

Julia

function code(x, s)
	t_0 = exp(Float32(Float32(-abs(x)) / s))
	t_1 = Float32(Float32(1.0) + t_0)
	return Float32(t_0 / Float32(Float32(s * t_1) * t_1))
end

MATLAB

function tmp = code(x, s)
	t_0 = exp((-abs(x) / s));
	t_1 = single(1.0) + t_0;
	tmp = t_0 / ((s * t_1) * t_1);
end

Alternative 1: 99.6% accurate, 1.1× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \begin{array}{l} t_0 := e^{\frac{-x\_m}{s}}\\ \frac{t\_0}{{\left(t\_0 + 1\right)}^{2} \cdot s} \end{array} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (let* ((t_0 (exp (/ (- x_m) s)))) (/ t_0 (* (pow (+ t_0 1.0) 2.0) s))))

C

x_m = fabs(x);
float code(float x_m, float s) {
	float t_0 = expf((-x_m / s));
	return t_0 / (powf((t_0 + 1.0f), 2.0f) * s);
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    real(4) :: t_0
    t_0 = exp((-x_m / s))
    code = t_0 / (((t_0 + 1.0e0) ** 2.0e0) * s)
end function

Julia

x_m = abs(x)
function code(x_m, s)
	t_0 = exp(Float32(Float32(-x_m) / s))
	return Float32(t_0 / Float32((Float32(t_0 + Float32(1.0)) ^ Float32(2.0)) * s))
end

MATLAB

x_m = abs(x);
function tmp = code(x_m, s)
	t_0 = exp((-x_m / s));
	tmp = t_0 / (((t_0 + single(1.0)) ^ single(2.0)) * s);
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing

3. Taylor expanded in x around 0

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

   1. *-commutative N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{{\left(1 + e^{-1 \cdot \frac{\left|x\right|}{s}}\right)}^{2} \cdot s}} \]

   2. lower-*.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{{\left(1 + e^{-1 \cdot \frac{\left|x\right|}{s}}\right)}^{2} \cdot s}} \]

   3. lower-pow.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{{\left(1 + e^{-1 \cdot \frac{\left|x\right|}{s}}\right)}^{2}} \cdot s} \]

   4. +-commutative N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\color{blue}{\left(e^{-1 \cdot \frac{\left|x\right|}{s}} + 1\right)}}^{2} \cdot s} \]

   5. lower-+.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\color{blue}{\left(e^{-1 \cdot \frac{\left|x\right|}{s}} + 1\right)}}^{2} \cdot s} \]

   6. lower-exp.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(\color{blue}{e^{-1 \cdot \frac{\left|x\right|}{s}}} + 1\right)}^{2} \cdot s} \]

   7. mul-1-neg N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\color{blue}{\mathsf{neg}\left(\frac{\left|x\right|}{s}\right)}} + 1\right)}^{2} \cdot s} \]

   8. distribute-neg-frac2 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\color{blue}{\frac{\left|x\right|}{\mathsf{neg}\left(s\right)}}} + 1\right)}^{2} \cdot s} \]

   9. lower-/.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\color{blue}{\frac{\left|x\right|}{\mathsf{neg}\left(s\right)}}} + 1\right)}^{2} \cdot s} \]

   10. lower-fabs.f32 N/A

       \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\frac{\color{blue}{\left|x\right|}}{\mathsf{neg}\left(s\right)}} + 1\right)}^{2} \cdot s} \]

   11. lower-neg.f32 99.4

       \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\frac{\left|x\right|}{\color{blue}{-s}}} + 1\right)}^{2} \cdot s} \]

5. Applied rewrites 99.4%

   \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{{\left(e^{\frac{\left|x\right|}{-s}} + 1\right)}^{2} \cdot s}} \]
6. Step-by-step derivation

7. Applied rewrites 96.7%

   \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]
8. Step-by-step derivation

   1. lift-fabs.f32 N/A

      \[\leadsto \frac{e^{\frac{-\color{blue}{\left|x\right|}}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]

   2. rem-sqrt-square-rev N/A

      \[\leadsto \frac{e^{\frac{-\color{blue}{\sqrt{x \cdot x}}}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]

   3. sqrt-prod N/A

      \[\leadsto \frac{e^{\frac{-\color{blue}{\sqrt{x} \cdot \sqrt{x}}}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]

   4. rem-square-sqrt 56.7

      \[\leadsto \frac{e^{\frac{-\color{blue}{x}}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]

9. Applied rewrites 56.7%

   \[\leadsto \frac{e^{\frac{\color{blue}{-x}}{s}}}{{\left(e^{\frac{x}{-s}} + 1\right)}^{2} \cdot s} \]

10. Final simplification 56.7%

    \[\leadsto \frac{e^{\frac{-x}{s}}}{{\left(e^{\frac{-x}{s}} + 1\right)}^{2} \cdot s} \]
11. Add Preprocessing

Alternative 2: 62.8% accurate, 0.9× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \begin{array}{l} t_0 := e^{\frac{-\left|x\_m\right|}{s}}\\ t_1 := 1 + t\_0\\ \mathbf{if}\;\frac{t\_0}{\left(s \cdot t\_1\right) \cdot t\_1} \leq 5.000000229068525 \cdot 10^{-19}:\\ \;\;\;\;\frac{\frac{\frac{\mathsf{fma}\left(x\_m \cdot x\_m, -0.0625, 0.25 \cdot \left(s \cdot s\right)\right)}{s}}{s}}{s}\\ \mathbf{else}:\\ \;\;\;\;\frac{\frac{0.0625}{s} \cdot \frac{x\_m \cdot x\_m}{s} - 0.25}{-s}\\ \end{array} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (let* ((t_0 (exp (/ (- (fabs x_m)) s))) (t_1 (+ 1.0 t_0)))
   (if (<= (/ t_0 (* (* s t_1) t_1)) 5.000000229068525e-19)
     (/ (/ (/ (fma (* x_m x_m) -0.0625 (* 0.25 (* s s))) s) s) s)
     (/ (- (* (/ 0.0625 s) (/ (* x_m x_m) s)) 0.25) (- s)))))

C

x_m = fabs(x);
float code(float x_m, float s) {
	float t_0 = expf((-fabsf(x_m) / s));
	float t_1 = 1.0f + t_0;
	float tmp;
	if ((t_0 / ((s * t_1) * t_1)) <= 5.000000229068525e-19f) {
		tmp = ((fmaf((x_m * x_m), -0.0625f, (0.25f * (s * s))) / s) / s) / s;
	} else {
		tmp = (((0.0625f / s) * ((x_m * x_m) / s)) - 0.25f) / -s;
	}
	return tmp;
}

Julia

x_m = abs(x)
function code(x_m, s)
	t_0 = exp(Float32(Float32(-abs(x_m)) / s))
	t_1 = Float32(Float32(1.0) + t_0)
	tmp = Float32(0.0)
	if (Float32(t_0 / Float32(Float32(s * t_1) * t_1)) <= Float32(5.000000229068525e-19))
		tmp = Float32(Float32(Float32(fma(Float32(x_m * x_m), Float32(-0.0625), Float32(Float32(0.25) * Float32(s * s))) / s) / s) / s);
	else
		tmp = Float32(Float32(Float32(Float32(Float32(0.0625) / s) * Float32(Float32(x_m * x_m) / s)) - Float32(0.25)) / Float32(-s));
	end
	return tmp
end

Derivation

1. Split input into 2 regimes

2. if (/.f32 (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)) (*.f32 (*.f32 s (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))) (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s))))) < 5.00000023e-19

   1. Initial program 99.9%

      \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
   2. Add Preprocessing
   3. Step-by-step derivation

      1. lift-/.f32 N/A

         \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      2. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      3. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

      4. associate-*l* N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

      5. *-commutative N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

      6. associate-/r* N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

      7. lower-/.f32 N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   4. Applied rewrites 80.5%

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

   5. Taylor expanded in x around 0

      \[\leadsto \frac{\color{blue}{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} + \frac{-1}{4} \cdot \frac{{x}^{2} \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}}}{s} \]
   6. Step-by-step derivation

      1. +-commutative N/A

         \[\leadsto \frac{\color{blue}{\frac{-1}{4} \cdot \frac{{x}^{2} \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}}{s} \]

      2. associate-/l* N/A

         \[\leadsto \frac{\frac{-1}{4} \cdot \color{blue}{\left({x}^{2} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}\right)} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      3. *-commutative N/A

         \[\leadsto \frac{\frac{-1}{4} \cdot \color{blue}{\left(\frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}} \cdot {x}^{2}\right)} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      4. associate-*l* N/A

         \[\leadsto \frac{\color{blue}{\left(\frac{-1}{4} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}\right) \cdot {x}^{2}} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      5. lower-fma.f32 N/A

         \[\leadsto \frac{\color{blue}{\mathsf{fma}\left(\frac{-1}{4} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}, {x}^{2}, e^{\mathsf{neg}\left(2 \cdot \log 2\right)}\right)}}{s} \]

   7. Applied rewrites 4.5%

      \[\leadsto \frac{\color{blue}{\mathsf{fma}\left(\frac{-0.0625}{s \cdot s}, x \cdot x, 0.25\right)}}{s} \]
   8. Step-by-step derivation

   9. Applied rewrites 4.5%

      \[\leadsto \frac{\mathsf{fma}\left(x, \color{blue}{\frac{-0.0625}{s \cdot s} \cdot x}, 0.25\right)}{s} \]
   10. Step-by-step derivation

   11. Applied rewrites 46.5%

       \[\leadsto \frac{\frac{\frac{\mathsf{fma}\left(x \cdot x, -0.0625, 0.25 \cdot \left(s \cdot s\right)\right)}{s}}{\color{blue}{s}}}{s} \]

   if 5.00000023e-19 < (/.f32 (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)) (*.f32 (*.f32 s (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))) (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))))

   1. Initial program 98.2%

      \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
   2. Add Preprocessing
   3. Step-by-step derivation

      1. lift-/.f32 N/A

         \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      2. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      3. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

      4. associate-*l* N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

      5. *-commutative N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

      6. associate-/r* N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

      7. lower-/.f32 N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   4. Applied rewrites 24.7%

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

   5. Taylor expanded in s around -inf

      \[\leadsto \color{blue}{-1 \cdot \frac{-1 \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)} + \left(\frac{-1}{3} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-3}{4} \cdot {x}^{3} + \left(\frac{1}{4} \cdot {x}^{3} + \frac{1}{2} \cdot {x}^{3}\right)\right)}{{s}^{3}} + \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-1}{4} \cdot {x}^{2} + \frac{1}{2} \cdot {x}^{2}\right)}{{s}^{2}}\right)}{s}} \]

   7. Applied rewrites 53.5%

      \[\leadsto \color{blue}{\frac{\mathsf{fma}\left(-0.08333333333333333, \frac{{x}^{3} \cdot 0}{{s}^{3}}, \frac{\frac{0.0625 \cdot \left(x \cdot x\right)}{s}}{s}\right) - 0.25}{-s}} \]

   8. Taylor expanded in x around 0

      \[\leadsto \frac{\frac{1}{16} \cdot \frac{{x}^{2}}{{s}^{2}} - \frac{1}{4}}{-s} \]
   9. Step-by-step derivation

   10. Applied rewrites 84.4%

       \[\leadsto \frac{\frac{0.0625}{s} \cdot \frac{x \cdot x}{s} - 0.25}{-s} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 3: 29.8% accurate, 0.9× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \begin{array}{l} t_0 := \frac{x\_m \cdot x\_m}{s}\\ t_1 := e^{\frac{-\left|x\_m\right|}{s}}\\ t_2 := 1 + t\_1\\ \mathbf{if}\;\frac{t\_1}{\left(s \cdot t\_2\right) \cdot t\_2} \leq 1.000000013351432 \cdot 10^{-10}:\\ \;\;\;\;\frac{\frac{-0.0625}{s} \cdot t\_0}{s}\\ \mathbf{else}:\\ \;\;\;\;\frac{\frac{0.0625}{s} \cdot t\_0 - 0.25}{-s}\\ \end{array} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (let* ((t_0 (/ (* x_m x_m) s))
        (t_1 (exp (/ (- (fabs x_m)) s)))
        (t_2 (+ 1.0 t_1)))
   (if (<= (/ t_1 (* (* s t_2) t_2)) 1.000000013351432e-10)
     (/ (* (/ -0.0625 s) t_0) s)
     (/ (- (* (/ 0.0625 s) t_0) 0.25) (- s)))))

C

x_m = fabs(x);
float code(float x_m, float s) {
	float t_0 = (x_m * x_m) / s;
	float t_1 = expf((-fabsf(x_m) / s));
	float t_2 = 1.0f + t_1;
	float tmp;
	if ((t_1 / ((s * t_2) * t_2)) <= 1.000000013351432e-10f) {
		tmp = ((-0.0625f / s) * t_0) / s;
	} else {
		tmp = (((0.0625f / s) * t_0) - 0.25f) / -s;
	}
	return tmp;
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    real(4) :: t_0
    real(4) :: t_1
    real(4) :: t_2
    real(4) :: tmp
    t_0 = (x_m * x_m) / s
    t_1 = exp((-abs(x_m) / s))
    t_2 = 1.0e0 + t_1
    if ((t_1 / ((s * t_2) * t_2)) <= 1.000000013351432e-10) then
        tmp = (((-0.0625e0) / s) * t_0) / s
    else
        tmp = (((0.0625e0 / s) * t_0) - 0.25e0) / -s
    end if
    code = tmp
end function

Julia

x_m = abs(x)
function code(x_m, s)
	t_0 = Float32(Float32(x_m * x_m) / s)
	t_1 = exp(Float32(Float32(-abs(x_m)) / s))
	t_2 = Float32(Float32(1.0) + t_1)
	tmp = Float32(0.0)
	if (Float32(t_1 / Float32(Float32(s * t_2) * t_2)) <= Float32(1.000000013351432e-10))
		tmp = Float32(Float32(Float32(Float32(-0.0625) / s) * t_0) / s);
	else
		tmp = Float32(Float32(Float32(Float32(Float32(0.0625) / s) * t_0) - Float32(0.25)) / Float32(-s));
	end
	return tmp
end

MATLAB

x_m = abs(x);
function tmp_2 = code(x_m, s)
	t_0 = (x_m * x_m) / s;
	t_1 = exp((-abs(x_m) / s));
	t_2 = single(1.0) + t_1;
	tmp = single(0.0);
	if ((t_1 / ((s * t_2) * t_2)) <= single(1.000000013351432e-10))
		tmp = ((single(-0.0625) / s) * t_0) / s;
	else
		tmp = (((single(0.0625) / s) * t_0) - single(0.25)) / -s;
	end
	tmp_2 = tmp;
end

Derivation

1. Split input into 2 regimes

2. if (/.f32 (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)) (*.f32 (*.f32 s (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))) (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s))))) < 1.00000001e-10

   1. Initial program 99.5%

      \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
   2. Add Preprocessing
   3. Step-by-step derivation

      1. lift-/.f32 N/A

         \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      2. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      3. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

      4. associate-*l* N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

      5. *-commutative N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

      6. associate-/r* N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

      7. lower-/.f32 N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   4. Applied rewrites 80.0%

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

   5. Taylor expanded in x around 0

      \[\leadsto \frac{\color{blue}{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} + \frac{-1}{4} \cdot \frac{{x}^{2} \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}}}{s} \]
   6. Step-by-step derivation

      1. +-commutative N/A

         \[\leadsto \frac{\color{blue}{\frac{-1}{4} \cdot \frac{{x}^{2} \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}}{s} \]

      2. associate-/l* N/A

         \[\leadsto \frac{\frac{-1}{4} \cdot \color{blue}{\left({x}^{2} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}\right)} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      3. *-commutative N/A

         \[\leadsto \frac{\frac{-1}{4} \cdot \color{blue}{\left(\frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}} \cdot {x}^{2}\right)} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      4. associate-*l* N/A

         \[\leadsto \frac{\color{blue}{\left(\frac{-1}{4} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}\right) \cdot {x}^{2}} + e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{s} \]

      5. lower-fma.f32 N/A

         \[\leadsto \frac{\color{blue}{\mathsf{fma}\left(\frac{-1}{4} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)}}{{s}^{2}}, {x}^{2}, e^{\mathsf{neg}\left(2 \cdot \log 2\right)}\right)}}{s} \]

   7. Applied rewrites 4.5%

      \[\leadsto \frac{\color{blue}{\mathsf{fma}\left(\frac{-0.0625}{s \cdot s}, x \cdot x, 0.25\right)}}{s} \]

   8. Taylor expanded in x around inf

      \[\leadsto \frac{\frac{-1}{16} \cdot \color{blue}{\frac{{x}^{2}}{{s}^{2}}}}{s} \]
   9. Step-by-step derivation

   10. Applied rewrites 8.2%

       \[\leadsto \frac{\frac{-0.0625}{s} \cdot \color{blue}{\frac{x \cdot x}{s}}}{s} \]

   if 1.00000001e-10 < (/.f32 (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)) (*.f32 (*.f32 s (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))) (+.f32 #s(literal 1 binary32) (exp.f32 (/.f32 (neg.f32 (fabs.f32 x)) s)))))

   1. Initial program 99.1%

      \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
   2. Add Preprocessing
   3. Step-by-step derivation

      1. lift-/.f32 N/A

         \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      2. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

      3. lift-*.f32 N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

      4. associate-*l* N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

      5. *-commutative N/A

         \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

      6. associate-/r* N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

      7. lower-/.f32 N/A

         \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   4. Applied rewrites 24.3%

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

   5. Taylor expanded in s around -inf

      \[\leadsto \color{blue}{-1 \cdot \frac{-1 \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)} + \left(\frac{-1}{3} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-3}{4} \cdot {x}^{3} + \left(\frac{1}{4} \cdot {x}^{3} + \frac{1}{2} \cdot {x}^{3}\right)\right)}{{s}^{3}} + \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-1}{4} \cdot {x}^{2} + \frac{1}{2} \cdot {x}^{2}\right)}{{s}^{2}}\right)}{s}} \]

   7. Applied rewrites 54.8%

      \[\leadsto \color{blue}{\frac{\mathsf{fma}\left(-0.08333333333333333, \frac{{x}^{3} \cdot 0}{{s}^{3}}, \frac{\frac{0.0625 \cdot \left(x \cdot x\right)}{s}}{s}\right) - 0.25}{-s}} \]

   8. Taylor expanded in x around 0

      \[\leadsto \frac{\frac{1}{16} \cdot \frac{{x}^{2}}{{s}^{2}} - \frac{1}{4}}{-s} \]
   9. Step-by-step derivation

   10. Applied rewrites 86.5%

       \[\leadsto \frac{\frac{0.0625}{s} \cdot \frac{x \cdot x}{s} - 0.25}{-s} \]
3. Recombined 2 regimes into one program.
4. Add Preprocessing

Alternative 4: 95.2% accurate, 1.5× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \begin{array}{l} t_0 := e^{\frac{-\left|x\_m\right|}{s}}\\ \frac{t\_0}{\left(2 \cdot s\right) \cdot \left(1 + t\_0\right)} \end{array} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (let* ((t_0 (exp (/ (- (fabs x_m)) s)))) (/ t_0 (* (* 2.0 s) (+ 1.0 t_0)))))

C

x_m = fabs(x);
float code(float x_m, float s) {
	float t_0 = expf((-fabsf(x_m) / s));
	return t_0 / ((2.0f * s) * (1.0f + t_0));
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    real(4) :: t_0
    t_0 = exp((-abs(x_m) / s))
    code = t_0 / ((2.0e0 * s) * (1.0e0 + t_0))
end function

Julia

x_m = abs(x)
function code(x_m, s)
	t_0 = exp(Float32(Float32(-abs(x_m)) / s))
	return Float32(t_0 / Float32(Float32(Float32(2.0) * s) * Float32(Float32(1.0) + t_0)))
end

MATLAB

x_m = abs(x);
function tmp = code(x_m, s)
	t_0 = exp((-abs(x_m) / s));
	tmp = t_0 / ((single(2.0) * s) * (single(1.0) + t_0));
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing

3. Taylor expanded in s around inf

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

   1. lower-*.f32 93.8

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(2 \cdot s\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

5. Applied rewrites 93.8%

   \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(2 \cdot s\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
6. Add Preprocessing

Alternative 5: 94.9% accurate, 1.6× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \frac{e^{\frac{-x\_m}{s} - \log 2 \cdot 2}}{s} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (/ (exp (- (/ (- x_m) s) (* (log 2.0) 2.0))) s))

C

x_m = fabs(x);
float code(float x_m, float s) {
	return expf(((-x_m / s) - (logf(2.0f) * 2.0f))) / s;
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    code = exp(((-x_m / s) - (log(2.0e0) * 2.0e0))) / s
end function

Julia

x_m = abs(x)
function code(x_m, s)
	return Float32(exp(Float32(Float32(Float32(-x_m) / s) - Float32(log(Float32(2.0)) * Float32(2.0)))) / s)
end

MATLAB

x_m = abs(x);
function tmp = code(x_m, s)
	tmp = exp(((-x_m / s) - (log(single(2.0)) * single(2.0)))) / s;
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing
3. Step-by-step derivation

   1. lift-/.f32 N/A

      \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

   2. lift-*.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

   3. lift-*.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

   4. associate-*l* N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

   5. *-commutative N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

   6. associate-/r* N/A

      \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   7. lower-/.f32 N/A

      \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

4. Applied rewrites 63.8%

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

5. Taylor expanded in x around 0

   \[\leadsto \frac{e^{\frac{-x}{s} - \color{blue}{\log 2} \cdot 2}}{s} \]
6. Step-by-step derivation

   1. lower-log.f32 52.3

      \[\leadsto \frac{e^{\frac{-x}{s} - \color{blue}{\log 2} \cdot 2}}{s} \]

7. Applied rewrites 52.3%

   \[\leadsto \frac{e^{\frac{-x}{s} - \color{blue}{\log 2} \cdot 2}}{s} \]
8. Add Preprocessing

Alternative 6: 94.9% accurate, 2.9× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \frac{e^{\frac{-x\_m}{s}}}{4 \cdot s} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s) :precision binary32 (/ (exp (/ (- x_m) s)) (* 4.0 s)))

C

x_m = fabs(x);
float code(float x_m, float s) {
	return expf((-x_m / s)) / (4.0f * s);
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    code = exp((-x_m / s)) / (4.0e0 * s)
end function

Julia

x_m = abs(x)
function code(x_m, s)
	return Float32(exp(Float32(Float32(-x_m) / s)) / Float32(Float32(4.0) * s))
end

MATLAB

x_m = abs(x);
function tmp = code(x_m, s)
	tmp = exp((-x_m / s)) / (single(4.0) * s);
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing

3. Taylor expanded in s around inf

   \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{4 \cdot s}} \]
4. Step-by-step derivation

   1. lower-*.f32 93.5

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{4 \cdot s}} \]

5. Applied rewrites 93.5%

   \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{4 \cdot s}} \]
6. Step-by-step derivation

   1. *-lft-identity N/A

      \[\leadsto \frac{\color{blue}{1 \cdot e^{\frac{-\left|x\right|}{s}}}}{4 \cdot s} \]

   2. lift-exp.f32 N/A

      \[\leadsto \frac{1 \cdot \color{blue}{e^{\frac{-\left|x\right|}{s}}}}{4 \cdot s} \]

   3. sinh-+-cosh-rev N/A

      \[\leadsto \frac{1 \cdot \color{blue}{\left(\cosh \left(\frac{-\left|x\right|}{s}\right) + \sinh \left(\frac{-\left|x\right|}{s}\right)\right)}}{4 \cdot s} \]

   4. distribute-rgt-in N/A

      \[\leadsto \frac{\color{blue}{\cosh \left(\frac{-\left|x\right|}{s}\right) \cdot 1 + \sinh \left(\frac{-\left|x\right|}{s}\right) \cdot 1}}{4 \cdot s} \]

   5. fp-cancel-sign-sub-inv N/A

      \[\leadsto \frac{\color{blue}{\cosh \left(\frac{-\left|x\right|}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}}{4 \cdot s} \]

   6. lift-/.f32 N/A

      \[\leadsto \frac{\cosh \color{blue}{\left(\frac{-\left|x\right|}{s}\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   7. lift-neg.f32 N/A

      \[\leadsto \frac{\cosh \left(\frac{\color{blue}{\mathsf{neg}\left(\left|x\right|\right)}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   8. distribute-frac-neg N/A

      \[\leadsto \frac{\cosh \color{blue}{\left(\mathsf{neg}\left(\frac{\left|x\right|}{s}\right)\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   9. cosh-neg N/A

      \[\leadsto \frac{\color{blue}{\cosh \left(\frac{\left|x\right|}{s}\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   10. lift-fabs.f32 N/A

       \[\leadsto \frac{\cosh \left(\frac{\color{blue}{\left|x\right|}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   11. rem-sqrt-square-rev N/A

       \[\leadsto \frac{\cosh \left(\frac{\color{blue}{\sqrt{x \cdot x}}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   12. sqrt-prod N/A

       \[\leadsto \frac{\cosh \left(\frac{\color{blue}{\sqrt{x} \cdot \sqrt{x}}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   13. rem-square-sqrt N/A

       \[\leadsto \frac{\cosh \left(\frac{\color{blue}{x}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   14. cosh-neg N/A

       \[\leadsto \frac{\color{blue}{\cosh \left(\mathsf{neg}\left(\frac{x}{s}\right)\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   15. distribute-frac-neg N/A

       \[\leadsto \frac{\cosh \color{blue}{\left(\frac{\mathsf{neg}\left(x\right)}{s}\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   16. lift-neg.f32 N/A

       \[\leadsto \frac{\cosh \left(\frac{\color{blue}{-x}}{s}\right) \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

   17. lift-/.f32 N/A

       \[\leadsto \frac{\cosh \color{blue}{\left(\frac{-x}{s}\right)} \cdot 1 - \left(\mathsf{neg}\left(\sinh \left(\frac{-\left|x\right|}{s}\right)\right)\right) \cdot 1}{4 \cdot s} \]

7. Applied rewrites 52.3%

   \[\leadsto \frac{\color{blue}{e^{\frac{-x}{s}}}}{4 \cdot s} \]
8. Add Preprocessing

Alternative 7: 27.5% accurate, 7.8× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \frac{\mathsf{fma}\left(\left|0.0625 \cdot x\_m\right|, \frac{\left|x\_m\right|}{s \cdot s}, 0\right) - 0.25}{-s} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s)
 :precision binary32
 (/ (- (fma (fabs (* 0.0625 x_m)) (/ (fabs x_m) (* s s)) 0.0) 0.25) (- s)))

C

x_m = fabs(x);
float code(float x_m, float s) {
	return (fmaf(fabsf((0.0625f * x_m)), (fabsf(x_m) / (s * s)), 0.0f) - 0.25f) / -s;
}

Julia

x_m = abs(x)
function code(x_m, s)
	return Float32(Float32(fma(abs(Float32(Float32(0.0625) * x_m)), Float32(abs(x_m) / Float32(s * s)), Float32(0.0)) - Float32(0.25)) / Float32(-s))
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing
3. Step-by-step derivation

   1. lift-/.f32 N/A

      \[\leadsto \color{blue}{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

   2. lift-*.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}} \]

   3. lift-*.f32 N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)} \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]

   4. associate-*l* N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{s \cdot \left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right)}} \]

   5. *-commutative N/A

      \[\leadsto \frac{e^{\frac{-\left|x\right|}{s}}}{\color{blue}{\left(\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot s}} \]

   6. associate-/r* N/A

      \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

   7. lower-/.f32 N/A

      \[\leadsto \color{blue}{\frac{\frac{e^{\frac{-\left|x\right|}{s}}}{\left(1 + e^{\frac{-\left|x\right|}{s}}\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)}}{s}} \]

4. Applied rewrites 63.8%

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

5. Taylor expanded in s around -inf

   \[\leadsto \color{blue}{-1 \cdot \frac{-1 \cdot e^{\mathsf{neg}\left(2 \cdot \log 2\right)} + \left(\frac{-1}{3} \cdot \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-3}{4} \cdot {x}^{3} + \left(\frac{1}{4} \cdot {x}^{3} + \frac{1}{2} \cdot {x}^{3}\right)\right)}{{s}^{3}} + \frac{e^{\mathsf{neg}\left(2 \cdot \log 2\right)} \cdot \left(\frac{-1}{4} \cdot {x}^{2} + \frac{1}{2} \cdot {x}^{2}\right)}{{s}^{2}}\right)}{s}} \]

7. Applied rewrites 18.7%

   \[\leadsto \color{blue}{\frac{\mathsf{fma}\left(-0.08333333333333333, \frac{{x}^{3} \cdot 0}{{s}^{3}}, \frac{\frac{0.0625 \cdot \left(x \cdot x\right)}{s}}{s}\right) - 0.25}{-s}} \]

9. Applied rewrites 27.9%

   \[\leadsto \frac{\mathsf{fma}\left(\left|0.0625 \cdot x\right|, \left|\frac{x}{s \cdot s}\right|, 0\right) - 0.25}{-s} \]

10. Final simplification 27.9%

    \[\leadsto \frac{\mathsf{fma}\left(\left|0.0625 \cdot x\right|, \frac{\left|x\right|}{s \cdot s}, 0\right) - 0.25}{-s} \]
11. Add Preprocessing

Alternative 8: 27.5% accurate, 31.1× speedup

Math

\[\begin{array}{l} x_m = \left|x\right| \\ \frac{0.25}{s} \end{array} \]

FPCore

x_m = (fabs.f32 x)
(FPCore (x_m s) :precision binary32 (/ 0.25 s))

C

x_m = fabs(x);
float code(float x_m, float s) {
	return 0.25f / s;
}

Fortran

x_m =     private
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(4) function code(x_m, s)
use fmin_fmax_functions
    real(4), intent (in) :: x_m
    real(4), intent (in) :: s
    code = 0.25e0 / s
end function

Julia

x_m = abs(x)
function code(x_m, s)
	return Float32(Float32(0.25) / s)
end

MATLAB

x_m = abs(x);
function tmp = code(x_m, s)
	tmp = single(0.25) / s;
end

Derivation

1. Initial program 99.4%

   \[\frac{e^{\frac{-\left|x\right|}{s}}}{\left(s \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)\right) \cdot \left(1 + e^{\frac{-\left|x\right|}{s}}\right)} \]
2. Add Preprocessing

3. Taylor expanded in s around inf

   \[\leadsto \color{blue}{\frac{\frac{1}{4}}{s}} \]
4. Step-by-step derivation

   1. lower-/.f32 27.9

      \[\leadsto \color{blue}{\frac{0.25}{s}} \]

5. Applied rewrites 27.9%

   \[\leadsto \color{blue}{\frac{0.25}{s}} \]
6. Add Preprocessing

Reproduce

herbie shell --seed 2024346 
(FPCore (x s)
  :name "Logistic distribution"
  :precision binary32
  :pre (and (<= 0.0 s) (<= s 1.0651631))
  (/ (exp (/ (- (fabs x)) s)) (* (* s (+ 1.0 (exp (/ (- (fabs x)) s)))) (+ 1.0 (exp (/ (- (fabs x)) s))))))