
(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))))
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);
}
real(4) function code(x, s)
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
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
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
\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}
Sampling outcomes in binary32 precision:
Herbie found 10 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(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))))
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);
}
real(4) function code(x, s)
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
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
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
\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 (x s) :precision binary32 (let* ((t_0 (exp (/ (fabs x) (- s))))) (/ t_0 (* (fma t_0 s s) (+ t_0 1.0)))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
return t_0 / (fmaf(t_0, s, s) * (t_0 + 1.0f));
}
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) return Float32(t_0 / Float32(fma(t_0, s, s) * Float32(t_0 + Float32(1.0)))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
\frac{t\_0}{\mathsf{fma}\left(t\_0, s, s\right) \cdot \left(t\_0 + 1\right)}
\end{array}
\end{array}
Initial program 99.8%
lift-fabs.f32N/A
remove-double-negN/A
lift-neg.f32N/A
remove-double-negN/A
frac-2negN/A
frac-2negN/A
lift-/.f32N/A
lift-exp.f32N/A
+-commutativeN/A
distribute-rgt-inN/A
*-lft-identityN/A
lower-fma.f3299.8
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-negN/A
lower-neg.f32N/A
lower-/.f3299.8
Applied rewrites99.8%
Final simplification99.8%
(FPCore (x s)
:precision binary32
(let* ((t_0 (exp (/ (fabs x) (- s)))) (t_1 (+ t_0 1.0)))
(if (<= (/ t_0 (* t_1 (* s t_1))) 0.0)
(/ t_0 s)
(/ (fma (/ x s) (/ (* x -0.0625) s) 0.25) s))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
float t_1 = t_0 + 1.0f;
float tmp;
if ((t_0 / (t_1 * (s * t_1))) <= 0.0f) {
tmp = t_0 / s;
} else {
tmp = fmaf((x / s), ((x * -0.0625f) / s), 0.25f) / s;
}
return tmp;
}
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) t_1 = Float32(t_0 + Float32(1.0)) tmp = Float32(0.0) if (Float32(t_0 / Float32(t_1 * Float32(s * t_1))) <= Float32(0.0)) tmp = Float32(t_0 / s); else tmp = Float32(fma(Float32(x / s), Float32(Float32(x * Float32(-0.0625)) / s), Float32(0.25)) / s); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
t_1 := t\_0 + 1\\
\mathbf{if}\;\frac{t\_0}{t\_1 \cdot \left(s \cdot t\_1\right)} \leq 0:\\
\;\;\;\;\frac{t\_0}{s}\\
\mathbf{else}:\\
\;\;\;\;\frac{\mathsf{fma}\left(\frac{x}{s}, \frac{x \cdot -0.0625}{s}, 0.25\right)}{s}\\
\end{array}
\end{array}
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))))) < 0.0Initial program 100.0%
lift-fabs.f32N/A
remove-double-negN/A
lift-neg.f32N/A
remove-double-negN/A
frac-2negN/A
frac-2negN/A
lift-/.f32N/A
lift-exp.f32N/A
+-commutativeN/A
distribute-rgt-inN/A
*-lft-identityN/A
lower-fma.f32100.0
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-negN/A
lower-neg.f32N/A
lower-/.f32100.0
Applied rewrites100.0%
Applied rewrites100.0%
Taylor expanded in s around 0
neg-mul-1N/A
lower-neg.f32N/A
lower-/.f32N/A
lower-fabs.f32100.0
Applied rewrites100.0%
if 0.0 < (/.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))))) Initial program 99.2%
Taylor expanded in s around inf
lower-/.f32N/A
Applied rewrites70.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-/.f32N/A
+-commutativeN/A
lift-/.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
times-fracN/A
lower-fma.f32N/A
lower-/.f32N/A
lower-/.f32N/A
lower-*.f3291.9
Applied rewrites91.9%
Final simplification97.8%
(FPCore (x s)
:precision binary32
(let* ((t_0 (exp (/ (fabs x) (- s)))) (t_1 (+ t_0 1.0)))
(if (<= (/ t_0 (* t_1 (* s t_1))) 200000000753664.0)
(/ (/ -0.0625 (- (* x (* x (/ -0.0625 (* s s)))) 0.25)) s)
(/ (fma (/ x s) (/ (* x -0.0625) s) 0.25) s))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
float t_1 = t_0 + 1.0f;
float tmp;
if ((t_0 / (t_1 * (s * t_1))) <= 200000000753664.0f) {
tmp = (-0.0625f / ((x * (x * (-0.0625f / (s * s)))) - 0.25f)) / s;
} else {
tmp = fmaf((x / s), ((x * -0.0625f) / s), 0.25f) / s;
}
return tmp;
}
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) t_1 = Float32(t_0 + Float32(1.0)) tmp = Float32(0.0) if (Float32(t_0 / Float32(t_1 * Float32(s * t_1))) <= Float32(200000000753664.0)) tmp = Float32(Float32(Float32(-0.0625) / Float32(Float32(x * Float32(x * Float32(Float32(-0.0625) / Float32(s * s)))) - Float32(0.25))) / s); else tmp = Float32(fma(Float32(x / s), Float32(Float32(x * Float32(-0.0625)) / s), Float32(0.25)) / s); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
t_1 := t\_0 + 1\\
\mathbf{if}\;\frac{t\_0}{t\_1 \cdot \left(s \cdot t\_1\right)} \leq 200000000753664:\\
\;\;\;\;\frac{\frac{-0.0625}{x \cdot \left(x \cdot \frac{-0.0625}{s \cdot s}\right) - 0.25}}{s}\\
\mathbf{else}:\\
\;\;\;\;\frac{\mathsf{fma}\left(\frac{x}{s}, \frac{x \cdot -0.0625}{s}, 0.25\right)}{s}\\
\end{array}
\end{array}
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))))) < 2.00000001e14Initial program 99.9%
Taylor expanded in s around inf
lower-/.f32N/A
Applied rewrites20.4%
Applied rewrites14.1%
Taylor expanded in x around 0
Applied rewrites88.4%
if 2.00000001e14 < (/.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))))) Initial program 98.6%
Taylor expanded in s around inf
lower-/.f32N/A
Applied rewrites29.5%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-/.f32N/A
+-commutativeN/A
lift-/.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
times-fracN/A
lower-fma.f32N/A
lower-/.f32N/A
lower-/.f32N/A
lower-*.f3283.2
Applied rewrites83.2%
Final simplification87.9%
(FPCore (x s) :precision binary32 (/ (exp (fma (/ -1.0 s) (fabs x) (* -2.0 (log1p (exp (/ (fabs x) (- s))))))) s))
float code(float x, float s) {
return expf(fmaf((-1.0f / s), fabsf(x), (-2.0f * log1pf(expf((fabsf(x) / -s)))))) / s;
}
function code(x, s) return Float32(exp(fma(Float32(Float32(-1.0) / s), abs(x), Float32(Float32(-2.0) * log1p(exp(Float32(abs(x) / Float32(-s))))))) / s) end
\begin{array}{l}
\\
\frac{e^{\mathsf{fma}\left(\frac{-1}{s}, \left|x\right|, -2 \cdot \mathsf{log1p}\left(e^{\frac{\left|x\right|}{-s}}\right)\right)}}{s}
\end{array}
Initial program 99.8%
lift-fabs.f32N/A
remove-double-negN/A
lift-neg.f32N/A
remove-double-negN/A
frac-2negN/A
frac-2negN/A
lift-/.f32N/A
lift-exp.f32N/A
+-commutativeN/A
distribute-rgt-inN/A
*-lft-identityN/A
lower-fma.f3299.8
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-negN/A
lower-neg.f32N/A
lower-/.f3299.8
Applied rewrites99.8%
Applied rewrites99.8%
lift-fabs.f32N/A
lift-neg.f32N/A
lift-/.f32N/A
lift-exp.f32N/A
lift-log1p.f32N/A
lift-fabs.f32N/A
lift-neg.f32N/A
lift-/.f32N/A
+-commutativeN/A
lift-/.f32N/A
clear-numN/A
associate-/r/N/A
lower-fma.f32N/A
Applied rewrites99.8%
(FPCore (x s) :precision binary32 (let* ((t_0 (/ (fabs x) (- s)))) (/ (exp (fma -2.0 (log1p (exp t_0)) t_0)) s)))
float code(float x, float s) {
float t_0 = fabsf(x) / -s;
return expf(fmaf(-2.0f, log1pf(expf(t_0)), t_0)) / s;
}
function code(x, s) t_0 = Float32(abs(x) / Float32(-s)) return Float32(exp(fma(Float32(-2.0), log1p(exp(t_0)), t_0)) / s) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{\left|x\right|}{-s}\\
\frac{e^{\mathsf{fma}\left(-2, \mathsf{log1p}\left(e^{t\_0}\right), t\_0\right)}}{s}
\end{array}
\end{array}
Initial program 99.8%
lift-fabs.f32N/A
remove-double-negN/A
lift-neg.f32N/A
remove-double-negN/A
frac-2negN/A
frac-2negN/A
lift-/.f32N/A
lift-exp.f32N/A
+-commutativeN/A
distribute-rgt-inN/A
*-lft-identityN/A
lower-fma.f3299.8
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-negN/A
lower-neg.f32N/A
lower-/.f3299.8
Applied rewrites99.8%
Applied rewrites99.8%
(FPCore (x s) :precision binary32 (let* ((t_0 (exp (/ (fabs x) (- s))))) (/ t_0 (* (* s (+ t_0 1.0)) (- 2.0 (/ (fabs x) s))))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
return t_0 / ((s * (t_0 + 1.0f)) * (2.0f - (fabsf(x) / s)));
}
real(4) function code(x, s)
real(4), intent (in) :: x
real(4), intent (in) :: s
real(4) :: t_0
t_0 = exp((abs(x) / -s))
code = t_0 / ((s * (t_0 + 1.0e0)) * (2.0e0 - (abs(x) / s)))
end function
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) return Float32(t_0 / Float32(Float32(s * Float32(t_0 + Float32(1.0))) * Float32(Float32(2.0) - Float32(abs(x) / s)))) end
function tmp = code(x, s) t_0 = exp((abs(x) / -s)); tmp = t_0 / ((s * (t_0 + single(1.0))) * (single(2.0) - (abs(x) / s))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
\frac{t\_0}{\left(s \cdot \left(t\_0 + 1\right)\right) \cdot \left(2 - \frac{\left|x\right|}{s}\right)}
\end{array}
\end{array}
Initial program 99.8%
Taylor expanded in s around inf
mul-1-negN/A
unsub-negN/A
lower--.f32N/A
lower-/.f32N/A
lower-fabs.f3296.8
Applied rewrites96.8%
Final simplification96.8%
(FPCore (x s) :precision binary32 (let* ((t_0 (exp (/ (fabs x) (- s))))) (/ t_0 (* (+ t_0 1.0) (fma s 2.0 (- (fabs x)))))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
return t_0 / ((t_0 + 1.0f) * fmaf(s, 2.0f, -fabsf(x)));
}
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) return Float32(t_0 / Float32(Float32(t_0 + Float32(1.0)) * fma(s, Float32(2.0), Float32(-abs(x))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
\frac{t\_0}{\left(t\_0 + 1\right) \cdot \mathsf{fma}\left(s, 2, -\left|x\right|\right)}
\end{array}
\end{array}
Initial program 99.8%
Taylor expanded in s around inf
mul-1-negN/A
unsub-negN/A
lower--.f32N/A
lower-/.f32N/A
lower-fabs.f3296.8
Applied rewrites96.8%
Taylor expanded in s around 0
+-commutativeN/A
associate-*r*N/A
associate-*r*N/A
distribute-rgt-outN/A
lower-*.f32N/A
+-commutativeN/A
lower-+.f32N/A
lower-exp.f32N/A
neg-mul-1N/A
lower-neg.f32N/A
lower-/.f32N/A
lower-fabs.f32N/A
*-commutativeN/A
lower-fma.f32N/A
mul-1-negN/A
lower-neg.f32N/A
lower-fabs.f3296.8
Applied rewrites96.8%
Final simplification96.8%
(FPCore (x s) :precision binary32 (let* ((t_0 (exp (/ (fabs x) (- s))))) (/ t_0 (* (fma t_0 s s) 2.0))))
float code(float x, float s) {
float t_0 = expf((fabsf(x) / -s));
return t_0 / (fmaf(t_0, s, s) * 2.0f);
}
function code(x, s) t_0 = exp(Float32(abs(x) / Float32(-s))) return Float32(t_0 / Float32(fma(t_0, s, s) * Float32(2.0))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := e^{\frac{\left|x\right|}{-s}}\\
\frac{t\_0}{\mathsf{fma}\left(t\_0, s, s\right) \cdot 2}
\end{array}
\end{array}
Initial program 99.8%
lift-fabs.f32N/A
remove-double-negN/A
lift-neg.f32N/A
remove-double-negN/A
frac-2negN/A
frac-2negN/A
lift-/.f32N/A
lift-exp.f32N/A
+-commutativeN/A
distribute-rgt-inN/A
*-lft-identityN/A
lower-fma.f3299.8
lift-/.f32N/A
lift-neg.f32N/A
distribute-frac-negN/A
lower-neg.f32N/A
lower-/.f3299.8
Applied rewrites99.8%
Taylor expanded in s around inf
Applied rewrites95.4%
Final simplification95.4%
(FPCore (x s) :precision binary32 (/ (exp (/ (fabs x) (- s))) (* s 4.0)))
float code(float x, float s) {
return expf((fabsf(x) / -s)) / (s * 4.0f);
}
real(4) function code(x, s)
real(4), intent (in) :: x
real(4), intent (in) :: s
code = exp((abs(x) / -s)) / (s * 4.0e0)
end function
function code(x, s) return Float32(exp(Float32(abs(x) / Float32(-s))) / Float32(s * Float32(4.0))) end
function tmp = code(x, s) tmp = exp((abs(x) / -s)) / (s * single(4.0)); end
\begin{array}{l}
\\
\frac{e^{\frac{\left|x\right|}{-s}}}{s \cdot 4}
\end{array}
Initial program 99.8%
Taylor expanded in s around inf
*-commutativeN/A
lower-*.f3295.1
Applied rewrites95.1%
Final simplification95.1%
(FPCore (x s) :precision binary32 (/ 0.25 s))
float code(float x, float s) {
return 0.25f / s;
}
real(4) function code(x, s)
real(4), intent (in) :: x
real(4), intent (in) :: s
code = 0.25e0 / s
end function
function code(x, s) return Float32(Float32(0.25) / s) end
function tmp = code(x, s) tmp = single(0.25) / s; end
\begin{array}{l}
\\
\frac{0.25}{s}
\end{array}
Initial program 99.8%
Taylor expanded in s around inf
lower-/.f3227.3
Applied rewrites27.3%
herbie shell --seed 2024219
(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))))))