
(FPCore (u s)
:precision binary32
(let* ((t_0 (/ 1.0 (+ 1.0 (exp (/ PI s))))))
(*
(- s)
(log
(-
(/ 1.0 (+ (* u (- (/ 1.0 (+ 1.0 (exp (/ (- PI) s)))) t_0)) t_0))
1.0)))))
float code(float u, float s) {
float t_0 = 1.0f / (1.0f + expf((((float) M_PI) / s)));
return -s * logf(((1.0f / ((u * ((1.0f / (1.0f + expf((-((float) M_PI) / s)))) - t_0)) + t_0)) - 1.0f));
}
function code(u, s) t_0 = Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / s)))) return Float32(Float32(-s) * log(Float32(Float32(Float32(1.0) / Float32(Float32(u * Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(-Float32(pi)) / s)))) - t_0)) + t_0)) - Float32(1.0)))) end
function tmp = code(u, s) t_0 = single(1.0) / (single(1.0) + exp((single(pi) / s))); tmp = -s * log(((single(1.0) / ((u * ((single(1.0) / (single(1.0) + exp((-single(pi) / s)))) - t_0)) + t_0)) - single(1.0))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{1 + e^{\frac{\pi}{s}}}\\
\left(-s\right) \cdot \log \left(\frac{1}{u \cdot \left(\frac{1}{1 + e^{\frac{-\pi}{s}}} - t\_0\right) + t\_0} - 1\right)
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 10 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (u s)
:precision binary32
(let* ((t_0 (/ 1.0 (+ 1.0 (exp (/ PI s))))))
(*
(- s)
(log
(-
(/ 1.0 (+ (* u (- (/ 1.0 (+ 1.0 (exp (/ (- PI) s)))) t_0)) t_0))
1.0)))))
float code(float u, float s) {
float t_0 = 1.0f / (1.0f + expf((((float) M_PI) / s)));
return -s * logf(((1.0f / ((u * ((1.0f / (1.0f + expf((-((float) M_PI) / s)))) - t_0)) + t_0)) - 1.0f));
}
function code(u, s) t_0 = Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / s)))) return Float32(Float32(-s) * log(Float32(Float32(Float32(1.0) / Float32(Float32(u * Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(-Float32(pi)) / s)))) - t_0)) + t_0)) - Float32(1.0)))) end
function tmp = code(u, s) t_0 = single(1.0) / (single(1.0) + exp((single(pi) / s))); tmp = -s * log(((single(1.0) / ((u * ((single(1.0) / (single(1.0) + exp((-single(pi) / s)))) - t_0)) + t_0)) - single(1.0))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{1 + e^{\frac{\pi}{s}}}\\
\left(-s\right) \cdot \log \left(\frac{1}{u \cdot \left(\frac{1}{1 + e^{\frac{-\pi}{s}}} - t\_0\right) + t\_0} - 1\right)
\end{array}
\end{array}
(FPCore (u s)
:precision binary32
(*
(- s)
(log
(+
-1.0
(/
1.0
(fma
(+ (/ 1.0 (+ 1.0 (exp (/ PI (- s))))) (/ -1.0 (+ 1.0 (exp (/ PI s)))))
u
(/ 1.0 (+ 1.0 (exp (* PI (/ 1.0 s)))))))))))
float code(float u, float s) {
return -s * logf((-1.0f + (1.0f / fmaf(((1.0f / (1.0f + expf((((float) M_PI) / -s)))) + (-1.0f / (1.0f + expf((((float) M_PI) / s))))), u, (1.0f / (1.0f + expf((((float) M_PI) * (1.0f / s)))))))));
}
function code(u, s) return Float32(Float32(-s) * log(Float32(Float32(-1.0) + Float32(Float32(1.0) / fma(Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / Float32(-s))))) + Float32(Float32(-1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / s))))), u, Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) * Float32(Float32(1.0) / s)))))))))) end
\begin{array}{l}
\\
\left(-s\right) \cdot \log \left(-1 + \frac{1}{\mathsf{fma}\left(\frac{1}{1 + e^{\frac{\pi}{-s}}} + \frac{-1}{1 + e^{\frac{\pi}{s}}}, u, \frac{1}{1 + e^{\pi \cdot \frac{1}{s}}}\right)}\right)
\end{array}
Initial program 99.0%
Applied egg-rr99.0%
lift-PI.f32N/A
div-invN/A
lift-/.f32N/A
*-commutativeN/A
lift-*.f3299.0
Applied egg-rr99.0%
Final simplification99.0%
(FPCore (u s)
:precision binary32
(let* ((t_0 (+ 1.0 (exp (/ PI s)))))
(*
(- s)
(log
(+
-1.0
(/
1.0
(fma
(+ (/ 1.0 (+ 1.0 (exp (/ PI (- s))))) (/ -1.0 t_0))
u
(/ 1.0 t_0))))))))
float code(float u, float s) {
float t_0 = 1.0f + expf((((float) M_PI) / s));
return -s * logf((-1.0f + (1.0f / fmaf(((1.0f / (1.0f + expf((((float) M_PI) / -s)))) + (-1.0f / t_0)), u, (1.0f / t_0)))));
}
function code(u, s) t_0 = Float32(Float32(1.0) + exp(Float32(Float32(pi) / s))) return Float32(Float32(-s) * log(Float32(Float32(-1.0) + Float32(Float32(1.0) / fma(Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / Float32(-s))))) + Float32(Float32(-1.0) / t_0)), u, Float32(Float32(1.0) / t_0)))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := 1 + e^{\frac{\pi}{s}}\\
\left(-s\right) \cdot \log \left(-1 + \frac{1}{\mathsf{fma}\left(\frac{1}{1 + e^{\frac{\pi}{-s}}} + \frac{-1}{t\_0}, u, \frac{1}{t\_0}\right)}\right)
\end{array}
\end{array}
Initial program 99.0%
Applied egg-rr99.0%
Final simplification99.0%
(FPCore (u s)
:precision binary32
(*
(- s)
(log
(+
-1.0
(/
1.0
(*
(+ (/ 1.0 (+ 1.0 (exp (/ PI (- s))))) (/ -1.0 (+ 1.0 (exp (/ PI s)))))
u))))))
float code(float u, float s) {
return -s * logf((-1.0f + (1.0f / (((1.0f / (1.0f + expf((((float) M_PI) / -s)))) + (-1.0f / (1.0f + expf((((float) M_PI) / s))))) * u))));
}
function code(u, s) return Float32(Float32(-s) * log(Float32(Float32(-1.0) + Float32(Float32(1.0) / Float32(Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / Float32(-s))))) + Float32(Float32(-1.0) / Float32(Float32(1.0) + exp(Float32(Float32(pi) / s))))) * u))))) end
function tmp = code(u, s) tmp = -s * log((single(-1.0) + (single(1.0) / (((single(1.0) / (single(1.0) + exp((single(pi) / -s)))) + (single(-1.0) / (single(1.0) + exp((single(pi) / s))))) * u)))); end
\begin{array}{l}
\\
\left(-s\right) \cdot \log \left(-1 + \frac{1}{\left(\frac{1}{1 + e^{\frac{\pi}{-s}}} + \frac{-1}{1 + e^{\frac{\pi}{s}}}\right) \cdot u}\right)
\end{array}
Initial program 99.0%
Taylor expanded in u around inf
lower-*.f32N/A
sub-negN/A
lower-+.f32N/A
lower-/.f32N/A
lower-+.f32N/A
lower-exp.f32N/A
mul-1-negN/A
distribute-neg-frac2N/A
mul-1-negN/A
lower-/.f32N/A
lower-PI.f32N/A
mul-1-negN/A
lower-neg.f32N/A
distribute-neg-fracN/A
Simplified97.7%
Final simplification97.7%
(FPCore (u s) :precision binary32 (* (- s) (log (fma (fma PI 0.25 (* PI (* u -0.5))) (/ 4.0 s) 1.0))))
float code(float u, float s) {
return -s * logf(fmaf(fmaf(((float) M_PI), 0.25f, (((float) M_PI) * (u * -0.5f))), (4.0f / s), 1.0f));
}
function code(u, s) return Float32(Float32(-s) * log(fma(fma(Float32(pi), Float32(0.25), Float32(Float32(pi) * Float32(u * Float32(-0.5)))), Float32(Float32(4.0) / s), Float32(1.0)))) end
\begin{array}{l}
\\
\left(-s\right) \cdot \log \left(\mathsf{fma}\left(\mathsf{fma}\left(\pi, 0.25, \pi \cdot \left(u \cdot -0.5\right)\right), \frac{4}{s}, 1\right)\right)
\end{array}
Initial program 99.0%
Taylor expanded in s around -inf
associate-*r/N/A
+-commutativeN/A
*-commutativeN/A
associate-/l*N/A
lower-fma.f32N/A
Simplified24.3%
Final simplification24.3%
(FPCore (u s)
:precision binary32
(let* ((t_0 (* PI (fma -0.5 u 0.25))))
(/
(fma
8.0
(* t_0 t_0)
(fma t_0 (* t_0 -8.0) (* s (fma u (* PI 2.0) (- PI)))))
s)))
float code(float u, float s) {
float t_0 = ((float) M_PI) * fmaf(-0.5f, u, 0.25f);
return fmaf(8.0f, (t_0 * t_0), fmaf(t_0, (t_0 * -8.0f), (s * fmaf(u, (((float) M_PI) * 2.0f), -((float) M_PI))))) / s;
}
function code(u, s) t_0 = Float32(Float32(pi) * fma(Float32(-0.5), u, Float32(0.25))) return Float32(fma(Float32(8.0), Float32(t_0 * t_0), fma(t_0, Float32(t_0 * Float32(-8.0)), Float32(s * fma(u, Float32(Float32(pi) * Float32(2.0)), Float32(-Float32(pi)))))) / s) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \pi \cdot \mathsf{fma}\left(-0.5, u, 0.25\right)\\
\frac{\mathsf{fma}\left(8, t\_0 \cdot t\_0, \mathsf{fma}\left(t\_0, t\_0 \cdot -8, s \cdot \mathsf{fma}\left(u, \pi \cdot 2, -\pi\right)\right)\right)}{s}
\end{array}
\end{array}
Initial program 99.0%
Taylor expanded in s around -inf
Simplified8.7%
add-log-expN/A
*-un-lft-identityN/A
lift-PI.f32N/A
exp-prodN/A
log-powN/A
lower-*.f32N/A
lower-log.f32N/A
exp-1-eN/A
lower-E.f326.9
Applied egg-rr6.9%
Taylor expanded in s around 0
Simplified13.3%
(FPCore (u s) :precision binary32 (/ (fma (* s -4.0) (* PI (fma -0.5 u 0.25)) 0.0) s))
float code(float u, float s) {
return fmaf((s * -4.0f), (((float) M_PI) * fmaf(-0.5f, u, 0.25f)), 0.0f) / s;
}
function code(u, s) return Float32(fma(Float32(s * Float32(-4.0)), Float32(Float32(pi) * fma(Float32(-0.5), u, Float32(0.25))), Float32(0.0)) / s) end
\begin{array}{l}
\\
\frac{\mathsf{fma}\left(s \cdot -4, \pi \cdot \mathsf{fma}\left(-0.5, u, 0.25\right), 0\right)}{s}
\end{array}
Initial program 99.0%
Taylor expanded in s around -inf
Simplified8.7%
lift-PI.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-PI.f32N/A
add-sqr-sqrtN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-PI.f32N/A
lower-sqrt.f32N/A
lift-PI.f32N/A
lower-sqrt.f328.7
Applied egg-rr8.7%
Taylor expanded in s around 0
lower-/.f32N/A
Simplified11.6%
Final simplification11.6%
(FPCore (u s) :precision binary32 (* (* PI (fma -0.5 u 0.25)) -4.0))
float code(float u, float s) {
return (((float) M_PI) * fmaf(-0.5f, u, 0.25f)) * -4.0f;
}
function code(u, s) return Float32(Float32(Float32(pi) * fma(Float32(-0.5), u, Float32(0.25))) * Float32(-4.0)) end
\begin{array}{l}
\\
\left(\pi \cdot \mathsf{fma}\left(-0.5, u, 0.25\right)\right) \cdot -4
\end{array}
Initial program 99.0%
Taylor expanded in s around -inf
Simplified8.7%
add-log-expN/A
*-un-lft-identityN/A
lift-PI.f32N/A
exp-prodN/A
log-powN/A
lower-*.f32N/A
lower-log.f32N/A
exp-1-eN/A
lower-E.f326.9
Applied egg-rr6.9%
Taylor expanded in u around -inf
Simplified9.0%
Taylor expanded in s around inf
lower-*.f32N/A
associate-*r*N/A
distribute-rgt-outN/A
lower-*.f32N/A
lower-PI.f32N/A
lower-fma.f3211.6
Simplified11.6%
Final simplification11.6%
(FPCore (u s) :precision binary32 (fma u (* PI 2.0) (- PI)))
float code(float u, float s) {
return fmaf(u, (((float) M_PI) * 2.0f), -((float) M_PI));
}
function code(u, s) return fma(u, Float32(Float32(pi) * Float32(2.0)), Float32(-Float32(pi))) end
\begin{array}{l}
\\
\mathsf{fma}\left(u, \pi \cdot 2, -\pi\right)
\end{array}
Initial program 99.0%
Applied egg-rr99.0%
lift-PI.f32N/A
div-invN/A
lift-/.f32N/A
*-commutativeN/A
lift-*.f3299.0
Applied egg-rr99.0%
Taylor expanded in s around -inf
cancel-sign-sub-invN/A
metadata-evalN/A
distribute-rgt-inN/A
distribute-rgt-out--N/A
metadata-evalN/A
associate-*r*N/A
associate-*l*N/A
metadata-evalN/A
associate-*r*N/A
*-commutativeN/A
*-commutativeN/A
associate-*l*N/A
metadata-evalN/A
*-commutativeN/A
lower-fma.f32N/A
Simplified11.6%
(FPCore (u s) :precision binary32 (- PI))
float code(float u, float s) {
return -((float) M_PI);
}
function code(u, s) return Float32(-Float32(pi)) end
function tmp = code(u, s) tmp = -single(pi); end
\begin{array}{l}
\\
-\pi
\end{array}
Initial program 99.0%
Taylor expanded in u around 0
mul-1-negN/A
lower-neg.f32N/A
lower-PI.f3211.2
Simplified11.2%
(FPCore (u s) :precision binary32 0.0)
float code(float u, float s) {
return 0.0f;
}
real(4) function code(u, s)
real(4), intent (in) :: u
real(4), intent (in) :: s
code = 0.0e0
end function
function code(u, s) return Float32(0.0) end
function tmp = code(u, s) tmp = single(0.0); end
\begin{array}{l}
\\
0
\end{array}
Initial program 99.0%
Taylor expanded in s around -inf
Simplified8.7%
Taylor expanded in s around 0
associate-*r/N/A
distribute-rgt-outN/A
metadata-evalN/A
mul0-rgtN/A
metadata-evalN/A
lower-/.f3210.3
Simplified10.3%
div010.3
Applied egg-rr10.3%
herbie shell --seed 2024208
(FPCore (u s)
:name "Sample trimmed logistic on [-pi, pi]"
:precision binary32
:pre (and (and (<= 2.328306437e-10 u) (<= u 1.0)) (and (<= 0.0 s) (<= s 1.0651631)))
(* (- s) (log (- (/ 1.0 (+ (* u (- (/ 1.0 (+ 1.0 (exp (/ (- PI) s)))) (/ 1.0 (+ 1.0 (exp (/ PI s)))))) (/ 1.0 (+ 1.0 (exp (/ PI s)))))) 1.0))))