
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))
float code(float cosTheta_i, float u1, float u2) {
return sqrtf(-logf((1.0f - u1))) * cosf(((2.0f * ((float) M_PI)) * u2));
}
function code(cosTheta_i, u1, u2) return Float32(sqrt(Float32(-log(Float32(Float32(1.0) - u1)))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))) end
function tmp = code(cosTheta_i, u1, u2) tmp = sqrt(-log((single(1.0) - u1))) * cos(((single(2.0) * single(pi)) * u2)); end
\begin{array}{l}
\\
\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 21 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))
float code(float cosTheta_i, float u1, float u2) {
return sqrtf(-logf((1.0f - u1))) * cosf(((2.0f * ((float) M_PI)) * u2));
}
function code(cosTheta_i, u1, u2) return Float32(sqrt(Float32(-log(Float32(Float32(1.0) - u1)))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))) end
function tmp = code(cosTheta_i, u1, u2) tmp = sqrt(-log((single(1.0) - u1))) * cos(((single(2.0) * single(pi)) * u2)); end
\begin{array}{l}
\\
\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)
\end{array}
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (* (* PI u2) (* PI u2))))
(*
(sqrt (- (log1p (- u1))))
(cos
(/
(fma (* PI u2) t_0 (* (* PI u2) t_0))
(fma (* PI u2) (* PI u2) (- t_0 t_0)))))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = (((float) M_PI) * u2) * (((float) M_PI) * u2);
return sqrtf(-log1pf(-u1)) * cosf((fmaf((((float) M_PI) * u2), t_0, ((((float) M_PI) * u2) * t_0)) / fmaf((((float) M_PI) * u2), (((float) M_PI) * u2), (t_0 - t_0))));
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(Float32(pi) * u2) * Float32(Float32(pi) * u2)) return Float32(sqrt(Float32(-log1p(Float32(-u1)))) * cos(Float32(fma(Float32(Float32(pi) * u2), t_0, Float32(Float32(Float32(pi) * u2) * t_0)) / fma(Float32(Float32(pi) * u2), Float32(Float32(pi) * u2), Float32(t_0 - t_0))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\pi \cdot u2\right) \cdot \left(\pi \cdot u2\right)\\
\sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \cos \left(\frac{\mathsf{fma}\left(\pi \cdot u2, t\_0, \left(\pi \cdot u2\right) \cdot t\_0\right)}{\mathsf{fma}\left(\pi \cdot u2, \pi \cdot u2, t\_0 - t\_0\right)}\right)
\end{array}
\end{array}
Initial program 62.0%
sub-negN/A
lower-log1p.f32N/A
lower-neg.f3298.9
Applied egg-rr98.9%
lift-PI.f32N/A
associate-*r*N/A
lift-*.f32N/A
count-2N/A
flip3-+N/A
lower-/.f32N/A
Applied egg-rr98.9%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(if (<= (cos (* u2 (* PI 2.0))) 0.9994000196456909)
(*
(sqrt
(/
(fma
(* u1 u1)
(fma u1 (fma u1 -0.3611111111111111 -0.3333333333333333) -0.25)
1.0)
(- u1 (* (* u1 u1) (fma u1 (fma u1 0.25 0.3333333333333333) 0.5)))))
(* u1 (cos (* (* PI u2) 2.0))))
(* (sqrt (- (log1p (- u1)))) (fma (* u2 u2) (* PI (* PI -2.0)) 1.0))))
float code(float cosTheta_i, float u1, float u2) {
float tmp;
if (cosf((u2 * (((float) M_PI) * 2.0f))) <= 0.9994000196456909f) {
tmp = sqrtf((fmaf((u1 * u1), fmaf(u1, fmaf(u1, -0.3611111111111111f, -0.3333333333333333f), -0.25f), 1.0f) / (u1 - ((u1 * u1) * fmaf(u1, fmaf(u1, 0.25f, 0.3333333333333333f), 0.5f))))) * (u1 * cosf(((((float) M_PI) * u2) * 2.0f)));
} else {
tmp = sqrtf(-log1pf(-u1)) * fmaf((u2 * u2), (((float) M_PI) * (((float) M_PI) * -2.0f)), 1.0f);
}
return tmp;
}
function code(cosTheta_i, u1, u2) tmp = Float32(0.0) if (cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) <= Float32(0.9994000196456909)) tmp = Float32(sqrt(Float32(fma(Float32(u1 * u1), fma(u1, fma(u1, Float32(-0.3611111111111111), Float32(-0.3333333333333333)), Float32(-0.25)), Float32(1.0)) / Float32(u1 - Float32(Float32(u1 * u1) * fma(u1, fma(u1, Float32(0.25), Float32(0.3333333333333333)), Float32(0.5)))))) * Float32(u1 * cos(Float32(Float32(Float32(pi) * u2) * Float32(2.0))))); else tmp = Float32(sqrt(Float32(-log1p(Float32(-u1)))) * fma(Float32(u2 * u2), Float32(Float32(pi) * Float32(Float32(pi) * Float32(-2.0))), Float32(1.0))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
\mathbf{if}\;\cos \left(u2 \cdot \left(\pi \cdot 2\right)\right) \leq 0.9994000196456909:\\
\;\;\;\;\sqrt{\frac{\mathsf{fma}\left(u1 \cdot u1, \mathsf{fma}\left(u1, \mathsf{fma}\left(u1, -0.3611111111111111, -0.3333333333333333\right), -0.25\right), 1\right)}{u1 - \left(u1 \cdot u1\right) \cdot \mathsf{fma}\left(u1, \mathsf{fma}\left(u1, 0.25, 0.3333333333333333\right), 0.5\right)}} \cdot \left(u1 \cdot \cos \left(\left(\pi \cdot u2\right) \cdot 2\right)\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{-\mathsf{log1p}\left(-u1\right)} \cdot \mathsf{fma}\left(u2 \cdot u2, \pi \cdot \left(\pi \cdot -2\right), 1\right)\\
\end{array}
\end{array}
if (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2)) < 0.99940002Initial program 58.2%
Taylor expanded in u1 around 0
+-commutativeN/A
distribute-lft-inN/A
associate-*r*N/A
unpow2N/A
*-rgt-identityN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
+-commutativeN/A
lower-fma.f32N/A
+-commutativeN/A
*-commutativeN/A
lower-fma.f3293.1
Simplified93.1%
lift-*.f32N/A
lift-fma.f32N/A
lift-fma.f32N/A
+-commutativeN/A
flip-+N/A
sqrt-divN/A
lower-/.f32N/A
Applied egg-rr92.8%
Taylor expanded in u1 around 0
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
+-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
sub-negN/A
metadata-evalN/A
lower-fma.f32N/A
sub-negN/A
*-commutativeN/A
metadata-evalN/A
lower-fma.f3293.0
Simplified93.0%
Taylor expanded in u2 around inf
*-commutativeN/A
lower-*.f32N/A
Simplified93.0%
if 0.99940002 < (cos.f32 (*.f32 (*.f32 #s(literal 2 binary32) (PI.f32)) u2)) Initial program 57.7%
sub-negN/A
lower-log1p.f32N/A
lower-neg.f3299.5
Applied egg-rr99.5%
Taylor expanded in u2 around 0
+-commutativeN/A
*-commutativeN/A
associate-*r*N/A
*-commutativeN/A
lower-fma.f32N/A
unpow2N/A
lower-*.f32N/A
*-commutativeN/A
unpow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-PI.f32N/A
lower-*.f32N/A
lower-PI.f3299.4
Simplified99.4%
Final simplification97.9%
herbie shell --seed 2024218
(FPCore (cosTheta_i u1 u2)
:name "Beckmann Sample, near normal, slope_x"
:precision binary32
:pre (and (and (and (> cosTheta_i 0.9999) (<= cosTheta_i 1.0)) (and (<= 2.328306437e-10 u1) (<= u1 1.0))) (and (<= 2.328306437e-10 u2) (<= u2 1.0)))
(* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))