
(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 10 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
(if (<= u1 0.03799999877810478)
(*
(sqrt
(-
(*
(- (* (- (* (- (* -0.25 u1) 0.3333333333333333) u1) 0.5) u1) 1.0)
u1)))
(cos (* (* 2.0 PI) u2)))
(* (sqrt (log (/ 1.0 (- 1.0 u1)))) (sin (fma (* PI u2) 2.0 (/ PI 2.0))))))
float code(float cosTheta_i, float u1, float u2) {
float tmp;
if (u1 <= 0.03799999877810478f) {
tmp = sqrtf(-(((((((-0.25f * u1) - 0.3333333333333333f) * u1) - 0.5f) * u1) - 1.0f) * u1)) * cosf(((2.0f * ((float) M_PI)) * u2));
} else {
tmp = sqrtf(logf((1.0f / (1.0f - u1)))) * sinf(fmaf((((float) M_PI) * u2), 2.0f, (((float) M_PI) / 2.0f)));
}
return tmp;
}
function code(cosTheta_i, u1, u2) tmp = Float32(0.0) if (u1 <= Float32(0.03799999877810478)) tmp = Float32(sqrt(Float32(-Float32(Float32(Float32(Float32(Float32(Float32(Float32(Float32(-0.25) * u1) - Float32(0.3333333333333333)) * u1) - Float32(0.5)) * u1) - Float32(1.0)) * u1))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))); else tmp = Float32(sqrt(log(Float32(Float32(1.0) / Float32(Float32(1.0) - u1)))) * sin(fma(Float32(Float32(pi) * u2), Float32(2.0), Float32(Float32(pi) / Float32(2.0))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
\mathbf{if}\;u1 \leq 0.03799999877810478:\\
\;\;\;\;\sqrt{-\left(\left(\left(-0.25 \cdot u1 - 0.3333333333333333\right) \cdot u1 - 0.5\right) \cdot u1 - 1\right) \cdot u1} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\log \left(\frac{1}{1 - u1}\right)} \cdot \sin \left(\mathsf{fma}\left(\pi \cdot u2, 2, \frac{\pi}{2}\right)\right)\\
\end{array}
\end{array}
if u1 < 0.0379999988Initial program 48.4%
Taylor expanded in u1 around 0
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
lower-*.f3298.8
Applied rewrites98.8%
if 0.0379999988 < u1 Initial program 98.2%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3298.3
Applied rewrites98.3%
lift-cos.f32N/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3298.4
Applied rewrites98.4%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(if (<= u1 0.03799999877810478)
(*
(sqrt
(-
(*
(+ (/ (fma (fma 0.3333333333333333 u1 0.5) u1 1.0) (pow u1 3.0)) 0.25)
(- (pow u1 4.0)))))
(cos (* (* 2.0 PI) u2)))
(* (sqrt (log (/ 1.0 (- 1.0 u1)))) (sin (fma (* PI u2) 2.0 (/ PI 2.0))))))
float code(float cosTheta_i, float u1, float u2) {
float tmp;
if (u1 <= 0.03799999877810478f) {
tmp = sqrtf(-(((fmaf(fmaf(0.3333333333333333f, u1, 0.5f), u1, 1.0f) / powf(u1, 3.0f)) + 0.25f) * -powf(u1, 4.0f))) * cosf(((2.0f * ((float) M_PI)) * u2));
} else {
tmp = sqrtf(logf((1.0f / (1.0f - u1)))) * sinf(fmaf((((float) M_PI) * u2), 2.0f, (((float) M_PI) / 2.0f)));
}
return tmp;
}
function code(cosTheta_i, u1, u2) tmp = Float32(0.0) if (u1 <= Float32(0.03799999877810478)) tmp = Float32(sqrt(Float32(-Float32(Float32(Float32(fma(fma(Float32(0.3333333333333333), u1, Float32(0.5)), u1, Float32(1.0)) / (u1 ^ Float32(3.0))) + Float32(0.25)) * Float32(-(u1 ^ Float32(4.0)))))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))); else tmp = Float32(sqrt(log(Float32(Float32(1.0) / Float32(Float32(1.0) - u1)))) * sin(fma(Float32(Float32(pi) * u2), Float32(2.0), Float32(Float32(pi) / Float32(2.0))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
\mathbf{if}\;u1 \leq 0.03799999877810478:\\
\;\;\;\;\sqrt{-\left(\frac{\mathsf{fma}\left(\mathsf{fma}\left(0.3333333333333333, u1, 0.5\right), u1, 1\right)}{{u1}^{3}} + 0.25\right) \cdot \left(-{u1}^{4}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\log \left(\frac{1}{1 - u1}\right)} \cdot \sin \left(\mathsf{fma}\left(\pi \cdot u2, 2, \frac{\pi}{2}\right)\right)\\
\end{array}
\end{array}
if u1 < 0.0379999988Initial program 48.4%
Taylor expanded in u1 around 0
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower--.f32N/A
lower-*.f3298.8
Applied rewrites98.8%
Taylor expanded in u1 around -inf
*-commutativeN/A
lower-*.f32N/A
Applied rewrites98.6%
Taylor expanded in u1 around 0
lower-/.f32N/A
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
+-commutativeN/A
lower-fma.f32N/A
lift-pow.f3298.6
Applied rewrites98.6%
if 0.0379999988 < u1 Initial program 98.2%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3298.3
Applied rewrites98.3%
lift-cos.f32N/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3298.4
Applied rewrites98.4%
Final simplification98.6%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (/ 1.0 (sqrt u1))) (t_1 (cos (* u2 (* PI 2.0)))))
(if (<= u1 0.029999999329447746)
(fma
(fma
(* 0.25 t_0)
t_1
(*
(fma
(* 0.5 (sqrt u1))
(* (- 0.25 (/ 0.0625 u1)) t_1)
(* (* 0.16666666666666666 t_0) t_1))
u1))
(* u1 u1)
(* (sin (fma (* 2.0 u2) (- PI) (/ PI 2.0))) (sqrt u1)))
(*
(sqrt (log (/ 1.0 (- 1.0 u1))))
(sin (fma (* PI u2) 2.0 (/ PI 2.0)))))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = 1.0f / sqrtf(u1);
float t_1 = cosf((u2 * (((float) M_PI) * 2.0f)));
float tmp;
if (u1 <= 0.029999999329447746f) {
tmp = fmaf(fmaf((0.25f * t_0), t_1, (fmaf((0.5f * sqrtf(u1)), ((0.25f - (0.0625f / u1)) * t_1), ((0.16666666666666666f * t_0) * t_1)) * u1)), (u1 * u1), (sinf(fmaf((2.0f * u2), -((float) M_PI), (((float) M_PI) / 2.0f))) * sqrtf(u1)));
} else {
tmp = sqrtf(logf((1.0f / (1.0f - u1)))) * sinf(fmaf((((float) M_PI) * u2), 2.0f, (((float) M_PI) / 2.0f)));
}
return tmp;
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(1.0) / sqrt(u1)) t_1 = cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) tmp = Float32(0.0) if (u1 <= Float32(0.029999999329447746)) tmp = fma(fma(Float32(Float32(0.25) * t_0), t_1, Float32(fma(Float32(Float32(0.5) * sqrt(u1)), Float32(Float32(Float32(0.25) - Float32(Float32(0.0625) / u1)) * t_1), Float32(Float32(Float32(0.16666666666666666) * t_0) * t_1)) * u1)), Float32(u1 * u1), Float32(sin(fma(Float32(Float32(2.0) * u2), Float32(-Float32(pi)), Float32(Float32(pi) / Float32(2.0)))) * sqrt(u1))); else tmp = Float32(sqrt(log(Float32(Float32(1.0) / Float32(Float32(1.0) - u1)))) * sin(fma(Float32(Float32(pi) * u2), Float32(2.0), Float32(Float32(pi) / Float32(2.0))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{\sqrt{u1}}\\
t_1 := \cos \left(u2 \cdot \left(\pi \cdot 2\right)\right)\\
\mathbf{if}\;u1 \leq 0.029999999329447746:\\
\;\;\;\;\mathsf{fma}\left(\mathsf{fma}\left(0.25 \cdot t\_0, t\_1, \mathsf{fma}\left(0.5 \cdot \sqrt{u1}, \left(0.25 - \frac{0.0625}{u1}\right) \cdot t\_1, \left(0.16666666666666666 \cdot t\_0\right) \cdot t\_1\right) \cdot u1\right), u1 \cdot u1, \sin \left(\mathsf{fma}\left(2 \cdot u2, -\pi, \frac{\pi}{2}\right)\right) \cdot \sqrt{u1}\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\log \left(\frac{1}{1 - u1}\right)} \cdot \sin \left(\mathsf{fma}\left(\pi \cdot u2, 2, \frac{\pi}{2}\right)\right)\\
\end{array}
\end{array}
if u1 < 0.0299999993Initial program 47.7%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3245.3
Applied rewrites45.3%
Taylor expanded in u1 around 0
Applied rewrites98.5%
lift-cos.f32N/A
cos-neg-revN/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lower-+.f32N/A
lower-neg.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3298.7
Applied rewrites98.7%
lift-+.f32N/A
lift-neg.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
distribute-rgt-neg-inN/A
lower-fma.f32N/A
lower-neg.f32N/A
lift-PI.f3298.7
Applied rewrites98.7%
if 0.0299999993 < u1 Initial program 97.9%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3297.9
Applied rewrites97.9%
lift-cos.f32N/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3297.9
Applied rewrites97.9%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (/ 1.0 (sqrt u1))) (t_1 (cos (* u2 (* PI 2.0)))))
(if (<= u1 0.029999999329447746)
(fma
(fma
(* 0.25 t_0)
t_1
(*
(fma
(* 0.5 (sqrt u1))
(* (- 0.25 (/ 0.0625 u1)) t_1)
(* (* 0.16666666666666666 t_0) t_1))
u1))
(* u1 u1)
(* (sin (fma (* 2.0 u2) (- PI) (/ PI 2.0))) (sqrt u1)))
(* (sqrt (log (/ 1.0 (- 1.0 u1)))) (cos (* (* 2.0 PI) u2))))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = 1.0f / sqrtf(u1);
float t_1 = cosf((u2 * (((float) M_PI) * 2.0f)));
float tmp;
if (u1 <= 0.029999999329447746f) {
tmp = fmaf(fmaf((0.25f * t_0), t_1, (fmaf((0.5f * sqrtf(u1)), ((0.25f - (0.0625f / u1)) * t_1), ((0.16666666666666666f * t_0) * t_1)) * u1)), (u1 * u1), (sinf(fmaf((2.0f * u2), -((float) M_PI), (((float) M_PI) / 2.0f))) * sqrtf(u1)));
} else {
tmp = sqrtf(logf((1.0f / (1.0f - u1)))) * cosf(((2.0f * ((float) M_PI)) * u2));
}
return tmp;
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(1.0) / sqrt(u1)) t_1 = cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) tmp = Float32(0.0) if (u1 <= Float32(0.029999999329447746)) tmp = fma(fma(Float32(Float32(0.25) * t_0), t_1, Float32(fma(Float32(Float32(0.5) * sqrt(u1)), Float32(Float32(Float32(0.25) - Float32(Float32(0.0625) / u1)) * t_1), Float32(Float32(Float32(0.16666666666666666) * t_0) * t_1)) * u1)), Float32(u1 * u1), Float32(sin(fma(Float32(Float32(2.0) * u2), Float32(-Float32(pi)), Float32(Float32(pi) / Float32(2.0)))) * sqrt(u1))); else tmp = Float32(sqrt(log(Float32(Float32(1.0) / Float32(Float32(1.0) - u1)))) * cos(Float32(Float32(Float32(2.0) * Float32(pi)) * u2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{\sqrt{u1}}\\
t_1 := \cos \left(u2 \cdot \left(\pi \cdot 2\right)\right)\\
\mathbf{if}\;u1 \leq 0.029999999329447746:\\
\;\;\;\;\mathsf{fma}\left(\mathsf{fma}\left(0.25 \cdot t\_0, t\_1, \mathsf{fma}\left(0.5 \cdot \sqrt{u1}, \left(0.25 - \frac{0.0625}{u1}\right) \cdot t\_1, \left(0.16666666666666666 \cdot t\_0\right) \cdot t\_1\right) \cdot u1\right), u1 \cdot u1, \sin \left(\mathsf{fma}\left(2 \cdot u2, -\pi, \frac{\pi}{2}\right)\right) \cdot \sqrt{u1}\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\log \left(\frac{1}{1 - u1}\right)} \cdot \cos \left(\left(2 \cdot \pi\right) \cdot u2\right)\\
\end{array}
\end{array}
if u1 < 0.0299999993Initial program 47.7%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3245.3
Applied rewrites45.3%
Taylor expanded in u1 around 0
Applied rewrites98.5%
lift-cos.f32N/A
cos-neg-revN/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lower-+.f32N/A
lower-neg.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3298.7
Applied rewrites98.7%
lift-+.f32N/A
lift-neg.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
distribute-rgt-neg-inN/A
lower-fma.f32N/A
lower-neg.f32N/A
lift-PI.f3298.7
Applied rewrites98.7%
if 0.0299999993 < u1 Initial program 97.9%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3297.9
Applied rewrites97.9%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (/ 1.0 (sqrt u1))) (t_1 (cos (* u2 (* PI 2.0)))))
(fma
(fma
(* 0.25 t_0)
t_1
(*
(fma
(* 0.5 (sqrt u1))
(* (- 0.25 (/ 0.0625 u1)) t_1)
(* (* 0.16666666666666666 t_0) t_1))
u1))
(* u1 u1)
(* (sin (fma (* 2.0 u2) (- PI) (/ PI 2.0))) (sqrt u1)))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = 1.0f / sqrtf(u1);
float t_1 = cosf((u2 * (((float) M_PI) * 2.0f)));
return fmaf(fmaf((0.25f * t_0), t_1, (fmaf((0.5f * sqrtf(u1)), ((0.25f - (0.0625f / u1)) * t_1), ((0.16666666666666666f * t_0) * t_1)) * u1)), (u1 * u1), (sinf(fmaf((2.0f * u2), -((float) M_PI), (((float) M_PI) / 2.0f))) * sqrtf(u1)));
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(1.0) / sqrt(u1)) t_1 = cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) return fma(fma(Float32(Float32(0.25) * t_0), t_1, Float32(fma(Float32(Float32(0.5) * sqrt(u1)), Float32(Float32(Float32(0.25) - Float32(Float32(0.0625) / u1)) * t_1), Float32(Float32(Float32(0.16666666666666666) * t_0) * t_1)) * u1)), Float32(u1 * u1), Float32(sin(fma(Float32(Float32(2.0) * u2), Float32(-Float32(pi)), Float32(Float32(pi) / Float32(2.0)))) * sqrt(u1))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{\sqrt{u1}}\\
t_1 := \cos \left(u2 \cdot \left(\pi \cdot 2\right)\right)\\
\mathsf{fma}\left(\mathsf{fma}\left(0.25 \cdot t\_0, t\_1, \mathsf{fma}\left(0.5 \cdot \sqrt{u1}, \left(0.25 - \frac{0.0625}{u1}\right) \cdot t\_1, \left(0.16666666666666666 \cdot t\_0\right) \cdot t\_1\right) \cdot u1\right), u1 \cdot u1, \sin \left(\mathsf{fma}\left(2 \cdot u2, -\pi, \frac{\pi}{2}\right)\right) \cdot \sqrt{u1}\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
Applied rewrites92.2%
lift-cos.f32N/A
cos-neg-revN/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lower-+.f32N/A
lower-neg.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3292.4
Applied rewrites92.4%
lift-+.f32N/A
lift-neg.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
distribute-rgt-neg-inN/A
lower-fma.f32N/A
lower-neg.f32N/A
lift-PI.f3292.4
Applied rewrites92.4%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (/ 1.0 (sqrt u1))) (t_1 (cos (* u2 (* PI 2.0)))))
(fma
(fma
(* 0.25 t_0)
t_1
(*
(fma
(* 0.5 (sqrt u1))
(* (- 0.25 (/ 0.0625 u1)) t_1)
(* (* 0.16666666666666666 t_0) t_1))
u1))
(* u1 u1)
(* (sin (+ (* (* 2.0 u2) (- PI)) (/ PI 2.0))) (sqrt u1)))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = 1.0f / sqrtf(u1);
float t_1 = cosf((u2 * (((float) M_PI) * 2.0f)));
return fmaf(fmaf((0.25f * t_0), t_1, (fmaf((0.5f * sqrtf(u1)), ((0.25f - (0.0625f / u1)) * t_1), ((0.16666666666666666f * t_0) * t_1)) * u1)), (u1 * u1), (sinf((((2.0f * u2) * -((float) M_PI)) + (((float) M_PI) / 2.0f))) * sqrtf(u1)));
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(1.0) / sqrt(u1)) t_1 = cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) return fma(fma(Float32(Float32(0.25) * t_0), t_1, Float32(fma(Float32(Float32(0.5) * sqrt(u1)), Float32(Float32(Float32(0.25) - Float32(Float32(0.0625) / u1)) * t_1), Float32(Float32(Float32(0.16666666666666666) * t_0) * t_1)) * u1)), Float32(u1 * u1), Float32(sin(Float32(Float32(Float32(Float32(2.0) * u2) * Float32(-Float32(pi))) + Float32(Float32(pi) / Float32(2.0)))) * sqrt(u1))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{\sqrt{u1}}\\
t_1 := \cos \left(u2 \cdot \left(\pi \cdot 2\right)\right)\\
\mathsf{fma}\left(\mathsf{fma}\left(0.25 \cdot t\_0, t\_1, \mathsf{fma}\left(0.5 \cdot \sqrt{u1}, \left(0.25 - \frac{0.0625}{u1}\right) \cdot t\_1, \left(0.16666666666666666 \cdot t\_0\right) \cdot t\_1\right) \cdot u1\right), u1 \cdot u1, \sin \left(\left(2 \cdot u2\right) \cdot \left(-\pi\right) + \frac{\pi}{2}\right) \cdot \sqrt{u1}\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
Applied rewrites92.2%
lift-cos.f32N/A
cos-neg-revN/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lower-+.f32N/A
lower-neg.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3292.4
Applied rewrites92.4%
Final simplification92.4%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (/ 1.0 (sqrt u1))) (t_1 (cos (* u2 (* PI 2.0)))))
(fma
(fma
(* 0.25 t_0)
t_1
(*
(fma
(* 0.5 (sqrt u1))
(* (- 0.25 (/ 0.0625 u1)) t_1)
(* (* 0.16666666666666666 t_0) t_1))
u1))
(* u1 u1)
(* t_1 (sqrt u1)))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = 1.0f / sqrtf(u1);
float t_1 = cosf((u2 * (((float) M_PI) * 2.0f)));
return fmaf(fmaf((0.25f * t_0), t_1, (fmaf((0.5f * sqrtf(u1)), ((0.25f - (0.0625f / u1)) * t_1), ((0.16666666666666666f * t_0) * t_1)) * u1)), (u1 * u1), (t_1 * sqrtf(u1)));
}
function code(cosTheta_i, u1, u2) t_0 = Float32(Float32(1.0) / sqrt(u1)) t_1 = cos(Float32(u2 * Float32(Float32(pi) * Float32(2.0)))) return fma(fma(Float32(Float32(0.25) * t_0), t_1, Float32(fma(Float32(Float32(0.5) * sqrt(u1)), Float32(Float32(Float32(0.25) - Float32(Float32(0.0625) / u1)) * t_1), Float32(Float32(Float32(0.16666666666666666) * t_0) * t_1)) * u1)), Float32(u1 * u1), Float32(t_1 * sqrt(u1))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \frac{1}{\sqrt{u1}}\\
t_1 := \cos \left(u2 \cdot \left(\pi \cdot 2\right)\right)\\
\mathsf{fma}\left(\mathsf{fma}\left(0.25 \cdot t\_0, t\_1, \mathsf{fma}\left(0.5 \cdot \sqrt{u1}, \left(0.25 - \frac{0.0625}{u1}\right) \cdot t\_1, \left(0.16666666666666666 \cdot t\_0\right) \cdot t\_1\right) \cdot u1\right), u1 \cdot u1, t\_1 \cdot \sqrt{u1}\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
Applied rewrites92.2%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (cos (* 2.0 (* u2 PI))))
(t_1 (/ 1.0 (sqrt u1)))
(t_2 (* t_1 t_0))
(t_3 (* t_1 (- t_0))))
(*
(pow u1 4.0)
(fma
-1.0
(/
(fma
-1.0
(fma 0.03125 t_3 (* 0.16666666666666666 t_2))
(/
(fma
-1.0
(- (* (pow u1 -1.5) (sin (fma 0.5 PI (* -2.0 (* u2 PI))))))
(* 0.25 t_2))
(- u1)))
u1)
(* -0.125 t_3)))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = cosf((2.0f * (u2 * ((float) M_PI))));
float t_1 = 1.0f / sqrtf(u1);
float t_2 = t_1 * t_0;
float t_3 = t_1 * -t_0;
return powf(u1, 4.0f) * fmaf(-1.0f, (fmaf(-1.0f, fmaf(0.03125f, t_3, (0.16666666666666666f * t_2)), (fmaf(-1.0f, -(powf(u1, -1.5f) * sinf(fmaf(0.5f, ((float) M_PI), (-2.0f * (u2 * ((float) M_PI)))))), (0.25f * t_2)) / -u1)) / u1), (-0.125f * t_3));
}
function code(cosTheta_i, u1, u2) t_0 = cos(Float32(Float32(2.0) * Float32(u2 * Float32(pi)))) t_1 = Float32(Float32(1.0) / sqrt(u1)) t_2 = Float32(t_1 * t_0) t_3 = Float32(t_1 * Float32(-t_0)) return Float32((u1 ^ Float32(4.0)) * fma(Float32(-1.0), Float32(fma(Float32(-1.0), fma(Float32(0.03125), t_3, Float32(Float32(0.16666666666666666) * t_2)), Float32(fma(Float32(-1.0), Float32(-Float32((u1 ^ Float32(-1.5)) * sin(fma(Float32(0.5), Float32(pi), Float32(Float32(-2.0) * Float32(u2 * Float32(pi))))))), Float32(Float32(0.25) * t_2)) / Float32(-u1))) / u1), Float32(Float32(-0.125) * t_3))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \cos \left(2 \cdot \left(u2 \cdot \pi\right)\right)\\
t_1 := \frac{1}{\sqrt{u1}}\\
t_2 := t\_1 \cdot t\_0\\
t_3 := t\_1 \cdot \left(-t\_0\right)\\
{u1}^{4} \cdot \mathsf{fma}\left(-1, \frac{\mathsf{fma}\left(-1, \mathsf{fma}\left(0.03125, t\_3, 0.16666666666666666 \cdot t\_2\right), \frac{\mathsf{fma}\left(-1, -{u1}^{-1.5} \cdot \sin \left(\mathsf{fma}\left(0.5, \pi, -2 \cdot \left(u2 \cdot \pi\right)\right)\right), 0.25 \cdot t\_2\right)}{-u1}\right)}{u1}, -0.125 \cdot t\_3\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
Applied rewrites92.2%
lift-cos.f32N/A
cos-neg-revN/A
sin-+PI/2-revN/A
lower-sin.f32N/A
lower-+.f32N/A
lower-neg.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-PI.f32N/A
lower-/.f32N/A
lift-PI.f3292.4
Applied rewrites92.4%
Taylor expanded in u1 around -inf
Applied rewrites91.7%
Final simplification91.7%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (cos (* (* 2.0 u2) PI))))
(*
(fma
(/ 0.16666666666666666 (sqrt u1))
(- t_0)
(/ (fma (/ 0.25 (sqrt u1)) t_0 (* (sqrt (pow u1 -3.0)) t_0)) (- u1)))
(- (pow u1 3.0)))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = cosf(((2.0f * u2) * ((float) M_PI)));
return fmaf((0.16666666666666666f / sqrtf(u1)), -t_0, (fmaf((0.25f / sqrtf(u1)), t_0, (sqrtf(powf(u1, -3.0f)) * t_0)) / -u1)) * -powf(u1, 3.0f);
}
function code(cosTheta_i, u1, u2) t_0 = cos(Float32(Float32(Float32(2.0) * u2) * Float32(pi))) return Float32(fma(Float32(Float32(0.16666666666666666) / sqrt(u1)), Float32(-t_0), Float32(fma(Float32(Float32(0.25) / sqrt(u1)), t_0, Float32(sqrt((u1 ^ Float32(-3.0))) * t_0)) / Float32(-u1))) * Float32(-(u1 ^ Float32(3.0)))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \cos \left(\left(2 \cdot u2\right) \cdot \pi\right)\\
\mathsf{fma}\left(\frac{0.16666666666666666}{\sqrt{u1}}, -t\_0, \frac{\mathsf{fma}\left(\frac{0.25}{\sqrt{u1}}, t\_0, \sqrt{{u1}^{-3}} \cdot t\_0\right)}{-u1}\right) \cdot \left(-{u1}^{3}\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites89.8%
Taylor expanded in u1 around -inf
Applied rewrites89.3%
Final simplification89.3%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (cos (* (* 2.0 u2) PI))))
(*
(fma
(/ 0.16666666666666666 (sqrt u1))
(- t_0)
(/ (fma (/ 0.25 (sqrt u1)) t_0 (* (sqrt (pow u1 -3.0)) t_0)) (- u1)))
(- (exp (* (log u1) 3.0))))))
float code(float cosTheta_i, float u1, float u2) {
float t_0 = cosf(((2.0f * u2) * ((float) M_PI)));
return fmaf((0.16666666666666666f / sqrtf(u1)), -t_0, (fmaf((0.25f / sqrtf(u1)), t_0, (sqrtf(powf(u1, -3.0f)) * t_0)) / -u1)) * -expf((logf(u1) * 3.0f));
}
function code(cosTheta_i, u1, u2) t_0 = cos(Float32(Float32(Float32(2.0) * u2) * Float32(pi))) return Float32(fma(Float32(Float32(0.16666666666666666) / sqrt(u1)), Float32(-t_0), Float32(fma(Float32(Float32(0.25) / sqrt(u1)), t_0, Float32(sqrt((u1 ^ Float32(-3.0))) * t_0)) / Float32(-u1))) * Float32(-exp(Float32(log(u1) * Float32(3.0))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \cos \left(\left(2 \cdot u2\right) \cdot \pi\right)\\
\mathsf{fma}\left(\frac{0.16666666666666666}{\sqrt{u1}}, -t\_0, \frac{\mathsf{fma}\left(\frac{0.25}{\sqrt{u1}}, t\_0, \sqrt{{u1}^{-3}} \cdot t\_0\right)}{-u1}\right) \cdot \left(-e^{\log u1 \cdot 3}\right)
\end{array}
\end{array}
Initial program 57.1%
lift-neg.f32N/A
lift--.f32N/A
lift-log.f32N/A
neg-logN/A
lower-log.f32N/A
lower-/.f32N/A
lift--.f3255.2
Applied rewrites55.2%
Taylor expanded in u1 around 0
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites89.8%
Taylor expanded in u1 around -inf
Applied rewrites89.3%
lift-pow.f32N/A
pow-to-expN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3282.3
Applied rewrites82.3%
Final simplification82.3%
herbie shell --seed 2025057
(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))))