
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 (PI)) u2))))
\begin{array}{l}
\\
\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u2\right)
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 7 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt (- (log (- 1.0 u1)))) (cos (* (* 2.0 (PI)) u2))))
\begin{array}{l}
\\
\sqrt{-\log \left(1 - u1\right)} \cdot \cos \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u2\right)
\end{array}
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (- (log (- 1.0 u1)))))
(if (<= t_0 0.00015500000154133886)
(* (pow (* u1 u1) 0.25) (cos (* (* 2.0 (PI)) u2)))
(*
(sqrt t_0)
(cos (* (* (* 2.0 u2) (cbrt (* (PI) (PI)))) (cbrt (PI))))))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := -\log \left(1 - u1\right)\\
\mathbf{if}\;t\_0 \leq 0.00015500000154133886:\\
\;\;\;\;{\left(u1 \cdot u1\right)}^{0.25} \cdot \cos \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u2\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{t\_0} \cdot \cos \left(\left(\left(2 \cdot u2\right) \cdot \sqrt[3]{\mathsf{PI}\left(\right) \cdot \mathsf{PI}\left(\right)}\right) \cdot \sqrt[3]{\mathsf{PI}\left(\right)}\right)\\
\end{array}
\end{array}
if (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) < 1.55000002e-4Initial program 39.7%
Applied rewrites52.7%
Taylor expanded in u1 around 0
unpow2N/A
lower-*.f3292.0
Applied rewrites92.0%
if 1.55000002e-4 < (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) Initial program 89.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lift-PI.f32N/A
add-cube-cbrtN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-PI.f32N/A
lower-cbrt.f32N/A
lift-PI.f32N/A
lower-cbrt.f3289.2
Applied rewrites89.2%
Taylor expanded in u2 around 0
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-cbrt.f32N/A
unpow2N/A
lower-*.f32N/A
lower-PI.f32N/A
lower-PI.f3289.2
Applied rewrites89.2%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (- (log (- 1.0 u1)))) (t_1 (cbrt (PI))))
(if (<= t_0 9.999999747378752e-5)
(*
(sqrt (+ (- (log1p (pow (- u1) 3.0))) (log1p (fma u1 u1 u1))))
(cos (* (* (* u2 2.0) (pow t_1 2.0)) t_1)))
(* (sqrt t_0) (cos (* (* (* 2.0 u2) (cbrt (* (PI) (PI)))) t_1))))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := -\log \left(1 - u1\right)\\
t_1 := \sqrt[3]{\mathsf{PI}\left(\right)}\\
\mathbf{if}\;t\_0 \leq 9.999999747378752 \cdot 10^{-5}:\\
\;\;\;\;\sqrt{\left(-\mathsf{log1p}\left({\left(-u1\right)}^{3}\right)\right) + \mathsf{log1p}\left(\mathsf{fma}\left(u1, u1, u1\right)\right)} \cdot \cos \left(\left(\left(u2 \cdot 2\right) \cdot {t\_1}^{2}\right) \cdot t\_1\right)\\
\mathbf{else}:\\
\;\;\;\;\sqrt{t\_0} \cdot \cos \left(\left(\left(2 \cdot u2\right) \cdot \sqrt[3]{\mathsf{PI}\left(\right) \cdot \mathsf{PI}\left(\right)}\right) \cdot t\_1\right)\\
\end{array}
\end{array}
if (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) < 9.99999975e-5Initial program 39.1%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lift-PI.f32N/A
add-cube-cbrtN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-PI.f32N/A
lower-cbrt.f32N/A
lift-PI.f32N/A
lower-cbrt.f3239.1
Applied rewrites39.1%
lift-log.f32N/A
lift--.f32N/A
flip3--N/A
log-divN/A
lower--.f32N/A
metadata-evalN/A
sub-negN/A
metadata-evalN/A
cube-negN/A
lift-neg.f32N/A
metadata-evalN/A
lower-log1p.f32N/A
lower-pow.f32N/A
metadata-evalN/A
lower-log1p.f32N/A
*-lft-identityN/A
lower-fma.f3231.1
Applied rewrites31.1%
if 9.99999975e-5 < (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) Initial program 88.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lift-PI.f32N/A
add-cube-cbrtN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-PI.f32N/A
lower-cbrt.f32N/A
lift-PI.f32N/A
lower-cbrt.f3288.6
Applied rewrites88.6%
Taylor expanded in u2 around 0
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-cbrt.f32N/A
unpow2N/A
lower-*.f32N/A
lower-PI.f32N/A
lower-PI.f3288.6
Applied rewrites88.6%
Final simplification52.9%
(FPCore (cosTheta_i u1 u2)
:precision binary32
(let* ((t_0 (- (log (- 1.0 u1)))) (t_1 (cos (* (* 2.0 (PI)) u2))))
(if (<= t_0 0.00015500000154133886)
(* (pow (* u1 u1) 0.25) t_1)
(* (sqrt t_0) t_1))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := -\log \left(1 - u1\right)\\
t_1 := \cos \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u2\right)\\
\mathbf{if}\;t\_0 \leq 0.00015500000154133886:\\
\;\;\;\;{\left(u1 \cdot u1\right)}^{0.25} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;\sqrt{t\_0} \cdot t\_1\\
\end{array}
\end{array}
if (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) < 1.55000002e-4Initial program 39.7%
Applied rewrites52.8%
Taylor expanded in u1 around 0
unpow2N/A
lower-*.f3292.0
Applied rewrites92.0%
if 1.55000002e-4 < (neg.f32 (log.f32 (-.f32 #s(literal 1 binary32) u1))) Initial program 89.2%
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt (- (* (fma -0.5 u1 1.0) u1) (* (- u1) u1))) (cos (* (* 2.0 (PI)) u2))))
\begin{array}{l}
\\
\sqrt{\mathsf{fma}\left(-0.5, u1, 1\right) \cdot u1 - \left(-u1\right) \cdot u1} \cdot \cos \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u2\right)
\end{array}
Initial program 57.9%
Applied rewrites42.0%
Taylor expanded in u1 around 0
*-commutativeN/A
lower-*.f32N/A
+-commutativeN/A
lower-fma.f3239.2
Applied rewrites38.8%
Taylor expanded in u1 around 0
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
mul-1-negN/A
lower-neg.f3256.7
Applied rewrites56.7%
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (cos (* (* (PI) u2) 2.0)) (sqrt u1)))
\begin{array}{l}
\\
\cos \left(\left(\mathsf{PI}\left(\right) \cdot u2\right) \cdot 2\right) \cdot \sqrt{u1}
\end{array}
Initial program 57.9%
Applied rewrites42.0%
Taylor expanded in u1 around 0
*-commutativeN/A
lower-*.f32N/A
lower-cos.f32N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-PI.f32N/A
lower-sqrt.f3277.4
Applied rewrites77.4%
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (pow (* (- u1) (- u1)) 0.25) 1.0))
float code(float cosTheta_i, float u1, float u2) {
return powf((-u1 * -u1), 0.25f) * 1.0f;
}
real(4) function code(costheta_i, u1, u2)
real(4), intent (in) :: costheta_i
real(4), intent (in) :: u1
real(4), intent (in) :: u2
code = ((-u1 * -u1) ** 0.25e0) * 1.0e0
end function
function code(cosTheta_i, u1, u2) return Float32((Float32(Float32(-u1) * Float32(-u1)) ^ Float32(0.25)) * Float32(1.0)) end
function tmp = code(cosTheta_i, u1, u2) tmp = ((-u1 * -u1) ^ single(0.25)) * single(1.0); end
\begin{array}{l}
\\
{\left(\left(-u1\right) \cdot \left(-u1\right)\right)}^{0.25} \cdot 1
\end{array}
Initial program 57.9%
Taylor expanded in u1 around 0
mul-1-negN/A
lower-neg.f3277.4
Applied rewrites77.4%
Taylor expanded in u2 around 0
Applied rewrites63.1%
lift-sqrt.f32N/A
pow1/2N/A
metadata-evalN/A
pow-sqrN/A
pow-prod-downN/A
lower-pow.f32N/A
lower-*.f3263.1
Applied rewrites63.1%
Final simplification63.1%
(FPCore (cosTheta_i u1 u2) :precision binary32 (* (sqrt u1) 1.0))
float code(float cosTheta_i, float u1, float u2) {
return sqrtf(u1) * 1.0f;
}
real(4) function code(costheta_i, u1, u2)
real(4), intent (in) :: costheta_i
real(4), intent (in) :: u1
real(4), intent (in) :: u2
code = sqrt(u1) * 1.0e0
end function
function code(cosTheta_i, u1, u2) return Float32(sqrt(u1) * Float32(1.0)) end
function tmp = code(cosTheta_i, u1, u2) tmp = sqrt(u1) * single(1.0); end
\begin{array}{l}
\\
\sqrt{u1} \cdot 1
\end{array}
Initial program 57.9%
Applied rewrites44.7%
Taylor expanded in u1 around 0
unpow2N/A
lower-*.f3277.4
Applied rewrites77.4%
Taylor expanded in u1 around 0
lower-sqrt.f3277.4
Applied rewrites77.4%
Taylor expanded in u2 around 0
Applied rewrites63.1%
herbie shell --seed 2024309
(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))))