
(FPCore (ux uy maxCos) :precision binary32 (let* ((t_0 (+ (- 1.0 ux) (* ux maxCos)))) (* (cos (* (* uy 2.0) PI)) (sqrt (- 1.0 (* t_0 t_0))))))
float code(float ux, float uy, float maxCos) {
float t_0 = (1.0f - ux) + (ux * maxCos);
return cosf(((uy * 2.0f) * ((float) M_PI))) * sqrtf((1.0f - (t_0 * t_0)));
}
function code(ux, uy, maxCos) t_0 = Float32(Float32(Float32(1.0) - ux) + Float32(ux * maxCos)) return Float32(cos(Float32(Float32(uy * Float32(2.0)) * Float32(pi))) * sqrt(Float32(Float32(1.0) - Float32(t_0 * t_0)))) end
function tmp = code(ux, uy, maxCos) t_0 = (single(1.0) - ux) + (ux * maxCos); tmp = cos(((uy * single(2.0)) * single(pi))) * sqrt((single(1.0) - (t_0 * t_0))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(1 - ux\right) + ux \cdot maxCos\\
\cos \left(\left(uy \cdot 2\right) \cdot \pi\right) \cdot \sqrt{1 - t\_0 \cdot t\_0}
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 17 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (ux uy maxCos) :precision binary32 (let* ((t_0 (+ (- 1.0 ux) (* ux maxCos)))) (* (cos (* (* uy 2.0) PI)) (sqrt (- 1.0 (* t_0 t_0))))))
float code(float ux, float uy, float maxCos) {
float t_0 = (1.0f - ux) + (ux * maxCos);
return cosf(((uy * 2.0f) * ((float) M_PI))) * sqrtf((1.0f - (t_0 * t_0)));
}
function code(ux, uy, maxCos) t_0 = Float32(Float32(Float32(1.0) - ux) + Float32(ux * maxCos)) return Float32(cos(Float32(Float32(uy * Float32(2.0)) * Float32(pi))) * sqrt(Float32(Float32(1.0) - Float32(t_0 * t_0)))) end
function tmp = code(ux, uy, maxCos) t_0 = (single(1.0) - ux) + (ux * maxCos); tmp = cos(((uy * single(2.0)) * single(pi))) * sqrt((single(1.0) - (t_0 * t_0))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(1 - ux\right) + ux \cdot maxCos\\
\cos \left(\left(uy \cdot 2\right) \cdot \pi\right) \cdot \sqrt{1 - t\_0 \cdot t\_0}
\end{array}
\end{array}
(FPCore (ux uy maxCos)
:precision binary32
(let* ((t_0 (cos (* uy PI))))
(*
(- (* t_0 t_0) (* (sin (* uy (* PI (log E)))) (sin (* uy PI))))
(sqrt
(*
ux
(fma maxCos -2.0 (fma (- 1.0 maxCos) (fma ux maxCos (- ux)) 2.0)))))))
float code(float ux, float uy, float maxCos) {
float t_0 = cosf((uy * ((float) M_PI)));
return ((t_0 * t_0) - (sinf((uy * (((float) M_PI) * logf(((float) M_E))))) * sinf((uy * ((float) M_PI))))) * sqrtf((ux * fmaf(maxCos, -2.0f, fmaf((1.0f - maxCos), fmaf(ux, maxCos, -ux), 2.0f))));
}
function code(ux, uy, maxCos) t_0 = cos(Float32(uy * Float32(pi))) return Float32(Float32(Float32(t_0 * t_0) - Float32(sin(Float32(uy * Float32(Float32(pi) * log(Float32(exp(1)))))) * sin(Float32(uy * Float32(pi))))) * sqrt(Float32(ux * fma(maxCos, Float32(-2.0), fma(Float32(Float32(1.0) - maxCos), fma(ux, maxCos, Float32(-ux)), Float32(2.0)))))) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \cos \left(uy \cdot \pi\right)\\
\left(t\_0 \cdot t\_0 - \sin \left(uy \cdot \left(\pi \cdot \log e\right)\right) \cdot \sin \left(uy \cdot \pi\right)\right) \cdot \sqrt{ux \cdot \mathsf{fma}\left(maxCos, -2, \mathsf{fma}\left(1 - maxCos, \mathsf{fma}\left(ux, maxCos, -ux\right), 2\right)\right)}
\end{array}
\end{array}
Initial program 57.5%
Taylor expanded in ux around 0
lower-*.f32N/A
cancel-sign-sub-invN/A
metadata-evalN/A
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
+-commutativeN/A
mul-1-negN/A
*-commutativeN/A
unpow2N/A
associate-*l*N/A
*-commutativeN/A
distribute-lft-neg-inN/A
neg-sub0N/A
associate-+l-N/A
neg-sub0N/A
mul-1-negN/A
+-commutativeN/A
Applied rewrites98.8%
lift-cos.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-PI.f32N/A
*-commutativeN/A
lift-PI.f32N/A
associate-*l*N/A
cos-2N/A
lower--.f32N/A
lower-*.f32N/A
lower-cos.f32N/A
lower-*.f32N/A
lower-cos.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-sin.f32N/A
lower-*.f32N/A
lower-sin.f32N/A
lower-*.f3298.7
Applied rewrites98.7%
lift-PI.f32N/A
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.f3298.9
Applied rewrites98.9%
(FPCore (ux uy maxCos)
:precision binary32
(*
(cos (* PI (* uy 2.0)))
(sqrt
(fma
(fma (- 1.0 maxCos) (fma ux maxCos (- ux)) (* maxCos -2.0))
ux
(* ux 2.0)))))
float code(float ux, float uy, float maxCos) {
return cosf((((float) M_PI) * (uy * 2.0f))) * sqrtf(fmaf(fmaf((1.0f - maxCos), fmaf(ux, maxCos, -ux), (maxCos * -2.0f)), ux, (ux * 2.0f)));
}
function code(ux, uy, maxCos) return Float32(cos(Float32(Float32(pi) * Float32(uy * Float32(2.0)))) * sqrt(fma(fma(Float32(Float32(1.0) - maxCos), fma(ux, maxCos, Float32(-ux)), Float32(maxCos * Float32(-2.0))), ux, Float32(ux * Float32(2.0))))) end
\begin{array}{l}
\\
\cos \left(\pi \cdot \left(uy \cdot 2\right)\right) \cdot \sqrt{\mathsf{fma}\left(\mathsf{fma}\left(1 - maxCos, \mathsf{fma}\left(ux, maxCos, -ux\right), maxCos \cdot -2\right), ux, ux \cdot 2\right)}
\end{array}
Initial program 57.5%
Taylor expanded in ux around 0
lower-*.f32N/A
cancel-sign-sub-invN/A
metadata-evalN/A
+-commutativeN/A
*-commutativeN/A
lower-fma.f32N/A
+-commutativeN/A
mul-1-negN/A
*-commutativeN/A
unpow2N/A
associate-*l*N/A
*-commutativeN/A
distribute-lft-neg-inN/A
neg-sub0N/A
associate-+l-N/A
neg-sub0N/A
mul-1-negN/A
+-commutativeN/A
Applied rewrites98.9%
Applied rewrites99.0%
Final simplification99.0%
herbie shell --seed 2024228
(FPCore (ux uy maxCos)
:name "UniformSampleCone, x"
:precision binary32
:pre (and (and (and (<= 2.328306437e-10 ux) (<= ux 1.0)) (and (<= 2.328306437e-10 uy) (<= uy 1.0))) (and (<= 0.0 maxCos) (<= maxCos 1.0)))
(* (cos (* (* uy 2.0) PI)) (sqrt (- 1.0 (* (+ (- 1.0 ux) (* ux maxCos)) (+ (- 1.0 ux) (* ux maxCos)))))))