
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0
(atan
(* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI)))))))
(t_1 (sin t_0))
(t_2 (cos t_0)))
(/
1.0
(sqrt
(+
1.0
(/
(*
(/
1.0
(+
(/ (* t_2 t_2) (* alphax alphax))
(/ (* t_1 t_1) (* alphay alphay))))
u0)
(- 1.0 u0)))))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan^{-1} \left(\frac{alphay}{alphax} \cdot \tan \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u1 + 0.5 \cdot \mathsf{PI}\left(\right)\right)\right)\\
t_1 := \sin t\_0\\
t_2 := \cos t\_0\\
\frac{1}{\sqrt{1 + \frac{\frac{1}{\frac{t\_2 \cdot t\_2}{alphax \cdot alphax} + \frac{t\_1 \cdot t\_1}{alphay \cdot alphay}} \cdot u0}{1 - u0}}}
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0
(atan
(* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI)))))))
(t_1 (sin t_0))
(t_2 (cos t_0)))
(/
1.0
(sqrt
(+
1.0
(/
(*
(/
1.0
(+
(/ (* t_2 t_2) (* alphax alphax))
(/ (* t_1 t_1) (* alphay alphay))))
u0)
(- 1.0 u0)))))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan^{-1} \left(\frac{alphay}{alphax} \cdot \tan \left(\left(2 \cdot \mathsf{PI}\left(\right)\right) \cdot u1 + 0.5 \cdot \mathsf{PI}\left(\right)\right)\right)\\
t_1 := \sin t\_0\\
t_2 := \cos t\_0\\
\frac{1}{\sqrt{1 + \frac{\frac{1}{\frac{t\_2 \cdot t\_2}{alphax \cdot alphax} + \frac{t\_1 \cdot t\_1}{alphay \cdot alphay}} \cdot u0}{1 - u0}}}
\end{array}
\end{array}
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0 (* (- 0.5 (* -2.0 u1)) (PI)))
(t_1 (atan (* (/ (sin t_0) (cos t_0)) (/ alphay alphax)))))
(sqrt
(/
1.0
(+
(/
(/
u0
(+
(/ (pow (sin t_1) 2.0) (* alphay alphay))
(/ (pow (cos t_1) 2.0) (* alphax alphax))))
(- 1.0 u0))
1.0)))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(0.5 - -2 \cdot u1\right) \cdot \mathsf{PI}\left(\right)\\
t_1 := \tan^{-1} \left(\frac{\sin t\_0}{\cos t\_0} \cdot \frac{alphay}{alphax}\right)\\
\sqrt{\frac{1}{\frac{\frac{u0}{\frac{{\sin t\_1}^{2}}{alphay \cdot alphay} + \frac{{\cos t\_1}^{2}}{alphax \cdot alphax}}}{1 - u0} + 1}}
\end{array}
\end{array}
Initial program 99.4%
Taylor expanded in u1 around 0
Applied rewrites98.4%
Applied rewrites98.6%
Taylor expanded in u1 around 0
Applied rewrites99.9%
Final simplification99.9%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0 (* (tan (+ (* (* (PI) 2.0) u1) (* (PI) 0.5))) (/ alphay alphax)))
(t_1 (sin (atan t_0))))
(/
1.0
(sqrt
(+
(/
(*
(/
1.0
(+
(/ (* t_1 t_1) (* alphay alphay))
(/ (/ 1.0 (+ (pow t_0 2.0) 1.0)) (* alphax alphax))))
u0)
(- 1.0 u0))
1.0)))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan \left(\left(\mathsf{PI}\left(\right) \cdot 2\right) \cdot u1 + \mathsf{PI}\left(\right) \cdot 0.5\right) \cdot \frac{alphay}{alphax}\\
t_1 := \sin \tan^{-1} t\_0\\
\frac{1}{\sqrt{\frac{\frac{1}{\frac{t\_1 \cdot t\_1}{alphay \cdot alphay} + \frac{\frac{1}{{t\_0}^{2} + 1}}{alphax \cdot alphax}} \cdot u0}{1 - u0} + 1}}
\end{array}
\end{array}
Initial program 99.4%
lift-*.f32N/A
lift-cos.f32N/A
lift-atan.f32N/A
cos-atanN/A
lift-cos.f32N/A
lift-atan.f32N/A
cos-atanN/A
Applied rewrites94.0%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3299.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3299.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3299.4
Applied rewrites99.4%
Final simplification99.4%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0 (* (- 0.5 (* -2.0 u1)) (PI))))
(sqrt
(/
1.0
(+
(*
(/ u0 (- 1.0 u0))
(/
(* alphay alphay)
(pow (sin (atan (* (/ (sin t_0) (cos t_0)) (/ alphay alphax)))) 2.0)))
1.0)))))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(0.5 - -2 \cdot u1\right) \cdot \mathsf{PI}\left(\right)\\
\sqrt{\frac{1}{\frac{u0}{1 - u0} \cdot \frac{alphay \cdot alphay}{{\sin \tan^{-1} \left(\frac{\sin t\_0}{\cos t\_0} \cdot \frac{alphay}{alphax}\right)}^{2}} + 1}}
\end{array}
\end{array}
Initial program 99.4%
Taylor expanded in u1 around 0
Applied rewrites98.4%
Applied rewrites98.6%
Taylor expanded in alphax around inf
Applied rewrites98.5%
Final simplification98.5%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0 (* (- 0.5 (* -2.0 u1)) (PI))))
(+
(/
(* (* (* alphay alphay) u0) -0.5)
(*
(- 1.0 u0)
(pow (sin (atan (* (/ (sin t_0) (cos t_0)) (/ alphay alphax)))) 2.0)))
1.0)))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(0.5 - -2 \cdot u1\right) \cdot \mathsf{PI}\left(\right)\\
\frac{\left(\left(alphay \cdot alphay\right) \cdot u0\right) \cdot -0.5}{\left(1 - u0\right) \cdot {\sin \tan^{-1} \left(\frac{\sin t\_0}{\cos t\_0} \cdot \frac{alphay}{alphax}\right)}^{2}} + 1
\end{array}
\end{array}
Initial program 99.4%
Taylor expanded in u1 around 0
Applied rewrites98.4%
Taylor expanded in alphay around 0
Applied rewrites35.1%
Applied rewrites58.3%
Taylor expanded in u1 around 0
Applied rewrites96.8%
Final simplification96.8%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(let* ((t_0 (* (- 0.5 (* -2.0 u1)) (PI))))
(+
(/
(* (* (* alphay alphay) u0) -0.5)
(pow (sin (atan (* (/ (sin t_0) (cos t_0)) (/ alphay alphax)))) 2.0))
1.0)))\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(0.5 - -2 \cdot u1\right) \cdot \mathsf{PI}\left(\right)\\
\frac{\left(\left(alphay \cdot alphay\right) \cdot u0\right) \cdot -0.5}{{\sin \tan^{-1} \left(\frac{\sin t\_0}{\cos t\_0} \cdot \frac{alphay}{alphax}\right)}^{2}} + 1
\end{array}
\end{array}
Initial program 99.4%
Taylor expanded in u1 around 0
Applied rewrites98.4%
Taylor expanded in alphay around 0
Applied rewrites34.5%
Applied rewrites57.7%
Taylor expanded in u0 around 0
Applied rewrites95.5%
Final simplification95.5%
(FPCore (u0 u1 alphax alphay) :precision binary32 1.0)
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f;
}
real(4) function code(u0, u1, alphax, alphay)
real(4), intent (in) :: u0
real(4), intent (in) :: u1
real(4), intent (in) :: alphax
real(4), intent (in) :: alphay
code = 1.0e0
end function
function code(u0, u1, alphax, alphay) return Float32(1.0) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0); end
\begin{array}{l}
\\
1
\end{array}
Initial program 99.4%
Taylor expanded in u0 around 0
Applied rewrites92.7%
herbie shell --seed 2024285
(FPCore (u0 u1 alphax alphay)
:name "Trowbridge-Reitz Sample, sample surface normal, cosTheta"
:precision binary32
:pre (and (and (and (and (<= 2.328306437e-10 u0) (<= u0 1.0)) (and (<= 2.328306437e-10 u1) (<= u1 0.5))) (and (<= 0.0001 alphax) (<= alphax 1.0))) (and (<= 0.0001 alphay) (<= alphay 1.0)))
(/ 1.0 (sqrt (+ 1.0 (/ (* (/ 1.0 (+ (/ (* (cos (atan (* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI))))))) (cos (atan (* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI)))))))) (* alphax alphax)) (/ (* (sin (atan (* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI))))))) (sin (atan (* (/ alphay alphax) (tan (+ (* (* 2.0 (PI)) u1) (* 0.5 (PI)))))))) (* alphay alphay)))) u0) (- 1.0 u0))))))