
(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)))))))
float code(float u0, float u1, float alphax, float alphay) {
float t_0 = atanf(((alphay / alphax) * tanf((((2.0f * ((float) M_PI)) * u1) + (0.5f * ((float) M_PI))))));
float t_1 = sinf(t_0);
float t_2 = cosf(t_0);
return 1.0f / sqrtf((1.0f + (((1.0f / (((t_2 * t_2) / (alphax * alphax)) + ((t_1 * t_1) / (alphay * alphay)))) * u0) / (1.0f - u0))));
}
function code(u0, u1, alphax, alphay) t_0 = atan(Float32(Float32(alphay / alphax) * tan(Float32(Float32(Float32(Float32(2.0) * Float32(pi)) * u1) + Float32(Float32(0.5) * Float32(pi)))))) t_1 = sin(t_0) t_2 = cos(t_0) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(Float32(Float32(Float32(1.0) / Float32(Float32(Float32(t_2 * t_2) / Float32(alphax * alphax)) + Float32(Float32(t_1 * t_1) / Float32(alphay * alphay)))) * u0) / Float32(Float32(1.0) - u0))))) end
function tmp = code(u0, u1, alphax, alphay) t_0 = atan(((alphay / alphax) * tan((((single(2.0) * single(pi)) * u1) + (single(0.5) * single(pi)))))); t_1 = sin(t_0); t_2 = cos(t_0); tmp = single(1.0) / sqrt((single(1.0) + (((single(1.0) / (((t_2 * t_2) / (alphax * alphax)) + ((t_1 * t_1) / (alphay * alphay)))) * u0) / (single(1.0) - u0)))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan^{-1} \left(\frac{alphay}{alphax} \cdot \tan \left(\left(2 \cdot \pi\right) \cdot u1 + 0.5 \cdot \pi\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 4 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)))))))
float code(float u0, float u1, float alphax, float alphay) {
float t_0 = atanf(((alphay / alphax) * tanf((((2.0f * ((float) M_PI)) * u1) + (0.5f * ((float) M_PI))))));
float t_1 = sinf(t_0);
float t_2 = cosf(t_0);
return 1.0f / sqrtf((1.0f + (((1.0f / (((t_2 * t_2) / (alphax * alphax)) + ((t_1 * t_1) / (alphay * alphay)))) * u0) / (1.0f - u0))));
}
function code(u0, u1, alphax, alphay) t_0 = atan(Float32(Float32(alphay / alphax) * tan(Float32(Float32(Float32(Float32(2.0) * Float32(pi)) * u1) + Float32(Float32(0.5) * Float32(pi)))))) t_1 = sin(t_0) t_2 = cos(t_0) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(Float32(Float32(Float32(1.0) / Float32(Float32(Float32(t_2 * t_2) / Float32(alphax * alphax)) + Float32(Float32(t_1 * t_1) / Float32(alphay * alphay)))) * u0) / Float32(Float32(1.0) - u0))))) end
function tmp = code(u0, u1, alphax, alphay) t_0 = atan(((alphay / alphax) * tan((((single(2.0) * single(pi)) * u1) + (single(0.5) * single(pi)))))); t_1 = sin(t_0); t_2 = cos(t_0); tmp = single(1.0) / sqrt((single(1.0) + (((single(1.0) / (((t_2 * t_2) / (alphax * alphax)) + ((t_1 * t_1) / (alphay * alphay)))) * u0) / (single(1.0) - u0)))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan^{-1} \left(\frac{alphay}{alphax} \cdot \tan \left(\left(2 \cdot \pi\right) \cdot u1 + 0.5 \cdot \pi\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 (tan (* PI (+ 0.5 (* 2.0 u1))))))
(/
1.0
(sqrt
(+
1.0
(/
(*
(/
1.0
(+
(/
(/ 1.0 (+ 1.0 (pow (* alphay (/ t_0 alphax)) 2.0)))
(* alphax alphax))
(/
(*
(sin
(atan
(* (/ alphay alphax) (tan (+ (* u1 (* PI 2.0)) (* PI 0.5))))))
(sin (atan (* t_0 (/ alphay alphax)))))
(* alphay alphay))))
u0)
(- 1.0 u0)))))))
float code(float u0, float u1, float alphax, float alphay) {
float t_0 = tanf((((float) M_PI) * (0.5f + (2.0f * u1))));
return 1.0f / sqrtf((1.0f + (((1.0f / (((1.0f / (1.0f + powf((alphay * (t_0 / alphax)), 2.0f))) / (alphax * alphax)) + ((sinf(atanf(((alphay / alphax) * tanf(((u1 * (((float) M_PI) * 2.0f)) + (((float) M_PI) * 0.5f)))))) * sinf(atanf((t_0 * (alphay / alphax))))) / (alphay * alphay)))) * u0) / (1.0f - u0))));
}
function code(u0, u1, alphax, alphay) t_0 = tan(Float32(Float32(pi) * Float32(Float32(0.5) + Float32(Float32(2.0) * u1)))) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(Float32(Float32(Float32(1.0) / Float32(Float32(Float32(Float32(1.0) / Float32(Float32(1.0) + (Float32(alphay * Float32(t_0 / alphax)) ^ Float32(2.0)))) / Float32(alphax * alphax)) + Float32(Float32(sin(atan(Float32(Float32(alphay / alphax) * tan(Float32(Float32(u1 * Float32(Float32(pi) * Float32(2.0))) + Float32(Float32(pi) * Float32(0.5))))))) * sin(atan(Float32(t_0 * Float32(alphay / alphax))))) / Float32(alphay * alphay)))) * u0) / Float32(Float32(1.0) - u0))))) end
function tmp = code(u0, u1, alphax, alphay) t_0 = tan((single(pi) * (single(0.5) + (single(2.0) * u1)))); tmp = single(1.0) / sqrt((single(1.0) + (((single(1.0) / (((single(1.0) / (single(1.0) + ((alphay * (t_0 / alphax)) ^ single(2.0)))) / (alphax * alphax)) + ((sin(atan(((alphay / alphax) * tan(((u1 * (single(pi) * single(2.0))) + (single(pi) * single(0.5))))))) * sin(atan((t_0 * (alphay / alphax))))) / (alphay * alphay)))) * u0) / (single(1.0) - u0)))); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \tan \left(\pi \cdot \left(0.5 + 2 \cdot u1\right)\right)\\
\frac{1}{\sqrt{1 + \frac{\frac{1}{\frac{\frac{1}{1 + {\left(alphay \cdot \frac{t\_0}{alphax}\right)}^{2}}}{alphax \cdot alphax} + \frac{\sin \tan^{-1} \left(\frac{alphay}{alphax} \cdot \tan \left(u1 \cdot \left(\pi \cdot 2\right) + \pi \cdot 0.5\right)\right) \cdot \sin \tan^{-1} \left(t\_0 \cdot \frac{alphay}{alphax}\right)}{alphay \cdot alphay}} \cdot u0}{1 - u0}}}
\end{array}
\end{array}
Initial program 99.5%
cos-atan99.5%
cos-atan99.5%
frac-times99.5%
metadata-eval99.5%
Applied egg-rr99.5%
associate-*l/99.5%
associate-*r/99.5%
*-commutative99.5%
*-commutative99.5%
fma-define99.5%
+-commutative99.5%
associate-*r*99.5%
*-commutative99.5%
distribute-rgt-out99.5%
*-commutative99.5%
Simplified99.5%
*-un-lft-identity99.5%
*-commutative99.5%
associate-*l*99.5%
fma-define99.5%
*-commutative99.5%
Applied egg-rr99.5%
*-rgt-identity99.5%
*-commutative99.5%
*-commutative99.5%
fma-define99.5%
+-commutative99.5%
associate-*r*99.5%
*-commutative99.5%
distribute-rgt-out99.5%
*-commutative99.5%
Simplified99.5%
Final simplification99.5%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(/
1.0
(sqrt
(+
1.0
(/
(* (* alphay alphay) u0)
(*
(- 1.0 u0)
(pow
(sin
(atan (/ (* alphay (tan (+ (* PI 0.5) (* 2.0 (* PI u1))))) alphax)))
2.0)))))))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f / sqrtf((1.0f + (((alphay * alphay) * u0) / ((1.0f - u0) * powf(sinf(atanf(((alphay * tanf(((((float) M_PI) * 0.5f) + (2.0f * (((float) M_PI) * u1))))) / alphax))), 2.0f)))));
}
function code(u0, u1, alphax, alphay) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(Float32(Float32(alphay * alphay) * u0) / Float32(Float32(Float32(1.0) - u0) * (sin(atan(Float32(Float32(alphay * tan(Float32(Float32(Float32(pi) * Float32(0.5)) + Float32(Float32(2.0) * Float32(Float32(pi) * u1))))) / alphax))) ^ Float32(2.0))))))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) / sqrt((single(1.0) + (((alphay * alphay) * u0) / ((single(1.0) - u0) * (sin(atan(((alphay * tan(((single(pi) * single(0.5)) + (single(2.0) * (single(pi) * u1))))) / alphax))) ^ single(2.0)))))); end
\begin{array}{l}
\\
\frac{1}{\sqrt{1 + \frac{\left(alphay \cdot alphay\right) \cdot u0}{\left(1 - u0\right) \cdot {\sin \tan^{-1} \left(\frac{alphay \cdot \tan \left(\pi \cdot 0.5 + 2 \cdot \left(\pi \cdot u1\right)\right)}{alphax}\right)}^{2}}}}
\end{array}
Initial program 99.5%
cos-atan99.5%
cos-atan99.5%
frac-times99.5%
metadata-eval99.5%
Applied egg-rr99.5%
associate-*l/99.5%
associate-*r/99.5%
*-commutative99.5%
*-commutative99.5%
fma-define99.5%
+-commutative99.5%
associate-*r*99.5%
*-commutative99.5%
distribute-rgt-out99.5%
*-commutative99.5%
Simplified99.5%
Taylor expanded in u1 around 0 98.0%
Taylor expanded in alphax around inf 98.0%
unpow298.0%
Applied egg-rr98.0%
Final simplification98.0%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(/
1.0
(+
1.0
(*
0.5
(/
(* (* alphay alphay) u0)
(*
(- 1.0 u0)
(pow
(sin
(atan (/ (* alphay (tan (+ (* PI 0.5) (* 2.0 (* PI u1))))) alphax)))
2.0)))))))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f / (1.0f + (0.5f * (((alphay * alphay) * u0) / ((1.0f - u0) * powf(sinf(atanf(((alphay * tanf(((((float) M_PI) * 0.5f) + (2.0f * (((float) M_PI) * u1))))) / alphax))), 2.0f)))));
}
function code(u0, u1, alphax, alphay) return Float32(Float32(1.0) / Float32(Float32(1.0) + Float32(Float32(0.5) * Float32(Float32(Float32(alphay * alphay) * u0) / Float32(Float32(Float32(1.0) - u0) * (sin(atan(Float32(Float32(alphay * tan(Float32(Float32(Float32(pi) * Float32(0.5)) + Float32(Float32(2.0) * Float32(Float32(pi) * u1))))) / alphax))) ^ Float32(2.0))))))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) / (single(1.0) + (single(0.5) * (((alphay * alphay) * u0) / ((single(1.0) - u0) * (sin(atan(((alphay * tan(((single(pi) * single(0.5)) + (single(2.0) * (single(pi) * u1))))) / alphax))) ^ single(2.0)))))); end
\begin{array}{l}
\\
\frac{1}{1 + 0.5 \cdot \frac{\left(alphay \cdot alphay\right) \cdot u0}{\left(1 - u0\right) \cdot {\sin \tan^{-1} \left(\frac{alphay \cdot \tan \left(\pi \cdot 0.5 + 2 \cdot \left(\pi \cdot u1\right)\right)}{alphax}\right)}^{2}}}
\end{array}
Initial program 99.5%
cos-atan99.5%
cos-atan99.5%
frac-times99.5%
metadata-eval99.5%
Applied egg-rr99.5%
associate-*l/99.5%
associate-*r/99.5%
*-commutative99.5%
*-commutative99.5%
fma-define99.5%
+-commutative99.5%
associate-*r*99.5%
*-commutative99.5%
distribute-rgt-out99.5%
*-commutative99.5%
Simplified99.5%
Taylor expanded in u1 around 0 98.0%
Taylor expanded in alphay around 0 97.3%
unpow298.0%
Applied egg-rr97.3%
Final simplification97.3%
(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 99.5%
cos-atan99.5%
cos-atan99.5%
frac-times99.5%
metadata-eval99.5%
Applied egg-rr99.5%
associate-*l/99.5%
associate-*r/99.5%
*-commutative99.5%
*-commutative99.5%
fma-define99.5%
+-commutative99.5%
associate-*r*99.5%
*-commutative99.5%
distribute-rgt-out99.5%
*-commutative99.5%
Simplified99.5%
Taylor expanded in u1 around 0 98.0%
Taylor expanded in alphay around 0 92.4%
metadata-eval92.4%
Applied egg-rr92.4%
herbie shell --seed 2024132
(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))))))