
(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 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)))))))
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
(/
1.0
(sqrt
(+
1.0
(/
(*
(/
1.0
(+
(/
(/
1.0
(+
1.0
(pow (* (/ alphay alphax) (tan (* PI (+ 0.5 (* 2.0 u1))))) 2.0)))
(* alphax alphax))
(/
(pow
(sin (atan (/ (* alphay (tan (* PI (- 0.5 (* u1 -2.0))))) alphax)))
2.0)
(* alphay alphay))))
u0)
(- 1.0 u0))))))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f / sqrtf((1.0f + (((1.0f / (((1.0f / (1.0f + powf(((alphay / alphax) * tanf((((float) M_PI) * (0.5f + (2.0f * u1))))), 2.0f))) / (alphax * alphax)) + (powf(sinf(atanf(((alphay * tanf((((float) M_PI) * (0.5f - (u1 * -2.0f))))) / alphax))), 2.0f) / (alphay * alphay)))) * u0) / (1.0f - u0))));
}
function code(u0, u1, alphax, alphay) 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(Float32(alphay / alphax) * tan(Float32(Float32(pi) * Float32(Float32(0.5) + Float32(Float32(2.0) * u1))))) ^ Float32(2.0)))) / Float32(alphax * alphax)) + Float32((sin(atan(Float32(Float32(alphay * tan(Float32(Float32(pi) * Float32(Float32(0.5) - Float32(u1 * Float32(-2.0)))))) / alphax))) ^ Float32(2.0)) / Float32(alphay * alphay)))) * u0) / Float32(Float32(1.0) - u0))))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) / sqrt((single(1.0) + (((single(1.0) / (((single(1.0) / (single(1.0) + (((alphay / alphax) * tan((single(pi) * (single(0.5) + (single(2.0) * u1))))) ^ single(2.0)))) / (alphax * alphax)) + ((sin(atan(((alphay * tan((single(pi) * (single(0.5) - (u1 * single(-2.0)))))) / alphax))) ^ single(2.0)) / (alphay * alphay)))) * u0) / (single(1.0) - u0)))); end
\begin{array}{l}
\\
\frac{1}{\sqrt{1 + \frac{\frac{1}{\frac{\frac{1}{1 + {\left(\frac{alphay}{alphax} \cdot \tan \left(\pi \cdot \left(0.5 + 2 \cdot u1\right)\right)\right)}^{2}}}{alphax \cdot alphax} + \frac{{\sin \tan^{-1} \left(\frac{alphay \cdot \tan \left(\pi \cdot \left(0.5 - u1 \cdot -2\right)\right)}{alphax}\right)}^{2}}{alphay \cdot alphay}} \cdot u0}{1 - u0}}}
\end{array}
Initial program 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
*-un-lft-identity99.3%
*-commutative99.3%
associate-*l*99.3%
fma-define99.3%
*-commutative99.3%
Applied egg-rr99.3%
*-rgt-identity99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
*-un-lft-identity99.3%
*-commutative99.3%
associate-*l*99.3%
fma-define99.3%
*-commutative99.3%
Applied egg-rr99.3%
*-rgt-identity99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in u1 around -inf 99.3%
Final simplification99.3%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(sqrt
(/
1.0
(+
1.0
(/
(* u0 (pow alphay 2.0))
(*
(- 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 sqrtf((1.0f / (1.0f + ((u0 * powf(alphay, 2.0f)) / ((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 sqrt(Float32(Float32(1.0) / Float32(Float32(1.0) + Float32(Float32(u0 * (alphay ^ Float32(2.0))) / 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 = sqrt((single(1.0) / (single(1.0) + ((u0 * (alphay ^ single(2.0))) / ((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}
\\
\sqrt{\frac{1}{1 + \frac{u0 \cdot {alphay}^{2}}{\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 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in alphay around 0 92.2%
Applied egg-rr92.2%
Taylor expanded in alphax around inf 97.5%
Final simplification97.5%
(FPCore (u0 u1 alphax alphay)
:precision binary32
(/
1.0
(sqrt
(+
1.0
(/
(*
u0
(/
(* alphay alphay)
(pow
(sin (atan (* alphay (/ (tan (* PI (+ 0.5 (* 2.0 u1)))) alphax))))
2.0)))
(- 1.0 u0))))))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f / sqrtf((1.0f + ((u0 * ((alphay * alphay) / powf(sinf(atanf((alphay * (tanf((((float) M_PI) * (0.5f + (2.0f * u1)))) / alphax)))), 2.0f))) / (1.0f - u0))));
}
function code(u0, u1, alphax, alphay) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(Float32(u0 * Float32(Float32(alphay * alphay) / (sin(atan(Float32(alphay * Float32(tan(Float32(Float32(pi) * Float32(Float32(0.5) + Float32(Float32(2.0) * u1)))) / alphax)))) ^ Float32(2.0)))) / Float32(Float32(1.0) - u0))))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) / sqrt((single(1.0) + ((u0 * ((alphay * alphay) / (sin(atan((alphay * (tan((single(pi) * (single(0.5) + (single(2.0) * u1)))) / alphax)))) ^ single(2.0)))) / (single(1.0) - u0)))); end
\begin{array}{l}
\\
\frac{1}{\sqrt{1 + \frac{u0 \cdot \frac{alphay \cdot alphay}{{\sin \tan^{-1} \left(alphay \cdot \frac{\tan \left(\pi \cdot \left(0.5 + 2 \cdot u1\right)\right)}{alphax}\right)}^{2}}}{1 - u0}}}
\end{array}
Initial program 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in alphay around 0 92.2%
Taylor expanded in alphax around inf 97.0%
associate-/l*97.0%
associate-*r*97.0%
*-commutative97.0%
distribute-rgt-out97.0%
*-commutative97.0%
Simplified97.0%
pow297.0%
Applied egg-rr97.0%
Final simplification97.0%
(FPCore (u0 u1 alphax alphay) :precision binary32 (/ 1.0 (sqrt (+ 1.0 (* u0 (pow alphax 2.0))))))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f / sqrtf((1.0f + (u0 * powf(alphax, 2.0f))));
}
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 / sqrt((1.0e0 + (u0 * (alphax ** 2.0e0))))
end function
function code(u0, u1, alphax, alphay) return Float32(Float32(1.0) / sqrt(Float32(Float32(1.0) + Float32(u0 * (alphax ^ Float32(2.0)))))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) / sqrt((single(1.0) + (u0 * (alphax ^ single(2.0))))); end
\begin{array}{l}
\\
\frac{1}{\sqrt{1 + u0 \cdot {alphax}^{2}}}
\end{array}
Initial program 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in alphay around 0 92.2%
Taylor expanded in alphax around 0 86.8%
Taylor expanded in u0 around 0 86.9%
Final simplification86.9%
(FPCore (u0 u1 alphax alphay) :precision binary32 (+ 1.0 (* (* u0 (pow alphax 2.0)) -0.5)))
float code(float u0, float u1, float alphax, float alphay) {
return 1.0f + ((u0 * powf(alphax, 2.0f)) * -0.5f);
}
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 + ((u0 * (alphax ** 2.0e0)) * (-0.5e0))
end function
function code(u0, u1, alphax, alphay) return Float32(Float32(1.0) + Float32(Float32(u0 * (alphax ^ Float32(2.0))) * Float32(-0.5))) end
function tmp = code(u0, u1, alphax, alphay) tmp = single(1.0) + ((u0 * (alphax ^ single(2.0))) * single(-0.5)); end
\begin{array}{l}
\\
1 + \left(u0 \cdot {alphax}^{2}\right) \cdot -0.5
\end{array}
Initial program 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in alphay around 0 92.2%
Taylor expanded in alphax around 0 86.8%
Taylor expanded in u0 around 0 86.9%
*-commutative86.9%
Simplified86.9%
Final simplification86.9%
(FPCore (u0 u1 alphax alphay) :precision binary32 (/ (sqrt (+ (/ 1.0 u0) -1.0)) alphax))
float code(float u0, float u1, float alphax, float alphay) {
return sqrtf(((1.0f / u0) + -1.0f)) / alphax;
}
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 = sqrt(((1.0e0 / u0) + (-1.0e0))) / alphax
end function
function code(u0, u1, alphax, alphay) return Float32(sqrt(Float32(Float32(Float32(1.0) / u0) + Float32(-1.0))) / alphax) end
function tmp = code(u0, u1, alphax, alphay) tmp = sqrt(((single(1.0) / u0) + single(-1.0))) / alphax; end
\begin{array}{l}
\\
\frac{\sqrt{\frac{1}{u0} + -1}}{alphax}
\end{array}
Initial program 99.2%
cos-atan99.3%
cos-atan99.3%
frac-times99.3%
metadata-eval99.3%
Applied egg-rr99.3%
fma-undefine99.3%
*-commutative99.3%
*-commutative99.3%
+-commutative99.3%
associate-*r*99.3%
distribute-rgt-out99.3%
Simplified99.3%
Taylor expanded in alphay around 0 92.2%
Taylor expanded in alphax around 0 86.8%
Taylor expanded in alphax around inf 16.8%
associate-*l/16.8%
*-lft-identity16.8%
div-sub16.8%
sub-neg16.8%
*-inverses16.8%
metadata-eval16.8%
Simplified16.8%
herbie shell --seed 2024163
(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))))))