
(FPCore (s u) :precision binary32 (* (* 3.0 s) (log (/ 1.0 (- 1.0 (/ (- u 0.25) 0.75))))))
float code(float s, float u) {
return (3.0f * s) * logf((1.0f / (1.0f - ((u - 0.25f) / 0.75f))));
}
real(4) function code(s, u)
real(4), intent (in) :: s
real(4), intent (in) :: u
code = (3.0e0 * s) * log((1.0e0 / (1.0e0 - ((u - 0.25e0) / 0.75e0))))
end function
function code(s, u) return Float32(Float32(Float32(3.0) * s) * log(Float32(Float32(1.0) / Float32(Float32(1.0) - Float32(Float32(u - Float32(0.25)) / Float32(0.75)))))) end
function tmp = code(s, u) tmp = (single(3.0) * s) * log((single(1.0) / (single(1.0) - ((u - single(0.25)) / single(0.75))))); end
\begin{array}{l}
\\
\left(3 \cdot s\right) \cdot \log \left(\frac{1}{1 - \frac{u - 0.25}{0.75}}\right)
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 8 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (s u) :precision binary32 (* (* 3.0 s) (log (/ 1.0 (- 1.0 (/ (- u 0.25) 0.75))))))
float code(float s, float u) {
return (3.0f * s) * logf((1.0f / (1.0f - ((u - 0.25f) / 0.75f))));
}
real(4) function code(s, u)
real(4), intent (in) :: s
real(4), intent (in) :: u
code = (3.0e0 * s) * log((1.0e0 / (1.0e0 - ((u - 0.25e0) / 0.75e0))))
end function
function code(s, u) return Float32(Float32(Float32(3.0) * s) * log(Float32(Float32(1.0) / Float32(Float32(1.0) - Float32(Float32(u - Float32(0.25)) / Float32(0.75)))))) end
function tmp = code(s, u) tmp = (single(3.0) * s) * log((single(1.0) / (single(1.0) - ((u - single(0.25)) / single(0.75))))); end
\begin{array}{l}
\\
\left(3 \cdot s\right) \cdot \log \left(\frac{1}{1 - \frac{u - 0.25}{0.75}}\right)
\end{array}
(FPCore (s u) :precision binary32 (fma (* 3.0 s) (log1p (fma u 1.3333333333333333 -0.3333333333333333)) (* (log1p (* (- 0.25 u) (fma u 1.7777777777777777 -0.4444444444444444))) (* s -3.0))))
float code(float s, float u) {
return fmaf((3.0f * s), log1pf(fmaf(u, 1.3333333333333333f, -0.3333333333333333f)), (log1pf(((0.25f - u) * fmaf(u, 1.7777777777777777f, -0.4444444444444444f))) * (s * -3.0f)));
}
function code(s, u) return fma(Float32(Float32(3.0) * s), log1p(fma(u, Float32(1.3333333333333333), Float32(-0.3333333333333333))), Float32(log1p(Float32(Float32(Float32(0.25) - u) * fma(u, Float32(1.7777777777777777), Float32(-0.4444444444444444)))) * Float32(s * Float32(-3.0)))) end
\begin{array}{l}
\\
\mathsf{fma}\left(3 \cdot s, \mathsf{log1p}\left(\mathsf{fma}\left(u, 1.3333333333333333, -0.3333333333333333\right)\right), \mathsf{log1p}\left(\left(0.25 - u\right) \cdot \mathsf{fma}\left(u, 1.7777777777777777, -0.4444444444444444\right)\right) \cdot \left(s \cdot -3\right)\right)
\end{array}
Initial program 96.2%
lift-log.f32N/A
lift-/.f32N/A
lift--.f32N/A
flip--N/A
clear-numN/A
log-divN/A
lower--.f32N/A
lower-log1p.f32N/A
lift-/.f32N/A
lift--.f32N/A
div-subN/A
sub-negN/A
div-invN/A
lower-fma.f32N/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
Applied rewrites98.2%
lift-*.f32N/A
lift--.f32N/A
sub-negN/A
distribute-lft-inN/A
lower-fma.f32N/A
lower-*.f32N/A
lower-neg.f3298.5
lift-*.f32N/A
*-commutativeN/A
lift--.f32N/A
sub-negN/A
metadata-evalN/A
distribute-neg-inN/A
+-commutativeN/A
lift-+.f32N/A
lower-*.f32N/A
Applied rewrites98.5%
Taylor expanded in s around 0
associate-*r*N/A
*-commutativeN/A
lower-*.f32N/A
lower-log1p.f32N/A
lower-*.f32N/A
lower--.f32N/A
sub-negN/A
*-commutativeN/A
metadata-evalN/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f3298.5
Applied rewrites98.5%
(FPCore (s u) :precision binary32 (* 3.0 (fma s (- (log1p (* (- 0.25 u) (fma u 1.7777777777777777 -0.4444444444444444)))) (* s (log1p (fma u 1.3333333333333333 -0.3333333333333333))))))
float code(float s, float u) {
return 3.0f * fmaf(s, -log1pf(((0.25f - u) * fmaf(u, 1.7777777777777777f, -0.4444444444444444f))), (s * log1pf(fmaf(u, 1.3333333333333333f, -0.3333333333333333f))));
}
function code(s, u) return Float32(Float32(3.0) * fma(s, Float32(-log1p(Float32(Float32(Float32(0.25) - u) * fma(u, Float32(1.7777777777777777), Float32(-0.4444444444444444))))), Float32(s * log1p(fma(u, Float32(1.3333333333333333), Float32(-0.3333333333333333)))))) end
\begin{array}{l}
\\
3 \cdot \mathsf{fma}\left(s, -\mathsf{log1p}\left(\left(0.25 - u\right) \cdot \mathsf{fma}\left(u, 1.7777777777777777, -0.4444444444444444\right)\right), s \cdot \mathsf{log1p}\left(\mathsf{fma}\left(u, 1.3333333333333333, -0.3333333333333333\right)\right)\right)
\end{array}
Initial program 95.8%
lift-log.f32N/A
lift-/.f32N/A
lift--.f32N/A
flip--N/A
clear-numN/A
log-divN/A
lower--.f32N/A
lower-log1p.f32N/A
lift-/.f32N/A
lift--.f32N/A
div-subN/A
sub-negN/A
div-invN/A
lower-fma.f32N/A
metadata-evalN/A
metadata-evalN/A
metadata-evalN/A
Applied rewrites98.2%
lift-*.f32N/A
lift--.f32N/A
sub-negN/A
distribute-lft-inN/A
lower-fma.f32N/A
lower-*.f32N/A
lower-neg.f3298.3
lift-*.f32N/A
*-commutativeN/A
lift--.f32N/A
sub-negN/A
metadata-evalN/A
distribute-neg-inN/A
+-commutativeN/A
lift-+.f32N/A
lower-*.f32N/A
Applied rewrites98.3%
Applied rewrites98.3%
herbie shell --seed 2024228
(FPCore (s u)
:name "Disney BSSRDF, sample scattering profile, upper"
:precision binary32
:pre (and (and (<= 0.0 s) (<= s 256.0)) (and (<= 0.25 u) (<= u 1.0)))
(* (* 3.0 s) (log (/ 1.0 (- 1.0 (/ (- u 0.25) 0.75))))))