
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_5)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = (Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : max(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(tmp) end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 6 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_5)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = (Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : max(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(tmp) end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor h) 2.0))
(t_1 (pow (floor w) 2.0))
(t_2
(fmax
(fma dX.u (* dX.u t_1) (* dX.v (* dX.v t_0)))
(fma dY.u (* t_1 dY.u) (* t_0 (* dY.v dY.v)))))
(t_3
(fabs
(fma
(floor h)
(* (floor w) (fma dX.u dY.v (* dX.v (- dY.u))))
0.0))))
(log2
(if (> (/ t_2 t_3) (floor maxAniso))
(/ (sqrt t_2) (floor maxAniso))
(*
t_3
(/
1.0
(sqrt
(fmax
(+ (pow (* dX.u (floor w)) 2.0) (pow (* dX.v (floor h)) 2.0))
(+ (pow (* (floor h) dY.v) 2.0) (pow (* (floor w) dY.u) 2.0))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(h), 2.0f);
float t_1 = powf(floorf(w), 2.0f);
float t_2 = fmaxf(fmaf(dX_46_u, (dX_46_u * t_1), (dX_46_v * (dX_46_v * t_0))), fmaf(dY_46_u, (t_1 * dY_46_u), (t_0 * (dY_46_v * dY_46_v))));
float t_3 = fabsf(fmaf(floorf(h), (floorf(w) * fmaf(dX_46_u, dY_46_v, (dX_46_v * -dY_46_u))), 0.0f));
float tmp;
if ((t_2 / t_3) > floorf(maxAniso)) {
tmp = sqrtf(t_2) / floorf(maxAniso);
} else {
tmp = t_3 * (1.0f / sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + powf((dX_46_v * floorf(h)), 2.0f)), (powf((floorf(h) * dY_46_v), 2.0f) + powf((floorf(w) * dY_46_u), 2.0f)))));
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ Float32(2.0) t_1 = floor(w) ^ Float32(2.0) t_2 = (fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0))) != fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0)))) ? fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_0 * Float32(dY_46_v * dY_46_v))) : ((fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_0 * Float32(dY_46_v * dY_46_v))) != fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_0 * Float32(dY_46_v * dY_46_v)))) ? fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0))) : max(fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0))), fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_0 * Float32(dY_46_v * dY_46_v))))) t_3 = abs(fma(floor(h), Float32(floor(w) * fma(dX_46_u, dY_46_v, Float32(dX_46_v * Float32(-dY_46_u)))), Float32(0.0))) tmp = Float32(0.0) if (Float32(t_2 / t_3) > floor(maxAniso)) tmp = Float32(sqrt(t_2) / floor(maxAniso)); else tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0)))) ? Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) : ((Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) != Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0)))) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))), Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))))))))); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(dX.u, dX.u \cdot t\_1, dX.v \cdot \left(dX.v \cdot t\_0\right)\right), \mathsf{fma}\left(dY.u, t\_1 \cdot dY.u, t\_0 \cdot \left(dY.v \cdot dY.v\right)\right)\right)\\
t_3 := \left|\mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor w\right\rfloor \cdot \mathsf{fma}\left(dX.u, dY.v, dX.v \cdot \left(-dY.u\right)\right), 0\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_3} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{t\_2}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in w around 0
Simplified77.2%
Applied egg-rr77.2%
Final simplification77.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* dX.u (floor w))))
(log2
(if (>
(/
(fmax (+ (* t_4 t_4) (* t_0 t_0)) t_3)
(fabs (fma (fma dX.u dY.v 0.0) (fma (floor w) (floor h) 0.0) 0.0)))
(floor maxAniso))
(/
(sqrt
(fmax (+ (pow t_4 2.0) (pow t_0 2.0)) (+ (pow t_2 2.0) (pow t_1 2.0))))
(floor maxAniso))
(/
(fabs (* (* dX.v dY.u) (* (floor w) (floor h))))
(sqrt (fmax (* dX.u (* dX.u (pow (floor w) 2.0))) t_3)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_v * floorf(h);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = dX_46_u * floorf(w);
float tmp;
if ((fmaxf(((t_4 * t_4) + (t_0 * t_0)), t_3) / fabsf(fmaf(fmaf(dX_46_u, dY_46_v, 0.0f), fmaf(floorf(w), floorf(h), 0.0f), 0.0f))) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf((powf(t_4, 2.0f) + powf(t_0, 2.0f)), (powf(t_2, 2.0f) + powf(t_1, 2.0f)))) / floorf(maxAniso);
} else {
tmp = fabsf(((dX_46_v * dY_46_u) * (floorf(w) * floorf(h)))) / sqrtf(fmaxf((dX_46_u * (dX_46_u * powf(floorf(w), 2.0f))), t_3));
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(dX_46_u * floor(w)) tmp = Float32(0.0) if (Float32(((Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) != Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) : max(Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)), t_3))) / abs(fma(fma(dX_46_u, dY_46_v, Float32(0.0)), fma(floor(w), floor(h), Float32(0.0)), Float32(0.0)))) > floor(maxAniso)) tmp = Float32(sqrt(((Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : ((Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : max(Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))), Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(Float32(dX_46_v * dY_46_u) * Float32(floor(w) * floor(h)))) / sqrt(((Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) != Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0))))) ? t_3 : ((t_3 != t_3) ? Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))) : max(Float32(dX_46_u * Float32(dX_46_u * (floor(w) ^ Float32(2.0)))), t_3))))); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := dX.u \cdot \left\lfloor w\right\rfloor \\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_0 \cdot t\_0, t\_3\right)}{\left|\mathsf{fma}\left(\mathsf{fma}\left(dX.u, dY.v, 0\right), \mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor h\right\rfloor , 0\right), 0\right)\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left({t\_4}^{2} + {t\_0}^{2}, {t\_2}^{2} + {t\_1}^{2}\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left(dX.v \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right|}{\sqrt{\mathsf{max}\left(dX.u \cdot \left(dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}\right), t\_3\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3276.9%
Simplified76.9%
Taylor expanded in dX.u around inf
+-rgt-identityN/A
associate-*r*N/A
accelerator-lowering-fma.f32N/A
+-rgt-identityN/A
accelerator-lowering-fma.f32N/A
--rgt-identityN/A
sub-negN/A
*-commutativeN/A
metadata-evalN/A
accelerator-lowering-fma.f32N/A
floor-lowering-floor.f32N/A
floor-lowering-floor.f3276.5%
Simplified76.5%
Taylor expanded in dX.u around 0
associate-*r*N/A
associate-*r*N/A
*-lowering-*.f32N/A
mul-1-negN/A
neg-sub0N/A
--lowering--.f32N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
floor-lowering-floor.f3275.1%
Simplified75.1%
sqrt-lowering-sqrt.f32N/A
fmax-lowering-fmax.f32N/A
Applied egg-rr75.1%
Final simplification75.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor w) 2.0))
(t_1 (pow (floor h) 2.0))
(t_2 (* t_0 dY.u))
(t_3
(fabs
(fma (floor h) (* (floor w) (fma dX.u dY.v (* dX.v (- dY.u)))) 0.0)))
(t_4 (fma dX.u (* dX.u t_0) (* dX.v (* dX.v t_1))))
(t_5
(/
(sqrt (fmax t_4 (fma dY.u t_2 (* t_1 (* dY.v dY.v)))))
(floor maxAniso)))
(t_6
(*
t_3
(/
1.0
(sqrt
(fmax
(+ (pow (* dX.u (floor w)) 2.0) (pow (* dX.v (floor h)) 2.0))
(+ (pow (* (floor h) dY.v) 2.0) (pow (* (floor w) dY.u) 2.0)))))))
(t_7
(log2
(if (> (/ (fmax t_4 (* dY.v (* t_1 dY.v))) t_3) (floor maxAniso))
t_5
t_6))))
(if (<= dY.v -6400.0)
t_7
(if (<= dY.v 80000.0)
(log2 (if (> (/ (fmax t_4 (* dY.u t_2)) t_3) (floor maxAniso)) t_5 t_6))
t_7))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(w), 2.0f);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = t_0 * dY_46_u;
float t_3 = fabsf(fmaf(floorf(h), (floorf(w) * fmaf(dX_46_u, dY_46_v, (dX_46_v * -dY_46_u))), 0.0f));
float t_4 = fmaf(dX_46_u, (dX_46_u * t_0), (dX_46_v * (dX_46_v * t_1)));
float t_5 = sqrtf(fmaxf(t_4, fmaf(dY_46_u, t_2, (t_1 * (dY_46_v * dY_46_v))))) / floorf(maxAniso);
float t_6 = t_3 * (1.0f / sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + powf((dX_46_v * floorf(h)), 2.0f)), (powf((floorf(h) * dY_46_v), 2.0f) + powf((floorf(w) * dY_46_u), 2.0f)))));
float tmp;
if ((fmaxf(t_4, (dY_46_v * (t_1 * dY_46_v))) / t_3) > floorf(maxAniso)) {
tmp = t_5;
} else {
tmp = t_6;
}
float t_7 = log2f(tmp);
float tmp_1;
if (dY_46_v <= -6400.0f) {
tmp_1 = t_7;
} else if (dY_46_v <= 80000.0f) {
float tmp_2;
if ((fmaxf(t_4, (dY_46_u * t_2)) / t_3) > floorf(maxAniso)) {
tmp_2 = t_5;
} else {
tmp_2 = t_6;
}
tmp_1 = log2f(tmp_2);
} else {
tmp_1 = t_7;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) ^ Float32(2.0) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(t_0 * dY_46_u) t_3 = abs(fma(floor(h), Float32(floor(w) * fma(dX_46_u, dY_46_v, Float32(dX_46_v * Float32(-dY_46_u)))), Float32(0.0))) t_4 = fma(dX_46_u, Float32(dX_46_u * t_0), Float32(dX_46_v * Float32(dX_46_v * t_1))) t_5 = Float32(sqrt(((t_4 != t_4) ? fma(dY_46_u, t_2, Float32(t_1 * Float32(dY_46_v * dY_46_v))) : ((fma(dY_46_u, t_2, Float32(t_1 * Float32(dY_46_v * dY_46_v))) != fma(dY_46_u, t_2, Float32(t_1 * Float32(dY_46_v * dY_46_v)))) ? t_4 : max(t_4, fma(dY_46_u, t_2, Float32(t_1 * Float32(dY_46_v * dY_46_v))))))) / floor(maxAniso)) t_6 = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0)))) ? Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) : ((Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) != Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0)))) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))), Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))))))))) tmp = Float32(0.0) if (Float32(((t_4 != t_4) ? Float32(dY_46_v * Float32(t_1 * dY_46_v)) : ((Float32(dY_46_v * Float32(t_1 * dY_46_v)) != Float32(dY_46_v * Float32(t_1 * dY_46_v))) ? t_4 : max(t_4, Float32(dY_46_v * Float32(t_1 * dY_46_v))))) / t_3) > floor(maxAniso)) tmp = t_5; else tmp = t_6; end t_7 = log2(tmp) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(-6400.0)) tmp_1 = t_7; elseif (dY_46_v <= Float32(80000.0)) tmp_2 = Float32(0.0) if (Float32(((t_4 != t_4) ? Float32(dY_46_u * t_2) : ((Float32(dY_46_u * t_2) != Float32(dY_46_u * t_2)) ? t_4 : max(t_4, Float32(dY_46_u * t_2)))) / t_3) > floor(maxAniso)) tmp_2 = t_5; else tmp_2 = t_6; end tmp_1 = log2(tmp_2); else tmp_1 = t_7; end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := t\_0 \cdot dY.u\\
t_3 := \left|\mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor w\right\rfloor \cdot \mathsf{fma}\left(dX.u, dY.v, dX.v \cdot \left(-dY.u\right)\right), 0\right)\right|\\
t_4 := \mathsf{fma}\left(dX.u, dX.u \cdot t\_0, dX.v \cdot \left(dX.v \cdot t\_1\right)\right)\\
t_5 := \frac{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(dY.u, t\_2, t\_1 \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}{\left\lfloor maxAniso\right\rfloor }\\
t_6 := t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}}\\
t_7 := \log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_4, dY.v \cdot \left(t\_1 \cdot dY.v\right)\right)}{t\_3} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{if}\;dY.v \leq -6400:\\
\;\;\;\;t\_7\\
\mathbf{elif}\;dY.v \leq 80000:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_4, dY.u \cdot t\_2\right)}{t\_3} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
if dY.v < -6400 or 8e4 < dY.v Initial program 76.7%
Taylor expanded in w around 0
Simplified76.7%
Applied egg-rr76.7%
Taylor expanded in dY.u around 0
unpow2N/A
associate-*l*N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3276.6%
Simplified76.6%
if -6400 < dY.v < 8e4Initial program 77.6%
Taylor expanded in w around 0
Simplified77.7%
Applied egg-rr77.7%
Taylor expanded in dY.u around inf
unpow2N/A
associate-*l*N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3277.0%
Simplified77.0%
Final simplification76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor h) 2.0))
(t_1 (pow (floor w) 2.0))
(t_2 (* t_1 dY.u))
(t_3 (fma dX.u (* dX.u t_1) (* dX.v (* dX.v t_0))))
(t_4
(fabs
(fma
(floor h)
(* (floor w) (fma dX.u dY.v (* dX.v (- dY.u))))
0.0))))
(log2
(if (> (/ (fmax t_3 (* dY.u t_2)) t_4) (floor maxAniso))
(/
(sqrt (fmax t_3 (fma dY.u t_2 (* t_0 (* dY.v dY.v)))))
(floor maxAniso))
(*
t_4
(/
1.0
(sqrt
(fmax
(+ (pow (* dX.u (floor w)) 2.0) (pow (* dX.v (floor h)) 2.0))
(+ (pow (* (floor h) dY.v) 2.0) (pow (* (floor w) dY.u) 2.0))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(h), 2.0f);
float t_1 = powf(floorf(w), 2.0f);
float t_2 = t_1 * dY_46_u;
float t_3 = fmaf(dX_46_u, (dX_46_u * t_1), (dX_46_v * (dX_46_v * t_0)));
float t_4 = fabsf(fmaf(floorf(h), (floorf(w) * fmaf(dX_46_u, dY_46_v, (dX_46_v * -dY_46_u))), 0.0f));
float tmp;
if ((fmaxf(t_3, (dY_46_u * t_2)) / t_4) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(t_3, fmaf(dY_46_u, t_2, (t_0 * (dY_46_v * dY_46_v))))) / floorf(maxAniso);
} else {
tmp = t_4 * (1.0f / sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + powf((dX_46_v * floorf(h)), 2.0f)), (powf((floorf(h) * dY_46_v), 2.0f) + powf((floorf(w) * dY_46_u), 2.0f)))));
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ Float32(2.0) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(t_1 * dY_46_u) t_3 = fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0))) t_4 = abs(fma(floor(h), Float32(floor(w) * fma(dX_46_u, dY_46_v, Float32(dX_46_v * Float32(-dY_46_u)))), Float32(0.0))) tmp = Float32(0.0) if (Float32(((t_3 != t_3) ? Float32(dY_46_u * t_2) : ((Float32(dY_46_u * t_2) != Float32(dY_46_u * t_2)) ? t_3 : max(t_3, Float32(dY_46_u * t_2)))) / t_4) > floor(maxAniso)) tmp = Float32(sqrt(((t_3 != t_3) ? fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))) : ((fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))) != fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v)))) ? t_3 : max(t_3, fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))))))) / floor(maxAniso)); else tmp = Float32(t_4 * Float32(Float32(1.0) / sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0)))) ? Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) : ((Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) != Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0)))) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))), Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))))))))); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := t\_1 \cdot dY.u\\
t_3 := \mathsf{fma}\left(dX.u, dX.u \cdot t\_1, dX.v \cdot \left(dX.v \cdot t\_0\right)\right)\\
t_4 := \left|\mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor w\right\rfloor \cdot \mathsf{fma}\left(dX.u, dY.v, dX.v \cdot \left(-dY.u\right)\right), 0\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_3, dY.u \cdot t\_2\right)}{t\_4} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(dY.u, t\_2, t\_0 \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot \frac{1}{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in w around 0
Simplified77.2%
Applied egg-rr77.2%
Taylor expanded in dY.u around inf
unpow2N/A
associate-*l*N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3264.4%
Simplified64.4%
Final simplification64.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (floor h) 2.0))
(t_1 (pow (floor w) 2.0))
(t_2 (* t_1 dY.u))
(t_3 (fma dX.u (* dX.u t_1) (* dX.v (* dX.v t_0)))))
(log2
(if (>
(/
(fmax t_3 (* dY.u t_2))
(fabs
(fma
(floor h)
(* (floor w) (fma dX.u dY.v (* dX.v (- dY.u))))
0.0)))
(floor maxAniso))
(/
(sqrt (fmax t_3 (fma dY.u t_2 (* t_0 (* dY.v dY.v)))))
(floor maxAniso))
(/
(fma
(floor w)
(* (floor h) (fma dX.v (- dY.u) (fma dX.u dY.v 0.0)))
0.0)
(sqrt
(fmax
(+ (pow (* dX.u (floor w)) 2.0) (pow (* dX.v (floor h)) 2.0))
(+ (pow (* (floor h) dY.v) 2.0) (pow (* (floor w) dY.u) 2.0)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(h), 2.0f);
float t_1 = powf(floorf(w), 2.0f);
float t_2 = t_1 * dY_46_u;
float t_3 = fmaf(dX_46_u, (dX_46_u * t_1), (dX_46_v * (dX_46_v * t_0)));
float tmp;
if ((fmaxf(t_3, (dY_46_u * t_2)) / fabsf(fmaf(floorf(h), (floorf(w) * fmaf(dX_46_u, dY_46_v, (dX_46_v * -dY_46_u))), 0.0f))) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(t_3, fmaf(dY_46_u, t_2, (t_0 * (dY_46_v * dY_46_v))))) / floorf(maxAniso);
} else {
tmp = fmaf(floorf(w), (floorf(h) * fmaf(dX_46_v, -dY_46_u, fmaf(dX_46_u, dY_46_v, 0.0f))), 0.0f) / sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + powf((dX_46_v * floorf(h)), 2.0f)), (powf((floorf(h) * dY_46_v), 2.0f) + powf((floorf(w) * dY_46_u), 2.0f))));
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) ^ Float32(2.0) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(t_1 * dY_46_u) t_3 = fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_0))) tmp = Float32(0.0) if (Float32(((t_3 != t_3) ? Float32(dY_46_u * t_2) : ((Float32(dY_46_u * t_2) != Float32(dY_46_u * t_2)) ? t_3 : max(t_3, Float32(dY_46_u * t_2)))) / abs(fma(floor(h), Float32(floor(w) * fma(dX_46_u, dY_46_v, Float32(dX_46_v * Float32(-dY_46_u)))), Float32(0.0)))) > floor(maxAniso)) tmp = Float32(sqrt(((t_3 != t_3) ? fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))) : ((fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))) != fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v)))) ? t_3 : max(t_3, fma(dY_46_u, t_2, Float32(t_0 * Float32(dY_46_v * dY_46_v))))))) / floor(maxAniso)); else tmp = Float32(fma(floor(w), Float32(floor(h) * fma(dX_46_v, Float32(-dY_46_u), fma(dX_46_u, dY_46_v, Float32(0.0)))), Float32(0.0)) / sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0)))) ? Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) : ((Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) != Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0)))) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))), Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0)))))))); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := t\_1 \cdot dY.u\\
t_3 := \mathsf{fma}\left(dX.u, dX.u \cdot t\_1, dX.v \cdot \left(dX.v \cdot t\_0\right)\right)\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_3, dY.u \cdot t\_2\right)}{\left|\mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor w\right\rfloor \cdot \mathsf{fma}\left(dX.u, dY.v, dX.v \cdot \left(-dY.u\right)\right), 0\right)\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(dY.u, t\_2, t\_0 \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor h\right\rfloor \cdot \mathsf{fma}\left(dX.v, -dY.u, \mathsf{fma}\left(dX.u, dY.v, 0\right)\right), 0\right)}{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in w around 0
Simplified77.2%
Applied egg-rr77.2%
Taylor expanded in dY.u around inf
unpow2N/A
associate-*l*N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3264.4%
Simplified64.4%
Applied egg-rr60.8%
Final simplification60.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* (floor w) dY.u) 2.0))
(t_1 (pow (floor w) 2.0))
(t_2 (+ (pow (* dX.u (floor w)) 2.0) (pow (* dX.v (floor h)) 2.0)))
(t_3 (pow (floor h) 2.0)))
(log2
(if (>
(/
(fmax t_2 t_0)
(fma
(floor w)
(* (floor h) (fma dX.v (- dY.u) (fma dX.u dY.v 0.0)))
0.0))
(floor maxAniso))
(/
(sqrt
(fmax
(fma dX.u (* dX.u t_1) (* dX.v (* dX.v t_3)))
(fma dY.u (* t_1 dY.u) (* t_3 (* dY.v dY.v)))))
(floor maxAniso))
(*
(fabs
(fma (floor h) (* (floor w) (fma dX.u dY.v (* dX.v (- dY.u)))) 0.0))
(/ 1.0 (sqrt (fmax t_2 (+ (pow (* (floor h) dY.v) 2.0) t_0)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf((floorf(w) * dY_46_u), 2.0f);
float t_1 = powf(floorf(w), 2.0f);
float t_2 = powf((dX_46_u * floorf(w)), 2.0f) + powf((dX_46_v * floorf(h)), 2.0f);
float t_3 = powf(floorf(h), 2.0f);
float tmp;
if ((fmaxf(t_2, t_0) / fmaf(floorf(w), (floorf(h) * fmaf(dX_46_v, -dY_46_u, fmaf(dX_46_u, dY_46_v, 0.0f))), 0.0f)) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(fmaf(dX_46_u, (dX_46_u * t_1), (dX_46_v * (dX_46_v * t_3))), fmaf(dY_46_u, (t_1 * dY_46_u), (t_3 * (dY_46_v * dY_46_v))))) / floorf(maxAniso);
} else {
tmp = fabsf(fmaf(floorf(h), (floorf(w) * fmaf(dX_46_u, dY_46_v, (dX_46_v * -dY_46_u))), 0.0f)) * (1.0f / sqrtf(fmaxf(t_2, (powf((floorf(h) * dY_46_v), 2.0f) + t_0))));
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) ^ Float32(2.0) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) t_3 = floor(h) ^ Float32(2.0) tmp = Float32(0.0) if (Float32(((t_2 != t_2) ? t_0 : ((t_0 != t_0) ? t_2 : max(t_2, t_0))) / fma(floor(w), Float32(floor(h) * fma(dX_46_v, Float32(-dY_46_u), fma(dX_46_u, dY_46_v, Float32(0.0)))), Float32(0.0))) > floor(maxAniso)) tmp = Float32(sqrt(((fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_3))) != fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_3)))) ? fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_3 * Float32(dY_46_v * dY_46_v))) : ((fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_3 * Float32(dY_46_v * dY_46_v))) != fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_3 * Float32(dY_46_v * dY_46_v)))) ? fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_3))) : max(fma(dX_46_u, Float32(dX_46_u * t_1), Float32(dX_46_v * Float32(dX_46_v * t_3))), fma(dY_46_u, Float32(t_1 * dY_46_u), Float32(t_3 * Float32(dY_46_v * dY_46_v))))))) / floor(maxAniso)); else tmp = Float32(abs(fma(floor(h), Float32(floor(w) * fma(dX_46_u, dY_46_v, Float32(dX_46_v * Float32(-dY_46_u)))), Float32(0.0))) * Float32(Float32(1.0) / sqrt(((t_2 != t_2) ? Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_0) : ((Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_0) != Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_0)) ? t_2 : max(t_2, Float32((Float32(floor(h) * dY_46_v) ^ Float32(2.0)) + t_0))))))); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_2, t\_0\right)}{\mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor h\right\rfloor \cdot \mathsf{fma}\left(dX.v, -dY.u, \mathsf{fma}\left(dX.u, dY.v, 0\right)\right), 0\right)} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u, dX.u \cdot t\_1, dX.v \cdot \left(dX.v \cdot t\_3\right)\right), \mathsf{fma}\left(dY.u, t\_1 \cdot dY.u, t\_3 \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\left|\mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor w\right\rfloor \cdot \mathsf{fma}\left(dX.u, dY.v, dX.v \cdot \left(-dY.u\right)\right), 0\right)\right| \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_2, {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2} + t\_0\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in w around 0
Simplified77.2%
Applied egg-rr77.2%
Taylor expanded in dY.u around inf
unpow2N/A
associate-*l*N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3264.4%
Simplified64.4%
Applied egg-rr39.1%
Final simplification39.1%
herbie shell --seed 2024193
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (LOD)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(log2 (if (> (/ (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u))))) (floor maxAniso)) (/ (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v))))) (floor maxAniso)) (/ (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u)))) (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))))))