
(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 5 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 (* (floor h) dY.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) dX.v))
(t_3 (- (* dX.u dY.v) (* dX.v dY.u)))
(t_4 (* (floor w) dY.u))
(t_5
(sqrt
(fmax
(+ (* (floor w) (* dX.u t_1)) (* (floor h) (* dX.v t_2)))
(+ (* (floor w) (* dY.u t_4)) (* (floor h) (* dY.v t_0)))))))
(log2
(if (>
(*
(/
(fmax
(+ (pow t_1 2.0) (pow t_2 2.0))
(+ (pow t_4 2.0) (pow t_0 2.0)))
(fabs (* (floor w) (floor h))))
(/ 1.0 (fabs t_3)))
(floor maxAniso))
(/ t_5 (floor maxAniso))
(/ (fabs (* (floor w) (* (floor h) t_3))) 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) * dY_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = (dX_46_u * dY_46_v) - (dX_46_v * dY_46_u);
float t_4 = floorf(w) * dY_46_u;
float t_5 = sqrtf(fmaxf(((floorf(w) * (dX_46_u * t_1)) + (floorf(h) * (dX_46_v * t_2))), ((floorf(w) * (dY_46_u * t_4)) + (floorf(h) * (dY_46_v * t_0)))));
float tmp;
if (((fmaxf((powf(t_1, 2.0f) + powf(t_2, 2.0f)), (powf(t_4, 2.0f) + powf(t_0, 2.0f))) / fabsf((floorf(w) * floorf(h)))) * (1.0f / fabsf(t_3))) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (floorf(h) * t_3))) / 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) * dY_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u)) t_4 = Float32(floor(w) * dY_46_u) t_5 = sqrt(((Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_2))) != Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_2)))) ? Float32(Float32(floor(w) * Float32(dY_46_u * t_4)) + Float32(floor(h) * Float32(dY_46_v * t_0))) : ((Float32(Float32(floor(w) * Float32(dY_46_u * t_4)) + Float32(floor(h) * Float32(dY_46_v * t_0))) != Float32(Float32(floor(w) * Float32(dY_46_u * t_4)) + Float32(floor(h) * Float32(dY_46_v * t_0)))) ? Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_2))) : max(Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_2))), Float32(Float32(floor(w) * Float32(dY_46_u * t_4)) + Float32(floor(h) * Float32(dY_46_v * t_0))))))) tmp = Float32(0.0) if (Float32(Float32(((Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : ((Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : max(Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))), Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))))) / abs(Float32(floor(w) * floor(h)))) * Float32(Float32(1.0) / abs(t_3))) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(floor(h) * t_3))) / 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) * dY_46_v; t_1 = floor(w) * dX_46_u; t_2 = floor(h) * dX_46_v; t_3 = (dX_46_u * dY_46_v) - (dX_46_v * dY_46_u); t_4 = floor(w) * dY_46_u; t_5 = sqrt(max(((floor(w) * (dX_46_u * t_1)) + (floor(h) * (dX_46_v * t_2))), ((floor(w) * (dY_46_u * t_4)) + (floor(h) * (dY_46_v * t_0))))); tmp = single(0.0); if (((max(((t_1 ^ single(2.0)) + (t_2 ^ single(2.0))), ((t_4 ^ single(2.0)) + (t_0 ^ single(2.0)))) / abs((floor(w) * floor(h)))) * (single(1.0) / abs(t_3))) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = abs((floor(w) * (floor(h) * t_3))) / t_5; end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dX.u \cdot dY.v - dX.v \cdot dY.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_5 := \sqrt{\mathsf{max}\left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_1\right) + \left\lfloor h\right\rfloor \cdot \left(dX.v \cdot t\_2\right), \left\lfloor w\right\rfloor \cdot \left(dY.u \cdot t\_4\right) + \left\lfloor h\right\rfloor \cdot \left(dY.v \cdot t\_0\right)\right)}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left({t\_1}^{2} + {t\_2}^{2}, {t\_4}^{2} + {t\_0}^{2}\right)}{\left|\left\lfloor w\right\rfloor \cdot \left\lfloor h\right\rfloor \right|} \cdot \frac{1}{\left|t\_3\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot t\_3\right)\right|}{t\_5}\\
\end{array}
\end{array}
\end{array}
Initial program 75.3%
Simplified75.3%
div-invN/A
associate-*r/N/A
associate-*r*N/A
fabs-mulN/A
times-fracN/A
Applied egg-rr75.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (- (* dX.u dY.v) (* dX.v dY.u))))
(t_1 (pow (* (floor w) dY.u) 2.0))
(t_2 (+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_3 (fmax t_2 (+ t_1 (pow (* (floor h) dY.v) 2.0)))))
(log2
(if (> (/ (/ t_3 (floor w)) (fabs t_0)) (floor maxAniso))
(/
(pow (fmax t_2 (+ t_1 (* dY.v (* dY.v (pow (floor h) 2.0))))) 0.5)
(floor maxAniso))
(/ (fabs (* (floor w) t_0)) (sqrt 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 = floorf(h) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u));
float t_1 = powf((floorf(w) * dY_46_u), 2.0f);
float t_2 = powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = fmaxf(t_2, (t_1 + powf((floorf(h) * dY_46_v), 2.0f)));
float tmp;
if (((t_3 / floorf(w)) / fabsf(t_0)) > floorf(maxAniso)) {
tmp = powf(fmaxf(t_2, (t_1 + (dY_46_v * (dY_46_v * powf(floorf(h), 2.0f))))), 0.5f) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * t_0)) / sqrtf(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(floor(h) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u))) t_1 = Float32(floor(w) * dY_46_u) ^ Float32(2.0) t_2 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_3 = (t_2 != t_2) ? Float32(t_1 + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) : ((Float32(t_1 + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) != Float32(t_1 + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) ? t_2 : max(t_2, Float32(t_1 + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) tmp = Float32(0.0) if (Float32(Float32(t_3 / floor(w)) / abs(t_0)) > floor(maxAniso)) tmp = Float32((((t_2 != t_2) ? Float32(t_1 + Float32(dY_46_v * Float32(dY_46_v * (floor(h) ^ Float32(2.0))))) : ((Float32(t_1 + Float32(dY_46_v * Float32(dY_46_v * (floor(h) ^ Float32(2.0))))) != Float32(t_1 + Float32(dY_46_v * Float32(dY_46_v * (floor(h) ^ Float32(2.0)))))) ? t_2 : max(t_2, Float32(t_1 + Float32(dY_46_v * Float32(dY_46_v * (floor(h) ^ Float32(2.0)))))))) ^ Float32(0.5)) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * t_0)) / sqrt(t_3)); 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_u * dY_46_v) - (dX_46_v * dY_46_u)); t_1 = (floor(w) * dY_46_u) ^ single(2.0); t_2 = ((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_3 = max(t_2, (t_1 + ((floor(h) * dY_46_v) ^ single(2.0)))); tmp = single(0.0); if (((t_3 / floor(w)) / abs(t_0)) > floor(maxAniso)) tmp = (max(t_2, (t_1 + (dY_46_v * (dY_46_v * (floor(h) ^ single(2.0)))))) ^ single(0.5)) / floor(maxAniso); else tmp = abs((floor(w) * t_0)) / sqrt(t_3); end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\\
t_1 := {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_2 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := \mathsf{max}\left(t\_2, t\_1 + {\left(\left\lfloor h\right\rfloor \cdot dY.v\right)}^{2}\right)\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\frac{t\_3}{\left\lfloor w\right\rfloor }}{\left|t\_0\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{{\left(\mathsf{max}\left(t\_2, t\_1 + dY.v \cdot \left(dY.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right)\right)\right)}^{0.5}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot t\_0\right|}{\sqrt{t\_3}}\\
\end{array}
\end{array}
\end{array}
Initial program 75.3%
Simplified75.3%
Applied egg-rr75.3%
unpow1/2N/A
sqrt-lowering-sqrt.f32N/A
pow2N/A
associate-*r*N/A
pow2N/A
associate-*r*N/A
pow2N/A
associate-*r*N/A
pow2N/A
associate-*r*N/A
Applied egg-rr75.3%
Applied egg-rr75.4%
metadata-evalN/A
pow-powN/A
pow-prod-downN/A
pow2N/A
associate-*r*N/A
unpow-prod-downN/A
unpow1N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3275.4%
Applied egg-rr75.4%
Final simplification75.4%
(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
(+ (pow t_3 2.0) (pow t_0 2.0))
(+ (pow t_1 2.0) (pow t_2 2.0)))))
(log2
(if (> (/ t_4 (* (floor w) (* (floor h) (* dX.u dY.v)))) (floor maxAniso))
(/
(sqrt (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(floor maxAniso))
(/ (fabs (* (floor w) (* dX.u t_2))) (pow (pow t_4 0.0625) 8.0))))))
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 = 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((powf(t_3, 2.0f) + powf(t_0, 2.0f)), (powf(t_1, 2.0f) + powf(t_2, 2.0f)));
float tmp;
if ((t_4 / (floorf(w) * (floorf(h) * (dX_46_u * dY_46_v)))) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (dX_46_u * t_2))) / powf(powf(t_4, 0.0625f), 8.0f);
}
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((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : max(Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))), Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))))) tmp = Float32(0.0) if (Float32(t_4 / Float32(floor(w) * Float32(floor(h) * Float32(dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = Float32(sqrt(((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)))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(dX_46_u * t_2))) / ((t_4 ^ Float32(0.0625)) ^ Float32(8.0))); 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 ^ single(2.0)) + (t_0 ^ single(2.0))), ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))); tmp = single(0.0); if ((t_4 / (floor(w) * (floor(h) * (dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))) / floor(maxAniso); else tmp = abs((floor(w) * (dX_46_u * t_2))) / ((t_4 ^ single(0.0625)) ^ single(8.0)); 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}^{2} + {t\_0}^{2}, {t\_1}^{2} + {t\_2}^{2}\right)\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{\left\lfloor w\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dX.u \cdot dY.v\right)\right)} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_2\right)\right|}{{\left({t\_4}^{0.0625}\right)}^{8}}\\
\end{array}
\end{array}
\end{array}
Initial program 75.3%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Applied egg-rr46.4%
Applied egg-rr46.4%
Final simplification46.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow (floor w) 2.0))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (pow t_3 2.0))
(t_5 (* (floor w) dX.u)))
(log2
(if (>
(/
(fmax (+ (pow t_5 2.0) (pow t_0 2.0)) (+ (pow t_2 2.0) t_4))
(* (floor w) (* (floor h) (* dX.u dY.v))))
(floor maxAniso))
(/
(sqrt (fmax (+ (* t_5 t_5) (* t_0 t_0)) (+ (* t_2 t_2) (* t_3 t_3))))
(floor maxAniso))
(/
(fabs (* (floor w) (* dX.u t_3)))
(sqrt
(fmax
(+ (* dX.u (* dX.u t_1)) (* dX.v (* dX.v (pow (floor h) 2.0))))
(+ t_4 (* t_1 (* dY.u dY.u))))))))))
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 = powf(floorf(w), 2.0f);
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = powf(t_3, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float tmp;
if ((fmaxf((powf(t_5, 2.0f) + powf(t_0, 2.0f)), (powf(t_2, 2.0f) + t_4)) / (floorf(w) * (floorf(h) * (dX_46_u * dY_46_v)))) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(((t_5 * t_5) + (t_0 * t_0)), ((t_2 * t_2) + (t_3 * t_3)))) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (dX_46_u * t_3))) / sqrtf(fmaxf(((dX_46_u * (dX_46_u * t_1)) + (dX_46_v * (dX_46_v * powf(floorf(h), 2.0f)))), (t_4 + (t_1 * (dY_46_u * dY_46_u)))));
}
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 = floor(w) ^ Float32(2.0) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (Float32(((Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_2 ^ Float32(2.0)) + t_4) : ((Float32((t_2 ^ Float32(2.0)) + t_4) != Float32((t_2 ^ Float32(2.0)) + t_4)) ? Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : max(Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))), Float32((t_2 ^ Float32(2.0)) + t_4)))) / Float32(floor(w) * Float32(floor(h) * Float32(dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = Float32(sqrt(((Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) != Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) : ((Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) != Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3))) ? Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) : max(Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(dX_46_u * t_3))) / sqrt(((Float32(Float32(dX_46_u * Float32(dX_46_u * t_1)) + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) != Float32(Float32(dX_46_u * Float32(dX_46_u * t_1)) + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0)))))) ? Float32(t_4 + Float32(t_1 * Float32(dY_46_u * dY_46_u))) : ((Float32(t_4 + Float32(t_1 * Float32(dY_46_u * dY_46_u))) != Float32(t_4 + Float32(t_1 * Float32(dY_46_u * dY_46_u)))) ? Float32(Float32(dX_46_u * Float32(dX_46_u * t_1)) + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))) : max(Float32(Float32(dX_46_u * Float32(dX_46_u * t_1)) + Float32(dX_46_v * Float32(dX_46_v * (floor(h) ^ Float32(2.0))))), Float32(t_4 + Float32(t_1 * Float32(dY_46_u * dY_46_u)))))))); 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) ^ single(2.0); t_2 = floor(w) * dY_46_u; t_3 = floor(h) * dY_46_v; t_4 = t_3 ^ single(2.0); t_5 = floor(w) * dX_46_u; tmp = single(0.0); if ((max(((t_5 ^ single(2.0)) + (t_0 ^ single(2.0))), ((t_2 ^ single(2.0)) + t_4)) / (floor(w) * (floor(h) * (dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = sqrt(max(((t_5 * t_5) + (t_0 * t_0)), ((t_2 * t_2) + (t_3 * t_3)))) / floor(maxAniso); else tmp = abs((floor(w) * (dX_46_u * t_3))) / sqrt(max(((dX_46_u * (dX_46_u * t_1)) + (dX_46_v * (dX_46_v * (floor(h) ^ single(2.0))))), (t_4 + (t_1 * (dY_46_u * dY_46_u))))); 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(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := {t\_3}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left({t\_5}^{2} + {t\_0}^{2}, {t\_2}^{2} + t\_4\right)}{\left\lfloor w\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dX.u \cdot dY.v\right)\right)} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(t\_5 \cdot t\_5 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_3\right)\right|}{\sqrt{\mathsf{max}\left(dX.u \cdot \left(dX.u \cdot t\_1\right) + dX.v \cdot \left(dX.v \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}\right), t\_4 + t\_1 \cdot \left(dY.u \cdot dY.u\right)\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 75.3%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Applied egg-rr46.4%
Taylor expanded in w around 0
sqrt-lowering-sqrt.f32N/A
fmax-lowering-fmax.f32N/A
Simplified46.4%
Final simplification46.4%
(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
(+ (pow t_3 2.0) (pow t_0 2.0))
(+ (pow t_1 2.0) (pow t_2 2.0)))))
(log2
(if (> (/ t_4 (* (floor w) (* (floor h) (* dX.u dY.v)))) (floor maxAniso))
(/
(sqrt (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(floor maxAniso))
(/ (fabs (* (floor w) (* dX.u t_2))) (sqrt t_4))))))
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((powf(t_3, 2.0f) + powf(t_0, 2.0f)), (powf(t_1, 2.0f) + powf(t_2, 2.0f)));
float tmp;
if ((t_4 / (floorf(w) * (floorf(h) * (dX_46_u * dY_46_v)))) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (dX_46_u * t_2))) / sqrtf(t_4);
}
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((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) ? Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : max(Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))), Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))))) tmp = Float32(0.0) if (Float32(t_4 / Float32(floor(w) * Float32(floor(h) * Float32(dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = Float32(sqrt(((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)))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(dX_46_u * t_2))) / sqrt(t_4)); 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 ^ single(2.0)) + (t_0 ^ single(2.0))), ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))); tmp = single(0.0); if ((t_4 / (floor(w) * (floor(h) * (dX_46_u * dY_46_v)))) > floor(maxAniso)) tmp = sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))) / floor(maxAniso); else tmp = abs((floor(w) * (dX_46_u * t_2))) / sqrt(t_4); 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}^{2} + {t\_0}^{2}, {t\_1}^{2} + {t\_2}^{2}\right)\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{\left\lfloor w\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dX.u \cdot dY.v\right)\right)} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_2\right)\right|}{\sqrt{t\_4}}\\
\end{array}
\end{array}
\end{array}
Initial program 75.3%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Taylor expanded in dX.u around inf
associate-*r*N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
associate-*r*N/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f3274.5%
Simplified74.5%
Applied egg-rr46.4%
pow2N/A
+-commutativeN/A
pow2N/A
*-commutativeN/A
pow2N/A
pow2N/A
sqrt-lowering-sqrt.f32N/A
Applied egg-rr46.4%
Final simplification46.4%
herbie shell --seed 2024288
(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)))))))))