
(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\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_2 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_3 := \left\lfloorw\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\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{t\_5}{\left\lfloormaxAniso\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\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_2 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_3 := \left\lfloorw\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\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{t\_5}{\left\lfloormaxAniso\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 (- (* dX.u dY.v) (* dX.v dY.u)))
(t_1
(fmax
(+ (pow (* (floor w) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0))
(+ (pow (* (floor w) dY.u) 2.0) (pow (* (floor h) dY.v) 2.0))))
(t_2 (pow t_1 0.5)))
(log2
(if (>
(/ (/ t_1 (fabs t_0)) (fabs (* (floor w) (floor h))))
(floor maxAniso))
(/ t_2 (floor maxAniso))
(/ (fabs (* (floor h) (* (floor w) t_0))) t_2)))))
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_u * dY_46_v) - (dX_46_v * dY_46_u);
float t_1 = fmaxf((powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f)), (powf((floorf(w) * dY_46_u), 2.0f) + powf((floorf(h) * dY_46_v), 2.0f)));
float t_2 = powf(t_1, 0.5f);
float tmp;
if (((t_1 / fabsf(t_0)) / fabsf((floorf(w) * floorf(h)))) > floorf(maxAniso)) {
tmp = t_2 / floorf(maxAniso);
} else {
tmp = fabsf((floorf(h) * (floorf(w) * t_0))) / t_2;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u)) t_1 = (Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) != Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) ? Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) : ((Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) != Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0)))) ? Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) : max(Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))), Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))))) t_2 = t_1 ^ Float32(0.5) tmp = Float32(0.0) if (Float32(Float32(t_1 / abs(t_0)) / abs(Float32(floor(w) * floor(h)))) > floor(maxAniso)) tmp = Float32(t_2 / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(h) * Float32(floor(w) * t_0))) / t_2); 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 = (dX_46_u * dY_46_v) - (dX_46_v * dY_46_u); t_1 = max((((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0))), (((floor(w) * dY_46_u) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0)))); t_2 = t_1 ^ single(0.5); tmp = single(0.0); if (((t_1 / abs(t_0)) / abs((floor(w) * floor(h)))) > floor(maxAniso)) tmp = t_2 / floor(maxAniso); else tmp = abs((floor(h) * (floor(w) * t_0))) / t_2; end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot dY.v - dX.v \cdot dY.u\\
t_1 := \mathsf{max}\left({\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloorh\right\rfloor \cdot dX.v\right)}^{2}, {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {\left(\left\lfloorh\right\rfloor \cdot dY.v\right)}^{2}\right)\\
t_2 := {t\_1}^{0.5}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\frac{t\_1}{\left|t\_0\right|}}{\left|\left\lfloorw\right\rfloor \cdot \left\lfloorh\right\rfloor\right|} > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{t\_2}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloorh\right\rfloor \cdot \left(\left\lfloorw\right\rfloor \cdot t\_0\right)\right|}{t\_2}\\
\end{array}
\end{array}
\end{array}
Initial program 79.8%
Simplified79.8%
Applied egg-rr79.8%
Applied egg-rr79.9%
Final simplification79.9%
(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 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dX.u))
(t_5
(fmax
(+ (pow t_4 2.0) (pow t_1 2.0))
(+ (pow t_2 2.0) (pow t_3 2.0)))))
(log2
(if (> (/ (/ t_5 (fabs (floor w))) (fabs t_0)) (floor maxAniso))
(/
(sqrt
(fmax
(+ (* (floor w) (* dX.u t_4)) (* (floor h) (* dX.v t_1)))
(+ (* (floor w) (* dY.u t_2)) (* (floor h) (* dY.v t_3)))))
(floor maxAniso))
(/ (fabs (* (floor w) t_0)) (sqrt 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_u * dY_46_v) - (dX_46_v * dY_46_u));
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = fmaxf((powf(t_4, 2.0f) + powf(t_1, 2.0f)), (powf(t_2, 2.0f) + powf(t_3, 2.0f)));
float tmp;
if (((t_5 / fabsf(floorf(w))) / fabsf(t_0)) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(((floorf(w) * (dX_46_u * t_4)) + (floorf(h) * (dX_46_v * t_1))), ((floorf(w) * (dY_46_u * t_2)) + (floorf(h) * (dY_46_v * t_3))))) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * t_0)) / sqrtf(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) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u))) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = (Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? Float32((t_2 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) : ((Float32((t_2 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) != Float32((t_2 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) ? Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : max(Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))), Float32((t_2 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))))) tmp = Float32(0.0) if (Float32(Float32(t_5 / abs(floor(w))) / abs(t_0)) > floor(maxAniso)) tmp = Float32(sqrt(((Float32(Float32(floor(w) * Float32(dX_46_u * t_4)) + Float32(floor(h) * Float32(dX_46_v * t_1))) != Float32(Float32(floor(w) * Float32(dX_46_u * t_4)) + Float32(floor(h) * Float32(dX_46_v * t_1)))) ? Float32(Float32(floor(w) * Float32(dY_46_u * t_2)) + Float32(floor(h) * Float32(dY_46_v * t_3))) : ((Float32(Float32(floor(w) * Float32(dY_46_u * t_2)) + Float32(floor(h) * Float32(dY_46_v * t_3))) != Float32(Float32(floor(w) * Float32(dY_46_u * t_2)) + Float32(floor(h) * Float32(dY_46_v * t_3)))) ? Float32(Float32(floor(w) * Float32(dX_46_u * t_4)) + Float32(floor(h) * Float32(dX_46_v * t_1))) : max(Float32(Float32(floor(w) * Float32(dX_46_u * t_4)) + Float32(floor(h) * Float32(dX_46_v * t_1))), Float32(Float32(floor(w) * Float32(dY_46_u * t_2)) + Float32(floor(h) * Float32(dY_46_v * t_3))))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * t_0)) / sqrt(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_u * dY_46_v) - (dX_46_v * dY_46_u)); t_1 = floor(h) * dX_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(h) * dY_46_v; t_4 = floor(w) * dX_46_u; t_5 = max(((t_4 ^ single(2.0)) + (t_1 ^ single(2.0))), ((t_2 ^ single(2.0)) + (t_3 ^ single(2.0)))); tmp = single(0.0); if (((t_5 / abs(floor(w))) / abs(t_0)) > floor(maxAniso)) tmp = sqrt(max(((floor(w) * (dX_46_u * t_4)) + (floor(h) * (dX_46_v * t_1))), ((floor(w) * (dY_46_u * t_2)) + (floor(h) * (dY_46_v * t_3))))) / floor(maxAniso); else tmp = abs((floor(w) * t_0)) / sqrt(t_5); end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorh\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\\
t_1 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_2 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_3 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_4 := \left\lfloorw\right\rfloor \cdot dX.u\\
t_5 := \mathsf{max}\left({t\_4}^{2} + {t\_1}^{2}, {t\_2}^{2} + {t\_3}^{2}\right)\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\frac{t\_5}{\left|\left\lfloorw\right\rfloor\right|}}{\left|t\_0\right|} > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(\left\lfloorw\right\rfloor \cdot \left(dX.u \cdot t\_4\right) + \left\lfloorh\right\rfloor \cdot \left(dX.v \cdot t\_1\right), \left\lfloorw\right\rfloor \cdot \left(dY.u \cdot t\_2\right) + \left\lfloorh\right\rfloor \cdot \left(dY.v \cdot t\_3\right)\right)}}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloorw\right\rfloor \cdot t\_0\right|}{\sqrt{t\_5}}\\
\end{array}
\end{array}
\end{array}
Initial program 79.8%
Simplified79.8%
fabs-mulN/A
associate-/r*N/A
/-lowering-/.f32N/A
Applied egg-rr79.8%
Applied egg-rr79.8%
(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) dX.u) 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_2 (* (floor h) dY.v))
(t_3 (fmax t_1 (+ t_0 (pow t_2 2.0))))
(t_4
(fabs (* (floor h) (* (floor w) (- (* dX.u dY.v) (* dX.v dY.u)))))))
(log2
(if (> (/ t_3 t_4) (floor maxAniso))
(/
(pow (fmax t_1 (+ t_0 (* dY.v (* (floor h) t_2)))) 0.5)
(floor maxAniso))
(/ t_4 (pow t_3 0.5))))))
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 = powf((floorf(w) * dY_46_u), 2.0f);
float t_1 = powf((floorf(w) * dX_46_u), 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = fmaxf(t_1, (t_0 + powf(t_2, 2.0f)));
float t_4 = fabsf((floorf(h) * (floorf(w) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u)))));
float tmp;
if ((t_3 / t_4) > floorf(maxAniso)) {
tmp = powf(fmaxf(t_1, (t_0 + (dY_46_v * (floorf(h) * t_2)))), 0.5f) / floorf(maxAniso);
} else {
tmp = t_4 / powf(t_3, 0.5f);
}
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 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = (t_1 != t_1) ? Float32(t_0 + (t_2 ^ Float32(2.0))) : ((Float32(t_0 + (t_2 ^ Float32(2.0))) != Float32(t_0 + (t_2 ^ Float32(2.0)))) ? t_1 : max(t_1, Float32(t_0 + (t_2 ^ Float32(2.0))))) t_4 = abs(Float32(floor(h) * Float32(floor(w) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u))))) tmp = Float32(0.0) if (Float32(t_3 / t_4) > floor(maxAniso)) tmp = Float32((((t_1 != t_1) ? Float32(t_0 + Float32(dY_46_v * Float32(floor(h) * t_2))) : ((Float32(t_0 + Float32(dY_46_v * Float32(floor(h) * t_2))) != Float32(t_0 + Float32(dY_46_v * Float32(floor(h) * t_2)))) ? t_1 : max(t_1, Float32(t_0 + Float32(dY_46_v * Float32(floor(h) * t_2)))))) ^ Float32(0.5)) / floor(maxAniso)); else tmp = Float32(t_4 / (t_3 ^ Float32(0.5))); 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(w) * dY_46_u) ^ single(2.0); t_1 = ((floor(w) * dX_46_u) ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_2 = floor(h) * dY_46_v; t_3 = max(t_1, (t_0 + (t_2 ^ single(2.0)))); t_4 = abs((floor(h) * (floor(w) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u))))); tmp = single(0.0); if ((t_3 / t_4) > floor(maxAniso)) tmp = (max(t_1, (t_0 + (dY_46_v * (floor(h) * t_2)))) ^ single(0.5)) / floor(maxAniso); else tmp = t_4 / (t_3 ^ single(0.5)); end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2}\\
t_1 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2} + {\left(\left\lfloorh\right\rfloor \cdot dX.v\right)}^{2}\\
t_2 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_3 := \mathsf{max}\left(t\_1, t\_0 + {t\_2}^{2}\right)\\
t_4 := \left|\left\lfloorh\right\rfloor \cdot \left(\left\lfloorw\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_3}{t\_4} > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{{\left(\mathsf{max}\left(t\_1, t\_0 + dY.v \cdot \left(\left\lfloorh\right\rfloor \cdot t\_2\right)\right)\right)}^{0.5}}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{{t\_3}^{0.5}}\\
\end{array}
\end{array}
\end{array}
Initial program 79.8%
Simplified79.8%
Applied egg-rr79.8%
metadata-evalN/A
pow-powN/A
unpow2N/A
associate-*r*N/A
unpow-prod-downN/A
unpow1N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
floor-lowering-floor.f3279.8%
Applied egg-rr79.8%
Final simplification79.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (+ (pow (* (floor w) dY.u) 2.0) (pow (* (floor h) dY.v) 2.0)))
(t_1 (pow (* (floor w) dX.u) 2.0))
(t_2 (fmax (+ t_1 (pow (* (floor h) dX.v) 2.0)) t_0))
(t_3
(fabs (* (floor h) (* (floor w) (- (* dX.u dY.v) (* dX.v dY.u)))))))
(log2
(if (> (/ t_2 t_3) (floor maxAniso))
(/ (pow t_2 0.5) (floor maxAniso))
(/ t_3 (sqrt (fmax t_1 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) + powf((floorf(h) * dY_46_v), 2.0f);
float t_1 = powf((floorf(w) * dX_46_u), 2.0f);
float t_2 = fmaxf((t_1 + powf((floorf(h) * dX_46_v), 2.0f)), t_0);
float t_3 = fabsf((floorf(h) * (floorf(w) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u)))));
float tmp;
if ((t_2 / t_3) > floorf(maxAniso)) {
tmp = powf(t_2, 0.5f) / floorf(maxAniso);
} else {
tmp = t_3 / sqrtf(fmaxf(t_1, 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((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (Float32(floor(h) * dY_46_v) ^ Float32(2.0))) t_1 = Float32(floor(w) * dX_46_u) ^ Float32(2.0) t_2 = (Float32(t_1 + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) != Float32(t_1 + (Float32(floor(h) * dX_46_v) ^ Float32(2.0)))) ? t_0 : ((t_0 != t_0) ? Float32(t_1 + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) : max(Float32(t_1 + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))), t_0)) t_3 = abs(Float32(floor(h) * Float32(floor(w) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u))))) tmp = Float32(0.0) if (Float32(t_2 / t_3) > floor(maxAniso)) tmp = Float32((t_2 ^ Float32(0.5)) / floor(maxAniso)); else tmp = Float32(t_3 / sqrt(((t_1 != t_1) ? t_0 : ((t_0 != t_0) ? t_1 : max(t_1, t_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(w) * dY_46_u) ^ single(2.0)) + ((floor(h) * dY_46_v) ^ single(2.0)); t_1 = (floor(w) * dX_46_u) ^ single(2.0); t_2 = max((t_1 + ((floor(h) * dX_46_v) ^ single(2.0))), t_0); t_3 = abs((floor(h) * (floor(w) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u))))); tmp = single(0.0); if ((t_2 / t_3) > floor(maxAniso)) tmp = (t_2 ^ single(0.5)) / floor(maxAniso); else tmp = t_3 / sqrt(max(t_1, t_0)); end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {\left(\left\lfloorh\right\rfloor \cdot dY.v\right)}^{2}\\
t_1 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2}\\
t_2 := \mathsf{max}\left(t\_1 + {\left(\left\lfloorh\right\rfloor \cdot dX.v\right)}^{2}, t\_0\right)\\
t_3 := \left|\left\lfloorh\right\rfloor \cdot \left(\left\lfloorw\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_3} > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{{t\_2}^{0.5}}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 79.8%
Simplified79.8%
Applied egg-rr79.8%
unpow1/2N/A
sqrt-lowering-sqrt.f32N/A
unpow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
fmax-lowering-fmax.f32N/A
Applied egg-rr79.8%
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.f3279.8%
Simplified79.8%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
fmax-lowering-fmax.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
unpow2N/A
associate-*r*N/A
+-lowering-+.f32N/A
Applied egg-rr79.8%
Final simplification79.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow t_0 2.0) (pow t_2 2.0)))
(t_4 (* (floor h) dX.v))
(t_5 (pow t_4 2.0))
(t_6 (- (* dX.u dY.v) (* dX.v dY.u))))
(log2
(if (>
(/ (/ (/ (fmax t_5 t_3) (floor w)) (floor h)) (fabs t_6))
(floor maxAniso))
(/
(sqrt
(fmax
(+ (* (floor w) (* dX.u t_1)) (* (floor h) (* dX.v t_4)))
(+ (* (floor w) (* dY.u t_0)) (* (floor h) (* dY.v t_2)))))
(floor maxAniso))
(/
(fabs (* (floor w) (* (floor h) t_6)))
(sqrt (fmax (+ (pow t_1 2.0) t_5) 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(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf(t_0, 2.0f) + powf(t_2, 2.0f);
float t_4 = floorf(h) * dX_46_v;
float t_5 = powf(t_4, 2.0f);
float t_6 = (dX_46_u * dY_46_v) - (dX_46_v * dY_46_u);
float tmp;
if ((((fmaxf(t_5, t_3) / floorf(w)) / floorf(h)) / fabsf(t_6)) > floorf(maxAniso)) {
tmp = sqrtf(fmaxf(((floorf(w) * (dX_46_u * t_1)) + (floorf(h) * (dX_46_v * t_4))), ((floorf(w) * (dY_46_u * t_0)) + (floorf(h) * (dY_46_v * t_2))))) / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (floorf(h) * t_6))) / sqrtf(fmaxf((powf(t_1, 2.0f) + t_5), 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(w) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = Float32(floor(h) * dX_46_v) t_5 = t_4 ^ Float32(2.0) t_6 = Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u)) tmp = Float32(0.0) if (Float32(Float32(Float32(((t_5 != t_5) ? t_3 : ((t_3 != t_3) ? t_5 : max(t_5, t_3))) / floor(w)) / floor(h)) / abs(t_6)) > floor(maxAniso)) tmp = Float32(sqrt(((Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_4))) != Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_4)))) ? Float32(Float32(floor(w) * Float32(dY_46_u * t_0)) + Float32(floor(h) * Float32(dY_46_v * t_2))) : ((Float32(Float32(floor(w) * Float32(dY_46_u * t_0)) + Float32(floor(h) * Float32(dY_46_v * t_2))) != Float32(Float32(floor(w) * Float32(dY_46_u * t_0)) + Float32(floor(h) * Float32(dY_46_v * t_2)))) ? Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_4))) : max(Float32(Float32(floor(w) * Float32(dX_46_u * t_1)) + Float32(floor(h) * Float32(dX_46_v * t_4))), Float32(Float32(floor(w) * Float32(dY_46_u * t_0)) + Float32(floor(h) * Float32(dY_46_v * t_2))))))) / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(floor(h) * t_6))) / sqrt(((Float32((t_1 ^ Float32(2.0)) + t_5) != Float32((t_1 ^ Float32(2.0)) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32((t_1 ^ Float32(2.0)) + t_5) : max(Float32((t_1 ^ Float32(2.0)) + t_5), 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(w) * dY_46_u; t_1 = floor(w) * dX_46_u; t_2 = floor(h) * dY_46_v; t_3 = (t_0 ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = floor(h) * dX_46_v; t_5 = t_4 ^ single(2.0); t_6 = (dX_46_u * dY_46_v) - (dX_46_v * dY_46_u); tmp = single(0.0); if ((((max(t_5, t_3) / floor(w)) / floor(h)) / abs(t_6)) > floor(maxAniso)) tmp = sqrt(max(((floor(w) * (dX_46_u * t_1)) + (floor(h) * (dX_46_v * t_4))), ((floor(w) * (dY_46_u * t_0)) + (floor(h) * (dY_46_v * t_2))))) / floor(maxAniso); else tmp = abs((floor(w) * (floor(h) * t_6))) / sqrt(max(((t_1 ^ single(2.0)) + t_5), t_3)); end tmp_2 = log2(tmp); end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_1 := \left\lfloorw\right\rfloor \cdot dX.u\\
t_2 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_3 := {t\_0}^{2} + {t\_2}^{2}\\
t_4 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_5 := {t\_4}^{2}\\
t_6 := dX.u \cdot dY.v - dX.v \cdot dY.u\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\frac{\frac{\mathsf{max}\left(t\_5, t\_3\right)}{\left\lfloorw\right\rfloor}}{\left\lfloorh\right\rfloor}}{\left|t\_6\right|} > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\frac{\sqrt{\mathsf{max}\left(\left\lfloorw\right\rfloor \cdot \left(dX.u \cdot t\_1\right) + \left\lfloorh\right\rfloor \cdot \left(dX.v \cdot t\_4\right), \left\lfloorw\right\rfloor \cdot \left(dY.u \cdot t\_0\right) + \left\lfloorh\right\rfloor \cdot \left(dY.v \cdot t\_2\right)\right)}}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloorw\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot t\_6\right)\right|}{\sqrt{\mathsf{max}\left({t\_1}^{2} + t\_5, t\_3\right)}}\\
\end{array}
\end{array}
\end{array}
Initial program 79.8%
Simplified79.8%
fabs-mulN/A
associate-/r*N/A
/-lowering-/.f32N/A
Applied egg-rr79.8%
Applied egg-rr79.8%
Taylor expanded in dX.u around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3271.2%
Simplified71.2%
Applied egg-rr71.3%
Final simplification71.3%
herbie shell --seed 2024155
(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)))))))))