
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) (* t_1 t_1))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_1) (* t_0 t_2))))
(t_7 (/ t_4 t_6))
(t_8 (> t_7 (floor maxAniso)))
(t_9 (if t_8 (/ t_5 (floor maxAniso)) (/ t_6 t_5)))
(t_10 (if t_8 (floor maxAniso) t_7)))
(if (< t_9 1.0) (fmax 1.0 (* t_10 t_9)) t_10)))
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(h) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_1) - (t_0 * t_2)));
float t_7 = t_4 / t_6;
int t_8 = t_7 > floorf(maxAniso);
float tmp;
if (t_8) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
float t_9 = tmp;
float tmp_1;
if (t_8) {
tmp_1 = floorf(maxAniso);
} else {
tmp_1 = t_7;
}
float t_10 = tmp_1;
float tmp_2;
if (t_9 < 1.0f) {
tmp_2 = fmaxf(1.0f, (t_10 * t_9));
} else {
tmp_2 = t_10;
}
return tmp_2;
}
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(h) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) 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_2 * t_2) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? 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_2 * t_2) + Float32(t_1 * t_1)))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_1) - Float32(t_0 * t_2))) t_7 = Float32(t_4 / t_6) t_8 = t_7 > floor(maxAniso) tmp = Float32(0.0) if (t_8) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end t_9 = tmp tmp_1 = Float32(0.0) if (t_8) tmp_1 = floor(maxAniso); else tmp_1 = t_7; end t_10 = tmp_1 tmp_2 = Float32(0.0) if (t_9 < Float32(1.0)) tmp_2 = (Float32(1.0) != Float32(1.0)) ? Float32(t_10 * t_9) : ((Float32(t_10 * t_9) != Float32(t_10 * t_9)) ? Float32(1.0) : max(Float32(1.0), Float32(t_10 * t_9))); else tmp_2 = t_10; end return tmp_2 end
function tmp_4 = 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(h) * dY_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_1) - (t_0 * t_2))); t_7 = t_4 / t_6; t_8 = t_7 > floor(maxAniso); tmp = single(0.0); if (t_8) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end t_9 = tmp; tmp_2 = single(0.0); if (t_8) tmp_2 = floor(maxAniso); else tmp_2 = t_7; end t_10 = tmp_2; tmp_3 = single(0.0); if (t_9 < single(1.0)) tmp_3 = max(single(1.0), (t_10 * t_9)); else tmp_3 = t_10; end tmp_4 = tmp_3; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_2 := \left\lfloorw\right\rfloor \cdot dY.u\\
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_2 \cdot t_2 + t_1 \cdot t_1\right)\\
t_5 := \sqrt{t_4}\\
t_6 := \left|t_3 \cdot t_1 - t_0 \cdot t_2\right|\\
t_7 := \frac{t_4}{t_6}\\
t_8 := t_7 > \left\lfloormaxAniso\right\rfloor\\
t_9 := \begin{array}{l}
\mathbf{if}\;t_8:\\
\;\;\;\;\frac{t_5}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t_6}{t_5}\\
\end{array}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t_8:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_7\\
\end{array}\\
\mathbf{if}\;t_9 < 1:\\
\;\;\;\;\mathsf{max}\left(1, t_10 \cdot t_9\right)\\
\mathbf{else}:\\
\;\;\;\;t_10\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 3 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 h) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) (* t_1 t_1))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_1) (* t_0 t_2))))
(t_7 (/ t_4 t_6))
(t_8 (> t_7 (floor maxAniso)))
(t_9 (if t_8 (/ t_5 (floor maxAniso)) (/ t_6 t_5)))
(t_10 (if t_8 (floor maxAniso) t_7)))
(if (< t_9 1.0) (fmax 1.0 (* t_10 t_9)) t_10)))
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(h) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_1) - (t_0 * t_2)));
float t_7 = t_4 / t_6;
int t_8 = t_7 > floorf(maxAniso);
float tmp;
if (t_8) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
float t_9 = tmp;
float tmp_1;
if (t_8) {
tmp_1 = floorf(maxAniso);
} else {
tmp_1 = t_7;
}
float t_10 = tmp_1;
float tmp_2;
if (t_9 < 1.0f) {
tmp_2 = fmaxf(1.0f, (t_10 * t_9));
} else {
tmp_2 = t_10;
}
return tmp_2;
}
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(h) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) 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_2 * t_2) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? 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_2 * t_2) + Float32(t_1 * t_1)))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_1) - Float32(t_0 * t_2))) t_7 = Float32(t_4 / t_6) t_8 = t_7 > floor(maxAniso) tmp = Float32(0.0) if (t_8) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end t_9 = tmp tmp_1 = Float32(0.0) if (t_8) tmp_1 = floor(maxAniso); else tmp_1 = t_7; end t_10 = tmp_1 tmp_2 = Float32(0.0) if (t_9 < Float32(1.0)) tmp_2 = (Float32(1.0) != Float32(1.0)) ? Float32(t_10 * t_9) : ((Float32(t_10 * t_9) != Float32(t_10 * t_9)) ? Float32(1.0) : max(Float32(1.0), Float32(t_10 * t_9))); else tmp_2 = t_10; end return tmp_2 end
function tmp_4 = 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(h) * dY_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(w) * dX_46_u; t_4 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_1) - (t_0 * t_2))); t_7 = t_4 / t_6; t_8 = t_7 > floor(maxAniso); tmp = single(0.0); if (t_8) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end t_9 = tmp; tmp_2 = single(0.0); if (t_8) tmp_2 = floor(maxAniso); else tmp_2 = t_7; end t_10 = tmp_2; tmp_3 = single(0.0); if (t_9 < single(1.0)) tmp_3 = max(single(1.0), (t_10 * t_9)); else tmp_3 = t_10; end tmp_4 = tmp_3; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_2 := \left\lfloorw\right\rfloor \cdot dY.u\\
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_2 \cdot t_2 + t_1 \cdot t_1\right)\\
t_5 := \sqrt{t_4}\\
t_6 := \left|t_3 \cdot t_1 - t_0 \cdot t_2\right|\\
t_7 := \frac{t_4}{t_6}\\
t_8 := t_7 > \left\lfloormaxAniso\right\rfloor\\
t_9 := \begin{array}{l}
\mathbf{if}\;t_8:\\
\;\;\;\;\frac{t_5}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t_6}{t_5}\\
\end{array}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t_8:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_7\\
\end{array}\\
\mathbf{if}\;t_9 < 1:\\
\;\;\;\;\mathsf{max}\left(1, t_10 \cdot t_9\right)\\
\mathbf{else}:\\
\;\;\;\;t_10\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (* (floor h) (- (* dX.u dY.v) (* dX.v dY.u)))))
(t_1 (fabs t_0))
(t_2
(fmax
(fma
(floor w)
(* (floor w) (* dX.u dX.u))
(* (floor h) (* (floor h) (* dX.v dX.v))))
(fma
(floor w)
(* (floor w) (* dY.u dY.u))
(* (floor h) (* (floor h) (* dY.v dY.v))))))
(t_3 (/ t_2 t_1))
(t_4 (> t_3 (floor maxAniso)))
(t_5 (sqrt t_2))
(t_6 (if t_4 (/ t_5 (floor maxAniso)) (/ t_1 t_5))))
(if (< t_6 1.0)
(fmax
1.0
(*
t_6
(if t_4
(floor maxAniso)
(*
(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)))
(/ 1.0 t_0)))))
(if t_4 (floor maxAniso) 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) * (floorf(h) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u)));
float t_1 = fabsf(t_0);
float t_2 = fmaxf(fmaf(floorf(w), (floorf(w) * (dX_46_u * dX_46_u)), (floorf(h) * (floorf(h) * (dX_46_v * dX_46_v)))), fmaf(floorf(w), (floorf(w) * (dY_46_u * dY_46_u)), (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v)))));
float t_3 = t_2 / t_1;
int t_4 = t_3 > floorf(maxAniso);
float t_5 = sqrtf(t_2);
float tmp;
if (t_4) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_1 / t_5;
}
float t_6 = tmp;
float tmp_2;
if (t_6 < 1.0f) {
float tmp_3;
if (t_4) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = 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))) * (1.0f / t_0);
}
tmp_2 = fmaxf(1.0f, (t_6 * tmp_3));
} else if (t_4) {
tmp_2 = floorf(maxAniso);
} else {
tmp_2 = t_3;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * Float32(floor(h) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u)))) t_1 = abs(t_0) t_2 = (fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) : max(fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))), fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))))) t_3 = Float32(t_2 / t_1) t_4 = t_3 > floor(maxAniso) t_5 = sqrt(t_2) tmp = Float32(0.0) if (t_4) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_1 / t_5); end t_6 = tmp tmp_2 = Float32(0.0) if (t_6 < Float32(1.0)) tmp_3 = Float32(0.0) if (t_4) tmp_3 = floor(maxAniso); else tmp_3 = Float32(((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)))))) * Float32(Float32(1.0) / t_0)); end tmp_2 = (Float32(1.0) != Float32(1.0)) ? Float32(t_6 * tmp_3) : ((Float32(t_6 * tmp_3) != Float32(t_6 * tmp_3)) ? Float32(1.0) : max(Float32(1.0), Float32(t_6 * tmp_3))); elseif (t_4) tmp_2 = floor(maxAniso); else tmp_2 = t_3; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorw\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\right)\\
t_1 := \left|t_0\right|\\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dX.u \cdot dX.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dX.v \cdot dX.v\right)\right)\right), \mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dY.u \cdot dY.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)\right)\\
t_3 := \frac{t_2}{t_1}\\
t_4 := t_3 > \left\lfloormaxAniso\right\rfloor\\
t_5 := \sqrt{t_2}\\
t_6 := \begin{array}{l}
\mathbf{if}\;t_4:\\
\;\;\;\;\frac{t_5}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t_1}{t_5}\\
\end{array}\\
\mathbf{if}\;t_6 < 1:\\
\;\;\;\;\mathsf{max}\left(1, t_6 \cdot \begin{array}{l}
\mathbf{if}\;t_4:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;\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) \cdot \frac{1}{t_0}\\
\end{array}\right)\\
\mathbf{elif}\;t_4:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_3\\
\end{array}
\end{array}
Initial program 97.2%
Simplified97.2%
Applied egg-rr97.2%
Final simplification97.2%
(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
(fmax
(fma
(floor w)
(* (floor w) (* dX.u dX.u))
(* (floor h) (* (floor h) (* dX.v dX.v))))
(fma
(floor w)
(* (floor w) (* dY.u dY.u))
(* (floor h) (* (floor h) (* dY.v dY.v))))))
(t_2 (* (floor w) t_0))
(t_3 (fabs t_2))
(t_4 (/ t_1 t_3))
(t_5
(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_6 (* t_5 (/ 1.0 t_2)))
(t_7 (> t_6 (floor maxAniso)))
(t_8 (> t_4 (floor maxAniso)))
(t_9 (sqrt t_1))
(t_10 (/ t_9 (floor maxAniso))))
(if (< (if t_8 t_10 (/ t_3 t_9)) 1.0)
(fmax
1.0
(*
(if t_7 t_10 (/ (floor w) (/ (sqrt t_5) t_0)))
(if t_7 (floor maxAniso) t_6)))
(if t_8 (floor maxAniso) 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_u * dY_46_v) - (dX_46_v * dY_46_u));
float t_1 = fmaxf(fmaf(floorf(w), (floorf(w) * (dX_46_u * dX_46_u)), (floorf(h) * (floorf(h) * (dX_46_v * dX_46_v)))), fmaf(floorf(w), (floorf(w) * (dY_46_u * dY_46_u)), (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v)))));
float t_2 = floorf(w) * t_0;
float t_3 = fabsf(t_2);
float t_4 = t_1 / t_3;
float t_5 = 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_6 = t_5 * (1.0f / t_2);
int t_7 = t_6 > floorf(maxAniso);
int t_8 = t_4 > floorf(maxAniso);
float t_9 = sqrtf(t_1);
float t_10 = t_9 / floorf(maxAniso);
float tmp;
if (t_8) {
tmp = t_10;
} else {
tmp = t_3 / t_9;
}
float tmp_3;
if (tmp < 1.0f) {
float tmp_4;
if (t_7) {
tmp_4 = t_10;
} else {
tmp_4 = floorf(w) / (sqrtf(t_5) / t_0);
}
float tmp_5;
if (t_7) {
tmp_5 = floorf(maxAniso);
} else {
tmp_5 = t_6;
}
tmp_3 = fmaxf(1.0f, (tmp_4 * tmp_5));
} else if (t_8) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_4;
}
return tmp_3;
}
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 = (fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) : max(fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))), fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))))) t_2 = Float32(floor(w) * t_0) t_3 = abs(t_2) t_4 = Float32(t_1 / t_3) t_5 = (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_6 = Float32(t_5 * Float32(Float32(1.0) / t_2)) t_7 = t_6 > floor(maxAniso) t_8 = t_4 > floor(maxAniso) t_9 = sqrt(t_1) t_10 = Float32(t_9 / floor(maxAniso)) tmp = Float32(0.0) if (t_8) tmp = t_10; else tmp = Float32(t_3 / t_9); end tmp_3 = Float32(0.0) if (tmp < Float32(1.0)) tmp_4 = Float32(0.0) if (t_7) tmp_4 = t_10; else tmp_4 = Float32(floor(w) / Float32(sqrt(t_5) / t_0)); end tmp_5 = Float32(0.0) if (t_7) tmp_5 = floor(maxAniso); else tmp_5 = t_6; end tmp_3 = (Float32(1.0) != Float32(1.0)) ? Float32(tmp_4 * tmp_5) : ((Float32(tmp_4 * tmp_5) != Float32(tmp_4 * tmp_5)) ? Float32(1.0) : max(Float32(1.0), Float32(tmp_4 * tmp_5))); elseif (t_8) tmp_3 = floor(maxAniso); else tmp_3 = t_4; end return tmp_3 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 := \mathsf{max}\left(\mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dX.u \cdot dX.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dX.v \cdot dX.v\right)\right)\right), \mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dY.u \cdot dY.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)\right)\\
t_2 := \left\lfloorw\right\rfloor \cdot t_0\\
t_3 := \left|t_2\right|\\
t_4 := \frac{t_1}{t_3}\\
t_5 := \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_6 := t_5 \cdot \frac{1}{t_2}\\
t_7 := t_6 > \left\lfloormaxAniso\right\rfloor\\
t_8 := t_4 > \left\lfloormaxAniso\right\rfloor\\
t_9 := \sqrt{t_1}\\
t_10 := \frac{t_9}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t_8:\\
\;\;\;\;t_10\\
\mathbf{else}:\\
\;\;\;\;\frac{t_3}{t_9}\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t_7:\\
\;\;\;\;t_10\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloorw\right\rfloor}{\frac{\sqrt{t_5}}{t_0}}\\
\end{array} \cdot \begin{array}{l}
\mathbf{if}\;t_7:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_6\\
\end{array}\right)\\
\mathbf{elif}\;t_8:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_4\\
\end{array}
\end{array}
Initial program 97.2%
Simplified97.2%
Applied egg-rr97.2%
Applied egg-rr96.6%
Applied egg-rr96.6%
Applied egg-rr97.2%
Final simplification97.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (* (floor h) (- (* dX.u dY.v) (* dX.v dY.u)))))
(t_1 (fabs t_0))
(t_2
(*
(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)))
(/ 1.0 t_0)))
(t_3 (> t_2 (floor maxAniso)))
(t_4
(fmax
(fma
(floor w)
(* (floor w) (* dX.u dX.u))
(* (floor h) (* (floor h) (* dX.v dX.v))))
(fma
(floor w)
(* (floor w) (* dY.u dY.u))
(* (floor h) (* (floor h) (* dY.v dY.v))))))
(t_5 (/ t_4 t_1))
(t_6 (sqrt t_4))
(t_7 (if t_3 (/ t_6 (floor maxAniso)) (/ t_1 t_6))))
(if (< t_7 1.0)
(fmax 1.0 (* (if t_3 (floor maxAniso) t_2) t_7))
(if (> t_5 (floor maxAniso)) (floor maxAniso) 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(w) * (floorf(h) * ((dX_46_u * dY_46_v) - (dX_46_v * dY_46_u)));
float t_1 = fabsf(t_0);
float t_2 = 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))) * (1.0f / t_0);
int t_3 = t_2 > floorf(maxAniso);
float t_4 = fmaxf(fmaf(floorf(w), (floorf(w) * (dX_46_u * dX_46_u)), (floorf(h) * (floorf(h) * (dX_46_v * dX_46_v)))), fmaf(floorf(w), (floorf(w) * (dY_46_u * dY_46_u)), (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v)))));
float t_5 = t_4 / t_1;
float t_6 = sqrtf(t_4);
float tmp;
if (t_3) {
tmp = t_6 / floorf(maxAniso);
} else {
tmp = t_1 / t_6;
}
float t_7 = tmp;
float tmp_2;
if (t_7 < 1.0f) {
float tmp_3;
if (t_3) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_2;
}
tmp_2 = fmaxf(1.0f, (tmp_3 * t_7));
} else if (t_5 > floorf(maxAniso)) {
tmp_2 = floorf(maxAniso);
} else {
tmp_2 = t_5;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * Float32(floor(h) * Float32(Float32(dX_46_u * dY_46_v) - Float32(dX_46_v * dY_46_u)))) t_1 = abs(t_0) t_2 = Float32(((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)))))) * Float32(Float32(1.0) / t_0)) t_3 = t_2 > floor(maxAniso) t_4 = (fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))) : max(fma(floor(w), Float32(floor(w) * Float32(dX_46_u * dX_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dX_46_v * dX_46_v)))), fma(floor(w), Float32(floor(w) * Float32(dY_46_u * dY_46_u)), Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))))) t_5 = Float32(t_4 / t_1) t_6 = sqrt(t_4) tmp = Float32(0.0) if (t_3) tmp = Float32(t_6 / floor(maxAniso)); else tmp = Float32(t_1 / t_6); end t_7 = tmp tmp_2 = Float32(0.0) if (t_7 < Float32(1.0)) tmp_3 = Float32(0.0) if (t_3) tmp_3 = floor(maxAniso); else tmp_3 = t_2; end tmp_2 = (Float32(1.0) != Float32(1.0)) ? Float32(tmp_3 * t_7) : ((Float32(tmp_3 * t_7) != Float32(tmp_3 * t_7)) ? Float32(1.0) : max(Float32(1.0), Float32(tmp_3 * t_7))); elseif (t_5 > floor(maxAniso)) tmp_2 = floor(maxAniso); else tmp_2 = t_5; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorw\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dX.u \cdot dY.v - dX.v \cdot dY.u\right)\right)\\
t_1 := \left|t_0\right|\\
t_2 := \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) \cdot \frac{1}{t_0}\\
t_3 := t_2 > \left\lfloormaxAniso\right\rfloor\\
t_4 := \mathsf{max}\left(\mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dX.u \cdot dX.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dX.v \cdot dX.v\right)\right)\right), \mathsf{fma}\left(\left\lfloorw\right\rfloor, \left\lfloorw\right\rfloor \cdot \left(dY.u \cdot dY.u\right), \left\lfloorh\right\rfloor \cdot \left(\left\lfloorh\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)\right)\\
t_5 := \frac{t_4}{t_1}\\
t_6 := \sqrt{t_4}\\
t_7 := \begin{array}{l}
\mathbf{if}\;t_3:\\
\;\;\;\;\frac{t_6}{\left\lfloormaxAniso\right\rfloor}\\
\mathbf{else}:\\
\;\;\;\;\frac{t_1}{t_6}\\
\end{array}\\
\mathbf{if}\;t_7 < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t_3:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_2\\
\end{array} \cdot t_7\right)\\
\mathbf{elif}\;t_5 > \left\lfloormaxAniso\right\rfloor:\\
\;\;\;\;\left\lfloormaxAniso\right\rfloor\\
\mathbf{else}:\\
\;\;\;\;t_5\\
\end{array}
\end{array}
Initial program 97.2%
Simplified97.2%
Applied egg-rr97.2%
Applied egg-rr96.6%
Applied egg-rr96.6%
Applied egg-rr73.4%
Final simplification73.4%
herbie shell --seed 2023301
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (ratio of anisotropy)"
: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))
(if (< (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))))))) 1.0) (fmax 1.0 (* (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)) (floor maxAniso) (/ (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)))))) (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))))))))) (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)) (floor maxAniso) (/ (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))))))))