
(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 = fmax(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}
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}
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 = fmax(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}
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}
(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 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_3 (* (floor h) dX.v))
(t_4 (* (* (- (floor w)) (floor w)) (* dY.u dY.u)))
(t_5 (+ (* t_1 t_1) (* t_3 t_3)))
(t_6
(fmax
t_5
(/
(- (pow t_2 3.0) (pow t_4 3.0))
(fma t_2 t_2 (fma t_4 t_4 (* t_2 t_4))))))
(t_7 (sqrt t_6))
(t_8 (* (floor h) dY.v))
(t_9 (+ (* t_0 t_0) (* t_8 t_8)))
(t_10 (fabs (- (* t_1 t_8) (* t_3 t_0))))
(t_11 (fmax t_5 t_9))
(t_12 (sqrt t_11))
(t_13
(fmax
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (* dX.v (floor h)) dX.v) (floor h)))
t_9))
(t_14 (sqrt t_13)))
(if (<=
(log2
(if (> (/ t_11 t_10) (floor maxAniso))
(/ t_12 (floor maxAniso))
(/ t_10 t_12)))
100.0)
(log2
(if (> (/ t_13 t_10) (floor maxAniso))
(/ t_14 (floor maxAniso))
(/ t_10 t_14)))
(log2
(if (> (/ t_6 t_10) (floor maxAniso))
(/ t_7 (floor maxAniso))
(/ t_10 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 = floorf(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_3 = floorf(h) * dX_46_v;
float t_4 = (-floorf(w) * floorf(w)) * (dY_46_u * dY_46_u);
float t_5 = (t_1 * t_1) + (t_3 * t_3);
float t_6 = fmaxf(t_5, ((powf(t_2, 3.0f) - powf(t_4, 3.0f)) / fmaf(t_2, t_2, fmaf(t_4, t_4, (t_2 * t_4)))));
float t_7 = sqrtf(t_6);
float t_8 = floorf(h) * dY_46_v;
float t_9 = (t_0 * t_0) + (t_8 * t_8);
float t_10 = fabsf(((t_1 * t_8) - (t_3 * t_0)));
float t_11 = fmaxf(t_5, t_9);
float t_12 = sqrtf(t_11);
float t_13 = fmaxf(fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), t_9);
float t_14 = sqrtf(t_13);
float tmp;
if ((t_11 / t_10) > floorf(maxAniso)) {
tmp = t_12 / floorf(maxAniso);
} else {
tmp = t_10 / t_12;
}
float tmp_2;
if (log2f(tmp) <= 100.0f) {
float tmp_3;
if ((t_13 / t_10) > floorf(maxAniso)) {
tmp_3 = t_14 / floorf(maxAniso);
} else {
tmp_3 = t_10 / t_14;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_6 / t_10) > floorf(maxAniso)) {
tmp_4 = t_7 / floorf(maxAniso);
} else {
tmp_4 = t_10 / t_7;
}
tmp_2 = log2f(tmp_4);
}
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) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(Float32(Float32(-floor(w)) * floor(w)) * Float32(dY_46_u * dY_46_u)) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_3 * t_3)) t_6 = fmax(t_5, Float32(Float32((t_2 ^ Float32(3.0)) - (t_4 ^ Float32(3.0))) / fma(t_2, t_2, fma(t_4, t_4, Float32(t_2 * t_4))))) t_7 = sqrt(t_6) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(Float32(t_0 * t_0) + Float32(t_8 * t_8)) t_10 = abs(Float32(Float32(t_1 * t_8) - Float32(t_3 * t_0))) t_11 = fmax(t_5, t_9) t_12 = sqrt(t_11) t_13 = fmax(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), t_9) t_14 = sqrt(t_13) tmp = Float32(0.0) if (Float32(t_11 / t_10) > floor(maxAniso)) tmp = Float32(t_12 / floor(maxAniso)); else tmp = Float32(t_10 / t_12); end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(100.0)) tmp_3 = Float32(0.0) if (Float32(t_13 / t_10) > floor(maxAniso)) tmp_3 = Float32(t_14 / floor(maxAniso)); else tmp_3 = Float32(t_10 / t_14); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_6 / t_10) > floor(maxAniso)) tmp_4 = Float32(t_7 / floor(maxAniso)); else tmp_4 = Float32(t_10 / t_7); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left(\left(-\left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dY.u \cdot dY.u\right)\\
t_5 := t\_1 \cdot t\_1 + t\_3 \cdot t\_3\\
t_6 := \mathsf{max}\left(t\_5, \frac{{t\_2}^{3} - {t\_4}^{3}}{\mathsf{fma}\left(t\_2, t\_2, \mathsf{fma}\left(t\_4, t\_4, t\_2 \cdot t\_4\right)\right)}\right)\\
t_7 := \sqrt{t\_6}\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_0 \cdot t\_0 + t\_8 \cdot t\_8\\
t_10 := \left|t\_1 \cdot t\_8 - t\_3 \cdot t\_0\right|\\
t_11 := \mathsf{max}\left(t\_5, t\_9\right)\\
t_12 := \sqrt{t\_11}\\
t_13 := \mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_9\right)\\
t_14 := \sqrt{t\_13}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_11}{t\_10} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_12}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_10}{t\_12}\\
\end{array} \leq 100:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_13}{t\_10} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_14}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_10}{t\_14}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_6}{t\_10} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_7}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_10}{t\_7}\\
\end{array}\\
\end{array}
if (log2.f32 (if.f32 (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) < 100Initial program 76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
if 100 < (log2.f32 (if.f32 (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) Initial program 76.7%
Applied rewrites59.5%
Applied rewrites48.4%
Applied rewrites55.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4 (* (* t_1 dY.v) (floor h)))
(t_5 (* (floor h) dX.v))
(t_6 (* (- t_0) (* dY.u dY.u)))
(t_7 (fabs (* -1.0 (* dX.v (* dY.u (* (floor h) (floor w)))))))
(t_8 (+ (* t_3 t_3) (* t_5 t_5)))
(t_9
(fmax
t_8
(/
(- (pow t_4 3.0) (pow t_6 3.0))
(+ (pow t_1 4.0) (fma t_6 t_6 (* t_4 t_6))))))
(t_10 (sqrt t_9))
(t_11 (* (floor h) dY.v))
(t_12 (+ (* t_2 t_2) (* t_11 t_11)))
(t_13 (fabs (- (* t_3 t_11) (* t_5 t_2))))
(t_14 (fmax t_8 t_12))
(t_15 (sqrt t_14))
(t_16
(fmax
(fma
(* t_0 dX.u)
dX.u
(* (* (* dX.v (floor h)) dX.v) (floor h)))
t_12))
(t_17 (sqrt t_16)))
(if (<=
(log2
(if (> (/ t_14 t_13) (floor maxAniso))
(/ t_15 (floor maxAniso))
(/ t_13 t_15)))
100.0)
(log2
(if (> (/ t_16 t_13) (floor maxAniso))
(/ t_17 (floor maxAniso))
(/ t_13 t_17)))
(log2
(if (> (/ t_9 t_7) (floor maxAniso))
(/ t_10 (floor maxAniso))
(/ t_7 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(w) * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_1 * dY_46_v) * floorf(h);
float t_5 = floorf(h) * dX_46_v;
float t_6 = -t_0 * (dY_46_u * dY_46_u);
float t_7 = fabsf((-1.0f * (dX_46_v * (dY_46_u * (floorf(h) * floorf(w))))));
float t_8 = (t_3 * t_3) + (t_5 * t_5);
float t_9 = fmaxf(t_8, ((powf(t_4, 3.0f) - powf(t_6, 3.0f)) / (powf(t_1, 4.0f) + fmaf(t_6, t_6, (t_4 * t_6)))));
float t_10 = sqrtf(t_9);
float t_11 = floorf(h) * dY_46_v;
float t_12 = (t_2 * t_2) + (t_11 * t_11);
float t_13 = fabsf(((t_3 * t_11) - (t_5 * t_2)));
float t_14 = fmaxf(t_8, t_12);
float t_15 = sqrtf(t_14);
float t_16 = fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), t_12);
float t_17 = sqrtf(t_16);
float tmp;
if ((t_14 / t_13) > floorf(maxAniso)) {
tmp = t_15 / floorf(maxAniso);
} else {
tmp = t_13 / t_15;
}
float tmp_2;
if (log2f(tmp) <= 100.0f) {
float tmp_3;
if ((t_16 / t_13) > floorf(maxAniso)) {
tmp_3 = t_17 / floorf(maxAniso);
} else {
tmp_3 = t_13 / t_17;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_9 / t_7) > floorf(maxAniso)) {
tmp_4 = t_10 / floorf(maxAniso);
} else {
tmp_4 = t_7 / t_10;
}
tmp_2 = log2f(tmp_4);
}
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) * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_1 * dY_46_v) * floor(h)) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(-t_0) * Float32(dY_46_u * dY_46_u)) t_7 = abs(Float32(Float32(-1.0) * Float32(dX_46_v * Float32(dY_46_u * Float32(floor(h) * floor(w)))))) t_8 = Float32(Float32(t_3 * t_3) + Float32(t_5 * t_5)) t_9 = fmax(t_8, Float32(Float32((t_4 ^ Float32(3.0)) - (t_6 ^ Float32(3.0))) / Float32((t_1 ^ Float32(4.0)) + fma(t_6, t_6, Float32(t_4 * t_6))))) t_10 = sqrt(t_9) t_11 = Float32(floor(h) * dY_46_v) t_12 = Float32(Float32(t_2 * t_2) + Float32(t_11 * t_11)) t_13 = abs(Float32(Float32(t_3 * t_11) - Float32(t_5 * t_2))) t_14 = fmax(t_8, t_12) t_15 = sqrt(t_14) t_16 = fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), t_12) t_17 = sqrt(t_16) tmp = Float32(0.0) if (Float32(t_14 / t_13) > floor(maxAniso)) tmp = Float32(t_15 / floor(maxAniso)); else tmp = Float32(t_13 / t_15); end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(100.0)) tmp_3 = Float32(0.0) if (Float32(t_16 / t_13) > floor(maxAniso)) tmp_3 = Float32(t_17 / floor(maxAniso)); else tmp_3 = Float32(t_13 / t_17); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_9 / t_7) > floor(maxAniso)) tmp_4 = Float32(t_10 / floor(maxAniso)); else tmp_4 = Float32(t_7 / t_10); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left(t\_1 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left(-t\_0\right) \cdot \left(dY.u \cdot dY.u\right)\\
t_7 := \left|-1 \cdot \left(dX.v \cdot \left(dY.u \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)\right|\\
t_8 := t\_3 \cdot t\_3 + t\_5 \cdot t\_5\\
t_9 := \mathsf{max}\left(t\_8, \frac{{t\_4}^{3} - {t\_6}^{3}}{{t\_1}^{4} + \mathsf{fma}\left(t\_6, t\_6, t\_4 \cdot t\_6\right)}\right)\\
t_10 := \sqrt{t\_9}\\
t_11 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_12 := t\_2 \cdot t\_2 + t\_11 \cdot t\_11\\
t_13 := \left|t\_3 \cdot t\_11 - t\_5 \cdot t\_2\right|\\
t_14 := \mathsf{max}\left(t\_8, t\_12\right)\\
t_15 := \sqrt{t\_14}\\
t_16 := \mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_12\right)\\
t_17 := \sqrt{t\_16}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_14}{t\_13} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_15}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_13}{t\_15}\\
\end{array} \leq 100:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_16}{t\_13} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_17}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_13}{t\_17}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_9}{t\_7} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_10}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_7}{t\_10}\\
\end{array}\\
\end{array}
if (log2.f32 (if.f32 (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) < 100Initial program 76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
if 100 < (log2.f32 (if.f32 (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) Initial program 76.7%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Applied rewrites64.2%
Applied rewrites48.2%
Applied rewrites52.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2
(fmax
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(+ (* t_0 t_0) (* t_1 t_1))))
(t_3 (sqrt t_2))
(t_4
(fabs
(- (* (* (floor w) dX.u) t_1) (* (* (floor h) dX.v) t_0)))))
(log2
(if (> (/ t_2 t_4) (floor maxAniso))
(/ t_3 (floor maxAniso))
(/ t_4 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(h) * dY_46_v;
float t_2 = fmaxf(fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((t_0 * t_0) + (t_1 * t_1)));
float t_3 = sqrtf(t_2);
float t_4 = fabsf((((floorf(w) * dX_46_u) * t_1) - ((floorf(h) * dX_46_v) * t_0)));
float tmp;
if ((t_2 / t_4) > floorf(maxAniso)) {
tmp = t_3 / floorf(maxAniso);
} else {
tmp = t_4 / 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(h) * dY_46_v) t_2 = fmax(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(Float32(floor(w) * dX_46_u) * t_1) - Float32(Float32(floor(h) * dX_46_v) * t_0))) tmp = Float32(0.0) if (Float32(t_2 / t_4) > floor(maxAniso)) tmp = Float32(t_3 / floor(maxAniso)); else tmp = Float32(t_4 / t_3); end return log2(tmp) end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot t\_1 - \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot t\_0\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_4} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_3}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_3}\\
\end{array}
\end{array}
Initial program 76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3276.7%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.7%
Applied rewrites76.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (fabs (* (floor h) (- (* t_0 dX.v) (* dY.v t_1)))))
(t_3
(fmax
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_1 dX.u) (floor w)))))
(t_4 (sqrt t_3)))
(log2
(if (> (/ t_3 t_2) (floor maxAniso))
(/ t_4 (floor maxAniso))
(/ t_2 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 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = fabsf((floorf(h) * ((t_0 * dX_46_v) - (dY_46_v * t_1))));
float t_3 = fmaxf(fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_1 * dX_46_u) * floorf(w))));
float t_4 = sqrtf(t_3);
float tmp;
if ((t_3 / t_2) > floorf(maxAniso)) {
tmp = t_4 / floorf(maxAniso);
} else {
tmp = t_2 / 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(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = abs(Float32(floor(h) * Float32(Float32(t_0 * dX_46_v) - Float32(dY_46_v * t_1)))) t_3 = fmax(fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_1 * dX_46_u) * floor(w)))) t_4 = sqrt(t_3) tmp = Float32(0.0) if (Float32(t_3 / t_2) > floor(maxAniso)) tmp = Float32(t_4 / floor(maxAniso)); else tmp = Float32(t_2 / t_4); end return log2(tmp) end
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left|\left\lfloor h\right\rfloor \cdot \left(t\_0 \cdot dX.v - dY.v \cdot t\_1\right)\right|\\
t_3 := \mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_4 := \sqrt{t\_3}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_3}{t\_2} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_4}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (fabs (* (* (* (- dX.v) (floor h)) (floor w)) dY.u)))
(t_1
(fmax
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* (* dY.u (floor w)) dY.u) (floor w)))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))))
(t_2 (sqrt t_1)))
(log2
(if (> (/ t_1 t_0) (floor maxAniso))
(/ t_2 (floor maxAniso))
(/ 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 = fabsf((((-dX_46_v * floorf(h)) * floorf(w)) * dY_46_u));
float t_1 = fmaxf(fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))));
float t_2 = sqrtf(t_1);
float tmp;
if ((t_1 / t_0) > floorf(maxAniso)) {
tmp = t_2 / floorf(maxAniso);
} else {
tmp = 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 = abs(Float32(Float32(Float32(Float32(-dX_46_v) * floor(h)) * floor(w)) * dY_46_u)) t_1 = fmax(fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)))) t_2 = sqrt(t_1) tmp = Float32(0.0) if (Float32(t_1 / t_0) > floor(maxAniso)) tmp = Float32(t_2 / floor(maxAniso)); else tmp = Float32(t_0 / t_2); end return log2(tmp) end
\begin{array}{l}
t_0 := \left|\left(\left(\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right|\\
t_1 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_2 := \sqrt{t\_1}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_1}{t\_0} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_2}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_2}\\
\end{array}
\end{array}
Initial program 76.7%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Applied rewrites75.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lift-neg.f32N/A
distribute-lft-neg-inN/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
distribute-lft-neg-inN/A
Applied rewrites75.8%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lift-neg.f32N/A
distribute-lft-neg-inN/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
distribute-lft-neg-inN/A
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (fabs (* (* (- dX.v) dY.u) (* (floor h) (floor w)))))
(t_1
(fmax
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* (* dY.u (floor w)) dY.u) (floor w)))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))))
(t_2 (sqrt t_1)))
(log2
(if (> (/ t_1 t_0) (floor maxAniso))
(/ t_2 (floor maxAniso))
(/ 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 = fabsf(((-dX_46_v * dY_46_u) * (floorf(h) * floorf(w))));
float t_1 = fmaxf(fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))));
float t_2 = sqrtf(t_1);
float tmp;
if ((t_1 / t_0) > floorf(maxAniso)) {
tmp = t_2 / floorf(maxAniso);
} else {
tmp = 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 = abs(Float32(Float32(Float32(-dX_46_v) * dY_46_u) * Float32(floor(h) * floor(w)))) t_1 = fmax(fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)))) t_2 = sqrt(t_1) tmp = Float32(0.0) if (Float32(t_1 / t_0) > floor(maxAniso)) tmp = Float32(t_2 / floor(maxAniso)); else tmp = Float32(t_0 / t_2); end return log2(tmp) end
\begin{array}{l}
t_0 := \left|\left(\left(-dX.v\right) \cdot dY.u\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right|\\
t_1 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_2 := \sqrt{t\_1}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_1}{t\_0} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_2}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_2}\\
\end{array}
\end{array}
Initial program 76.7%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.8%
Applied rewrites75.8%
Applied rewrites75.8%
herbie shell --seed 2025210
(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)))))))))