
(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 = fmax(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 = fmax(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\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
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\_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\lfloor maxAniso\right\rfloor \\
t_9 := \begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;\left\lfloor maxAniso\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 2 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 = fmax(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 = fmax(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\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
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\_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\lfloor maxAniso\right\rfloor \\
t_9 := \begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;\left\lfloor maxAniso\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 (pow (* dX.u (floor w)) 2.0))
(t_1 (pow (floor h) 2.0))
(t_2 (pow (* dY.u (floor w)) 2.0))
(t_3 (pow (floor w) 2.0))
(t_4 (fma (* t_3 dX.u) dX.u (* (* t_1 dX.v) dX.v)))
(t_5 (pow (* dX.v (floor h)) 2.0))
(t_6 (* (floor w) (floor h)))
(t_7 (fabs (* (fma dY.u dX.v (* (- dY.v) dX.u)) t_6)))
(t_8 (pow (* dY.v (floor h)) 2.0))
(t_9 (* t_3 dY.u))
(t_10 (fmax t_4 (fma t_9 dY.u (* (* t_1 dY.v) dY.v))))
(t_11 (/ (sqrt t_10) (floor maxAniso)))
(t_12 (/ t_10 t_7))
(t_13 (> t_12 (floor maxAniso))))
(if (<
(if t_13 t_11 (* (/ 1.0 (sqrt (fmax (+ t_5 t_0) (+ t_8 t_2)))) t_7))
1.0)
(fmax
1.0
(*
(if t_13 (floor maxAniso) t_12)
(if (> (/ (fmax t_4 (* t_9 dY.u)) t_7) (floor maxAniso))
t_11
(* (sqrt (/ 1.0 t_10)) t_7))))
(if (>
(/
(/
(fmax (+ t_0 t_5) (+ t_2 t_8))
(fabs (fma dX.u dY.v (* (- dY.u) dX.v))))
(fabs t_6))
(floor maxAniso))
(floor maxAniso)
t_12))))
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((dX_46_u * floorf(w)), 2.0f);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = powf((dY_46_u * floorf(w)), 2.0f);
float t_3 = powf(floorf(w), 2.0f);
float t_4 = fmaf((t_3 * dX_46_u), dX_46_u, ((t_1 * dX_46_v) * dX_46_v));
float t_5 = powf((dX_46_v * floorf(h)), 2.0f);
float t_6 = floorf(w) * floorf(h);
float t_7 = fabsf((fmaf(dY_46_u, dX_46_v, (-dY_46_v * dX_46_u)) * t_6));
float t_8 = powf((dY_46_v * floorf(h)), 2.0f);
float t_9 = t_3 * dY_46_u;
float t_10 = fmaxf(t_4, fmaf(t_9, dY_46_u, ((t_1 * dY_46_v) * dY_46_v)));
float t_11 = sqrtf(t_10) / floorf(maxAniso);
float t_12 = t_10 / t_7;
int t_13 = t_12 > floorf(maxAniso);
float tmp;
if (t_13) {
tmp = t_11;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + t_0), (t_8 + t_2)))) * t_7;
}
float tmp_3;
if (tmp < 1.0f) {
float tmp_4;
if (t_13) {
tmp_4 = floorf(maxAniso);
} else {
tmp_4 = t_12;
}
float tmp_5;
if ((fmaxf(t_4, (t_9 * dY_46_u)) / t_7) > floorf(maxAniso)) {
tmp_5 = t_11;
} else {
tmp_5 = sqrtf((1.0f / t_10)) * t_7;
}
tmp_3 = fmaxf(1.0f, (tmp_4 * tmp_5));
} else if (((fmaxf((t_0 + t_5), (t_2 + t_8)) / fabsf(fmaf(dX_46_u, dY_46_v, (-dY_46_u * dX_46_v)))) / fabsf(t_6)) > floorf(maxAniso)) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_12;
}
return tmp_3;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_3 = floor(w) ^ Float32(2.0) t_4 = fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(Float32(t_1 * dX_46_v) * dX_46_v)) t_5 = Float32(dX_46_v * floor(h)) ^ Float32(2.0) t_6 = Float32(floor(w) * floor(h)) t_7 = abs(Float32(fma(dY_46_u, dX_46_v, Float32(Float32(-dY_46_v) * dX_46_u)) * t_6)) t_8 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_9 = Float32(t_3 * dY_46_u) t_10 = fmax(t_4, fma(t_9, dY_46_u, Float32(Float32(t_1 * dY_46_v) * dY_46_v))) t_11 = Float32(sqrt(t_10) / floor(maxAniso)) t_12 = Float32(t_10 / t_7) t_13 = t_12 > floor(maxAniso) tmp = Float32(0.0) if (t_13) tmp = t_11; else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_5 + t_0), Float32(t_8 + t_2)))) * t_7); end tmp_3 = Float32(0.0) if (tmp < Float32(1.0)) tmp_4 = Float32(0.0) if (t_13) tmp_4 = floor(maxAniso); else tmp_4 = t_12; end tmp_5 = Float32(0.0) if (Float32(fmax(t_4, Float32(t_9 * dY_46_u)) / t_7) > floor(maxAniso)) tmp_5 = t_11; else tmp_5 = Float32(sqrt(Float32(Float32(1.0) / t_10)) * t_7); end tmp_3 = fmax(Float32(1.0), Float32(tmp_4 * tmp_5)); elseif (Float32(Float32(fmax(Float32(t_0 + t_5), Float32(t_2 + t_8)) / abs(fma(dX_46_u, dY_46_v, Float32(Float32(-dY_46_u) * dX_46_v)))) / abs(t_6)) > floor(maxAniso)) tmp_3 = floor(maxAniso); else tmp_3 = t_12; end return tmp_3 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := \mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, \left(t\_1 \cdot dX.v\right) \cdot dX.v\right)\\
t_5 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_6 := \left\lfloor w\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_7 := \left|\mathsf{fma}\left(dY.u, dX.v, \left(-dY.v\right) \cdot dX.u\right) \cdot t\_6\right|\\
t_8 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_9 := t\_3 \cdot dY.u\\
t_10 := \mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_9, dY.u, \left(t\_1 \cdot dY.v\right) \cdot dY.v\right)\right)\\
t_11 := \frac{\sqrt{t\_10}}{\left\lfloor maxAniso\right\rfloor }\\
t_12 := \frac{t\_10}{t\_7}\\
t_13 := t\_12 > \left\lfloor maxAniso\right\rfloor \\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_13:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0, t\_8 + t\_2\right)}} \cdot t\_7\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t\_13:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_12\\
\end{array} \cdot \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_4, t\_9 \cdot dY.u\right)}{t\_7} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\frac{1}{t\_10}} \cdot t\_7\\
\end{array}\right)\\
\mathbf{elif}\;\frac{\frac{\mathsf{max}\left(t\_0 + t\_5, t\_2 + t\_8\right)}{\left|\mathsf{fma}\left(dX.u, dY.v, \left(-dY.u\right) \cdot dX.v\right)\right|}}{\left|t\_6\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_12\\
\end{array}
\end{array}
Initial program 98.4%
Taylor expanded in w around 0
Applied rewrites98.4%
Applied rewrites98.4%
Applied rewrites99.6%
Taylor expanded in dY.u around inf
Applied rewrites99.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* dX.u (floor w)) 2.0))
(t_1 (pow (* dY.v (floor h)) 2.0))
(t_2 (pow (floor h) 2.0))
(t_3 (pow (* dY.u (floor w)) 2.0))
(t_4 (pow (floor w) 2.0))
(t_5 (fma (* t_4 dX.u) dX.u (* (* t_2 dX.v) dX.v)))
(t_6
(fabs (* (fma dY.u dX.v (* (- dY.v) dX.u)) (* (floor w) (floor h)))))
(t_7 (pow (* dX.v (floor h)) 2.0))
(t_8 (* t_4 dY.u))
(t_9 (fmax t_5 (fma t_8 dY.u (* (* t_2 dY.v) dY.v))))
(t_10 (/ (sqrt t_9) (floor maxAniso)))
(t_11 (/ t_9 t_6))
(t_12 (> t_11 (floor maxAniso)))
(t_13 (if t_12 (floor maxAniso) t_11)))
(if (<
(if t_12 t_10 (* (/ 1.0 (sqrt (fmax (+ t_7 t_0) (+ t_1 t_3)))) t_6))
1.0)
(fmax
1.0
(*
t_13
(if (> (/ (fmax t_5 (* t_8 dY.u)) t_6) (floor maxAniso))
t_10
(* (/ 1.0 (sqrt (fmax (+ t_0 t_7) (+ t_3 t_1)))) t_6))))
t_13)))
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((dX_46_u * floorf(w)), 2.0f);
float t_1 = powf((dY_46_v * floorf(h)), 2.0f);
float t_2 = powf(floorf(h), 2.0f);
float t_3 = powf((dY_46_u * floorf(w)), 2.0f);
float t_4 = powf(floorf(w), 2.0f);
float t_5 = fmaf((t_4 * dX_46_u), dX_46_u, ((t_2 * dX_46_v) * dX_46_v));
float t_6 = fabsf((fmaf(dY_46_u, dX_46_v, (-dY_46_v * dX_46_u)) * (floorf(w) * floorf(h))));
float t_7 = powf((dX_46_v * floorf(h)), 2.0f);
float t_8 = t_4 * dY_46_u;
float t_9 = fmaxf(t_5, fmaf(t_8, dY_46_u, ((t_2 * dY_46_v) * dY_46_v)));
float t_10 = sqrtf(t_9) / floorf(maxAniso);
float t_11 = t_9 / t_6;
int t_12 = t_11 > floorf(maxAniso);
float tmp;
if (t_12) {
tmp = floorf(maxAniso);
} else {
tmp = t_11;
}
float t_13 = tmp;
float tmp_1;
if (t_12) {
tmp_1 = t_10;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf((t_7 + t_0), (t_1 + t_3)))) * t_6;
}
float tmp_3;
if (tmp_1 < 1.0f) {
float tmp_4;
if ((fmaxf(t_5, (t_8 * dY_46_u)) / t_6) > floorf(maxAniso)) {
tmp_4 = t_10;
} else {
tmp_4 = (1.0f / sqrtf(fmaxf((t_0 + t_7), (t_3 + t_1)))) * t_6;
}
tmp_3 = fmaxf(1.0f, (t_13 * tmp_4));
} else {
tmp_3 = t_13;
}
return tmp_3;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_1 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_2 = floor(h) ^ Float32(2.0) t_3 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_4 = floor(w) ^ Float32(2.0) t_5 = fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(Float32(t_2 * dX_46_v) * dX_46_v)) t_6 = abs(Float32(fma(dY_46_u, dX_46_v, Float32(Float32(-dY_46_v) * dX_46_u)) * Float32(floor(w) * floor(h)))) t_7 = Float32(dX_46_v * floor(h)) ^ Float32(2.0) t_8 = Float32(t_4 * dY_46_u) t_9 = fmax(t_5, fma(t_8, dY_46_u, Float32(Float32(t_2 * dY_46_v) * dY_46_v))) t_10 = Float32(sqrt(t_9) / floor(maxAniso)) t_11 = Float32(t_9 / t_6) t_12 = t_11 > floor(maxAniso) tmp = Float32(0.0) if (t_12) tmp = floor(maxAniso); else tmp = t_11; end t_13 = tmp tmp_1 = Float32(0.0) if (t_12) tmp_1 = t_10; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_7 + t_0), Float32(t_1 + t_3)))) * t_6); end tmp_3 = Float32(0.0) if (tmp_1 < Float32(1.0)) tmp_4 = Float32(0.0) if (Float32(fmax(t_5, Float32(t_8 * dY_46_u)) / t_6) > floor(maxAniso)) tmp_4 = t_10; else tmp_4 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_0 + t_7), Float32(t_3 + t_1)))) * t_6); end tmp_3 = fmax(Float32(1.0), Float32(t_13 * tmp_4)); else tmp_3 = t_13; end return tmp_3 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := \mathsf{fma}\left(t\_4 \cdot dX.u, dX.u, \left(t\_2 \cdot dX.v\right) \cdot dX.v\right)\\
t_6 := \left|\mathsf{fma}\left(dY.u, dX.v, \left(-dY.v\right) \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right|\\
t_7 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_8 := t\_4 \cdot dY.u\\
t_9 := \mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_8, dY.u, \left(t\_2 \cdot dY.v\right) \cdot dY.v\right)\right)\\
t_10 := \frac{\sqrt{t\_9}}{\left\lfloor maxAniso\right\rfloor }\\
t_11 := \frac{t\_9}{t\_6}\\
t_12 := t\_11 > \left\lfloor maxAniso\right\rfloor \\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_12:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_12:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_7 + t\_0, t\_1 + t\_3\right)}} \cdot t\_6\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, t\_13 \cdot \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_5, t\_8 \cdot dY.u\right)}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 + t\_7, t\_3 + t\_1\right)}} \cdot t\_6\\
\end{array}\right)\\
\mathbf{else}:\\
\;\;\;\;t\_13\\
\end{array}
\end{array}
Initial program 98.4%
Taylor expanded in w around 0
Applied rewrites98.4%
Applied rewrites98.4%
Applied rewrites98.4%
Taylor expanded in dY.u around inf
Applied rewrites98.4%
herbie shell --seed 2024360
(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))))))))