
(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 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 = 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 (fma dY.u dX.v (* (- dX.u) dY.v)))
(t_1 (* (floor h) (floor w)))
(t_2 (fabs (* t_0 t_1)))
(t_3 (pow (floor w) 2.0))
(t_4 (pow (floor h) 2.0))
(t_5 (fma (* t_4 dX.v) dX.v (* (* t_3 dX.u) dX.u)))
(t_6 (fmax t_5 (fma (* t_4 dY.v) dY.v (* (* t_3 dY.u) dY.u))))
(t_7 (/ (sqrt t_6) (floor maxAniso)))
(t_8 (/ t_6 t_2))
(t_9 (> t_8 (floor maxAniso)))
(t_10 (* (sqrt (/ 1.0 t_6)) t_2)))
(if (< (if t_9 t_7 t_10) 1.0)
(fmax
1.0
(*
(if t_9 (floor maxAniso) t_8)
(if (> (/ (fmax t_5 (pow (* (floor w) dY.u) 2.0)) t_2) (floor maxAniso))
t_7
t_10)))
(if (>
(/
(/
(fmax
(+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0))
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(fabs t_0))
(fabs t_1))
(floor maxAniso))
(floor maxAniso)
t_8))))
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 = fmaf(dY_46_u, dX_46_v, (-dX_46_u * dY_46_v));
float t_1 = floorf(h) * floorf(w);
float t_2 = fabsf((t_0 * t_1));
float t_3 = powf(floorf(w), 2.0f);
float t_4 = powf(floorf(h), 2.0f);
float t_5 = fmaf((t_4 * dX_46_v), dX_46_v, ((t_3 * dX_46_u) * dX_46_u));
float t_6 = fmaxf(t_5, fmaf((t_4 * dY_46_v), dY_46_v, ((t_3 * dY_46_u) * dY_46_u)));
float t_7 = sqrtf(t_6) / floorf(maxAniso);
float t_8 = t_6 / t_2;
int t_9 = t_8 > floorf(maxAniso);
float t_10 = sqrtf((1.0f / t_6)) * t_2;
float tmp;
if (t_9) {
tmp = t_7;
} else {
tmp = t_10;
}
float tmp_3;
if (tmp < 1.0f) {
float tmp_4;
if (t_9) {
tmp_4 = floorf(maxAniso);
} else {
tmp_4 = t_8;
}
float tmp_5;
if ((fmaxf(t_5, powf((floorf(w) * dY_46_u), 2.0f)) / t_2) > floorf(maxAniso)) {
tmp_5 = t_7;
} else {
tmp_5 = t_10;
}
tmp_3 = fmaxf(1.0f, (tmp_4 * tmp_5));
} else if (((fmaxf((powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) / fabsf(t_0)) / fabsf(t_1)) > floorf(maxAniso)) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_8;
}
return tmp_3;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(dY_46_u, dX_46_v, Float32(Float32(-dX_46_u) * dY_46_v)) t_1 = Float32(floor(h) * floor(w)) t_2 = abs(Float32(t_0 * t_1)) t_3 = floor(w) ^ Float32(2.0) t_4 = floor(h) ^ Float32(2.0) t_5 = fma(Float32(t_4 * dX_46_v), dX_46_v, Float32(Float32(t_3 * dX_46_u) * dX_46_u)) t_6 = fmax(t_5, fma(Float32(t_4 * dY_46_v), dY_46_v, Float32(Float32(t_3 * dY_46_u) * dY_46_u))) t_7 = Float32(sqrt(t_6) / floor(maxAniso)) t_8 = Float32(t_6 / t_2) t_9 = t_8 > floor(maxAniso) t_10 = Float32(sqrt(Float32(Float32(1.0) / t_6)) * t_2) tmp = Float32(0.0) if (t_9) tmp = t_7; else tmp = t_10; end tmp_3 = Float32(0.0) if (tmp < Float32(1.0)) tmp_4 = Float32(0.0) if (t_9) tmp_4 = floor(maxAniso); else tmp_4 = t_8; end tmp_5 = Float32(0.0) if (Float32(fmax(t_5, (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) / t_2) > floor(maxAniso)) tmp_5 = t_7; else tmp_5 = t_10; end tmp_3 = fmax(Float32(1.0), Float32(tmp_4 * tmp_5)); elseif (Float32(Float32(fmax(Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) / abs(t_0)) / abs(t_1)) > floor(maxAniso)) tmp_3 = floor(maxAniso); else tmp_3 = t_8; end return tmp_3 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(dY.u, dX.v, \left(-dX.u\right) \cdot dY.v\right)\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left|t\_0 \cdot t\_1\right|\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_5 := \mathsf{fma}\left(t\_4 \cdot dX.v, dX.v, \left(t\_3 \cdot dX.u\right) \cdot dX.u\right)\\
t_6 := \mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_4 \cdot dY.v, dY.v, \left(t\_3 \cdot dY.u\right) \cdot dY.u\right)\right)\\
t_7 := \frac{\sqrt{t\_6}}{\left\lfloor maxAniso\right\rfloor }\\
t_8 := \frac{t\_6}{t\_2}\\
t_9 := t\_8 > \left\lfloor maxAniso\right\rfloor \\
t_10 := \sqrt{\frac{1}{t\_6}} \cdot t\_2\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_9:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t\_9:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array} \cdot \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(t\_5, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}{t\_2} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}\right)\\
\mathbf{elif}\;\frac{\frac{\mathsf{max}\left({\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\right)}{\left|t\_0\right|}}{\left|t\_1\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}
\end{array}
Initial program 97.7%
Taylor expanded in w around 0
Applied rewrites97.8%
Applied rewrites99.3%
Taylor expanded in dY.u around inf
Applied rewrites99.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (fma dY.u dX.v (* (- dX.u) dY.v)))
(t_1 (* (floor h) (floor w)))
(t_2 (fabs (* t_0 t_1)))
(t_3 (pow (floor w) 2.0))
(t_4 (pow (floor h) 2.0))
(t_5
(fmax
(fma (* t_4 dX.v) dX.v (* (* t_3 dX.u) dX.u))
(fma (* t_4 dY.v) dY.v (* (* t_3 dY.u) dY.u))))
(t_6 (/ (sqrt t_5) (floor maxAniso)))
(t_7 (/ t_5 t_2))
(t_8 (> t_7 (floor maxAniso)))
(t_9 (* (sqrt (/ 1.0 t_5)) t_2)))
(if (< (if t_8 t_6 t_9) 1.0)
(fmax
1.0
(*
(if t_8 (floor maxAniso) t_7)
(if (>
(/
(fmax (pow (* (floor w) dX.u) 2.0) (pow (* (floor w) dY.u) 2.0))
t_2)
(floor maxAniso))
t_6
t_9)))
(if (>
(/
(/
(fmax
(+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0))
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(fabs t_0))
(fabs t_1))
(floor maxAniso))
(floor maxAniso)
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 = fmaf(dY_46_u, dX_46_v, (-dX_46_u * dY_46_v));
float t_1 = floorf(h) * floorf(w);
float t_2 = fabsf((t_0 * t_1));
float t_3 = powf(floorf(w), 2.0f);
float t_4 = powf(floorf(h), 2.0f);
float t_5 = fmaxf(fmaf((t_4 * dX_46_v), dX_46_v, ((t_3 * dX_46_u) * dX_46_u)), fmaf((t_4 * dY_46_v), dY_46_v, ((t_3 * dY_46_u) * dY_46_u)));
float t_6 = sqrtf(t_5) / floorf(maxAniso);
float t_7 = t_5 / t_2;
int t_8 = t_7 > floorf(maxAniso);
float t_9 = sqrtf((1.0f / t_5)) * t_2;
float tmp;
if (t_8) {
tmp = t_6;
} else {
tmp = t_9;
}
float tmp_3;
if (tmp < 1.0f) {
float tmp_4;
if (t_8) {
tmp_4 = floorf(maxAniso);
} else {
tmp_4 = t_7;
}
float tmp_5;
if ((fmaxf(powf((floorf(w) * dX_46_u), 2.0f), powf((floorf(w) * dY_46_u), 2.0f)) / t_2) > floorf(maxAniso)) {
tmp_5 = t_6;
} else {
tmp_5 = t_9;
}
tmp_3 = fmaxf(1.0f, (tmp_4 * tmp_5));
} else if (((fmaxf((powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) / fabsf(t_0)) / fabsf(t_1)) > floorf(maxAniso)) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_7;
}
return tmp_3;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(dY_46_u, dX_46_v, Float32(Float32(-dX_46_u) * dY_46_v)) t_1 = Float32(floor(h) * floor(w)) t_2 = abs(Float32(t_0 * t_1)) t_3 = floor(w) ^ Float32(2.0) t_4 = floor(h) ^ Float32(2.0) t_5 = fmax(fma(Float32(t_4 * dX_46_v), dX_46_v, Float32(Float32(t_3 * dX_46_u) * dX_46_u)), fma(Float32(t_4 * dY_46_v), dY_46_v, Float32(Float32(t_3 * dY_46_u) * dY_46_u))) t_6 = Float32(sqrt(t_5) / floor(maxAniso)) t_7 = Float32(t_5 / t_2) t_8 = t_7 > floor(maxAniso) t_9 = Float32(sqrt(Float32(Float32(1.0) / t_5)) * t_2) tmp = Float32(0.0) if (t_8) tmp = t_6; else tmp = t_9; end tmp_3 = Float32(0.0) if (tmp < Float32(1.0)) tmp_4 = Float32(0.0) if (t_8) tmp_4 = floor(maxAniso); else tmp_4 = t_7; end tmp_5 = Float32(0.0) if (Float32(fmax((Float32(floor(w) * dX_46_u) ^ Float32(2.0)), (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) / t_2) > floor(maxAniso)) tmp_5 = t_6; else tmp_5 = t_9; end tmp_3 = fmax(Float32(1.0), Float32(tmp_4 * tmp_5)); elseif (Float32(Float32(fmax(Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) / abs(t_0)) / abs(t_1)) > floor(maxAniso)) tmp_3 = floor(maxAniso); else tmp_3 = t_7; end return tmp_3 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(dY.u, dX.v, \left(-dX.u\right) \cdot dY.v\right)\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left|t\_0 \cdot t\_1\right|\\
t_3 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_5 := \mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.v, dX.v, \left(t\_3 \cdot dX.u\right) \cdot dX.u\right), \mathsf{fma}\left(t\_4 \cdot dY.v, dY.v, \left(t\_3 \cdot dY.u\right) \cdot dY.u\right)\right)\\
t_6 := \frac{\sqrt{t\_5}}{\left\lfloor maxAniso\right\rfloor }\\
t_7 := \frac{t\_5}{t\_2}\\
t_8 := t\_7 > \left\lfloor maxAniso\right\rfloor \\
t_9 := \sqrt{\frac{1}{t\_5}} \cdot t\_2\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t\_8:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array} \cdot \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left({\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}{t\_2} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\right)\\
\mathbf{elif}\;\frac{\frac{\mathsf{max}\left({\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\right)}{\left|t\_0\right|}}{\left|t\_1\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
Initial program 97.7%
Taylor expanded in w around 0
Applied rewrites97.8%
Applied rewrites99.3%
Taylor expanded in dY.u around inf
Applied rewrites99.3%
Taylor expanded in dX.u around inf
Applied rewrites99.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fabs (* (fma dY.u dX.v (* (- dX.u) dY.v)) (* (floor h) (floor w)))))
(t_1 (pow (floor w) 2.0))
(t_2 (pow (floor h) 2.0))
(t_3
(fmax
(fma (* t_2 dX.v) dX.v (* (* t_1 dX.u) dX.u))
(fma (* t_2 dY.v) dY.v (* (* t_1 dY.u) dY.u))))
(t_4 (/ (sqrt t_3) (floor maxAniso)))
(t_5 (/ t_3 t_0))
(t_6 (> t_5 (floor maxAniso)))
(t_7 (if t_6 (floor maxAniso) t_5)))
(if (<
(if t_6
t_4
(*
(/
1.0
(sqrt
(fmax
(+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0))
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))))
t_0))
1.0)
(fmax
1.0
(*
t_7
(if (>
(/
(fmax (pow (* (floor h) dX.v) 2.0) (pow (* (floor w) dY.u) 2.0))
t_0)
(floor maxAniso))
t_4
(* (sqrt (/ 1.0 t_3)) t_0))))
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 = fabsf((fmaf(dY_46_u, dX_46_v, (-dX_46_u * dY_46_v)) * (floorf(h) * floorf(w))));
float t_1 = powf(floorf(w), 2.0f);
float t_2 = powf(floorf(h), 2.0f);
float t_3 = fmaxf(fmaf((t_2 * dX_46_v), dX_46_v, ((t_1 * dX_46_u) * dX_46_u)), fmaf((t_2 * dY_46_v), dY_46_v, ((t_1 * dY_46_u) * dY_46_u)));
float t_4 = sqrtf(t_3) / floorf(maxAniso);
float t_5 = t_3 / t_0;
int t_6 = t_5 > floorf(maxAniso);
float tmp;
if (t_6) {
tmp = floorf(maxAniso);
} else {
tmp = t_5;
}
float t_7 = tmp;
float tmp_1;
if (t_6) {
tmp_1 = t_4;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf((powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))))) * t_0;
}
float tmp_3;
if (tmp_1 < 1.0f) {
float tmp_4;
if ((fmaxf(powf((floorf(h) * dX_46_v), 2.0f), powf((floorf(w) * dY_46_u), 2.0f)) / t_0) > floorf(maxAniso)) {
tmp_4 = t_4;
} else {
tmp_4 = sqrtf((1.0f / t_3)) * t_0;
}
tmp_3 = fmaxf(1.0f, (t_7 * tmp_4));
} else {
tmp_3 = t_7;
}
return tmp_3;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = abs(Float32(fma(dY_46_u, dX_46_v, Float32(Float32(-dX_46_u) * dY_46_v)) * Float32(floor(h) * floor(w)))) t_1 = floor(w) ^ Float32(2.0) t_2 = floor(h) ^ Float32(2.0) t_3 = fmax(fma(Float32(t_2 * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)), fma(Float32(t_2 * dY_46_v), dY_46_v, Float32(Float32(t_1 * dY_46_u) * dY_46_u))) t_4 = Float32(sqrt(t_3) / floor(maxAniso)) t_5 = Float32(t_3 / t_0) t_6 = t_5 > floor(maxAniso) tmp = Float32(0.0) if (t_6) tmp = floor(maxAniso); else tmp = t_5; end t_7 = tmp tmp_1 = Float32(0.0) if (t_6) tmp_1 = t_4; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))))) * t_0); end tmp_3 = Float32(0.0) if (tmp_1 < Float32(1.0)) tmp_4 = Float32(0.0) if (Float32(fmax((Float32(floor(h) * dX_46_v) ^ Float32(2.0)), (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) / t_0) > floor(maxAniso)) tmp_4 = t_4; else tmp_4 = Float32(sqrt(Float32(Float32(1.0) / t_3)) * t_0); end tmp_3 = fmax(Float32(1.0), Float32(t_7 * tmp_4)); else tmp_3 = t_7; end return tmp_3 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left|\mathsf{fma}\left(dY.u, dX.v, \left(-dX.u\right) \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right|\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := \mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.v, dX.v, \left(t\_1 \cdot dX.u\right) \cdot dX.u\right), \mathsf{fma}\left(t\_2 \cdot dY.v, dY.v, \left(t\_1 \cdot dY.u\right) \cdot dY.u\right)\right)\\
t_4 := \frac{\sqrt{t\_3}}{\left\lfloor maxAniso\right\rfloor }\\
t_5 := \frac{t\_3}{t\_0}\\
t_6 := t\_5 > \left\lfloor maxAniso\right\rfloor \\
t_7 := \begin{array}{l}
\mathbf{if}\;t\_6:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_5\\
\end{array}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_6:\\
\;\;\;\;t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left({\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\right)}} \cdot t\_0\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, t\_7 \cdot \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}, {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\right)}{t\_0} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_4\\
\mathbf{else}:\\
\;\;\;\;\sqrt{\frac{1}{t\_3}} \cdot t\_0\\
\end{array}\right)\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
Initial program 97.7%
Taylor expanded in w around 0
Applied rewrites97.8%
Applied rewrites97.8%
Taylor expanded in dY.u around inf
Applied rewrites97.8%
Taylor expanded in dX.u around 0
Applied rewrites97.8%
herbie shell --seed 2025024
(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))))))))