
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
: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))
(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}
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}
Herbie found 3 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
: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))
(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}
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}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
: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))
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dX.u (floor w)))
(t_2 (* dY.u (floor w)))
(t_3 (* dX.v (floor h)))
(t_4 (fma t_3 t_3 (* t_1 t_1)))
(t_5
(*
(/
(fmax
(fma (* (* dY.u dY.u) (floor w)) (floor w) (* t_0 t_0))
t_4)
(fabs (- (* dY.u t_3) (* (* dY.v dX.u) (floor h)))))
(/ 1.0 (fabs (floor w)))))
(t_6 (> t_5 (floor maxAniso)))
(t_7 (if t_6 (floor maxAniso) t_5))
(t_8 (sqrt (fmax (fma t_0 t_0 (* t_2 t_2)) t_4)))
(t_9
(if t_6
(/ t_8 (floor maxAniso))
(/
(fabs (* (floor h) (- (* dY.v t_1) (* t_2 dX.v))))
t_8))))
(if (< t_9 1.0) (fmax (* t_7 t_9) 1.0) 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 = dY_46_v * floorf(h);
float t_1 = dX_46_u * floorf(w);
float t_2 = dY_46_u * floorf(w);
float t_3 = dX_46_v * floorf(h);
float t_4 = fmaf(t_3, t_3, (t_1 * t_1));
float t_5 = (fmaxf(fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), (t_0 * t_0)), t_4) / fabsf(((dY_46_u * t_3) - ((dY_46_v * dX_46_u) * floorf(h))))) * (1.0f / fabsf(floorf(w)));
int t_6 = t_5 > floorf(maxAniso);
float tmp;
if (t_6) {
tmp = floorf(maxAniso);
} else {
tmp = t_5;
}
float t_7 = tmp;
float t_8 = sqrtf(fmaxf(fmaf(t_0, t_0, (t_2 * t_2)), t_4));
float tmp_1;
if (t_6) {
tmp_1 = t_8 / floorf(maxAniso);
} else {
tmp_1 = fabsf((floorf(h) * ((dY_46_v * t_1) - (t_2 * dX_46_v)))) / t_8;
}
float t_9 = tmp_1;
float tmp_2;
if (t_9 < 1.0f) {
tmp_2 = fmaxf((t_7 * t_9), 1.0f);
} else {
tmp_2 = t_7;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(dX_46_u * floor(w)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(dX_46_v * floor(h)) t_4 = fma(t_3, t_3, Float32(t_1 * t_1)) t_5 = Float32(Float32(fmax(fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(t_0 * t_0)), t_4) / abs(Float32(Float32(dY_46_u * t_3) - Float32(Float32(dY_46_v * dX_46_u) * floor(h))))) * Float32(Float32(1.0) / abs(floor(w)))) t_6 = t_5 > floor(maxAniso) tmp = Float32(0.0) if (t_6) tmp = floor(maxAniso); else tmp = t_5; end t_7 = tmp t_8 = sqrt(fmax(fma(t_0, t_0, Float32(t_2 * t_2)), t_4)) tmp_1 = Float32(0.0) if (t_6) tmp_1 = Float32(t_8 / floor(maxAniso)); else tmp_1 = Float32(abs(Float32(floor(h) * Float32(Float32(dY_46_v * t_1) - Float32(t_2 * dX_46_v)))) / t_8); end t_9 = tmp_1 tmp_2 = Float32(0.0) if (t_9 < Float32(1.0)) tmp_2 = fmax(Float32(t_7 * t_9), Float32(1.0)); else tmp_2 = t_7; end return tmp_2 end
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_3, t\_3, t\_1 \cdot t\_1\right)\\
t_5 := \frac{\mathsf{max}\left(\mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_0 \cdot t\_0\right), t\_4\right)}{\left|dY.u \cdot t\_3 - \left(dY.v \cdot dX.u\right) \cdot \left\lfloor h\right\rfloor \right|} \cdot \frac{1}{\left|\left\lfloor w\right\rfloor \right|}\\
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}\\
t_8 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_2 \cdot t\_2\right), t\_4\right)}\\
t_9 := \begin{array}{l}
\mathbf{if}\;t\_6:\\
\;\;\;\;\frac{t\_8}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor h\right\rfloor \cdot \left(dY.v \cdot t\_1 - t\_2 \cdot dX.v\right)\right|}{t\_8}\\
\end{array}\\
\mathbf{if}\;t\_9 < 1:\\
\;\;\;\;\mathsf{max}\left(t\_7 \cdot t\_9, 1\right)\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
Initial program 98.0%
Applied rewrites98.0%
Applied rewrites98.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
: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))
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dX.v))
(t_3 (* (* dY.v dX.u) (floor h)))
(t_4 (fabs (* t_3 (floor w))))
(t_5 (* (floor h) dY.v))
(t_6 (* dY.v (floor h)))
(t_7 (fma (* (* dY.u dY.u) (floor w)) (floor w) (* t_6 t_6)))
(t_8 (* (floor w) dX.u))
(t_9
(sqrt
(fmax
(+ (* t_8 t_8) (* t_2 t_2))
(+ (* t_1 t_1) (* t_5 t_5)))))
(t_10 (* dX.v (floor h)))
(t_11
(*
(/ 1.0 (fabs (floor w)))
(/ (fmax t_7 (fma t_10 t_10 (* t_0 t_0))) (fabs t_3))))
(t_12 (> t_11 (floor maxAniso)))
(t_13
(fmax
t_7
(fma
(* (* (floor h) (floor h)) dX.v)
dX.v
(* (* (* (floor w) (floor w)) dX.u) dX.u))))
(t_14 (sqrt t_13))
(t_15 (/ t_13 t_4))
(t_16 (> t_15 (floor maxAniso))))
(if (<
(if t_12
(/ t_9 (floor maxAniso))
(/ (fabs (* (floor w) (* dX.u t_6))) t_9))
1.0)
(fmax
1.0
(*
(if t_16 (floor maxAniso) t_15)
(if t_16 (/ t_14 (floor maxAniso)) (/ t_4 t_14))))
(if t_12 (floor maxAniso) t_11))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_u * floorf(w);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = (dY_46_v * dX_46_u) * floorf(h);
float t_4 = fabsf((t_3 * floorf(w)));
float t_5 = floorf(h) * dY_46_v;
float t_6 = dY_46_v * floorf(h);
float t_7 = fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), (t_6 * t_6));
float t_8 = floorf(w) * dX_46_u;
float t_9 = sqrtf(fmaxf(((t_8 * t_8) + (t_2 * t_2)), ((t_1 * t_1) + (t_5 * t_5))));
float t_10 = dX_46_v * floorf(h);
float t_11 = (1.0f / fabsf(floorf(w))) * (fmaxf(t_7, fmaf(t_10, t_10, (t_0 * t_0))) / fabsf(t_3));
int t_12 = t_11 > floorf(maxAniso);
float t_13 = fmaxf(t_7, fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, (((floorf(w) * floorf(w)) * dX_46_u) * dX_46_u)));
float t_14 = sqrtf(t_13);
float t_15 = t_13 / t_4;
int t_16 = t_15 > floorf(maxAniso);
float tmp;
if (t_12) {
tmp = t_9 / floorf(maxAniso);
} else {
tmp = fabsf((floorf(w) * (dX_46_u * t_6))) / t_9;
}
float tmp_3;
if (tmp < 1.0f) {
float tmp_4;
if (t_16) {
tmp_4 = floorf(maxAniso);
} else {
tmp_4 = t_15;
}
float tmp_5;
if (t_16) {
tmp_5 = t_14 / floorf(maxAniso);
} else {
tmp_5 = t_4 / t_14;
}
tmp_3 = fmaxf(1.0f, (tmp_4 * tmp_5));
} else if (t_12) {
tmp_3 = floorf(maxAniso);
} else {
tmp_3 = t_11;
}
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)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(Float32(dY_46_v * dX_46_u) * floor(h)) t_4 = abs(Float32(t_3 * floor(w))) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(dY_46_v * floor(h)) t_7 = fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(t_6 * t_6)) t_8 = Float32(floor(w) * dX_46_u) t_9 = sqrt(fmax(Float32(Float32(t_8 * t_8) + Float32(t_2 * t_2)), Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)))) t_10 = Float32(dX_46_v * floor(h)) t_11 = Float32(Float32(Float32(1.0) / abs(floor(w))) * Float32(fmax(t_7, fma(t_10, t_10, Float32(t_0 * t_0))) / abs(t_3))) t_12 = t_11 > floor(maxAniso) t_13 = fmax(t_7, fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, Float32(Float32(Float32(floor(w) * floor(w)) * dX_46_u) * dX_46_u))) t_14 = sqrt(t_13) t_15 = Float32(t_13 / t_4) t_16 = t_15 > floor(maxAniso) tmp = Float32(0.0) if (t_12) tmp = Float32(t_9 / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(w) * Float32(dX_46_u * t_6))) / t_9); end tmp_3 = Float32(0.0) if (tmp < Float32(1.0)) tmp_4 = Float32(0.0) if (t_16) tmp_4 = floor(maxAniso); else tmp_4 = t_15; end tmp_5 = Float32(0.0) if (t_16) tmp_5 = Float32(t_14 / floor(maxAniso)); else tmp_5 = Float32(t_4 / t_14); end tmp_3 = fmax(Float32(1.0), Float32(tmp_4 * tmp_5)); elseif (t_12) tmp_3 = floor(maxAniso); else tmp_3 = t_11; end return tmp_3 end
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left(dY.v \cdot dX.u\right) \cdot \left\lfloor h\right\rfloor \\
t_4 := \left|t\_3 \cdot \left\lfloor w\right\rfloor \right|\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_7 := \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_6 \cdot t\_6\right)\\
t_8 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_9 := \sqrt{\mathsf{max}\left(t\_8 \cdot t\_8 + t\_2 \cdot t\_2, t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right)}\\
t_10 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_11 := \frac{1}{\left|\left\lfloor w\right\rfloor \right|} \cdot \frac{\mathsf{max}\left(t\_7, \mathsf{fma}\left(t\_10, t\_10, t\_0 \cdot t\_0\right)\right)}{\left|t\_3\right|}\\
t_12 := t\_11 > \left\lfloor maxAniso\right\rfloor \\
t_13 := \mathsf{max}\left(t\_7, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, \left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot dX.u\right)\right)\\
t_14 := \sqrt{t\_13}\\
t_15 := \frac{t\_13}{t\_4}\\
t_16 := t\_15 > \left\lfloor maxAniso\right\rfloor \\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_12:\\
\;\;\;\;\frac{t\_9}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot t\_6\right)\right|}{t\_9}\\
\end{array} < 1:\\
\;\;\;\;\mathsf{max}\left(1, \begin{array}{l}
\mathbf{if}\;t\_16:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array} \cdot \begin{array}{l}
\mathbf{if}\;t\_16:\\
\;\;\;\;\frac{t\_14}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_14}\\
\end{array}\right)\\
\mathbf{elif}\;t\_12:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}
Initial program 98.0%
Applied rewrites98.0%
Taylor expanded in dX.u around inf
Applied rewrites97.8%
Applied rewrites98.3%
Applied rewrites98.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
: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))
(let* ((t_0 (fabs (* (* (* dY.v dX.u) (floor h)) (floor w))))
(t_1 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (* dX.u (floor w)))
(t_4
(fmax
(fma (* (* dY.u dY.u) (floor w)) (floor w) (* t_1 t_1))
(fma t_2 t_2 (* t_3 t_3))))
(t_5 (sqrt t_4))
(t_6 (if 0 (/ t_5 (floor maxAniso)) (/ t_0 t_5)))
(t_7 (if 0 (floor maxAniso) (/ t_4 t_0))))
(if (< t_6 1.0) (fmax (* t_7 t_6) 1.0) 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((((dY_46_v * dX_46_u) * floorf(h)) * floorf(w)));
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = dX_46_u * floorf(w);
float t_4 = fmaxf(fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), (t_1 * t_1)), fmaf(t_2, t_2, (t_3 * t_3)));
float t_5 = sqrtf(t_4);
float tmp;
if (0.0f) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_0 / t_5;
}
float t_6 = tmp;
float tmp_1;
if (0.0f) {
tmp_1 = floorf(maxAniso);
} else {
tmp_1 = t_4 / t_0;
}
float t_7 = tmp_1;
float tmp_2;
if (t_6 < 1.0f) {
tmp_2 = fmaxf((t_7 * t_6), 1.0f);
} else {
tmp_2 = t_7;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = abs(Float32(Float32(Float32(dY_46_v * dX_46_u) * floor(h)) * floor(w))) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(dX_46_u * floor(w)) t_4 = fmax(fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(t_1 * t_1)), fma(t_2, t_2, Float32(t_3 * t_3))) t_5 = sqrt(t_4) tmp = Float32(0.0) if (Float32(0.0)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_0 / t_5); end t_6 = tmp tmp_1 = Float32(0.0) if (Float32(0.0)) tmp_1 = floor(maxAniso); else tmp_1 = Float32(t_4 / t_0); end t_7 = tmp_1 tmp_2 = Float32(0.0) if (t_6 < Float32(1.0)) tmp_2 = fmax(Float32(t_7 * t_6), Float32(1.0)); else tmp_2 = t_7; end return tmp_2 end
\begin{array}{l}
t_0 := \left|\left(\left(dY.v \cdot dX.u\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right|\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_4 := \mathsf{max}\left(\mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_1 \cdot t\_1\right), \mathsf{fma}\left(t\_2, t\_2, t\_3 \cdot t\_3\right)\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \begin{array}{l}
\mathbf{if}\;0:\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_5}\\
\end{array}\\
t_7 := \begin{array}{l}
\mathbf{if}\;0:\\
\;\;\;\;\left\lfloor maxAniso\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_0}\\
\end{array}\\
\mathbf{if}\;t\_6 < 1:\\
\;\;\;\;\mathsf{max}\left(t\_7 \cdot t\_6, 1\right)\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
Initial program 98.0%
Applied rewrites98.0%
Taylor expanded in dX.u around inf
Applied rewrites97.8%
Applied rewrites97.8%
Taylor expanded in undef-var around zero
Applied rewrites97.5%
herbie shell --seed 2026084
(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))))))))