
(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 w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
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(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return 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(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return 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(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\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 w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 7 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 w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
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(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return 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(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return 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(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\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 w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(t_1 (* (floor w) dX.u))
(t_2 (+ (pow t_1 2.0) (pow (* (floor h) dX.v) 2.0)))
(t_3 (sqrt (fmax t_2 t_0))))
(if (>= t_2 t_0) (* t_1 (/ 1.0 t_3)) (/ (floor w) (/ t_3 dY.u)))))
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((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f);
float t_1 = floorf(w) * dX_46_u;
float t_2 = powf(t_1, 2.0f) + powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = sqrtf(fmaxf(t_2, t_0));
float tmp;
if (t_2 >= t_0) {
tmp = t_1 * (1.0f / t_3);
} else {
tmp = floorf(w) / (t_3 / dY_46_u);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32((t_1 ^ Float32(2.0)) + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_3 = sqrt(((t_2 != t_2) ? t_0 : ((t_0 != t_0) ? t_2 : max(t_2, t_0)))) tmp = Float32(0.0) if (t_2 >= t_0) tmp = Float32(t_1 * Float32(Float32(1.0) / t_3)); else tmp = Float32(floor(w) / Float32(t_3 / dY_46_u)); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = ((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)); t_1 = floor(w) * dX_46_u; t_2 = (t_1 ^ single(2.0)) + ((floor(h) * dX_46_v) ^ single(2.0)); t_3 = sqrt(max(t_2, t_0)); tmp = single(0.0); if (t_2 >= t_0) tmp = t_1 * (single(1.0) / t_3); else tmp = floor(w) / (t_3 / dY_46_u); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := {t\_1}^{2} + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_0\right)}\\
\mathbf{if}\;t\_2 \geq t\_0:\\
\;\;\;\;t\_1 \cdot \frac{1}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\frac{t\_3}{dY.u}}\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites77.6%
Applied rewrites77.7%
lift-+.f32N/A
+-commutativeN/A
lower-+.f3277.7
lift-*.f32N/A
pow2N/A
lift-pow.f3277.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3277.7
lift-*.f32N/A
pow2N/A
lift-pow.f3277.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3277.7
lift-+.f32N/A
Applied rewrites77.7%
Applied rewrites77.7%
Final simplification77.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* (floor h) dX.v))
(t_2 (* dY.u (floor w)))
(t_3 (+ (* t_0 t_0) (* t_2 t_2)))
(t_4 (+ (pow t_2 2.0) (pow t_0 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (/ 1.0 (sqrt (fmax (+ (* t_1 t_1) (* t_5 t_5)) t_3))))
(t_7 (* t_6 t_5)))
(if (<= dX.v 2.299999952316284)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_4) t_7 (* t_6 t_2))
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) t_3)
t_7
(*
(/ 1.0 (/ 1.0 (pow (fmax (+ (pow t_5 2.0) (pow t_1 2.0)) t_4) -0.5)))
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 = dY_46_v * floorf(h);
float t_1 = floorf(h) * dX_46_v;
float t_2 = dY_46_u * floorf(w);
float t_3 = (t_0 * t_0) + (t_2 * t_2);
float t_4 = powf(t_2, 2.0f) + powf(t_0, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = 1.0f / sqrtf(fmaxf(((t_1 * t_1) + (t_5 * t_5)), t_3));
float t_7 = t_6 * t_5;
float tmp_1;
if (dX_46_v <= 2.299999952316284f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_4) {
tmp_2 = t_7;
} else {
tmp_2 = t_6 * t_2;
}
tmp_1 = tmp_2;
} else if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= t_3) {
tmp_1 = t_7;
} else {
tmp_1 = (1.0f / (1.0f / powf(fmaxf((powf(t_5, 2.0f) + powf(t_1, 2.0f)), t_4), -0.5f))) * t_2;
}
return tmp_1;
}
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(floor(h) * dX_46_v) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) t_4 = Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) != Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5))) ? t_3 : ((t_3 != t_3) ? Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) : max(Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)), t_3))))) t_7 = Float32(t_6 * t_5) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(2.299999952316284)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_2 = t_7; else tmp_2 = Float32(t_6 * t_2); end tmp_1 = tmp_2; elseif (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= t_3) tmp_1 = t_7; else tmp_1 = Float32(Float32(Float32(1.0) / Float32(Float32(1.0) / (((Float32((t_5 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_5 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? t_4 : ((t_4 != t_4) ? Float32((t_5 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : max(Float32((t_5 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))), t_4))) ^ Float32(-0.5)))) * t_2); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = dY_46_v * floor(h); t_1 = floor(h) * dX_46_v; t_2 = dY_46_u * floor(w); t_3 = (t_0 * t_0) + (t_2 * t_2); t_4 = (t_2 ^ single(2.0)) + (t_0 ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = single(1.0) / sqrt(max(((t_1 * t_1) + (t_5 * t_5)), t_3)); t_7 = t_6 * t_5; tmp_2 = single(0.0); if (dX_46_v <= single(2.299999952316284)) tmp_3 = single(0.0); if ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_3 = t_7; else tmp_3 = t_6 * t_2; end tmp_2 = tmp_3; elseif ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= t_3) tmp_2 = t_7; else tmp_2 = (single(1.0) / (single(1.0) / (max(((t_5 ^ single(2.0)) + (t_1 ^ single(2.0))), t_4) ^ single(-0.5)))) * t_2; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := t\_0 \cdot t\_0 + t\_2 \cdot t\_2\\
t_4 := {t\_2}^{2} + {t\_0}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_1 \cdot t\_1 + t\_5 \cdot t\_5, t\_3\right)}}\\
t_7 := t\_6 \cdot t\_5\\
\mathbf{if}\;dX.v \leq 2.299999952316284:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_4:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_2\\
\end{array}\\
\mathbf{elif}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_3:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{1}{{\left(\mathsf{max}\left({t\_5}^{2} + {t\_1}^{2}, t\_4\right)\right)}^{-0.5}}} \cdot t\_2\\
\end{array}
\end{array}
if dX.v < 2.29999995Initial program 76.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.9
Applied rewrites62.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3262.9
Applied rewrites62.9%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites62.9%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.7
Applied rewrites69.7%
if 2.29999995 < dX.v Initial program 80.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3279.2
Applied rewrites79.2%
Applied rewrites79.2%
Final simplification72.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* t_0 t_0))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.u (floor w)))
(t_4 (+ (pow t_3 2.0) (pow t_0 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (+ (* t_2 t_2) (* t_5 t_5)))
(t_7 (/ 1.0 (sqrt (fmax t_6 (+ t_1 (* t_3 t_3))))))
(t_8 (* t_7 t_5)))
(if (<= dX.v 2.299999952316284)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_4) t_8 (* t_7 t_3))
(if (>= (pow t_2 2.0) t_4)
t_8
(* (/ 1.0 (sqrt (fmax t_6 (+ (* (* t_3 dY.u) (floor w)) t_1)))) 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 = dY_46_v * floorf(h);
float t_1 = t_0 * t_0;
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_u * floorf(w);
float t_4 = powf(t_3, 2.0f) + powf(t_0, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = (t_2 * t_2) + (t_5 * t_5);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, (t_1 + (t_3 * t_3))));
float t_8 = t_7 * t_5;
float tmp_1;
if (dX_46_v <= 2.299999952316284f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_4) {
tmp_2 = t_8;
} else {
tmp_2 = t_7 * t_3;
}
tmp_1 = tmp_2;
} else if (powf(t_2, 2.0f) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf(t_6, (((t_3 * dY_46_u) * floorf(w)) + t_1)))) * t_3;
}
return tmp_1;
}
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(t_0 * t_0) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_u * floor(w)) t_4 = Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) t_7 = Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(t_1 + Float32(t_3 * t_3)) : ((Float32(t_1 + Float32(t_3 * t_3)) != Float32(t_1 + Float32(t_3 * t_3))) ? t_6 : max(t_6, Float32(t_1 + Float32(t_3 * t_3))))))) t_8 = Float32(t_7 * t_5) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(2.299999952316284)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_2 = t_8; else tmp_2 = Float32(t_7 * t_3); end tmp_1 = tmp_2; elseif ((t_2 ^ Float32(2.0)) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(Float32(t_3 * dY_46_u) * floor(w)) + t_1) : ((Float32(Float32(Float32(t_3 * dY_46_u) * floor(w)) + t_1) != Float32(Float32(Float32(t_3 * dY_46_u) * floor(w)) + t_1)) ? t_6 : max(t_6, Float32(Float32(Float32(t_3 * dY_46_u) * floor(w)) + t_1)))))) * t_3); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = dY_46_v * floor(h); t_1 = t_0 * t_0; t_2 = floor(h) * dX_46_v; t_3 = dY_46_u * floor(w); t_4 = (t_3 ^ single(2.0)) + (t_0 ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = (t_2 * t_2) + (t_5 * t_5); t_7 = single(1.0) / sqrt(max(t_6, (t_1 + (t_3 * t_3)))); t_8 = t_7 * t_5; tmp_2 = single(0.0); if (dX_46_v <= single(2.299999952316284)) tmp_3 = single(0.0); if ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_3 = t_8; else tmp_3 = t_7 * t_3; end tmp_2 = tmp_3; elseif ((t_2 ^ single(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = (single(1.0) / sqrt(max(t_6, (((t_3 * dY_46_u) * floor(w)) + t_1)))) * t_3; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := {t\_3}^{2} + {t\_0}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_2 \cdot t\_2 + t\_5 \cdot t\_5\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_1 + t\_3 \cdot t\_3\right)}}\\
t_8 := t\_7 \cdot t\_5\\
\mathbf{if}\;dX.v \leq 2.299999952316284:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;{t\_2}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, \left(t\_3 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor + t\_1\right)}} \cdot t\_3\\
\end{array}
\end{array}
if dX.v < 2.29999995Initial program 76.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.9
Applied rewrites62.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3262.9
Applied rewrites62.9%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites62.9%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3269.7
Applied rewrites69.7%
if 2.29999995 < dX.v Initial program 80.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3279.2
Applied rewrites79.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3279.2
Applied rewrites79.2%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites79.2%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3279.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3279.2
Applied rewrites79.2%
Final simplification72.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor w) 2.0))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.v (floor h)))
(t_4 (pow t_3 2.0))
(t_5 (* (floor w) dX.u))
(t_6
(/
1.0
(sqrt
(fmax (+ (* t_2 t_2) (* t_5 t_5)) (+ (* t_3 t_3) (* t_0 t_0))))))
(t_7 (>= (pow t_2 2.0) (+ (pow t_0 2.0) t_4)))
(t_8
(/
1.0
(sqrt
(fmax
(fma (* (pow (floor h) 2.0) dX.v) dX.v (* (* t_1 dX.u) dX.u))
(fma (* t_1 dY.u) dY.u t_4))))))
(if (<= dX.v 1200000000.0)
(if t_7 (* t_8 t_5) (* t_6 t_0))
(if t_7 (* t_6 t_5) (* t_8 t_0)))))
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 = powf(floorf(w), 2.0f);
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_v * floorf(h);
float t_4 = powf(t_3, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = 1.0f / sqrtf(fmaxf(((t_2 * t_2) + (t_5 * t_5)), ((t_3 * t_3) + (t_0 * t_0))));
int t_7 = powf(t_2, 2.0f) >= (powf(t_0, 2.0f) + t_4);
float t_8 = 1.0f / sqrtf(fmaxf(fmaf((powf(floorf(h), 2.0f) * dX_46_v), dX_46_v, ((t_1 * dX_46_u) * dX_46_u)), fmaf((t_1 * dY_46_u), dY_46_u, t_4)));
float tmp_1;
if (dX_46_v <= 1200000000.0f) {
float tmp_2;
if (t_7) {
tmp_2 = t_8 * t_5;
} else {
tmp_2 = t_6 * t_0;
}
tmp_1 = tmp_2;
} else if (t_7) {
tmp_1 = t_6 * t_5;
} else {
tmp_1 = t_8 * t_0;
}
return tmp_1;
}
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 = floor(w) ^ Float32(2.0) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_v * floor(h)) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) != Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : ((Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) : max(Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)), Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))))))) t_7 = (t_2 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + t_4) t_8 = Float32(Float32(1.0) / sqrt(((fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)) != fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u))) ? fma(Float32(t_1 * dY_46_u), dY_46_u, t_4) : ((fma(Float32(t_1 * dY_46_u), dY_46_u, t_4) != fma(Float32(t_1 * dY_46_u), dY_46_u, t_4)) ? fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)) : max(fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)), fma(Float32(t_1 * dY_46_u), dY_46_u, t_4)))))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(1200000000.0)) tmp_2 = Float32(0.0) if (t_7) tmp_2 = Float32(t_8 * t_5); else tmp_2 = Float32(t_6 * t_0); end tmp_1 = tmp_2; elseif (t_7) tmp_1 = Float32(t_6 * t_5); else tmp_1 = Float32(t_8 * t_0); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {t\_3}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_5 \cdot t\_5, t\_3 \cdot t\_3 + t\_0 \cdot t\_0\right)}}\\
t_7 := {t\_2}^{2} \geq {t\_0}^{2} + t\_4\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v, dX.v, \left(t\_1 \cdot dX.u\right) \cdot dX.u\right), \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, t\_4\right)\right)}}\\
\mathbf{if}\;dX.v \leq 1200000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7:\\
\;\;\;\;t\_8 \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_0\\
\end{array}\\
\mathbf{elif}\;t\_7:\\
\;\;\;\;t\_6 \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_0\\
\end{array}
\end{array}
if dX.v < 1.2e9Initial program 78.2%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3265.8
Applied rewrites65.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3265.8
Applied rewrites65.8%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.v around 0
Applied rewrites48.2%
if 1.2e9 < dX.v Initial program 74.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3274.0
Applied rewrites74.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3274.0
Applied rewrites74.0%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites74.0%
Taylor expanded in dY.v around 0
Applied rewrites69.3%
Final simplification51.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (+ (* t_2 t_2) (* t_1 t_1)))
(t_4 (* (floor h) dX.v))
(t_5 (* t_4 t_4)))
(if (>= (pow t_4 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_0 t_0)) t_3))) t_0)
(*
(/ 1.0 (sqrt (fmax (+ t_5 (* (* (pow (floor w) 2.0) dX.u) dX.u)) t_3)))
t_1))))
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) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = (t_2 * t_2) + (t_1 * t_1);
float t_4 = floorf(h) * dX_46_v;
float t_5 = t_4 * t_4;
float tmp;
if (powf(t_4, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (t_0 * t_0)), t_3))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + ((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u)), t_3))) * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_0 * t_0)) != Float32(t_5 + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_0 * t_0)) : max(Float32(t_5 + Float32(t_0 * t_0)), t_3))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u)) != Float32(t_5 + Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u)) : max(Float32(t_5 + Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u)), t_3))))) * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = dY_46_v * floor(h); t_3 = (t_2 * t_2) + (t_1 * t_1); t_4 = floor(h) * dX_46_v; t_5 = t_4 * t_4; tmp = single(0.0); if ((t_4 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_5 + (t_0 * t_0)), t_3))) * t_0; else tmp = (single(1.0) / sqrt(max((t_5 + (((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u)), t_3))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2 + t\_1 \cdot t\_1\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0 \cdot t\_0, t\_3\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + \left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u, t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.9
Applied rewrites66.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3266.9
Applied rewrites66.9%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites66.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f3266.9
Applied rewrites66.9%
Final simplification66.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) dX.v))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) (* t_1 t_1)))))))
(if (>= (pow t_3 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* t_4 t_0)
(* t_4 t_1))))
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) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) * dX_46_v;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1))));
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * t_0;
} else {
tmp = t_4 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : max(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_4 * t_0); else tmp = Float32(t_4 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = dY_46_v * floor(h); t_3 = floor(h) * dX_46_v; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1)))); tmp = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * t_0; else tmp = t_4 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.9
Applied rewrites66.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3266.9
Applied rewrites66.9%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites66.9%
Final simplification66.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (pow (floor w) 2.0))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.v (floor h)))
(t_4 (pow t_3 2.0))
(t_5 (* (floor w) dX.u)))
(if (>= (pow t_2 2.0) (+ (pow t_0 2.0) t_4))
(*
(/
1.0
(sqrt
(fmax
(fma (* (pow (floor h) 2.0) dX.v) dX.v (* (* t_1 dX.u) dX.u))
(fma (* t_1 dY.u) dY.u t_4))))
t_5)
(*
(/
1.0
(sqrt (fmax (+ (* t_2 t_2) (* t_5 t_5)) (+ (* t_3 t_3) (* t_0 t_0)))))
t_0))))
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 = powf(floorf(w), 2.0f);
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_v * floorf(h);
float t_4 = powf(t_3, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float tmp;
if (powf(t_2, 2.0f) >= (powf(t_0, 2.0f) + t_4)) {
tmp = (1.0f / sqrtf(fmaxf(fmaf((powf(floorf(h), 2.0f) * dX_46_v), dX_46_v, ((t_1 * dX_46_u) * dX_46_u)), fmaf((t_1 * dY_46_u), dY_46_u, t_4)))) * t_5;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_2 * t_2) + (t_5 * t_5)), ((t_3 * t_3) + (t_0 * t_0))))) * t_0;
}
return 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 = floor(w) ^ Float32(2.0) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_v * floor(h)) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + t_4)) tmp = Float32(Float32(Float32(1.0) / sqrt(((fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)) != fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u))) ? fma(Float32(t_1 * dY_46_u), dY_46_u, t_4) : ((fma(Float32(t_1 * dY_46_u), dY_46_u, t_4) != fma(Float32(t_1 * dY_46_u), dY_46_u, t_4)) ? fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)) : max(fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, Float32(Float32(t_1 * dX_46_u) * dX_46_u)), fma(Float32(t_1 * dY_46_u), dY_46_u, t_4)))))) * t_5); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) != Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : ((Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) : max(Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)), Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))))))) * t_0); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {t\_3}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;{t\_2}^{2} \geq {t\_0}^{2} + t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v, dX.v, \left(t\_1 \cdot dX.u\right) \cdot dX.u\right), \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, t\_4\right)\right)}} \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_5 \cdot t\_5, t\_3 \cdot t\_3 + t\_0 \cdot t\_0\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 77.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.9
Applied rewrites66.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3266.9
Applied rewrites66.9%
lift-+.f32N/A
+-commutativeN/A
lower-+.f32N/A
Applied rewrites66.9%
Taylor expanded in dY.v around 0
Applied rewrites43.7%
Final simplification43.9%
herbie shell --seed 2024276
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, u)"
: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 (>= (+ (* (* (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 (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 w) dX.u)) (* (/ 1.0 (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 w) dY.u))))