
(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_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = 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_0;
} else {
tmp = t_6 * t_4;
}
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_0); else tmp = Float32(t_6 * t_4); 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
Sampling outcomes in binary32 precision:
Herbie found 5 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_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = 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_0;
} else {
tmp = t_6 * t_4;
}
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_0); else tmp = Float32(t_6 * t_4); 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow t_2 2.0) (pow t_1 2.0)))
(t_4 (* dX.v (floor h)))
(t_5 (pow t_4 2.0)))
(if (>= (+ (* t_4 t_4) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))
(/ 1.0 (/ (sqrt (fmax (+ (pow t_0 2.0) t_5) t_3)) t_4))
(* (/ 1.0 (pow (fmax (+ (exp (/ 0.0 0.0)) t_5) t_3) 0.5)) 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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf(t_2, 2.0f) + powf(t_1, 2.0f);
float t_4 = dX_46_v * floorf(h);
float t_5 = powf(t_4, 2.0f);
float tmp;
if (((t_4 * t_4) + (t_0 * t_0)) >= ((t_1 * t_1) + (t_2 * t_2))) {
tmp = 1.0f / (sqrtf(fmaxf((powf(t_0, 2.0f) + t_5), t_3)) / t_4);
} else {
tmp = (1.0f / powf(fmaxf((expf((0.0f / 0.0f)) + t_5), t_3), 0.5f)) * t_1;
}
return tmp;
}
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(dY_46_v * floor(h)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((t_2 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = Float32(dX_46_v * floor(h)) t_5 = t_4 ^ Float32(2.0) tmp = Float32(0.0) if (Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) >= Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((t_0 ^ Float32(2.0)) + t_5) != Float32((t_0 ^ Float32(2.0)) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32((t_0 ^ Float32(2.0)) + t_5) : max(Float32((t_0 ^ Float32(2.0)) + t_5), t_3)))) / t_4)); else tmp = Float32(Float32(Float32(1.0) / (((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5) : max(Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5), t_3))) ^ Float32(0.5))) * 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = (t_2 ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = dX_46_v * floor(h); t_5 = t_4 ^ single(2.0); tmp = single(0.0); if (((t_4 * t_4) + (t_0 * t_0)) >= ((t_1 * t_1) + (t_2 * t_2))) tmp = single(1.0) / (sqrt(max(((t_0 ^ single(2.0)) + t_5), t_3)) / t_4); else tmp = (single(1.0) / (max((exp((single(0.0) / single(0.0))) + t_5), t_3) ^ single(0.5))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {t\_2}^{2} + {t\_1}^{2}\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := {t\_4}^{2}\\
\mathbf{if}\;t\_4 \cdot t\_4 + t\_0 \cdot t\_0 \geq t\_1 \cdot t\_1 + t\_2 \cdot t\_2:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({t\_0}^{2} + t\_5, t\_3\right)}}{t\_4}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{{\left(\mathsf{max}\left(e^{\frac{0}{0}} + t\_5, t\_3\right)\right)}^{0.5}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 78.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites78.4%
Applied rewrites78.4%
Final simplification78.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dY.u (floor w)))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (+ (pow t_1 2.0) (pow t_0 2.0)))
(t_4 (* dX.v (floor h)))
(t_5 (pow t_4 2.0))
(t_6 (* (/ 1.0 (pow (fmax (+ (exp (/ 0.0 0.0)) t_5) t_3) 0.5)) t_0))
(t_7
(/
1.0
(/ (sqrt (fmax (+ (pow (* dX.u (floor w)) 2.0) t_5) t_3)) t_4))))
(if (<= dX.u 12.0)
(if (>= t_5 t_2) t_7 t_6)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_2) t_7 t_6))))
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 = dY_46_u * floorf(w);
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = powf(t_1, 2.0f) + powf(t_0, 2.0f);
float t_4 = dX_46_v * floorf(h);
float t_5 = powf(t_4, 2.0f);
float t_6 = (1.0f / powf(fmaxf((expf((0.0f / 0.0f)) + t_5), t_3), 0.5f)) * t_0;
float t_7 = 1.0f / (sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + t_5), t_3)) / t_4);
float tmp_1;
if (dX_46_u <= 12.0f) {
float tmp_2;
if (t_5 >= t_2) {
tmp_2 = t_7;
} else {
tmp_2 = t_6;
}
tmp_1 = tmp_2;
} else if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_2) {
tmp_1 = t_7;
} else {
tmp_1 = t_6;
}
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(dY_46_u * floor(w)) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32((t_1 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = Float32(dX_46_v * floor(h)) t_5 = t_4 ^ Float32(2.0) t_6 = Float32(Float32(Float32(1.0) / (((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5) : max(Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_5), t_3))) ^ Float32(0.5))) * t_0) t_7 = Float32(Float32(1.0) / Float32(sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_5) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_5) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_5), t_3)))) / t_4)) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(12.0)) tmp_2 = Float32(0.0) if (t_5 >= t_2) tmp_2 = t_7; else tmp_2 = t_6; end tmp_1 = tmp_2; elseif (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_2) tmp_1 = t_7; else tmp_1 = t_6; 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 = dY_46_u * floor(w); t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = (t_1 ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = dX_46_v * floor(h); t_5 = t_4 ^ single(2.0); t_6 = (single(1.0) / (max((exp((single(0.0) / single(0.0))) + t_5), t_3) ^ single(0.5))) * t_0; t_7 = single(1.0) / (sqrt(max((((dX_46_u * floor(w)) ^ single(2.0)) + t_5), t_3)) / t_4); tmp_2 = single(0.0); if (dX_46_u <= single(12.0)) tmp_3 = single(0.0); if (t_5 >= t_2) tmp_3 = t_7; else tmp_3 = t_6; end tmp_2 = tmp_3; elseif ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_2) tmp_2 = t_7; else tmp_2 = t_6; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := {t\_1}^{2} + {t\_0}^{2}\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := {t\_4}^{2}\\
t_6 := \frac{1}{{\left(\mathsf{max}\left(e^{\frac{0}{0}} + t\_5, t\_3\right)\right)}^{0.5}} \cdot t\_0\\
t_7 := \frac{1}{\frac{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + t\_5, t\_3\right)}}{t\_4}}\\
\mathbf{if}\;dX.u \leq 12:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_2:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{elif}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_2:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.u < 12Initial program 79.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites79.4%
Applied rewrites79.4%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3272.3
Applied rewrites72.3%
if 12 < dX.u Initial program 75.1%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites75.2%
Applied rewrites75.2%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3273.6
Applied rewrites73.6%
Final simplification72.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.v (floor h)))
(t_2 (pow t_1 2.0))
(t_3 (* dY.v (floor h)))
(t_4 (+ (pow t_0 2.0) (pow t_3 2.0))))
(if (>= t_2 (+ (* t_3 t_3) (* t_0 t_0)))
(/ 1.0 (/ (sqrt (fmax (+ (pow (* dX.u (floor w)) 2.0) t_2) t_4)) t_1))
(* (/ 1.0 (pow (fmax (+ (exp (/ 0.0 0.0)) t_2) t_4) 0.5)) 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_u * floorf(w);
float t_1 = dX_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f);
float t_3 = dY_46_v * floorf(h);
float t_4 = powf(t_0, 2.0f) + powf(t_3, 2.0f);
float tmp;
if (t_2 >= ((t_3 * t_3) + (t_0 * t_0))) {
tmp = 1.0f / (sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + t_2), t_4)) / t_1);
} else {
tmp = (1.0f / powf(fmaxf((expf((0.0f / 0.0f)) + t_2), t_4), 0.5f)) * t_3;
}
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 = Float32(dX_46_v * floor(h)) t_2 = t_1 ^ Float32(2.0) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32((t_0 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) tmp = Float32(0.0) if (t_2 >= Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2)) ? t_4 : ((t_4 != t_4) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2), t_4)))) / t_1)); else tmp = Float32(Float32(Float32(1.0) / (((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2)) ? t_4 : ((t_4 != t_4) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2) : max(Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2), t_4))) ^ Float32(0.5))) * t_3); 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); t_1 = dX_46_v * floor(h); t_2 = t_1 ^ single(2.0); t_3 = dY_46_v * floor(h); t_4 = (t_0 ^ single(2.0)) + (t_3 ^ single(2.0)); tmp = single(0.0); if (t_2 >= ((t_3 * t_3) + (t_0 * t_0))) tmp = single(1.0) / (sqrt(max((((dX_46_u * floor(w)) ^ single(2.0)) + t_2), t_4)) / t_1); else tmp = (single(1.0) / (max((exp((single(0.0) / single(0.0))) + t_2), t_4) ^ single(0.5))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2}\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {t\_0}^{2} + {t\_3}^{2}\\
\mathbf{if}\;t\_2 \geq t\_3 \cdot t\_3 + t\_0 \cdot t\_0:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + t\_2, t\_4\right)}}{t\_1}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{{\left(\mathsf{max}\left(e^{\frac{0}{0}} + t\_2, t\_4\right)\right)}^{0.5}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 78.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites78.4%
Applied rewrites78.4%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3269.2
Applied rewrites69.2%
Final simplification69.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (+ (pow (* dY.u (floor w)) 2.0) (pow t_0 2.0)))
(t_2 (* dX.v (floor h)))
(t_3 (pow t_2 2.0))
(t_4
(/
1.0
(/ (sqrt (fmax (+ (pow (* dX.u (floor w)) 2.0) t_3) t_1)) t_2)))
(t_5 (* (/ 1.0 (pow (fmax (+ (exp (/ 0.0 0.0)) t_3) t_1) 0.5)) t_0)))
(if (<= dY.u 1400000.0)
(if (>= t_3 (* (* (pow (floor h) 2.0) dY.v) dY.v)) t_4 t_5)
(if (>= t_3 (* (* (pow (floor w) 2.0) dY.u) dY.u)) t_4 t_5))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_v * floorf(h);
float t_1 = powf((dY_46_u * floorf(w)), 2.0f) + powf(t_0, 2.0f);
float t_2 = dX_46_v * floorf(h);
float t_3 = powf(t_2, 2.0f);
float t_4 = 1.0f / (sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + t_3), t_1)) / t_2);
float t_5 = (1.0f / powf(fmaxf((expf((0.0f / 0.0f)) + t_3), t_1), 0.5f)) * t_0;
float tmp_1;
if (dY_46_u <= 1400000.0f) {
float tmp_2;
if (t_3 >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp_2 = t_4;
} else {
tmp_2 = t_5;
}
tmp_1 = tmp_2;
} else if (t_3 >= ((powf(floorf(w), 2.0f) * dY_46_u) * dY_46_u)) {
tmp_1 = t_4;
} else {
tmp_1 = t_5;
}
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((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(dX_46_v * floor(h)) t_3 = t_2 ^ Float32(2.0) t_4 = Float32(Float32(1.0) / Float32(sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_3) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_3)) ? t_1 : ((t_1 != t_1) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_3) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_3), t_1)))) / t_2)) t_5 = Float32(Float32(Float32(1.0) / (((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_3) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_3)) ? t_1 : ((t_1 != t_1) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_3) : max(Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_3), t_1))) ^ Float32(0.5))) * t_0) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(1400000.0)) tmp_2 = Float32(0.0) if (t_3 >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp_2 = t_4; else tmp_2 = t_5; end tmp_1 = tmp_2; elseif (t_3 >= Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u)) tmp_1 = t_4; else tmp_1 = t_5; 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 = ((dY_46_u * floor(w)) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = dX_46_v * floor(h); t_3 = t_2 ^ single(2.0); t_4 = single(1.0) / (sqrt(max((((dX_46_u * floor(w)) ^ single(2.0)) + t_3), t_1)) / t_2); t_5 = (single(1.0) / (max((exp((single(0.0) / single(0.0))) + t_3), t_1) ^ single(0.5))) * t_0; tmp_2 = single(0.0); if (dY_46_u <= single(1400000.0)) tmp_3 = single(0.0); if (t_3 >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp_3 = t_4; else tmp_3 = t_5; end tmp_2 = tmp_3; elseif (t_3 >= (((floor(w) ^ single(2.0)) * dY_46_u) * dY_46_u)) tmp_2 = t_4; else tmp_2 = t_5; 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(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {t\_0}^{2}\\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := {t\_2}^{2}\\
t_4 := \frac{1}{\frac{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + t\_3, t\_1\right)}}{t\_2}}\\
t_5 := \frac{1}{{\left(\mathsf{max}\left(e^{\frac{0}{0}} + t\_3, t\_1\right)\right)}^{0.5}} \cdot t\_0\\
\mathbf{if}\;dY.u \leq 1400000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_3 \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_5\\
\end{array}\\
\mathbf{elif}\;t\_3 \geq \left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dY.u\right) \cdot dY.u:\\
\;\;\;\;t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_5\\
\end{array}
\end{array}
if dY.u < 1.4e6Initial program 79.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites79.4%
Applied rewrites79.4%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3270.4
Applied rewrites70.4%
Taylor expanded in dY.v around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.3
Applied rewrites66.3%
if 1.4e6 < dY.u Initial program 73.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites73.1%
Applied rewrites73.1%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3262.3
Applied rewrites62.3%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.3
Applied rewrites62.3%
Final simplification65.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dX.v (floor h)))
(t_2 (pow t_1 2.0))
(t_3 (+ (pow (* dY.u (floor w)) 2.0) (pow t_0 2.0))))
(if (>= t_2 (* (* (pow (floor h) 2.0) dY.v) dY.v))
(/ 1.0 (/ (sqrt (fmax (+ (pow (* dX.u (floor w)) 2.0) t_2) t_3)) t_1))
(* (/ 1.0 (pow (fmax (+ (exp (/ 0.0 0.0)) t_2) t_3) 0.5)) 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_v * floorf(h);
float t_1 = dX_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f);
float t_3 = powf((dY_46_u * floorf(w)), 2.0f) + powf(t_0, 2.0f);
float tmp;
if (t_2 >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp = 1.0f / (sqrtf(fmaxf((powf((dX_46_u * floorf(w)), 2.0f) + t_2), t_3)) / t_1);
} else {
tmp = (1.0f / powf(fmaxf((expf((0.0f / 0.0f)) + t_2), t_3), 0.5f)) * 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_v * floor(h)) t_1 = Float32(dX_46_v * floor(h)) t_2 = t_1 ^ Float32(2.0) t_3 = Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) tmp = Float32(0.0) if (t_2 >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2) != Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2)) ? t_3 : ((t_3 != t_3) ? Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2) : max(Float32((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) + t_2), t_3)))) / t_1)); else tmp = Float32(Float32(Float32(1.0) / (((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2)) ? t_3 : ((t_3 != t_3) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2) : max(Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_2), t_3))) ^ Float32(0.5))) * t_0); 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_v * floor(h); t_1 = dX_46_v * floor(h); t_2 = t_1 ^ single(2.0); t_3 = ((dY_46_u * floor(w)) ^ single(2.0)) + (t_0 ^ single(2.0)); tmp = single(0.0); if (t_2 >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp = single(1.0) / (sqrt(max((((dX_46_u * floor(w)) ^ single(2.0)) + t_2), t_3)) / t_1); else tmp = (single(1.0) / (max((exp((single(0.0) / single(0.0))) + t_2), t_3) ^ single(0.5))) * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2}\\
t_3 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {t\_0}^{2}\\
\mathbf{if}\;t\_2 \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + t\_2, t\_3\right)}}{t\_1}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{{\left(\mathsf{max}\left(e^{\frac{0}{0}} + t\_2, t\_3\right)\right)}^{0.5}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 78.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites78.4%
Applied rewrites78.4%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3269.2
Applied rewrites69.2%
Taylor expanded in dY.v around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3263.7
Applied rewrites63.7%
Final simplification63.7%
herbie shell --seed 2024244
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, v)"
: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 h) dX.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 h) dY.v))))