
(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 4 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 (* dY.v (floor h)))
(t_1 (* dY.u (floor w)))
(t_2 (* dX.v (floor h)))
(t_3 (+ (pow t_2 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_4 (sqrt (fmax t_3 (+ (pow t_0 2.0) (pow t_1 2.0))))))
(if (>= t_3 (+ (* t_0 t_0) (* t_1 t_1)))
(* (/ 1.0 t_4) t_2)
(/ 1.0 (/ t_4 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 = dY_46_u * floorf(w);
float t_2 = dX_46_v * floorf(h);
float t_3 = powf(t_2, 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_4 = sqrtf(fmaxf(t_3, (powf(t_0, 2.0f) + powf(t_1, 2.0f))));
float tmp;
if (t_3 >= ((t_0 * t_0) + (t_1 * t_1))) {
tmp = (1.0f / t_4) * t_2;
} else {
tmp = 1.0f / (t_4 / 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(dY_46_u * floor(w)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32((t_2 ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) : ((Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) != Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) ? t_3 : max(t_3, Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))))))) tmp = Float32(0.0) if (t_3 >= Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) tmp = Float32(Float32(Float32(1.0) / t_4) * t_2); else tmp = Float32(Float32(1.0) / Float32(t_4 / 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 = dY_46_u * floor(w); t_2 = dX_46_v * floor(h); t_3 = (t_2 ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_4 = sqrt(max(t_3, ((t_0 ^ single(2.0)) + (t_1 ^ single(2.0))))); tmp = single(0.0); if (t_3 >= ((t_0 * t_0) + (t_1 * t_1))) tmp = (single(1.0) / t_4) * t_2; else tmp = single(1.0) / (t_4 / 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 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := {t\_2}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, {t\_0}^{2} + {t\_1}^{2}\right)}\\
\mathbf{if}\;t\_3 \geq t\_0 \cdot t\_0 + t\_1 \cdot t\_1:\\
\;\;\;\;\frac{1}{t\_4} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{t\_4}{t\_0}}\\
\end{array}
\end{array}
Initial program 71.7%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites71.9%
Applied rewrites71.9%
lift-*.f32N/A
pow2N/A
lift-pow.f3271.9
Applied rewrites71.9%
lift-*.f32N/A
pow2N/A
lift-pow.f3271.9
Applied rewrites71.9%
Final simplification71.9%
(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 (* t_2 t_2))
(t_4 (* dX.v (floor h)))
(t_5 (+ (* t_4 t_4) (* t_0 t_0)))
(t_6 (+ (pow t_4 2.0) (pow t_0 2.0)))
(t_7 (/ 1.0 (sqrt (fmax t_5 (+ (* t_1 t_1) t_3)))))
(t_8 (* t_7 t_1)))
(if (<= dY.u 6000000.0)
(if (>= t_6 (* (* (pow (floor h) 2.0) dY.v) dY.v)) (* t_7 t_4) t_8)
(if (>= t_6 (* (* (pow (floor w) 2.0) dY.u) dY.u))
(* (/ 1.0 (sqrt (fmax t_5 (+ (pow (exp 2.0) (log t_1)) t_3)))) t_4)
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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = t_2 * t_2;
float t_4 = dX_46_v * floorf(h);
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float t_6 = powf(t_4, 2.0f) + powf(t_0, 2.0f);
float t_7 = 1.0f / sqrtf(fmaxf(t_5, ((t_1 * t_1) + t_3)));
float t_8 = t_7 * t_1;
float tmp_1;
if (dY_46_u <= 6000000.0f) {
float tmp_2;
if (t_6 >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp_2 = t_7 * t_4;
} else {
tmp_2 = t_8;
}
tmp_1 = tmp_2;
} else if (t_6 >= ((powf(floorf(w), 2.0f) * dY_46_u) * dY_46_u)) {
tmp_1 = (1.0f / sqrtf(fmaxf(t_5, (powf(expf(2.0f), logf(t_1)) + t_3)))) * t_4;
} else {
tmp_1 = t_8;
}
return tmp_1;
}
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 * t_2) t_4 = Float32(dX_46_v * floor(h)) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) t_6 = Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_7 = Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_1 * t_1) + t_3) : ((Float32(Float32(t_1 * t_1) + t_3) != Float32(Float32(t_1 * t_1) + t_3)) ? t_5 : max(t_5, Float32(Float32(t_1 * t_1) + t_3)))))) t_8 = Float32(t_7 * t_1) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(6000000.0)) tmp_2 = Float32(0.0) if (t_6 >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp_2 = Float32(t_7 * t_4); else tmp_2 = t_8; end tmp_1 = tmp_2; elseif (t_6 >= Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u)) tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32((exp(Float32(2.0)) ^ log(t_1)) + t_3) : ((Float32((exp(Float32(2.0)) ^ log(t_1)) + t_3) != Float32((exp(Float32(2.0)) ^ log(t_1)) + t_3)) ? t_5 : max(t_5, Float32((exp(Float32(2.0)) ^ log(t_1)) + t_3)))))) * t_4); else tmp_1 = t_8; 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = t_2 * t_2; t_4 = dX_46_v * floor(h); t_5 = (t_4 * t_4) + (t_0 * t_0); t_6 = (t_4 ^ single(2.0)) + (t_0 ^ single(2.0)); t_7 = single(1.0) / sqrt(max(t_5, ((t_1 * t_1) + t_3))); t_8 = t_7 * t_1; tmp_2 = single(0.0); if (dY_46_u <= single(6000000.0)) tmp_3 = single(0.0); if (t_6 >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp_3 = t_7 * t_4; else tmp_3 = t_8; end tmp_2 = tmp_3; elseif (t_6 >= (((floor(w) ^ single(2.0)) * dY_46_u) * dY_46_u)) tmp_2 = (single(1.0) / sqrt(max(t_5, ((exp(single(2.0)) ^ log(t_1)) + t_3)))) * t_4; else tmp_2 = t_8; end tmp_4 = tmp_2; 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 \cdot t\_2\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
t_6 := {t\_4}^{2} + {t\_0}^{2}\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_1 \cdot t\_1 + t\_3\right)}}\\
t_8 := t\_7 \cdot t\_1\\
\mathbf{if}\;dY.u \leq 6000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;t\_7 \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq \left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dY.u\right) \cdot dY.u:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, {\left(e^{2}\right)}^{\log t\_1} + t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}
\end{array}
if dY.u < 6e6Initial program 74.8%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.3
Applied rewrites62.3%
lift-*.f32N/A
pow2N/A
lower-pow.f3262.3
Applied rewrites62.3%
lift-*.f32N/A
pow2N/A
lift-pow.f3262.3
Applied rewrites62.3%
Taylor expanded in dY.v around inf
unpow2N/A
unpow2N/A
unswap-sqrN/A
*-commutativeN/A
associate-*r*N/A
associate-*r*N/A
unpow2N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.7
Applied rewrites68.7%
if 6e6 < dY.u Initial program 56.2%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3253.9
Applied rewrites53.9%
lift-*.f32N/A
pow2N/A
lower-pow.f3253.9
Applied rewrites53.9%
lift-*.f32N/A
pow2N/A
lift-pow.f3253.9
Applied rewrites53.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
pow-to-expN/A
*-commutativeN/A
exp-prodN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-pow.f32N/A
lower-exp.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3256.2
Applied rewrites56.2%
Final simplification66.6%
(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 (* dX.v (floor h)))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))))
(t_5 (* t_4 t_1))
(t_6 (+ (pow t_3 2.0) (pow t_0 2.0)))
(t_7 (* t_4 t_3)))
(if (<= dY.u 6000000.0)
(if (>= t_6 (* (* (pow (floor h) 2.0) dY.v) dY.v)) t_7 t_5)
(if (>= t_6 (pow t_2 2.0)) t_7 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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = dX_46_v * floorf(h);
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))));
float t_5 = t_4 * t_1;
float t_6 = powf(t_3, 2.0f) + powf(t_0, 2.0f);
float t_7 = t_4 * t_3;
float tmp_1;
if (dY_46_u <= 6000000.0f) {
float tmp_2;
if (t_6 >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp_2 = t_7;
} else {
tmp_2 = t_5;
}
tmp_1 = tmp_2;
} else if (t_6 >= powf(t_2, 2.0f)) {
tmp_1 = t_7;
} 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(dX_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(dX_46_v * floor(h)) 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_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? 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_1 * t_1) + Float32(t_2 * t_2))))))) t_5 = Float32(t_4 * t_1) t_6 = Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_7 = Float32(t_4 * t_3) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(6000000.0)) tmp_2 = Float32(0.0) if (t_6 >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp_2 = t_7; else tmp_2 = t_5; end tmp_1 = tmp_2; elseif (t_6 >= (t_2 ^ Float32(2.0))) tmp_1 = t_7; 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = dX_46_v * floor(h); t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))); t_5 = t_4 * t_1; t_6 = (t_3 ^ single(2.0)) + (t_0 ^ single(2.0)); t_7 = t_4 * t_3; tmp_2 = single(0.0); if (dY_46_u <= single(6000000.0)) tmp_3 = single(0.0); if (t_6 >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp_3 = t_7; else tmp_3 = t_5; end tmp_2 = tmp_3; elseif (t_6 >= (t_2 ^ single(2.0))) tmp_2 = t_7; else tmp_2 = t_5; end tmp_4 = tmp_2; 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 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
t_5 := t\_4 \cdot t\_1\\
t_6 := {t\_3}^{2} + {t\_0}^{2}\\
t_7 := t\_4 \cdot t\_3\\
\mathbf{if}\;dY.u \leq 6000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_5\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq {t\_2}^{2}:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_5\\
\end{array}
\end{array}
if dY.u < 6e6Initial program 74.8%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.3
Applied rewrites62.3%
lift-*.f32N/A
pow2N/A
lower-pow.f3262.3
Applied rewrites62.3%
lift-*.f32N/A
pow2N/A
lift-pow.f3262.3
Applied rewrites62.3%
Taylor expanded in dY.v around inf
unpow2N/A
unpow2N/A
unswap-sqrN/A
*-commutativeN/A
associate-*r*N/A
associate-*r*N/A
unpow2N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3268.7
Applied rewrites68.7%
if 6e6 < dY.u Initial program 56.2%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3253.9
Applied rewrites53.9%
lift-*.f32N/A
pow2N/A
lower-pow.f3253.9
Applied rewrites53.9%
lift-*.f32N/A
pow2N/A
lift-pow.f3253.9
Applied rewrites53.9%
Applied rewrites53.9%
Final simplification66.3%
(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 (* dX.v (floor h)))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))))))
(if (>= (+ (pow t_3 2.0) (pow t_0 2.0)) (pow t_2 2.0))
(* t_4 t_3)
(* 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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = dX_46_v * floorf(h);
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))));
float tmp;
if ((powf(t_3, 2.0f) + powf(t_0, 2.0f)) >= powf(t_2, 2.0f)) {
tmp = t_4 * t_3;
} 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(dX_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(dX_46_v * floor(h)) 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_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? 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_1 * t_1) + Float32(t_2 * t_2))))))) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= (t_2 ^ Float32(2.0))) tmp = Float32(t_4 * t_3); 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 = dX_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = dX_46_v * floor(h); t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))); tmp = single(0.0); if (((t_3 ^ single(2.0)) + (t_0 ^ single(2.0))) >= (t_2 ^ single(2.0))) tmp = t_4 * t_3; else tmp = t_4 * 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 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;{t\_3}^{2} + {t\_0}^{2} \geq {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 71.7%
Taylor expanded in dY.v around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3260.9
Applied rewrites60.9%
lift-*.f32N/A
pow2N/A
lower-pow.f3260.9
Applied rewrites60.9%
lift-*.f32N/A
pow2N/A
lift-pow.f3260.9
Applied rewrites60.9%
Applied rewrites60.9%
Final simplification60.9%
herbie shell --seed 2024268
(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))))