
(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 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_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 (* dX.u (floor w)) 2.0))
(t_1 (+ t_0 (pow (* (floor h) dX.v) 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow (* dY.v (floor h)) 2.0) (pow t_2 2.0))))
(if (>= t_1 t_3)
(*
(/
(floor w)
(sqrt (fmax (+ (* (* (pow (floor h) 2.0) dX.v) dX.v) t_0) t_3)))
dX.u)
(* t_2 (/ 1.0 (sqrt (fmax 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 = powf((dX_46_u * floorf(w)), 2.0f);
float t_1 = t_0 + powf((floorf(h) * dX_46_v), 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_2, 2.0f);
float tmp;
if (t_1 >= t_3) {
tmp = (floorf(w) / sqrtf(fmaxf((((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) + t_0), t_3))) * dX_46_u;
} else {
tmp = t_2 * (1.0f / sqrtf(fmaxf(t_1, t_3)));
}
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)) ^ Float32(2.0) t_1 = Float32(t_0 + (Float32(floor(h) * dX_46_v) ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(floor(w) / sqrt(((Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_0) != Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_0)) ? t_3 : ((t_3 != t_3) ? Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_0) : max(Float32(Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) + t_0), t_3))))) * dX_46_u); else tmp = Float32(t_2 * Float32(Float32(1.0) / sqrt(((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, 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 = (dX_46_u * floor(w)) ^ single(2.0); t_1 = t_0 + ((floor(h) * dX_46_v) ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_2 ^ single(2.0)); tmp = single(0.0); if (t_1 >= t_3) tmp = (floor(w) / sqrt(max(((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) + t_0), t_3))) * dX_46_u; else tmp = t_2 * (single(1.0) / sqrt(max(t_1, t_3))); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := t\_0 + {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_2}^{2}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v + t\_0, t\_3\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\end{array}
\end{array}
Initial program 71.0%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
*-rgt-identityN/A
Applied rewrites71.1%
lift-pow.f32N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
*-commutativeN/A
pow2N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3271.1
Applied rewrites71.1%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-+.f3271.1
Applied rewrites71.1%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
lift-+.f3271.1
Applied rewrites71.1%
Final simplification71.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dX.v))
(t_4 (* dY.v (floor h)))
(t_5 (+ (pow t_4 2.0) (pow t_0 2.0)))
(t_6 (+ (* t_4 t_4) (* t_0 t_0)))
(t_7 (/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_2) t_6))))
(t_8 (* t_7 t_0))
(t_9 (pow (floor h) 2.0)))
(if (<= dX.v 10000.0)
(if (>= (pow t_1 2.0) t_5) (* t_7 t_1) t_8)
(if (>= (* (* t_9 dX.v) dX.v) t_5)
(* (/ 1.0 (sqrt (fmax (+ (* (* dX.v dX.v) t_9) t_2) t_6))) t_1)
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 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dX_46_v;
float t_4 = dY_46_v * floorf(h);
float t_5 = powf(t_4, 2.0f) + powf(t_0, 2.0f);
float t_6 = (t_4 * t_4) + (t_0 * t_0);
float t_7 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_2), t_6));
float t_8 = t_7 * t_0;
float t_9 = powf(floorf(h), 2.0f);
float tmp_1;
if (dX_46_v <= 10000.0f) {
float tmp_2;
if (powf(t_1, 2.0f) >= t_5) {
tmp_2 = t_7 * t_1;
} else {
tmp_2 = t_8;
}
tmp_1 = tmp_2;
} else if (((t_9 * dX_46_v) * dX_46_v) >= t_5) {
tmp_1 = (1.0f / sqrtf(fmaxf((((dX_46_v * dX_46_v) * t_9) + t_2), t_6))) * t_1;
} 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(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(dY_46_v * floor(h)) t_5 = Float32((t_4 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_6 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) t_7 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + t_2) != Float32(Float32(t_3 * t_3) + t_2)) ? t_6 : ((t_6 != t_6) ? Float32(Float32(t_3 * t_3) + t_2) : max(Float32(Float32(t_3 * t_3) + t_2), t_6))))) t_8 = Float32(t_7 * t_0) t_9 = floor(h) ^ Float32(2.0) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(10000.0)) tmp_2 = Float32(0.0) if ((t_1 ^ Float32(2.0)) >= t_5) tmp_2 = Float32(t_7 * t_1); else tmp_2 = t_8; end tmp_1 = tmp_2; elseif (Float32(Float32(t_9 * dX_46_v) * dX_46_v) >= t_5) tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(dX_46_v * dX_46_v) * t_9) + t_2) != Float32(Float32(Float32(dX_46_v * dX_46_v) * t_9) + t_2)) ? t_6 : ((t_6 != t_6) ? Float32(Float32(Float32(dX_46_v * dX_46_v) * t_9) + t_2) : max(Float32(Float32(Float32(dX_46_v * dX_46_v) * t_9) + t_2), t_6))))) * t_1); 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 = dY_46_u * floor(w); t_1 = dX_46_u * floor(w); t_2 = t_1 * t_1; t_3 = floor(h) * dX_46_v; t_4 = dY_46_v * floor(h); t_5 = (t_4 ^ single(2.0)) + (t_0 ^ single(2.0)); t_6 = (t_4 * t_4) + (t_0 * t_0); t_7 = single(1.0) / sqrt(max(((t_3 * t_3) + t_2), t_6)); t_8 = t_7 * t_0; t_9 = floor(h) ^ single(2.0); tmp_2 = single(0.0); if (dX_46_v <= single(10000.0)) tmp_3 = single(0.0); if ((t_1 ^ single(2.0)) >= t_5) tmp_3 = t_7 * t_1; else tmp_3 = t_8; end tmp_2 = tmp_3; elseif (((t_9 * dX_46_v) * dX_46_v) >= t_5) tmp_2 = (single(1.0) / sqrt(max((((dX_46_v * dX_46_v) * t_9) + t_2), t_6))) * t_1; else tmp_2 = t_8; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_5 := {t\_4}^{2} + {t\_0}^{2}\\
t_6 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_2, t\_6\right)}}\\
t_8 := t\_7 \cdot t\_0\\
t_9 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;dX.v \leq 10000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_1}^{2} \geq t\_5:\\
\;\;\;\;t\_7 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}\\
\mathbf{elif}\;\left(t\_9 \cdot dX.v\right) \cdot dX.v \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(dX.v \cdot dX.v\right) \cdot t\_9 + t\_2, t\_6\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}
\end{array}
if dX.v < 1e4Initial program 73.0%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
Taylor expanded in dX.u around inf
exp-to-powN/A
*-commutativeN/A
exp-to-powN/A
*-commutativeN/A
exp-sumN/A
distribute-lft-outN/A
*-commutativeN/A
exp-prodN/A
prod-expN/A
rem-exp-logN/A
rem-exp-logN/A
*-commutativeN/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3267.8
Applied rewrites67.8%
if 1e4 < dX.v Initial program 62.8%
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.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lower-pow.f3260.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3260.8
Applied rewrites60.8%
Final simplification66.5%
(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_1 t_1))
(t_3 (* dX.u (floor w)))
(t_4 (+ (pow t_0 2.0) (pow t_1 2.0)))
(t_5 (* (floor h) dX.v))
(t_6 (+ (* t_5 t_5) (* t_3 t_3)))
(t_7 (/ 1.0 (sqrt (fmax t_6 (+ (* t_0 t_0) t_2)))))
(t_8 (* t_7 t_3))
(t_9 (pow (floor h) 2.0)))
(if (<= dX.v 10000.0)
(if (>= (pow t_3 2.0) t_4) t_8 (* t_7 t_1))
(if (>= (* (* t_9 dX.v) dX.v) t_4)
t_8
(* (/ 1.0 (sqrt (fmax t_6 (+ (* (* t_9 dY.v) dY.v) t_2)))) 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 = dY_46_v * floorf(h);
float t_1 = dY_46_u * floorf(w);
float t_2 = t_1 * t_1;
float t_3 = dX_46_u * floorf(w);
float t_4 = powf(t_0, 2.0f) + powf(t_1, 2.0f);
float t_5 = floorf(h) * dX_46_v;
float t_6 = (t_5 * t_5) + (t_3 * t_3);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, ((t_0 * t_0) + t_2)));
float t_8 = t_7 * t_3;
float t_9 = powf(floorf(h), 2.0f);
float tmp_1;
if (dX_46_v <= 10000.0f) {
float tmp_2;
if (powf(t_3, 2.0f) >= t_4) {
tmp_2 = t_8;
} else {
tmp_2 = t_7 * t_1;
}
tmp_1 = tmp_2;
} else if (((t_9 * dX_46_v) * dX_46_v) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf(t_6, (((t_9 * dY_46_v) * dY_46_v) + t_2)))) * t_1;
}
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(t_1 * t_1) t_3 = Float32(dX_46_u * floor(w)) t_4 = Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(t_5 * t_5) + Float32(t_3 * t_3)) t_7 = Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(t_0 * t_0) + t_2) : ((Float32(Float32(t_0 * t_0) + t_2) != Float32(Float32(t_0 * t_0) + t_2)) ? t_6 : max(t_6, Float32(Float32(t_0 * t_0) + t_2)))))) t_8 = Float32(t_7 * t_3) t_9 = floor(h) ^ Float32(2.0) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(10000.0)) tmp_2 = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= t_4) tmp_2 = t_8; else tmp_2 = Float32(t_7 * t_1); end tmp_1 = tmp_2; elseif (Float32(Float32(t_9 * dX_46_v) * dX_46_v) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(Float32(t_9 * dY_46_v) * dY_46_v) + t_2) : ((Float32(Float32(Float32(t_9 * dY_46_v) * dY_46_v) + t_2) != Float32(Float32(Float32(t_9 * dY_46_v) * dY_46_v) + t_2)) ? t_6 : max(t_6, Float32(Float32(Float32(t_9 * dY_46_v) * dY_46_v) + t_2)))))) * t_1); 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_1 * t_1; t_3 = dX_46_u * floor(w); t_4 = (t_0 ^ single(2.0)) + (t_1 ^ single(2.0)); t_5 = floor(h) * dX_46_v; t_6 = (t_5 * t_5) + (t_3 * t_3); t_7 = single(1.0) / sqrt(max(t_6, ((t_0 * t_0) + t_2))); t_8 = t_7 * t_3; t_9 = floor(h) ^ single(2.0); tmp_2 = single(0.0); if (dX_46_v <= single(10000.0)) tmp_3 = single(0.0); if ((t_3 ^ single(2.0)) >= t_4) tmp_3 = t_8; else tmp_3 = t_7 * t_1; end tmp_2 = tmp_3; elseif (((t_9 * dX_46_v) * dX_46_v) >= t_4) tmp_2 = t_8; else tmp_2 = (single(1.0) / sqrt(max(t_6, (((t_9 * dY_46_v) * dY_46_v) + t_2)))) * t_1; 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\_1 \cdot t\_1\\
t_3 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_4 := {t\_0}^{2} + {t\_1}^{2}\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_5 \cdot t\_5 + t\_3 \cdot t\_3\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_0 \cdot t\_0 + t\_2\right)}}\\
t_8 := t\_7 \cdot t\_3\\
t_9 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;dX.v \leq 10000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_3}^{2} \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;\left(t\_9 \cdot dX.v\right) \cdot dX.v \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, \left(t\_9 \cdot dY.v\right) \cdot dY.v + t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
if dX.v < 1e4Initial program 73.0%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
Taylor expanded in dX.u around inf
exp-to-powN/A
*-commutativeN/A
exp-to-powN/A
*-commutativeN/A
exp-sumN/A
distribute-lft-outN/A
*-commutativeN/A
exp-prodN/A
prod-expN/A
rem-exp-logN/A
rem-exp-logN/A
*-commutativeN/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3267.8
Applied rewrites67.8%
if 1e4 < dX.v Initial program 62.8%
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.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lower-pow.f3260.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3260.8
Applied rewrites60.8%
Final simplification66.5%
(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.u (floor w)))
(t_3 (+ (pow t_0 2.0) (pow t_1 2.0)))
(t_4 (* (floor h) dX.v))
(t_5
(/
1.0
(sqrt
(fmax (+ (* t_4 t_4) (* t_2 t_2)) (+ (* t_0 t_0) (* t_1 t_1))))))
(t_6 (* t_5 t_1))
(t_7 (* t_5 t_2)))
(if (<= dX.v 10000.0)
(if (>= (pow t_2 2.0) t_3) t_7 t_6)
(if (>= (pow t_4 2.0) t_3) 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 = dX_46_u * floorf(w);
float t_3 = powf(t_0, 2.0f) + powf(t_1, 2.0f);
float t_4 = floorf(h) * dX_46_v;
float t_5 = 1.0f / sqrtf(fmaxf(((t_4 * t_4) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1))));
float t_6 = t_5 * t_1;
float t_7 = t_5 * t_2;
float tmp_1;
if (dX_46_v <= 10000.0f) {
float tmp_2;
if (powf(t_2, 2.0f) >= t_3) {
tmp_2 = t_7;
} else {
tmp_2 = t_6;
}
tmp_1 = tmp_2;
} else if (powf(t_4, 2.0f) >= t_3) {
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(dX_46_u * floor(w)) t_3 = Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)) != Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2))) ? Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) : ((Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) != Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) ? Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)) : max(Float32(Float32(t_4 * t_4) + Float32(t_2 * t_2)), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))))))) t_6 = Float32(t_5 * t_1) t_7 = Float32(t_5 * t_2) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(10000.0)) tmp_2 = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= t_3) tmp_2 = t_7; else tmp_2 = t_6; end tmp_1 = tmp_2; elseif ((t_4 ^ Float32(2.0)) >= t_3) 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 = dX_46_u * floor(w); t_3 = (t_0 ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = floor(h) * dX_46_v; t_5 = single(1.0) / sqrt(max(((t_4 * t_4) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1)))); t_6 = t_5 * t_1; t_7 = t_5 * t_2; tmp_2 = single(0.0); if (dX_46_v <= single(10000.0)) tmp_3 = single(0.0); if ((t_2 ^ single(2.0)) >= t_3) tmp_3 = t_7; else tmp_3 = t_6; end tmp_2 = tmp_3; elseif ((t_4 ^ single(2.0)) >= t_3) 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 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {t\_0}^{2} + {t\_1}^{2}\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_2 \cdot t\_2, t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}}\\
t_6 := t\_5 \cdot t\_1\\
t_7 := t\_5 \cdot t\_2\\
\mathbf{if}\;dX.v \leq 10000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_2}^{2} \geq t\_3:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{elif}\;{t\_4}^{2} \geq t\_3:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.v < 1e4Initial program 73.0%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
Taylor expanded in dX.u around inf
exp-to-powN/A
*-commutativeN/A
exp-to-powN/A
*-commutativeN/A
exp-sumN/A
distribute-lft-outN/A
*-commutativeN/A
exp-prodN/A
prod-expN/A
rem-exp-logN/A
rem-exp-logN/A
*-commutativeN/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3267.8
Applied rewrites67.8%
if 1e4 < dX.v Initial program 62.8%
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.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3260.8
Applied rewrites60.8%
lift-*.f32N/A
pow2N/A
lower-pow.f3260.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.8
Applied rewrites60.8%
Applied rewrites60.8%
Final simplification66.5%
(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 (* (floor h) dX.v))
(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_1 2.0) (pow t_2 2.0)))
(* t_4 t_0)
(* t_4 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 = dX_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = floorf(h) * dX_46_v;
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_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * t_0;
} else {
tmp = t_4 * t_2;
}
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(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_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 ((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_2); 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 = floor(h) * dX_46_v; 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_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * t_0; else tmp = t_4 * t_2; 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 := \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\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\end{array}
\end{array}
Initial program 71.0%
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.f3261.4
Applied rewrites61.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.4
Applied rewrites61.4%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.4
Applied rewrites61.4%
Applied rewrites61.4%
Final simplification61.4%
herbie shell --seed 2024273
(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))))