
(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 (* (floor w) dX.u) 2.0)))
(t_4 (sqrt (fmax t_3 (+ (pow t_1 2.0) (pow t_0 2.0))))))
(if (>= t_3 (+ (* t_0 t_0) (* t_1 t_1)))
(/ (/ t_2 1.0) t_4)
(* (/ 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((floorf(w) * dX_46_u), 2.0f);
float t_4 = sqrtf(fmaxf(t_3, (powf(t_1, 2.0f) + powf(t_0, 2.0f))));
float tmp;
if (t_3 >= ((t_0 * t_0) + (t_1 * t_1))) {
tmp = (t_2 / 1.0f) / t_4;
} 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(floor(w) * dX_46_u) ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? Float32((t_1 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : ((Float32((t_1 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_1 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? t_3 : max(t_3, Float32((t_1 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))))))) tmp = Float32(0.0) if (t_3 >= Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))) tmp = Float32(Float32(t_2 / Float32(1.0)) / t_4); else tmp = Float32(Float32(Float32(1.0) / 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)) + ((floor(w) * dX_46_u) ^ single(2.0)); t_4 = sqrt(max(t_3, ((t_1 ^ single(2.0)) + (t_0 ^ single(2.0))))); tmp = single(0.0); if (t_3 >= ((t_0 * t_0) + (t_1 * t_1))) tmp = (t_2 / single(1.0)) / t_4; 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(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, {t\_1}^{2} + {t\_0}^{2}\right)}\\
\mathbf{if}\;t\_3 \geq t\_0 \cdot t\_0 + t\_1 \cdot t\_1:\\
\;\;\;\;\frac{\frac{t\_2}{1}}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_4} \cdot t\_0\\
\end{array}
\end{array}
Initial program 70.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites70.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3270.8
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3270.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.8
Applied rewrites70.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3270.8
Applied rewrites70.8%
lift-*.f32N/A
pow2N/A
lift-pow.f3270.8
Applied rewrites70.8%
Final simplification70.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (* (pow (floor w) 2.0) dX.u) dX.u))
(t_2 (* dY.u (floor w)))
(t_3 (* dX.v (floor h)))
(t_4 (* t_3 t_3))
(t_5 (* dY.v (floor h)))
(t_6 (+ (* t_5 t_5) (* t_2 t_2)))
(t_7 (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_4) t_6))))
(t_8 (+ (pow t_2 2.0) (pow t_5 2.0)))
(t_9 (* t_7 t_5)))
(if (<= dX.u 5.0)
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) t_8)
(* (/ 1.0 (sqrt (fmax (+ t_4 t_1) t_6))) t_3)
t_9)
(if (>= t_1 t_8) (* t_7 t_3) t_9))))
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 = (powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u;
float t_2 = dY_46_u * floorf(w);
float t_3 = dX_46_v * floorf(h);
float t_4 = t_3 * t_3;
float t_5 = dY_46_v * floorf(h);
float t_6 = (t_5 * t_5) + (t_2 * t_2);
float t_7 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_4), t_6));
float t_8 = powf(t_2, 2.0f) + powf(t_5, 2.0f);
float t_9 = t_7 * t_5;
float tmp_1;
if (dX_46_u <= 5.0f) {
float tmp_2;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= t_8) {
tmp_2 = (1.0f / sqrtf(fmaxf((t_4 + t_1), t_6))) * t_3;
} else {
tmp_2 = t_9;
}
tmp_1 = tmp_2;
} else if (t_1 >= t_8) {
tmp_1 = t_7 * t_3;
} else {
tmp_1 = t_9;
}
return tmp_1;
}
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(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(dX_46_v * floor(h)) t_4 = Float32(t_3 * t_3) t_5 = Float32(dY_46_v * floor(h)) t_6 = Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) t_7 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + t_4) != Float32(Float32(t_0 * t_0) + t_4)) ? t_6 : ((t_6 != t_6) ? Float32(Float32(t_0 * t_0) + t_4) : max(Float32(Float32(t_0 * t_0) + t_4), t_6))))) t_8 = Float32((t_2 ^ Float32(2.0)) + (t_5 ^ Float32(2.0))) t_9 = Float32(t_7 * t_5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(5.0)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= t_8) tmp_2 = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_4 + t_1) != Float32(t_4 + t_1)) ? t_6 : ((t_6 != t_6) ? Float32(t_4 + t_1) : max(Float32(t_4 + t_1), t_6))))) * t_3); else tmp_2 = t_9; end tmp_1 = tmp_2; elseif (t_1 >= t_8) tmp_1 = Float32(t_7 * t_3); else tmp_1 = t_9; 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 = floor(w) * dX_46_u; t_1 = ((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u; t_2 = dY_46_u * floor(w); t_3 = dX_46_v * floor(h); t_4 = t_3 * t_3; t_5 = dY_46_v * floor(h); t_6 = (t_5 * t_5) + (t_2 * t_2); t_7 = single(1.0) / sqrt(max(((t_0 * t_0) + t_4), t_6)); t_8 = (t_2 ^ single(2.0)) + (t_5 ^ single(2.0)); t_9 = t_7 * t_5; tmp_2 = single(0.0); if (dX_46_u <= single(5.0)) tmp_3 = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= t_8) tmp_3 = (single(1.0) / sqrt(max((t_4 + t_1), t_6))) * t_3; else tmp_3 = t_9; end tmp_2 = tmp_3; elseif (t_1 >= t_8) tmp_2 = t_7 * t_3; else tmp_2 = t_9; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := t\_3 \cdot t\_3\\
t_5 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_6 := t\_5 \cdot t\_5 + t\_2 \cdot t\_2\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_4, t\_6\right)}}\\
t_8 := {t\_2}^{2} + {t\_5}^{2}\\
t_9 := t\_7 \cdot t\_5\\
\mathbf{if}\;dX.u \leq 5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_8:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + t\_1, t\_6\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{elif}\;t\_1 \geq t\_8:\\
\;\;\;\;t\_7 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}
\end{array}
if dX.u < 5Initial program 71.2%
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.f3266.0
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3266.0
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3266.0
Applied rewrites66.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-*.f3266.0
Applied rewrites66.0%
if 5 < dX.u Initial program 68.9%
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.f3246.4
Applied rewrites46.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3246.4
Applied rewrites46.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3246.4
Applied rewrites46.4%
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.f3268.9
Applied rewrites68.9%
Final simplification66.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(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 w) dX.u))
(t_5 (* t_0 t_0)))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow t_1 2.0) (pow t_2 2.0)))
(*
(/ 1.0 (sqrt (fmax (+ t_5 (* (* (pow (floor w) 2.0) dX.u) dX.u)) t_3)))
t_0)
(* (/ 1.0 (sqrt (fmax (+ (* t_4 t_4) t_5) t_3))) 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_v * floorf(h);
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(w) * dX_46_u;
float t_5 = t_0 * t_0;
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_5 + ((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u)), t_3))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_4 * t_4) + t_5), t_3))) * 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_v * floor(h)) 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(w) * dX_46_u) t_5 = Float32(t_0 * t_0) tmp = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) 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_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_4 * t_4) + t_5) != Float32(Float32(t_4 * t_4) + t_5)) ? t_3 : ((t_3 != t_3) ? Float32(Float32(t_4 * t_4) + t_5) : max(Float32(Float32(t_4 * t_4) + t_5), t_3))))) * 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_v * floor(h); 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(w) * dX_46_u; t_5 = t_0 * t_0; tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_5 + (((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u)), t_3))) * t_0; else tmp = (single(1.0) / sqrt(max(((t_4 * t_4) + t_5), t_3))) * t_2; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
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 w\right\rfloor \cdot dX.u\\
t_5 := t\_0 \cdot t\_0\\
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;\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\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_5, t\_3\right)}} \cdot t\_2\\
\end{array}
\end{array}
Initial program 70.6%
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.f3261.1
Applied rewrites61.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.1
Applied rewrites61.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.1
Applied rewrites61.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-*.f3261.2
Applied rewrites61.2%
Final simplification61.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_2 t_2)) (+ (* t_1 t_1) (* t_0 t_0)))))))
(if (>= (pow t_2 2.0) (+ (pow t_0 2.0) (pow t_1 2.0)))
(* t_4 t_2)
(* 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 = dY_46_u * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_2 * t_2)), ((t_1 * t_1) + (t_0 * t_0))));
float tmp;
if (powf(t_2, 2.0f) >= (powf(t_0, 2.0f) + powf(t_1, 2.0f))) {
tmp = t_4 * t_2;
} 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(dY_46_u * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)) != Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2))) ? Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) : ((Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) != Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0))) ? Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)) : max(Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)), Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0))))))) tmp = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(t_4 * t_2); 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 = dY_46_u * floor(w); t_1 = dY_46_v * floor(h); t_2 = dX_46_v * floor(h); t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_2 * t_2)), ((t_1 * t_1) + (t_0 * t_0)))); tmp = single(0.0); if ((t_2 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = t_4 * t_2; else tmp = t_4 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_2 \cdot t\_2, t\_1 \cdot t\_1 + t\_0 \cdot t\_0\right)}}\\
\mathbf{if}\;{t\_2}^{2} \geq {t\_0}^{2} + {t\_1}^{2}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 70.6%
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.f3261.1
Applied rewrites61.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.1
Applied rewrites61.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3261.1
Applied rewrites61.1%
Applied rewrites61.1%
Final simplification61.1%
herbie shell --seed 2024271
(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))))