
(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 (* (floor h) dX.v))
(t_1 (* dY.v (floor h)))
(t_2 (+ (pow t_1 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_3 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ t_0 t_4) (/ t_1 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 = dY_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_3 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 / t_4;
} else {
tmp = t_1 / 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(dY_46_v * floor(h)) t_2 = Float32((t_1 ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_3 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? t_2 : ((t_2 != t_2) ? t_3 : max(t_3, t_2)))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 / t_4); else tmp = Float32(t_1 / 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 = dY_46_v * floor(h); t_2 = (t_1 ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_3 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = sqrt(max(t_3, t_2)); tmp = single(0.0); if (t_3 >= t_2) tmp = t_0 / t_4; else tmp = t_1 / 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 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites77.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.7
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.7
lift-+.f32N/A
+-commutativeN/A
Applied rewrites77.7%
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-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-+.f3277.7
Applied rewrites77.7%
Applied rewrites77.8%
Final simplification77.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.v (floor h)))
(t_2 (+ (pow t_1 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_3 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (* t_0 (/ 1.0 t_4)) (/ t_1 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 = dY_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_3 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 * (1.0f / t_4);
} else {
tmp = t_1 / 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(dY_46_v * floor(h)) t_2 = Float32((t_1 ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_3 = Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = sqrt(((t_3 != t_3) ? t_2 : ((t_2 != t_2) ? t_3 : max(t_3, t_2)))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 * Float32(Float32(1.0) / t_4)); else tmp = Float32(t_1 / 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 = dY_46_v * floor(h); t_2 = (t_1 ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_3 = ((floor(w) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = sqrt(max(t_3, t_2)); tmp = single(0.0); if (t_3 >= t_2) tmp = t_0 * (single(1.0) / t_4); else tmp = t_1 / 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 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;t\_0 \cdot \frac{1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites77.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.7
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3277.7
lift-+.f32N/A
+-commutativeN/A
Applied rewrites77.7%
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-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lift-+.f3277.7
Applied rewrites77.7%
lift-/.f32N/A
clear-numN/A
associate-/r/N/A
metadata-evalN/A
neg-mul-1N/A
neg-sub0N/A
lower--.f3277.7
lift-*.f32N/A
*-commutativeN/A
lower-*.f3277.7
Applied rewrites77.7%
Final simplification77.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* t_0 t_0))
(t_2 (* dY.u (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5 (+ (* t_4 t_4) (* t_3 t_3))))
(if (>= (pow t_3 2.0) (+ (pow t_0 2.0) (pow t_2 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ t_1 (* t_2 t_2))))) t_4)
(*
(/
1.0
(sqrt (fmax t_5 (+ (* (* (* dY.u dY.u) (floor w)) (floor w)) t_1))))
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 = t_0 * t_0;
float t_2 = dY_46_u * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = (t_4 * t_4) + (t_3 * t_3);
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_0, 2.0f) + powf(t_2, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(t_5, (t_1 + (t_2 * t_2))))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((((dY_46_u * dY_46_u) * floorf(w)) * floorf(w)) + t_1)))) * 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(t_0 * t_0) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_3 * t_3)) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_1 + Float32(t_2 * t_2)) : ((Float32(t_1 + Float32(t_2 * t_2)) != Float32(t_1 + Float32(t_2 * t_2))) ? t_5 : max(t_5, Float32(t_1 + Float32(t_2 * t_2))))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * floor(w)) * floor(w)) + t_1) : ((Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * floor(w)) * floor(w)) + t_1) != Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * floor(w)) * floor(w)) + t_1)) ? t_5 : max(t_5, Float32(Float32(Float32(Float32(dY_46_u * dY_46_u) * floor(w)) * floor(w)) + t_1)))))) * 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 = t_0 * t_0; t_2 = dY_46_u * floor(w); t_3 = floor(w) * dX_46_u; t_4 = floor(h) * dX_46_v; t_5 = (t_4 * t_4) + (t_3 * t_3); tmp = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(t_5, (t_1 + (t_2 * t_2))))) * t_4; else tmp = (single(1.0) / sqrt(max(t_5, ((((dY_46_u * dY_46_u) * floor(w)) * floor(w)) + t_1)))) * 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 := t\_0 \cdot t\_0\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_4 \cdot t\_4 + t\_3 \cdot t\_3\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_0}^{2} + {t\_2}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_1 + t\_2 \cdot t\_2\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor + t\_1\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 77.6%
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.f3264.3
Applied rewrites64.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3264.3
Applied rewrites64.3%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites64.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3264.3
Applied rewrites64.3%
Final simplification64.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.u (floor w)))
(t_2 (* (floor w) dX.u))
(t_3 (* dY.v (floor h)))
(t_4 (+ (* t_3 t_3) (* t_1 t_1)))
(t_5 (* t_0 t_0)))
(if (>= (pow t_2 2.0) (+ (pow t_3 2.0) (pow t_1 2.0)))
(*
(/
1.0
(sqrt (fmax (+ (* (* (* dX.u dX.u) (floor w)) (floor w)) t_5) t_4)))
t_0)
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_2 t_2)) t_4))) 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 = floorf(h) * dX_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = floorf(w) * dX_46_u;
float t_3 = dY_46_v * floorf(h);
float t_4 = (t_3 * t_3) + (t_1 * t_1);
float t_5 = t_0 * t_0;
float tmp;
if (powf(t_2, 2.0f) >= (powf(t_3, 2.0f) + powf(t_1, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(((((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)) + t_5), t_4))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (t_2 * t_2)), t_4))) * t_3;
}
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(dY_46_u * floor(w)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)) t_5 = Float32(t_0 * t_0) tmp = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= Float32((t_3 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w)) + t_5) != Float32(Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w)) + t_5)) ? t_4 : ((t_4 != t_4) ? Float32(Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w)) + t_5) : max(Float32(Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w)) + t_5), t_4))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_2 * t_2)) != Float32(t_5 + Float32(t_2 * t_2))) ? t_4 : ((t_4 != t_4) ? Float32(t_5 + Float32(t_2 * t_2)) : max(Float32(t_5 + Float32(t_2 * t_2)), t_4))))) * 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 = floor(h) * dX_46_v; t_1 = dY_46_u * floor(w); t_2 = floor(w) * dX_46_u; t_3 = dY_46_v * floor(h); t_4 = (t_3 * t_3) + (t_1 * t_1); t_5 = t_0 * t_0; tmp = single(0.0); if ((t_2 ^ single(2.0)) >= ((t_3 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(((((dX_46_u * dX_46_u) * floor(w)) * floor(w)) + t_5), t_4))) * t_0; else tmp = (single(1.0) / sqrt(max((t_5 + (t_2 * t_2)), t_4))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := t\_3 \cdot t\_3 + t\_1 \cdot t\_1\\
t_5 := t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_2}^{2} \geq {t\_3}^{2} + {t\_1}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor + t\_5, t\_4\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_2 \cdot t\_2, t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 77.6%
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.f3264.3
Applied rewrites64.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3264.3
Applied rewrites64.3%
lift-*.f32N/A
pow2N/A
lift-pow.f32N/A
Applied rewrites64.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3264.3
Applied rewrites64.3%
Final simplification64.3%
herbie shell --seed 2024331
(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))))