
(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 7 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 (* (floor w) dX.u))
(t_1 (pow t_0 2.0))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (* dY.u (floor w)))
(t_4 (+ (pow (* dY.v (floor h)) 2.0) (pow t_3 2.0))))
(if (>= (+ t_2 t_1) t_4)
(* (/ 1.0 (sqrt (fmax (+ t_1 t_2) t_4))) t_0)
(/
1.0
(/
(sqrt
(fmax
(+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0))
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(w) * dX_46_u;
float t_1 = powf(t_0, 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = dY_46_u * floorf(w);
float t_4 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_3, 2.0f);
float tmp;
if ((t_2 + t_1) >= t_4) {
tmp = (1.0f / sqrtf(fmaxf((t_1 + t_2), t_4))) * t_0;
} else {
tmp = 1.0f / (sqrtf(fmaxf((powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), 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(w) * dX_46_u) t_1 = t_0 ^ Float32(2.0) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = Float32(dY_46_u * floor(w)) t_4 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) tmp = Float32(0.0) if (Float32(t_2 + t_1) >= t_4) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_1 + t_2) != Float32(t_1 + t_2)) ? t_4 : ((t_4 != t_4) ? Float32(t_1 + t_2) : max(Float32(t_1 + t_2), t_4))))) * t_0); else tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) != Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0)))) ? t_4 : ((t_4 != t_4) ? Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) : max(Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), 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(w) * dX_46_u; t_1 = t_0 ^ single(2.0); t_2 = (floor(h) * dX_46_v) ^ single(2.0); t_3 = dY_46_u * floor(w); t_4 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_3 ^ single(2.0)); tmp = single(0.0); if ((t_2 + t_1) >= t_4) tmp = (single(1.0) / sqrt(max((t_1 + t_2), t_4))) * t_0; else tmp = single(1.0) / (sqrt(max((((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0))), t_4)) / t_3); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {t\_0}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_3}^{2}\\
\mathbf{if}\;t\_2 + t\_1 \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1 + t\_2, t\_4\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left({\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, t\_4\right)}}{t\_3}}\\
\end{array}
\end{array}
Initial program 74.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites74.7%
Applied rewrites74.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lower-+.f3274.7
lift-*.f32N/A
pow2N/A
lift-pow.f3274.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3274.7
Applied rewrites74.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (pow t_0 2.0))
(t_2 (pow (* (floor h) dX.v) 2.0))
(t_3 (+ t_2 t_1))
(t_4 (* dY.u (floor w)))
(t_5 (+ (pow (* dY.v (floor h)) 2.0) (pow t_4 2.0))))
(if (>= t_3 t_5)
(* (/ 1.0 (sqrt (fmax (+ t_1 t_2) t_5))) t_0)
(/ t_4 (sqrt (fmax t_3 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 = floorf(w) * dX_46_u;
float t_1 = powf(t_0, 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f);
float t_3 = t_2 + t_1;
float t_4 = dY_46_u * floorf(w);
float t_5 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_4, 2.0f);
float tmp;
if (t_3 >= t_5) {
tmp = (1.0f / sqrtf(fmaxf((t_1 + t_2), t_5))) * t_0;
} else {
tmp = t_4 / sqrtf(fmaxf(t_3, t_5));
}
return tmp;
}
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 = t_0 ^ Float32(2.0) t_2 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_3 = Float32(t_2 + t_1) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_4 ^ Float32(2.0))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_1 + t_2) != Float32(t_1 + t_2)) ? t_5 : ((t_5 != t_5) ? Float32(t_1 + t_2) : max(Float32(t_1 + t_2), t_5))))) * t_0); else tmp = Float32(t_4 / sqrt(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(t_3, t_5))))); 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(w) * dX_46_u; t_1 = t_0 ^ single(2.0); t_2 = (floor(h) * dX_46_v) ^ single(2.0); t_3 = t_2 + t_1; t_4 = dY_46_u * floor(w); t_5 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_4 ^ single(2.0)); tmp = single(0.0); if (t_3 >= t_5) tmp = (single(1.0) / sqrt(max((t_1 + t_2), t_5))) * t_0; else tmp = t_4 / sqrt(max(t_3, t_5)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {t\_0}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_3 := t\_2 + t\_1\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_4}^{2}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1 + t\_2, t\_5\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\end{array}
\end{array}
Initial program 74.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites74.7%
Applied rewrites74.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lower-+.f3274.7
lift-*.f32N/A
pow2N/A
lift-pow.f3274.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3274.7
Applied rewrites74.7%
Applied rewrites74.7%
Final simplification74.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (pow t_0 2.0))
(t_2 (+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_3 (pow (* (floor h) dX.v) 2.0))
(t_4 (+ t_3 t_1)))
(if (>= t_4 t_2)
(* (/ 1.0 (sqrt (fmax (+ t_1 t_3) t_2))) t_0)
(* (/ (floor w) (sqrt (fmax t_4 t_2))) dY.u))))
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(t_0, 2.0f);
float t_2 = powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_3 = powf((floorf(h) * dX_46_v), 2.0f);
float t_4 = t_3 + t_1;
float tmp;
if (t_4 >= t_2) {
tmp = (1.0f / sqrtf(fmaxf((t_1 + t_3), t_2))) * t_0;
} else {
tmp = (floorf(w) / sqrtf(fmaxf(t_4, t_2))) * dY_46_u;
}
return tmp;
}
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 = t_0 ^ Float32(2.0) t_2 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_3 = Float32(floor(h) * dX_46_v) ^ Float32(2.0) t_4 = Float32(t_3 + t_1) tmp = Float32(0.0) if (t_4 >= t_2) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_1 + t_3) != Float32(t_1 + t_3)) ? t_2 : ((t_2 != t_2) ? Float32(t_1 + t_3) : max(Float32(t_1 + t_3), t_2))))) * t_0); else tmp = Float32(Float32(floor(w) / sqrt(((t_4 != t_4) ? t_2 : ((t_2 != t_2) ? t_4 : max(t_4, t_2))))) * dY_46_u); 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(w) * dX_46_u; t_1 = t_0 ^ single(2.0); t_2 = ((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_3 = (floor(h) * dX_46_v) ^ single(2.0); t_4 = t_3 + t_1; tmp = single(0.0); if (t_4 >= t_2) tmp = (single(1.0) / sqrt(max((t_1 + t_3), t_2))) * t_0; else tmp = (floor(w) / sqrt(max(t_4, t_2))) * dY_46_u; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {t\_0}^{2}\\
t_2 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2}\\
t_4 := t\_3 + t\_1\\
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1 + t\_3, t\_2\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_4, t\_2\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 74.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites74.7%
Applied rewrites74.7%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
lower-+.f3274.7
lift-*.f32N/A
pow2N/A
lift-pow.f3274.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3274.7
Applied rewrites74.7%
Applied rewrites74.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* t_1 t_1))
(t_3 (* (floor h) dY.v))
(t_4 (* t_3 t_3))
(t_5 (* (floor w) dY.u))
(t_6 (+ (* t_5 t_5) t_4)))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) t_4))
(* (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_2) t_6))) t_0)
(*
(/ 1.0 (sqrt (fmax (+ (exp (* 2.0 (* 2.0 (* (log t_0) 0.5)))) t_2) t_6)))
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 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_3 * t_3;
float t_5 = floorf(w) * dY_46_u;
float t_6 = (t_5 * t_5) + t_4;
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf((dY_46_u * floorf(w)), 2.0f) + t_4)) {
tmp = (1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_2), t_6))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf((expf((2.0f * (2.0f * (logf(t_0) * 0.5f)))) + t_2), t_6))) * t_5;
}
return tmp;
}
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(floor(h) * dX_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(Float32(t_5 * t_5) + t_4) tmp = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + t_4)) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + t_2) != Float32(Float32(t_0 * t_0) + t_2)) ? t_6 : ((t_6 != t_6) ? Float32(Float32(t_0 * t_0) + t_2) : max(Float32(Float32(t_0 * t_0) + t_2), t_6))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(exp(Float32(Float32(2.0) * Float32(Float32(2.0) * Float32(log(t_0) * Float32(0.5))))) + t_2) != Float32(exp(Float32(Float32(2.0) * Float32(Float32(2.0) * Float32(log(t_0) * Float32(0.5))))) + t_2)) ? t_6 : ((t_6 != t_6) ? Float32(exp(Float32(Float32(2.0) * Float32(Float32(2.0) * Float32(log(t_0) * Float32(0.5))))) + t_2) : max(Float32(exp(Float32(Float32(2.0) * Float32(Float32(2.0) * Float32(log(t_0) * Float32(0.5))))) + t_2), t_6))))) * t_5); 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(w) * dX_46_u; t_1 = floor(h) * dX_46_v; t_2 = t_1 * t_1; t_3 = floor(h) * dY_46_v; t_4 = t_3 * t_3; t_5 = floor(w) * dY_46_u; t_6 = (t_5 * t_5) + t_4; tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (((dY_46_u * floor(w)) ^ single(2.0)) + t_4)) tmp = (single(1.0) / sqrt(max(((t_0 * t_0) + t_2), t_6))) * t_0; else tmp = (single(1.0) / sqrt(max((exp((single(2.0) * (single(2.0) * (log(t_0) * single(0.5))))) + t_2), t_6))) * t_5; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := t\_5 \cdot t\_5 + t\_4\\
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_2, t\_6\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(e^{2 \cdot \left(2 \cdot \left(\log t\_0 \cdot 0.5\right)\right)} + t\_2, t\_6\right)}} \cdot t\_5\\
\end{array}
\end{array}
Initial program 74.6%
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.f3258.1
Applied rewrites58.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3258.1
Applied rewrites58.1%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lift-log.f32N/A
*-commutativeN/A
pow-expN/A
lift-exp.f32N/A
sqr-powN/A
pow-sqrN/A
pow-to-expN/A
lower-exp.f32N/A
lift-exp.f32N/A
rem-log-expN/A
lower-*.f32N/A
lower-*.f32N/A
div-invN/A
metadata-evalN/A
lower-*.f3259.3
Applied rewrites59.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor w) dY.u))
(t_4 (+ (* t_3 t_3) (* t_1 t_1)))
(t_5 (* t_0 t_0)))
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) t_4)
(* (/ 1.0 (sqrt (fmax (+ (* t_2 t_2) t_5) t_4))) t_2)
(*
(/
1.0
(sqrt (fmax (+ (/ (* dX.u (pow (floor w) 2.0)) (/ 1.0 dX.u)) t_5) 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 = floorf(h) * dY_46_v;
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(w) * dY_46_u;
float t_4 = (t_3 * t_3) + (t_1 * t_1);
float t_5 = t_0 * t_0;
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= t_4) {
tmp = (1.0f / sqrtf(fmaxf(((t_2 * t_2) + t_5), t_4))) * t_2;
} else {
tmp = (1.0f / sqrtf(fmaxf((((dX_46_u * powf(floorf(w), 2.0f)) / (1.0f / dX_46_u)) + t_5), 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(floor(h) * dY_46_v) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(w) * dY_46_u) 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 (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= t_4) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_2 * t_2) + t_5) != Float32(Float32(t_2 * t_2) + t_5)) ? t_4 : ((t_4 != t_4) ? Float32(Float32(t_2 * t_2) + t_5) : max(Float32(Float32(t_2 * t_2) + t_5), t_4))))) * t_2); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(dX_46_u * (floor(w) ^ Float32(2.0))) / Float32(Float32(1.0) / dX_46_u)) + t_5) != Float32(Float32(Float32(dX_46_u * (floor(w) ^ Float32(2.0))) / Float32(Float32(1.0) / dX_46_u)) + t_5)) ? t_4 : ((t_4 != t_4) ? Float32(Float32(Float32(dX_46_u * (floor(w) ^ Float32(2.0))) / Float32(Float32(1.0) / dX_46_u)) + t_5) : max(Float32(Float32(Float32(dX_46_u * (floor(w) ^ Float32(2.0))) / Float32(Float32(1.0) / dX_46_u)) + t_5), 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 = floor(h) * dY_46_v; t_2 = floor(w) * dX_46_u; t_3 = floor(w) * dY_46_u; t_4 = (t_3 * t_3) + (t_1 * t_1); t_5 = t_0 * t_0; tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= t_4) tmp = (single(1.0) / sqrt(max(((t_2 * t_2) + t_5), t_4))) * t_2; else tmp = (single(1.0) / sqrt(max((((dX_46_u * (floor(w) ^ single(2.0))) / (single(1.0) / dX_46_u)) + t_5), 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 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := t\_3 \cdot t\_3 + t\_1 \cdot t\_1\\
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\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_5, t\_4\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\frac{dX.u \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}}{\frac{1}{dX.u}} + t\_5, t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 74.6%
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.f3258.1
Applied rewrites58.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
remove-double-divN/A
un-div-invN/A
lower-/.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
pow2N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-/.f3258.1
Applied rewrites58.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor h) dX.v))
(t_4 (* t_3 t_3))
(t_5 (* (floor w) dX.u)))
(if (>=
(pow t_3 2.0)
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(* (/ 1.0 (sqrt (fmax (+ (* t_5 t_5) t_4) t_2))) t_5)
(*
(/ 1.0 (sqrt (fmax (+ (* (* (pow (floor w) 2.0) dX.u) dX.u) t_4) t_2)))
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 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(h) * dX_46_v;
float t_4 = t_3 * t_3;
float t_5 = floorf(w) * dX_46_u;
float tmp;
if (powf(t_3, 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(((t_5 * t_5) + t_4), t_2))) * t_5;
} else {
tmp = (1.0f / sqrtf(fmaxf((((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) + t_4), t_2))) * t_0;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_5 * t_5) + t_4) != Float32(Float32(t_5 * t_5) + t_4)) ? t_2 : ((t_2 != t_2) ? Float32(Float32(t_5 * t_5) + t_4) : max(Float32(Float32(t_5 * t_5) + t_4), t_2))))) * t_5); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_4) != Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_4)) ? t_2 : ((t_2 != t_2) ? Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_4) : max(Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) + t_4), t_2))))) * 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 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = floor(h) * dX_46_v; t_4 = t_3 * t_3; t_5 = floor(w) * dX_46_u; tmp = single(0.0); if ((t_3 ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(((t_5 * t_5) + t_4), t_2))) * t_5; else tmp = (single(1.0) / sqrt(max(((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) + t_4), t_2))) * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;{t\_3}^{2} \geq {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 \cdot t\_5 + t\_4, t\_2\right)}} \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u + t\_4, t\_2\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 74.6%
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.f3258.1
Applied rewrites58.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3258.1
Applied rewrites58.1%
lift-+.f32N/A
+-commutativeN/A
Applied rewrites58.1%
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-*.f3258.1
Applied rewrites58.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dX.u))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_2 t_2)) (+ (* t_0 t_0) (* t_1 t_1)))))))
(if (>=
(pow t_2 2.0)
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(* t_4 t_3)
(* 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 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1))));
float tmp;
if (powf(t_2, 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) {
tmp = t_4 * t_3;
} else {
tmp = 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(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) 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_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_3 * t_3) + Float32(t_2 * t_2)) : max(Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2)), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if ((t_2 ^ Float32(2.0)) >= Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0)))) tmp = Float32(t_4 * t_3); else tmp = 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 = floor(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_2 * t_2)), ((t_0 * t_0) + (t_1 * t_1)))); tmp = single(0.0); if ((t_2 ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)))) tmp = t_4 * t_3; else tmp = t_4 * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
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\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_2}^{2} \geq {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_0\\
\end{array}
\end{array}
Initial program 74.6%
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.f3258.1
Applied rewrites58.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3258.1
Applied rewrites58.1%
lift-+.f32N/A
+-commutativeN/A
Applied rewrites58.1%
herbie shell --seed 2024305
(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))))