
(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 6 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 (* (floor w) dX.u) 2.0))
(t_1 (pow (* dY.u (floor w)) 2.0))
(t_2 (+ t_1 (pow (* dY.v (floor h)) 2.0)))
(t_3 (pow (* dX.v (floor h)) 2.0)))
(if (>= (+ t_3 t_0) t_2)
(/
(/ (* (- dX.u) (floor w)) -1.0)
(/
(sqrt
(fmax (+ (pow (* (floor h) dX.v) 2.0) t_0) (+ (exp (/ 0.0 0.0)) t_1)))
1.0))
(* (/ 1.0 (sqrt (fmax (+ t_0 t_3) t_2))) (* (floor w) 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 = powf((floorf(w) * dX_46_u), 2.0f);
float t_1 = powf((dY_46_u * floorf(w)), 2.0f);
float t_2 = t_1 + powf((dY_46_v * floorf(h)), 2.0f);
float t_3 = powf((dX_46_v * floorf(h)), 2.0f);
float tmp;
if ((t_3 + t_0) >= t_2) {
tmp = ((-dX_46_u * floorf(w)) / -1.0f) / (sqrtf(fmaxf((powf((floorf(h) * dX_46_v), 2.0f) + t_0), (expf((0.0f / 0.0f)) + t_1))) / 1.0f);
} else {
tmp = (1.0f / sqrtf(fmaxf((t_0 + t_3), t_2))) * (floorf(w) * 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) ^ Float32(2.0) t_1 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_2 = Float32(t_1 + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_3 = Float32(dX_46_v * floor(h)) ^ Float32(2.0) tmp = Float32(0.0) if (Float32(t_3 + t_0) >= t_2) tmp = Float32(Float32(Float32(Float32(-dX_46_u) * floor(w)) / Float32(-1.0)) / Float32(sqrt(((Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + t_0) != Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + t_0)) ? Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_1) : ((Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_1) != Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_1)) ? Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + t_0) : max(Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + t_0), Float32(exp(Float32(Float32(0.0) / Float32(0.0))) + t_1))))) / Float32(1.0))); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_0 + t_3) != Float32(t_0 + t_3)) ? t_2 : ((t_2 != t_2) ? Float32(t_0 + t_3) : max(Float32(t_0 + t_3), t_2))))) * Float32(floor(w) * 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) ^ single(2.0); t_1 = (dY_46_u * floor(w)) ^ single(2.0); t_2 = t_1 + ((dY_46_v * floor(h)) ^ single(2.0)); t_3 = (dX_46_v * floor(h)) ^ single(2.0); tmp = single(0.0); if ((t_3 + t_0) >= t_2) tmp = ((-dX_46_u * floor(w)) / single(-1.0)) / (sqrt(max((((floor(h) * dX_46_v) ^ single(2.0)) + t_0), (exp((single(0.0) / single(0.0))) + t_1))) / single(1.0)); else tmp = (single(1.0) / sqrt(max((t_0 + t_3), t_2))) * (floor(w) * dY_46_u); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := t\_1 + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
\mathbf{if}\;t\_3 + t\_0 \geq t\_2:\\
\;\;\;\;\frac{\frac{\left(-dX.u\right) \cdot \left\lfloor w\right\rfloor }{-1}}{\frac{\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + t\_0, e^{\frac{0}{0}} + t\_1\right)}}{1}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 + t\_3, t\_2\right)}} \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
\end{array}
\end{array}
Initial program 80.0%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites80.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3280.1
lift-*.f32N/A
pow2N/A
lower-pow.f3280.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3280.1
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3280.1
lift-*.f32N/A
Applied rewrites80.1%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f32N/A
+-commutativeN/A
lower-+.f3280.1
lift-*.f32N/A
pow2N/A
lift-pow.f3280.1
Applied rewrites80.1%
Applied rewrites80.1%
(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 w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (* (floor h) dX.v)))
(if (>=
(pow t_3 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(*
(/ 1.0 (sqrt (fmax (+ t_4 (* (pow (floor h) 2.0) (* dX.v dX.v))) t_2)))
t_3)
(* (/ 1.0 (sqrt (fmax (+ t_4 (* t_5 t_5)) 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(w) * dX_46_u;
float t_4 = t_3 * t_3;
float t_5 = floorf(h) * dX_46_v;
float tmp;
if (powf(t_3, 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_4 + (powf(floorf(h), 2.0f) * (dX_46_v * dX_46_v))), t_2))) * t_3;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_4 + (t_5 * t_5)), 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(w) * dX_46_u) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_4 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))) != Float32(t_4 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v)))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))) : max(Float32(t_4 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))), t_2))))) * t_3); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_4 + Float32(t_5 * t_5)) != Float32(t_4 + Float32(t_5 * t_5))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32(t_5 * t_5)) : max(Float32(t_4 + Float32(t_5 * t_5)), 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(w) * dX_46_u; t_4 = t_3 * t_3; t_5 = floor(h) * dX_46_v; tmp = single(0.0); if ((t_3 ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_4 + ((floor(h) ^ single(2.0)) * (dX_46_v * dX_46_v))), t_2))) * t_3; else tmp = (single(1.0) / sqrt(max((t_4 + (t_5 * t_5)), 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 w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;{t\_3}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot \left(dX.v \cdot dX.v\right), t\_2\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 + t\_5 \cdot t\_5, t\_2\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 80.0%
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.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3268.2
Applied rewrites68.2%
(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 h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>=
(pow t_4 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_0 t_0)) t_3))) t_4)
(*
(/ 1.0 (sqrt (fmax (+ t_5 (* (* (pow (floor h) 2.0) dX.v) dX.v)) t_3)))
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(h) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if (powf(t_4, 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (t_0 * t_0)), t_3))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + ((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v)), t_3))) * 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(h) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_0 * t_0)) != Float32(t_5 + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_0 * t_0)) : max(Float32(t_5 + Float32(t_0 * t_0)), t_3))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v)) != Float32(t_5 + Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v)) : max(Float32(t_5 + Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v)), t_3))))) * 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(h) * dY_46_v; t_3 = (t_1 * t_1) + (t_2 * t_2); t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if ((t_4 ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_5 + (t_0 * t_0)), t_3))) * t_4; else tmp = (single(1.0) / sqrt(max((t_5 + (((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v)), t_3))) * 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 h\right\rfloor \cdot dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0 \cdot t\_0, t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v, t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 80.0%
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.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites68.2%
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
lift-*.f3268.2
Applied rewrites68.2%
(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 h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>=
(pow t_4 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_0 t_0)) t_3))) t_4)
(* (/ 1.0 (sqrt (fmax (+ t_5 (* (* t_0 dX.v) (floor h))) t_3))) 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(h) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if (powf(t_4, 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (t_0 * t_0)), t_3))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + ((t_0 * dX_46_v) * floorf(h))), t_3))) * 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(h) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(t_0 * t_0)) != Float32(t_5 + Float32(t_0 * t_0))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(t_0 * t_0)) : max(Float32(t_5 + Float32(t_0 * t_0)), t_3))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_5 + Float32(Float32(t_0 * dX_46_v) * floor(h))) != Float32(t_5 + Float32(Float32(t_0 * dX_46_v) * floor(h)))) ? t_3 : ((t_3 != t_3) ? Float32(t_5 + Float32(Float32(t_0 * dX_46_v) * floor(h))) : max(Float32(t_5 + Float32(Float32(t_0 * dX_46_v) * floor(h))), t_3))))) * 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(h) * dY_46_v; t_3 = (t_1 * t_1) + (t_2 * t_2); t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if ((t_4 ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_5 + (t_0 * t_0)), t_3))) * t_4; else tmp = (single(1.0) / sqrt(max((t_5 + ((t_0 * dX_46_v) * floor(h))), t_3))) * 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 h\right\rfloor \cdot dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{t\_4}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0 \cdot t\_0, t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + \left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 80.0%
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.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites68.2%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3268.2
Applied rewrites68.2%
(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 h) dY.v))
(t_3 (* (floor w) dX.u))
(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 (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* t_4 t_3)
(* t_4 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
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((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = t_4 * t_3;
} else {
tmp = t_4 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_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_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((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(t_4 * t_3); else tmp = Float32(t_4 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = floor(w) * dX_46_u; 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)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = t_4 * t_3; else tmp = t_4 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.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\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;{t\_3}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 80.0%
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.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites68.2%
(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 (* t_1 t_1))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_4 t_4) (* t_0 t_0))))
(if (>=
(pow t_4 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ t_2 (* t_3 t_3))))) t_4)
(*
(/
1.0
(sqrt (fmax t_5 (+ t_2 (* (floor h) (* (floor h) (* dY.v dY.v)))))))
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 = t_1 * t_1;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float tmp;
if (powf(t_4, 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(t_5, (t_2 + (t_3 * t_3))))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, (t_2 + (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v))))))) * 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(t_1 * t_1) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) tmp = Float32(0.0) if ((t_4 ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + Float32(t_3 * t_3)) : ((Float32(t_2 + Float32(t_3 * t_3)) != Float32(t_2 + Float32(t_3 * t_3))) ? t_5 : max(t_5, Float32(t_2 + Float32(t_3 * t_3))))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(t_2 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((Float32(t_2 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != Float32(t_2 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? t_5 : max(t_5, Float32(t_2 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))))))) * 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 = t_1 * t_1; t_3 = floor(h) * dY_46_v; t_4 = floor(w) * dX_46_u; t_5 = (t_4 * t_4) + (t_0 * t_0); tmp = single(0.0); if ((t_4 ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(t_5, (t_2 + (t_3 * t_3))))) * t_4; else tmp = (single(1.0) / sqrt(max(t_5, (t_2 + (floor(h) * (floor(h) * (dY_46_v * dY_46_v))))))) * 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 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
\mathbf{if}\;{t\_4}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + t\_3 \cdot t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + \left\lfloor h\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 80.0%
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.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3268.2
Applied rewrites68.2%
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites68.2%
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-*.f3268.1
Applied rewrites68.1%
herbie shell --seed 2024313
(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))))