
(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 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_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 (* dX.v (floor h)))
(t_2 (pow t_1 2.0))
(t_3 (* dY.v (floor h)))
(t_4 (+ (pow t_3 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_5 (* (floor w) dY.u))
(t_6 (pow (* dX.u (floor w)) 2.0))
(t_7 (* (floor h) dY.v)))
(if (>= (+ t_6 (* t_0 t_0)) (+ (* t_5 t_5) (* t_7 t_7)))
(/ 1.0 (/ (sqrt (fmax (+ t_6 t_2) t_4)) t_1))
(/ 1.0 (/ (sqrt (fmax (+ t_2 t_6) 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 = dX_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f);
float t_3 = dY_46_v * floorf(h);
float t_4 = powf(t_3, 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_5 = floorf(w) * dY_46_u;
float t_6 = powf((dX_46_u * floorf(w)), 2.0f);
float t_7 = floorf(h) * dY_46_v;
float tmp;
if ((t_6 + (t_0 * t_0)) >= ((t_5 * t_5) + (t_7 * t_7))) {
tmp = 1.0f / (sqrtf(fmaxf((t_6 + t_2), t_4)) / t_1);
} else {
tmp = 1.0f / (sqrtf(fmaxf((t_2 + t_6), 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(dX_46_v * floor(h)) t_2 = t_1 ^ Float32(2.0) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32((t_3 ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_7 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (Float32(t_6 + Float32(t_0 * t_0)) >= Float32(Float32(t_5 * t_5) + Float32(t_7 * t_7))) tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32(t_6 + t_2) != Float32(t_6 + t_2)) ? t_4 : ((t_4 != t_4) ? Float32(t_6 + t_2) : max(Float32(t_6 + t_2), t_4)))) / t_1)); else tmp = Float32(Float32(1.0) / Float32(sqrt(((Float32(t_2 + t_6) != Float32(t_2 + t_6)) ? t_4 : ((t_4 != t_4) ? Float32(t_2 + t_6) : max(Float32(t_2 + t_6), 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 = dX_46_v * floor(h); t_2 = t_1 ^ single(2.0); t_3 = dY_46_v * floor(h); t_4 = (t_3 ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_5 = floor(w) * dY_46_u; t_6 = (dX_46_u * floor(w)) ^ single(2.0); t_7 = floor(h) * dY_46_v; tmp = single(0.0); if ((t_6 + (t_0 * t_0)) >= ((t_5 * t_5) + (t_7 * t_7))) tmp = single(1.0) / (sqrt(max((t_6 + t_2), t_4)) / t_1); else tmp = single(1.0) / (sqrt(max((t_2 + t_6), 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 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2}\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {t\_3}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;t\_6 + t\_0 \cdot t\_0 \geq t\_5 \cdot t\_5 + t\_7 \cdot t\_7:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(t\_6 + t\_2, t\_4\right)}}{t\_1}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(t\_2 + t\_6, t\_4\right)}}{t\_3}}\\
\end{array}
\end{array}
Initial program 75.5%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites75.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites75.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3275.7
Applied rewrites75.7%
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (pow (* dX.v (floor h)) 2.0))
(t_1 (* dY.v (floor h)))
(t_2 (pow t_1 2.0))
(t_3 (pow (* dY.u (floor w)) 2.0))
(t_4 (+ t_2 t_3))
(t_5 (pow (* dX.u (floor w)) 2.0))
(t_6 (+ t_0 t_5)))
(if (>= (+ t_5 t_0) t_4)
(* (/ 1.0 (sqrt (fmax t_6 (+ t_3 t_2)))) (* (floor h) dX.v))
(/ 1.0 (/ (sqrt (fmax t_6 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 = powf((dX_46_v * floorf(h)), 2.0f);
float t_1 = dY_46_v * floorf(h);
float t_2 = powf(t_1, 2.0f);
float t_3 = powf((dY_46_u * floorf(w)), 2.0f);
float t_4 = t_2 + t_3;
float t_5 = powf((dX_46_u * floorf(w)), 2.0f);
float t_6 = t_0 + t_5;
float tmp;
if ((t_5 + t_0) >= t_4) {
tmp = (1.0f / sqrtf(fmaxf(t_6, (t_3 + t_2)))) * (floorf(h) * dX_46_v);
} else {
tmp = 1.0f / (sqrtf(fmaxf(t_6, 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(dX_46_v * floor(h)) ^ Float32(2.0) t_1 = Float32(dY_46_v * floor(h)) t_2 = t_1 ^ Float32(2.0) t_3 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_4 = Float32(t_2 + t_3) t_5 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_6 = Float32(t_0 + t_5) tmp = Float32(0.0) if (Float32(t_5 + t_0) >= t_4) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(t_3 + t_2) : ((Float32(t_3 + t_2) != Float32(t_3 + t_2)) ? t_6 : max(t_6, Float32(t_3 + t_2)))))) * Float32(floor(h) * dX_46_v)); else tmp = Float32(Float32(1.0) / Float32(sqrt(((t_6 != t_6) ? t_4 : ((t_4 != t_4) ? t_6 : max(t_6, 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 = (dX_46_v * floor(h)) ^ single(2.0); t_1 = dY_46_v * floor(h); t_2 = t_1 ^ single(2.0); t_3 = (dY_46_u * floor(w)) ^ single(2.0); t_4 = t_2 + t_3; t_5 = (dX_46_u * floor(w)) ^ single(2.0); t_6 = t_0 + t_5; tmp = single(0.0); if ((t_5 + t_0) >= t_4) tmp = (single(1.0) / sqrt(max(t_6, (t_3 + t_2)))) * (floor(h) * dX_46_v); else tmp = single(1.0) / (sqrt(max(t_6, t_4)) / t_1); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {t\_1}^{2}\\
t_3 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_4 := t\_2 + t\_3\\
t_5 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_6 := t\_0 + t\_5\\
\mathbf{if}\;t\_5 + t\_0 \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_3 + t\_2\right)}} \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right)\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(t\_6, t\_4\right)}}{t\_1}}\\
\end{array}
\end{array}
Initial program 75.5%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
clear-numN/A
lower-/.f32N/A
Applied rewrites75.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
Applied rewrites75.7%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3275.7
Applied rewrites75.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3275.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3275.7
lift-+.f32N/A
+-commutativeN/A
lower-+.f3275.7
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3275.7
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
Applied rewrites75.7%
(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 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* t_3 t_3))
(t_5 (+ (* t_0 t_0) (* t_1 t_1))))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_2 t_2) t_4)))) t_1)
(*
(/ 1.0 (sqrt (fmax t_5 (+ (* (pow (floor w) 2.0) (* dY.u dY.u)) 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 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_3 * t_3;
float t_5 = (t_0 * t_0) + (t_1 * t_1);
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (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_2) + t_4)))) * t_1;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((powf(floorf(w), 2.0f) * (dY_46_u * dY_46_u)) + 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 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) 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)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_2 * t_2) + t_4) : ((Float32(Float32(t_2 * t_2) + t_4) != Float32(Float32(t_2 * t_2) + t_4)) ? t_5 : max(t_5, Float32(Float32(t_2 * t_2) + t_4)))))) * t_1); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)) + t_4) : ((Float32(Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)) + t_4) != Float32(Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)) + t_4)) ? t_5 : max(t_5, Float32(Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)) + 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 = floor(h) * dX_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(h) * dY_46_v; t_4 = t_3 * t_3; t_5 = (t_0 * t_0) + (t_1 * t_1); tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (((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_2) + t_4)))) * t_1; else tmp = (single(1.0) / sqrt(max(t_5, (((floor(w) ^ single(2.0)) * (dY_46_u * dY_46_u)) + 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 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
\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} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 \cdot t\_2 + t\_4\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, {\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot \left(dY.u \cdot dY.u\right) + t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 75.5%
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.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3265.1
Applied rewrites65.1%
(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 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* t_3 t_3))
(t_5 (+ (* t_0 t_0) (* t_1 t_1))))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_2 t_2) t_4)))) t_1)
(*
(/ 1.0 (sqrt (fmax t_5 (+ (* (* (pow (floor w) 2.0) dY.u) dY.u) 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 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_3 * t_3;
float t_5 = (t_0 * t_0) + (t_1 * t_1);
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (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_2) + t_4)))) * t_1;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, (((powf(floorf(w), 2.0f) * dY_46_u) * dY_46_u) + 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 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) 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)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_2 * t_2) + t_4) : ((Float32(Float32(t_2 * t_2) + t_4) != Float32(Float32(t_2 * t_2) + t_4)) ? t_5 : max(t_5, Float32(Float32(t_2 * t_2) + t_4)))))) * t_1); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u) + t_4) : ((Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u) + t_4) != Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u) + t_4)) ? t_5 : max(t_5, Float32(Float32(Float32((floor(w) ^ Float32(2.0)) * dY_46_u) * dY_46_u) + 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 = floor(h) * dX_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(h) * dY_46_v; t_4 = t_3 * t_3; t_5 = (t_0 * t_0) + (t_1 * t_1); tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (((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_2) + t_4)))) * t_1; else tmp = (single(1.0) / sqrt(max(t_5, ((((floor(w) ^ single(2.0)) * dY_46_u) * dY_46_u) + 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 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
\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} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 \cdot t\_2 + t\_4\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dY.u\right) \cdot dY.u + t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 75.5%
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.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.1
Applied rewrites65.1%
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
lift-*.f3265.1
Applied rewrites65.1%
(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 w) dY.u))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_3 t_3) (* t_4 t_4))))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_2) t_5))) t_1)
(*
(/
1.0
(sqrt (fmax (+ (* (floor w) (* (floor w) (* dX.u dX.u))) t_2) t_5)))
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(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = t_1 * t_1;
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_3 * t_3) + (t_4 * t_4);
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_2), t_5))) * t_1;
} else {
tmp = (1.0f / sqrtf(fmaxf(((floorf(w) * (floorf(w) * (dX_46_u * dX_46_u))) + t_2), t_5))) * 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(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_3 * t_3) + Float32(t_4 * 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)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + t_2) != Float32(Float32(t_0 * t_0) + t_2)) ? t_5 : ((t_5 != t_5) ? Float32(Float32(t_0 * t_0) + t_2) : max(Float32(Float32(t_0 * t_0) + t_2), t_5))))) * t_1); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_2) != Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_2)) ? t_5 : ((t_5 != t_5) ? Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_2) : max(Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_2), t_5))))) * 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(w) * dX_46_u; t_1 = floor(h) * dX_46_v; t_2 = t_1 * t_1; t_3 = floor(w) * dY_46_u; t_4 = floor(h) * dY_46_v; t_5 = (t_3 * t_3) + (t_4 * 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)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(((t_0 * t_0) + t_2), t_5))) * t_1; else tmp = (single(1.0) / sqrt(max(((floor(w) * (floor(w) * (dX_46_u * dX_46_u))) + t_2), t_5))) * t_4; 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 w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_3 \cdot t\_3 + t\_4 \cdot 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} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_2, t\_5\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot dX.u\right)\right) + t\_2, t\_5\right)}} \cdot t\_4\\
\end{array}
\end{array}
Initial program 75.5%
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.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f3265.1
Applied rewrites65.1%
(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 (* dX.v (floor h)) 2.0)
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(* t_4 t_0)
(* t_4 t_2))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = 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((dX_46_v * floorf(h)), 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))) {
tmp = t_4 * t_0;
} else {
tmp = t_4 * t_2;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(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 ((Float32(dX_46_v * floor(h)) ^ 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_0); else tmp = Float32(t_4 * t_2); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = 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 (((dX_46_v * floor(h)) ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)))) tmp = t_4 * t_0; else tmp = t_4 * t_2; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \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}\;{\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\end{array}
\end{array}
Initial program 75.5%
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.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.1
Applied rewrites65.1%
lift-+.f32N/A
+-commutativeN/A
lift-+.f32N/A
Applied rewrites65.1%
herbie shell --seed 2024318
(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))))