
(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
(+
(* (floor w) (* (floor w) (* dX.u dX.u)))
(* (floor h) (* dX.v t_0))))
(t_2 (* (floor h) dY.v))
(t_3
(+
(* (floor w) (* (floor w) (* dY.u dY.u)))
(* (floor h) (* dY.v t_2))))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ t_2 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) * (floorf(w) * (dX_46_u * dX_46_u))) + (floorf(h) * (dX_46_v * t_0));
float t_2 = floorf(h) * dY_46_v;
float t_3 = (floorf(w) * (floorf(w) * (dY_46_u * dY_46_u))) + (floorf(h) * (dY_46_v * t_2));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = t_2 / 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(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + Float32(floor(h) * Float32(dX_46_v * t_0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u))) + Float32(floor(h) * Float32(dY_46_v * t_2))) t_4 = sqrt(((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3)))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(t_2 / 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) * (floor(w) * (dX_46_u * dX_46_u))) + (floor(h) * (dX_46_v * t_0)); t_2 = floor(h) * dY_46_v; t_3 = (floor(w) * (floor(w) * (dY_46_u * dY_46_u))) + (floor(h) * (dY_46_v * t_2)); t_4 = sqrt(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = t_2 / 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 \left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot dX.u\right)\right) + \left\lfloor h\right\rfloor \cdot \left(dX.v \cdot t\_0\right)\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right)\right) + \left\lfloor h\right\rfloor \cdot \left(dY.v \cdot t\_2\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 72.7%
Simplified73.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (+ (pow t_0 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_2 (+ (pow (* (floor h) dX.v) 2.0) (pow (* (floor w) dX.u) 2.0)))
(t_3 (pow (fmax t_2 t_1) 0.5)))
(if (>= t_2 t_1) (* dX.v (/ (floor h) t_3)) (/ t_0 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) * dY_46_v;
float t_1 = powf(t_0, 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f);
float t_3 = powf(fmaxf(t_2, t_1), 0.5f);
float tmp;
if (t_2 >= t_1) {
tmp = dX_46_v * (floorf(h) / t_3);
} else {
tmp = t_0 / 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) * dY_46_v) t_1 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_2 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) t_3 = ((t_2 != t_2) ? t_1 : ((t_1 != t_1) ? t_2 : max(t_2, t_1))) ^ Float32(0.5) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(dX_46_v * Float32(floor(h) / t_3)); else tmp = Float32(t_0 / 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) * dY_46_v; t_1 = (t_0 ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_2 = ((floor(h) * dX_46_v) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0)); t_3 = max(t_2, t_1) ^ single(0.5); tmp = single(0.0); if (t_2 >= t_1) tmp = dX_46_v * (floor(h) / t_3); else tmp = t_0 / t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := {t\_0}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_3 := {\left(\mathsf{max}\left(t\_2, t\_1\right)\right)}^{0.5}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;dX.v \cdot \frac{\left\lfloor h\right\rfloor }{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_3}\\
\end{array}
\end{array}
Initial program 72.7%
Simplified73.0%
Applied egg-rr72.7%
associate-/r/N/A
*-lowering-*.f32N/A
Applied egg-rr72.8%
Final simplification72.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow (floor w) 2.0))
(t_2 (pow t_0 2.0))
(t_3 (* dX.v t_0))
(t_4 (* (floor h) dY.v))
(t_5 (+ (pow t_4 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_6
(+
(* (floor w) (* (floor w) (* dY.u dY.u)))
(* (floor h) (* dY.v t_4))))
(t_7 (* (floor w) (* (floor w) (* dX.u dX.u)))))
(if (<= dX.u 3.3999999686784577e-6)
(if (>= (* (* dX.v dX.v) (pow (floor h) 2.0)) t_6)
(/ t_0 (sqrt (fmax (+ t_7 (+ t_2 (* t_3 0.0))) t_6)))
(/ t_4 (sqrt (fmax (+ t_7 (* (floor h) t_3)) t_6))))
(if (>= (* dX.u (* dX.u t_1)) t_5)
(/
(floor h)
(/ (pow (fmax (+ t_2 (pow (* (floor w) dX.u) 2.0)) t_5) 0.5) dX.v))
(/ t_4 (pow (fmax (* (* dX.u dX.u) t_1) t_5) 0.5))))))
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(h) * dX_46_v;
float t_1 = powf(floorf(w), 2.0f);
float t_2 = powf(t_0, 2.0f);
float t_3 = dX_46_v * t_0;
float t_4 = floorf(h) * dY_46_v;
float t_5 = powf(t_4, 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_6 = (floorf(w) * (floorf(w) * (dY_46_u * dY_46_u))) + (floorf(h) * (dY_46_v * t_4));
float t_7 = floorf(w) * (floorf(w) * (dX_46_u * dX_46_u));
float tmp_1;
if (dX_46_u <= 3.3999999686784577e-6f) {
float tmp_2;
if (((dX_46_v * dX_46_v) * powf(floorf(h), 2.0f)) >= t_6) {
tmp_2 = t_0 / sqrtf(fmaxf((t_7 + (t_2 + (t_3 * 0.0f))), t_6));
} else {
tmp_2 = t_4 / sqrtf(fmaxf((t_7 + (floorf(h) * t_3)), t_6));
}
tmp_1 = tmp_2;
} else if ((dX_46_u * (dX_46_u * t_1)) >= t_5) {
tmp_1 = floorf(h) / (powf(fmaxf((t_2 + powf((floorf(w) * dX_46_u), 2.0f)), t_5), 0.5f) / dX_46_v);
} else {
tmp_1 = t_4 / powf(fmaxf(((dX_46_u * dX_46_u) * t_1), t_5), 0.5f);
}
return tmp_1;
}
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 = floor(w) ^ Float32(2.0) t_2 = t_0 ^ Float32(2.0) t_3 = Float32(dX_46_v * t_0) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32((t_4 ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_6 = Float32(Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u))) + Float32(floor(h) * Float32(dY_46_v * t_4))) t_7 = Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(3.3999999686784577e-6)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_v * dX_46_v) * (floor(h) ^ Float32(2.0))) >= t_6) tmp_2 = Float32(t_0 / sqrt(((Float32(t_7 + Float32(t_2 + Float32(t_3 * Float32(0.0)))) != Float32(t_7 + Float32(t_2 + Float32(t_3 * Float32(0.0))))) ? t_6 : ((t_6 != t_6) ? Float32(t_7 + Float32(t_2 + Float32(t_3 * Float32(0.0)))) : max(Float32(t_7 + Float32(t_2 + Float32(t_3 * Float32(0.0)))), t_6))))); else tmp_2 = Float32(t_4 / sqrt(((Float32(t_7 + Float32(floor(h) * t_3)) != Float32(t_7 + Float32(floor(h) * t_3))) ? t_6 : ((t_6 != t_6) ? Float32(t_7 + Float32(floor(h) * t_3)) : max(Float32(t_7 + Float32(floor(h) * t_3)), t_6))))); end tmp_1 = tmp_2; elseif (Float32(dX_46_u * Float32(dX_46_u * t_1)) >= t_5) tmp_1 = Float32(floor(h) / Float32((((Float32(t_2 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) != Float32(t_2 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0)))) ? t_5 : ((t_5 != t_5) ? Float32(t_2 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) : max(Float32(t_2 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_5))) ^ Float32(0.5)) / dX_46_v)); else tmp_1 = Float32(t_4 / (((Float32(Float32(dX_46_u * dX_46_u) * t_1) != Float32(Float32(dX_46_u * dX_46_u) * t_1)) ? t_5 : ((t_5 != t_5) ? Float32(Float32(dX_46_u * dX_46_u) * t_1) : max(Float32(Float32(dX_46_u * dX_46_u) * t_1), t_5))) ^ Float32(0.5))); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) ^ single(2.0); t_2 = t_0 ^ single(2.0); t_3 = dX_46_v * t_0; t_4 = floor(h) * dY_46_v; t_5 = (t_4 ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_6 = (floor(w) * (floor(w) * (dY_46_u * dY_46_u))) + (floor(h) * (dY_46_v * t_4)); t_7 = floor(w) * (floor(w) * (dX_46_u * dX_46_u)); tmp_2 = single(0.0); if (dX_46_u <= single(3.3999999686784577e-6)) tmp_3 = single(0.0); if (((dX_46_v * dX_46_v) * (floor(h) ^ single(2.0))) >= t_6) tmp_3 = t_0 / sqrt(max((t_7 + (t_2 + (t_3 * single(0.0)))), t_6)); else tmp_3 = t_4 / sqrt(max((t_7 + (floor(h) * t_3)), t_6)); end tmp_2 = tmp_3; elseif ((dX_46_u * (dX_46_u * t_1)) >= t_5) tmp_2 = floor(h) / ((max((t_2 + ((floor(w) * dX_46_u) ^ single(2.0))), t_5) ^ single(0.5)) / dX_46_v); else tmp_2 = t_4 / (max(((dX_46_u * dX_46_u) * t_1), t_5) ^ single(0.5)); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := {t\_0}^{2}\\
t_3 := dX.v \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := {t\_4}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_6 := \left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right)\right) + \left\lfloor h\right\rfloor \cdot \left(dY.v \cdot t\_4\right)\\
t_7 := \left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot dX.u\right)\right)\\
\mathbf{if}\;dX.u \leq 3.3999999686784577 \cdot 10^{-6}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.v \cdot dX.v\right) \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2} \geq t\_6:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_7 + \left(t\_2 + t\_3 \cdot 0\right), t\_6\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_7 + \left\lfloor h\right\rfloor \cdot t\_3, t\_6\right)}}\\
\end{array}\\
\mathbf{elif}\;dX.u \cdot \left(dX.u \cdot t\_1\right) \geq t\_5:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\frac{{\left(\mathsf{max}\left(t\_2 + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_5\right)\right)}^{0.5}}{dX.v}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{{\left(\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot t\_1, t\_5\right)\right)}^{0.5}}\\
\end{array}
\end{array}
if dX.u < 3.39999997e-6Initial program 72.4%
Simplified72.8%
Taylor expanded in dX.u around 0
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3263.8%
Simplified63.8%
associate-*r*N/A
/-rgt-identityN/A
frac-2negN/A
metadata-evalN/A
distribute-rgt-neg-inN/A
associate-*l/N/A
associate-*l*N/A
frac-2negN/A
metadata-evalN/A
/-rgt-identityN/A
*-commutativeN/A
+-lft-identityN/A
distribute-lft-inN/A
*-commutativeN/A
/-rgt-identityN/A
metadata-evalN/A
frac-2negN/A
associate-*l*N/A
Applied egg-rr64.1%
if 3.39999997e-6 < dX.u Initial program 73.1%
Simplified73.3%
Applied egg-rr72.5%
Taylor expanded in dX.v around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.9%
Simplified68.9%
Taylor expanded in dX.v around 0
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3271.5%
Simplified71.5%
Final simplification66.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (+ (pow t_0 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_2 (pow (floor w) 2.0))
(t_3 (* (floor h) dX.v))
(t_4 (pow t_3 2.0))
(t_5 (pow (fmax (+ t_4 (pow (* (floor w) dX.u) 2.0)) t_1) 0.5)))
(if (<= dX.u 3.3999999686784577e-6)
(if (>= t_4 t_1) (/ t_3 t_5) (/ t_0 t_5))
(if (>= (* dX.u (* dX.u t_2)) t_1)
(/ (floor h) (/ t_5 dX.v))
(/ t_0 (pow (fmax (* (* dX.u dX.u) t_2) t_1) 0.5))))))
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(h) * dY_46_v;
float t_1 = powf(t_0, 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_2 = powf(floorf(w), 2.0f);
float t_3 = floorf(h) * dX_46_v;
float t_4 = powf(t_3, 2.0f);
float t_5 = powf(fmaxf((t_4 + powf((floorf(w) * dX_46_u), 2.0f)), t_1), 0.5f);
float tmp_1;
if (dX_46_u <= 3.3999999686784577e-6f) {
float tmp_2;
if (t_4 >= t_1) {
tmp_2 = t_3 / t_5;
} else {
tmp_2 = t_0 / t_5;
}
tmp_1 = tmp_2;
} else if ((dX_46_u * (dX_46_u * t_2)) >= t_1) {
tmp_1 = floorf(h) / (t_5 / dX_46_v);
} else {
tmp_1 = t_0 / powf(fmaxf(((dX_46_u * dX_46_u) * t_2), t_1), 0.5f);
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dY_46_v) t_1 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_2 = floor(w) ^ Float32(2.0) t_3 = Float32(floor(h) * dX_46_v) t_4 = t_3 ^ Float32(2.0) t_5 = ((Float32(t_4 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) != Float32(t_4 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0)))) ? t_1 : ((t_1 != t_1) ? Float32(t_4 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) : max(Float32(t_4 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_1))) ^ Float32(0.5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(3.3999999686784577e-6)) tmp_2 = Float32(0.0) if (t_4 >= t_1) tmp_2 = Float32(t_3 / t_5); else tmp_2 = Float32(t_0 / t_5); end tmp_1 = tmp_2; elseif (Float32(dX_46_u * Float32(dX_46_u * t_2)) >= t_1) tmp_1 = Float32(floor(h) / Float32(t_5 / dX_46_v)); else tmp_1 = Float32(t_0 / (((Float32(Float32(dX_46_u * dX_46_u) * t_2) != Float32(Float32(dX_46_u * dX_46_u) * t_2)) ? t_1 : ((t_1 != t_1) ? Float32(Float32(dX_46_u * dX_46_u) * t_2) : max(Float32(Float32(dX_46_u * dX_46_u) * t_2), t_1))) ^ Float32(0.5))); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dY_46_v; t_1 = (t_0 ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_2 = floor(w) ^ single(2.0); t_3 = floor(h) * dX_46_v; t_4 = t_3 ^ single(2.0); t_5 = max((t_4 + ((floor(w) * dX_46_u) ^ single(2.0))), t_1) ^ single(0.5); tmp_2 = single(0.0); if (dX_46_u <= single(3.3999999686784577e-6)) tmp_3 = single(0.0); if (t_4 >= t_1) tmp_3 = t_3 / t_5; else tmp_3 = t_0 / t_5; end tmp_2 = tmp_3; elseif ((dX_46_u * (dX_46_u * t_2)) >= t_1) tmp_2 = floor(h) / (t_5 / dX_46_v); else tmp_2 = t_0 / (max(((dX_46_u * dX_46_u) * t_2), t_1) ^ single(0.5)); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := {t\_0}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_2 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := {t\_3}^{2}\\
t_5 := {\left(\mathsf{max}\left(t\_4 + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_1\right)\right)}^{0.5}\\
\mathbf{if}\;dX.u \leq 3.3999999686784577 \cdot 10^{-6}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_1:\\
\;\;\;\;\frac{t\_3}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_5}\\
\end{array}\\
\mathbf{elif}\;dX.u \cdot \left(dX.u \cdot t\_2\right) \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\frac{t\_5}{dX.v}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{{\left(\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot t\_2, t\_1\right)\right)}^{0.5}}\\
\end{array}
\end{array}
if dX.u < 3.39999997e-6Initial program 72.4%
Simplified72.8%
Applied egg-rr72.8%
clear-numN/A
div-invN/A
associate-/l*N/A
associate-/r*N/A
/-lowering-/.f32N/A
Applied egg-rr72.4%
Taylor expanded in dX.v around inf
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3263.5%
Simplified63.5%
Applied egg-rr63.8%
if 3.39999997e-6 < dX.u Initial program 73.1%
Simplified73.3%
Applied egg-rr72.5%
Taylor expanded in dX.v around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3268.9%
Simplified68.9%
Taylor expanded in dX.v around 0
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3271.5%
Simplified71.5%
Final simplification66.7%
(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 (* (floor h) dY.v))
(t_2 (* (floor h) dX.v))
(t_3 (+ (pow t_1 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_4 (pow t_2 2.0))
(t_5 (pow (fmax (+ t_4 t_0) t_3) 0.5))
(t_6 (/ t_1 t_5)))
(if (<= dX.u 10000.0)
(if (>= t_4 t_3) (/ t_2 t_5) t_6)
(if (>= t_0 t_3) (* (floor h) (/ dX.v t_5)) t_6))))
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 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * dX_46_v;
float t_3 = powf(t_1, 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_4 = powf(t_2, 2.0f);
float t_5 = powf(fmaxf((t_4 + t_0), t_3), 0.5f);
float t_6 = t_1 / t_5;
float tmp_1;
if (dX_46_u <= 10000.0f) {
float tmp_2;
if (t_4 >= t_3) {
tmp_2 = t_2 / t_5;
} else {
tmp_2 = t_6;
}
tmp_1 = tmp_2;
} else if (t_0 >= t_3) {
tmp_1 = floorf(h) * (dX_46_v / t_5);
} else {
tmp_1 = t_6;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) ^ Float32(2.0) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32((t_1 ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_4 = t_2 ^ Float32(2.0) t_5 = ((Float32(t_4 + t_0) != Float32(t_4 + t_0)) ? t_3 : ((t_3 != t_3) ? Float32(t_4 + t_0) : max(Float32(t_4 + t_0), t_3))) ^ Float32(0.5) t_6 = Float32(t_1 / t_5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(10000.0)) tmp_2 = Float32(0.0) if (t_4 >= t_3) tmp_2 = Float32(t_2 / t_5); else tmp_2 = t_6; end tmp_1 = tmp_2; elseif (t_0 >= t_3) tmp_1 = Float32(floor(h) * Float32(dX_46_v / t_5)); else tmp_1 = t_6; end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = (floor(w) * dX_46_u) ^ single(2.0); t_1 = floor(h) * dY_46_v; t_2 = floor(h) * dX_46_v; t_3 = (t_1 ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_4 = t_2 ^ single(2.0); t_5 = max((t_4 + t_0), t_3) ^ single(0.5); t_6 = t_1 / t_5; tmp_2 = single(0.0); if (dX_46_u <= single(10000.0)) tmp_3 = single(0.0); if (t_4 >= t_3) tmp_3 = t_2 / t_5; else tmp_3 = t_6; end tmp_2 = tmp_3; elseif (t_0 >= t_3) tmp_2 = floor(h) * (dX_46_v / t_5); else tmp_2 = t_6; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := {t\_1}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_4 := {t\_2}^{2}\\
t_5 := {\left(\mathsf{max}\left(t\_4 + t\_0, t\_3\right)\right)}^{0.5}\\
t_6 := \frac{t\_1}{t\_5}\\
\mathbf{if}\;dX.u \leq 10000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_3:\\
\;\;\;\;\frac{t\_2}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_3:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.u < 1e4Initial program 75.9%
Simplified76.3%
Applied egg-rr76.2%
clear-numN/A
div-invN/A
associate-/l*N/A
associate-/r*N/A
/-lowering-/.f32N/A
Applied egg-rr75.9%
Taylor expanded in dX.v around inf
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3267.4%
Simplified67.4%
Applied egg-rr67.7%
if 1e4 < dX.u Initial program 63.6%
Simplified63.7%
Applied egg-rr62.5%
Taylor expanded in dX.v around 0
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3261.4%
Simplified61.4%
associate-*r*N/A
*-commutativeN/A
unpow2N/A
associate-*r*N/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow2N/A
unpow-prod-downN/A
pow-lowering-pow.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
unpow2N/A
associate-*r*N/A
+-commutativeN/A
+-lowering-+.f32N/A
Applied egg-rr61.4%
clear-numN/A
associate-/r/N/A
clear-numN/A
*-lowering-*.f32N/A
Applied egg-rr62.5%
Final simplification66.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor h) dX.v))
(t_2 (+ (pow t_0 2.0) (pow (* (floor w) dY.u) 2.0)))
(t_3 (pow t_1 2.0))
(t_4 (pow (fmax (+ t_3 (pow (* (floor w) dX.u) 2.0)) t_2) 0.5)))
(if (>= t_3 t_2) (/ t_1 t_4) (/ t_0 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) * dY_46_v;
float t_1 = floorf(h) * dX_46_v;
float t_2 = powf(t_0, 2.0f) + powf((floorf(w) * dY_46_u), 2.0f);
float t_3 = powf(t_1, 2.0f);
float t_4 = powf(fmaxf((t_3 + powf((floorf(w) * dX_46_u), 2.0f)), t_2), 0.5f);
float tmp;
if (t_3 >= t_2) {
tmp = t_1 / t_4;
} else {
tmp = t_0 / 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) * dY_46_v) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32((t_0 ^ Float32(2.0)) + (Float32(floor(w) * dY_46_u) ^ Float32(2.0))) t_3 = t_1 ^ Float32(2.0) t_4 = ((Float32(t_3 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) != Float32(t_3 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0)))) ? t_2 : ((t_2 != t_2) ? Float32(t_3 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) : max(Float32(t_3 + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_2))) ^ Float32(0.5) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_1 / t_4); else tmp = Float32(t_0 / 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) * dY_46_v; t_1 = floor(h) * dX_46_v; t_2 = (t_0 ^ single(2.0)) + ((floor(w) * dY_46_u) ^ single(2.0)); t_3 = t_1 ^ single(2.0); t_4 = max((t_3 + ((floor(w) * dX_46_u) ^ single(2.0))), t_2) ^ single(0.5); tmp = single(0.0); if (t_3 >= t_2) tmp = t_1 / t_4; else tmp = t_0 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := {t\_0}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dY.u\right)}^{2}\\
t_3 := {t\_1}^{2}\\
t_4 := {\left(\mathsf{max}\left(t\_3 + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_2\right)\right)}^{0.5}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\end{array}
\end{array}
Initial program 72.7%
Simplified73.0%
Applied egg-rr72.7%
clear-numN/A
div-invN/A
associate-/l*N/A
associate-/r*N/A
/-lowering-/.f32N/A
Applied egg-rr72.6%
Taylor expanded in dX.v around inf
*-lowering-*.f32N/A
unpow2N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3260.9%
Simplified60.9%
Applied egg-rr61.1%
Final simplification61.1%
herbie shell --seed 2024164
(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))))