
(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\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_2 := \left\lfloorw\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloorh\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\lfloorh\right\rfloor \cdot dX.v\\
t_1 := \left\lfloorw\right\rfloor \cdot dY.u\\
t_2 := \left\lfloorw\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloorh\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 (+ (pow (* (floor w) dX.u) 2.0) (pow t_0 2.0)))
(t_2 (* (floor h) dY.v))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) (pow t_2 2.0)))
(t_4 (pow (fmax t_1 t_3) 0.5)))
(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 = powf((floorf(w) * dX_46_u), 2.0f) + powf(t_0, 2.0f);
float t_2 = floorf(h) * dY_46_v;
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_2, 2.0f);
float t_4 = powf(fmaxf(t_1, t_3), 0.5f);
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) * dX_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = ((t_1 != t_1) ? t_3 : ((t_3 != t_3) ? t_1 : max(t_1, t_3))) ^ Float32(0.5) 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) * dX_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = floor(h) * dY_46_v; t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = max(t_1, t_3) ^ single(0.5); 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\lfloorh\right\rfloor \cdot dX.v\\
t_1 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2} + {t\_0}^{2}\\
t_2 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_3 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {t\_2}^{2}\\
t_4 := {\left(\mathsf{max}\left(t\_1, t\_3\right)\right)}^{0.5}\\
\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 78.2%
Simplified78.3%
Applied egg-rr78.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (pow (* (floor w) dX.u) 2.0))
(t_2 (pow (floor w) 2.0))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) (pow t_0 2.0)))
(t_4 (* (floor h) dX.v))
(t_5 (pow t_4 2.0))
(t_6 (pow (fmax (+ t_1 t_5) t_3) 0.5))
(t_7 (/ t_0 t_6))
(t_8 (pow (floor h) 2.0)))
(if (<= dX.u 186.0)
(if (>= t_5 t_3) (/ t_4 t_6) t_7)
(if (>= t_1 t_3)
(*
t_4
(sqrt
(/
1.0
(fmax
(+ (* dX.v (* dX.v t_8)) (* t_2 (* dX.u dX.u)))
(+ (* dY.u (* dY.u t_2)) (* t_8 (* dY.v dY.v)))))))
t_7))))
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((floorf(w) * dX_46_u), 2.0f);
float t_2 = powf(floorf(w), 2.0f);
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + powf(t_0, 2.0f);
float t_4 = floorf(h) * dX_46_v;
float t_5 = powf(t_4, 2.0f);
float t_6 = powf(fmaxf((t_1 + t_5), t_3), 0.5f);
float t_7 = t_0 / t_6;
float t_8 = powf(floorf(h), 2.0f);
float tmp_1;
if (dX_46_u <= 186.0f) {
float tmp_2;
if (t_5 >= t_3) {
tmp_2 = t_4 / t_6;
} else {
tmp_2 = t_7;
}
tmp_1 = tmp_2;
} else if (t_1 >= t_3) {
tmp_1 = t_4 * sqrtf((1.0f / fmaxf(((dX_46_v * (dX_46_v * t_8)) + (t_2 * (dX_46_u * dX_46_u))), ((dY_46_u * (dY_46_u * t_2)) + (t_8 * (dY_46_v * dY_46_v))))));
} else {
tmp_1 = t_7;
}
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(floor(w) * dX_46_u) ^ Float32(2.0) t_2 = floor(w) ^ Float32(2.0) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_4 = Float32(floor(h) * dX_46_v) t_5 = t_4 ^ Float32(2.0) t_6 = ((Float32(t_1 + t_5) != Float32(t_1 + t_5)) ? t_3 : ((t_3 != t_3) ? Float32(t_1 + t_5) : max(Float32(t_1 + t_5), t_3))) ^ Float32(0.5) t_7 = Float32(t_0 / t_6) t_8 = floor(h) ^ Float32(2.0) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(186.0)) tmp_2 = Float32(0.0) if (t_5 >= t_3) tmp_2 = Float32(t_4 / t_6); else tmp_2 = t_7; end tmp_1 = tmp_2; elseif (t_1 >= t_3) tmp_1 = Float32(t_4 * sqrt(Float32(Float32(1.0) / ((Float32(Float32(dX_46_v * Float32(dX_46_v * t_8)) + Float32(t_2 * Float32(dX_46_u * dX_46_u))) != Float32(Float32(dX_46_v * Float32(dX_46_v * t_8)) + Float32(t_2 * Float32(dX_46_u * dX_46_u)))) ? Float32(Float32(dY_46_u * Float32(dY_46_u * t_2)) + Float32(t_8 * Float32(dY_46_v * dY_46_v))) : ((Float32(Float32(dY_46_u * Float32(dY_46_u * t_2)) + Float32(t_8 * Float32(dY_46_v * dY_46_v))) != Float32(Float32(dY_46_u * Float32(dY_46_u * t_2)) + Float32(t_8 * Float32(dY_46_v * dY_46_v)))) ? Float32(Float32(dX_46_v * Float32(dX_46_v * t_8)) + Float32(t_2 * Float32(dX_46_u * dX_46_u))) : max(Float32(Float32(dX_46_v * Float32(dX_46_v * t_8)) + Float32(t_2 * Float32(dX_46_u * dX_46_u))), Float32(Float32(dY_46_u * Float32(dY_46_u * t_2)) + Float32(t_8 * Float32(dY_46_v * dY_46_v))))))))); else tmp_1 = t_7; 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 = (floor(w) * dX_46_u) ^ single(2.0); t_2 = floor(w) ^ single(2.0); t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_4 = floor(h) * dX_46_v; t_5 = t_4 ^ single(2.0); t_6 = max((t_1 + t_5), t_3) ^ single(0.5); t_7 = t_0 / t_6; t_8 = floor(h) ^ single(2.0); tmp_2 = single(0.0); if (dX_46_u <= single(186.0)) tmp_3 = single(0.0); if (t_5 >= t_3) tmp_3 = t_4 / t_6; else tmp_3 = t_7; end tmp_2 = tmp_3; elseif (t_1 >= t_3) tmp_2 = t_4 * sqrt((single(1.0) / max(((dX_46_v * (dX_46_v * t_8)) + (t_2 * (dX_46_u * dX_46_u))), ((dY_46_u * (dY_46_u * t_2)) + (t_8 * (dY_46_v * dY_46_v)))))); else tmp_2 = t_7; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_1 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2}\\
t_2 := {\left(\left\lfloorw\right\rfloor\right)}^{2}\\
t_3 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {t\_0}^{2}\\
t_4 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_5 := {t\_4}^{2}\\
t_6 := {\left(\mathsf{max}\left(t\_1 + t\_5, t\_3\right)\right)}^{0.5}\\
t_7 := \frac{t\_0}{t\_6}\\
t_8 := {\left(\left\lfloorh\right\rfloor\right)}^{2}\\
\mathbf{if}\;dX.u \leq 186:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;\frac{t\_4}{t\_6}\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}\\
\mathbf{elif}\;t\_1 \geq t\_3:\\
\;\;\;\;t\_4 \cdot \sqrt{\frac{1}{\mathsf{max}\left(dX.v \cdot \left(dX.v \cdot t\_8\right) + t\_2 \cdot \left(dX.u \cdot dX.u\right), dY.u \cdot \left(dY.u \cdot t\_2\right) + t\_8 \cdot \left(dY.v \cdot dY.v\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
if dX.u < 186Initial program 79.1%
Simplified79.3%
Applied egg-rr79.5%
Taylor expanded in dX.u 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.f3270.1%
Simplified70.1%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
unpow2N/A
swap-sqrN/A
associate-*r*N/A
Applied egg-rr70.1%
if 186 < dX.u Initial program 75.4%
Simplified75.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3270.7%
Simplified70.7%
Applied egg-rr71.0%
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3271.0%
Applied egg-rr71.0%
Taylor expanded in h around 0
*-lowering-*.f32N/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
sqrt-lowering-sqrt.f32N/A
*-inversesN/A
/-lowering-/.f32N/A
*-inversesN/A
Simplified71.1%
Final simplification70.4%
(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 (pow t_1 2.0))
(t_3 (+ (pow (* (floor w) dY.u) 2.0) t_2))
(t_4 (* (floor h) dX.v))
(t_5 (pow t_4 2.0))
(t_6 (+ t_0 t_5))
(t_7 (pow (fmax t_6 t_3) 0.5))
(t_8 (/ t_4 t_7)))
(if (<= dX.u 186.0)
(if (>= t_5 t_3) t_8 (/ t_1 t_7))
(if (>= t_0 t_3)
t_8
(/
t_1
(pow (fmax t_6 (+ t_2 (* (pow (floor w) 2.0) (* dY.u dY.u)))) 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 = powf((floorf(w) * dX_46_u), 2.0f);
float t_1 = floorf(h) * dY_46_v;
float t_2 = powf(t_1, 2.0f);
float t_3 = powf((floorf(w) * dY_46_u), 2.0f) + t_2;
float t_4 = floorf(h) * dX_46_v;
float t_5 = powf(t_4, 2.0f);
float t_6 = t_0 + t_5;
float t_7 = powf(fmaxf(t_6, t_3), 0.5f);
float t_8 = t_4 / t_7;
float tmp_1;
if (dX_46_u <= 186.0f) {
float tmp_2;
if (t_5 >= t_3) {
tmp_2 = t_8;
} else {
tmp_2 = t_1 / t_7;
}
tmp_1 = tmp_2;
} else if (t_0 >= t_3) {
tmp_1 = t_8;
} else {
tmp_1 = t_1 / powf(fmaxf(t_6, (t_2 + (powf(floorf(w), 2.0f) * (dY_46_u * dY_46_u)))), 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(w) * dX_46_u) ^ Float32(2.0) t_1 = Float32(floor(h) * dY_46_v) t_2 = t_1 ^ Float32(2.0) t_3 = Float32((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + t_2) t_4 = Float32(floor(h) * dX_46_v) t_5 = t_4 ^ Float32(2.0) t_6 = Float32(t_0 + t_5) t_7 = ((t_6 != t_6) ? t_3 : ((t_3 != t_3) ? t_6 : max(t_6, t_3))) ^ Float32(0.5) t_8 = Float32(t_4 / t_7) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(186.0)) tmp_2 = Float32(0.0) if (t_5 >= t_3) tmp_2 = t_8; else tmp_2 = Float32(t_1 / t_7); end tmp_1 = tmp_2; elseif (t_0 >= t_3) tmp_1 = t_8; else tmp_1 = Float32(t_1 / (((t_6 != t_6) ? Float32(t_2 + Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u))) : ((Float32(t_2 + Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u))) != Float32(t_2 + Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)))) ? t_6 : max(t_6, Float32(t_2 + Float32((floor(w) ^ Float32(2.0)) * Float32(dY_46_u * dY_46_u)))))) ^ 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(w) * dX_46_u) ^ single(2.0); t_1 = floor(h) * dY_46_v; t_2 = t_1 ^ single(2.0); t_3 = ((floor(w) * dY_46_u) ^ single(2.0)) + t_2; t_4 = floor(h) * dX_46_v; t_5 = t_4 ^ single(2.0); t_6 = t_0 + t_5; t_7 = max(t_6, t_3) ^ single(0.5); t_8 = t_4 / t_7; tmp_2 = single(0.0); if (dX_46_u <= single(186.0)) tmp_3 = single(0.0); if (t_5 >= t_3) tmp_3 = t_8; else tmp_3 = t_1 / t_7; end tmp_2 = tmp_3; elseif (t_0 >= t_3) tmp_2 = t_8; else tmp_2 = t_1 / (max(t_6, (t_2 + ((floor(w) ^ single(2.0)) * (dY_46_u * dY_46_u)))) ^ single(0.5)); end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_2 := {t\_1}^{2}\\
t_3 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + t\_2\\
t_4 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_5 := {t\_4}^{2}\\
t_6 := t\_0 + t\_5\\
t_7 := {\left(\mathsf{max}\left(t\_6, t\_3\right)\right)}^{0.5}\\
t_8 := \frac{t\_4}{t\_7}\\
\mathbf{if}\;dX.u \leq 186:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_7}\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_3:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{{\left(\mathsf{max}\left(t\_6, t\_2 + {\left(\left\lfloorw\right\rfloor\right)}^{2} \cdot \left(dY.u \cdot dY.u\right)\right)\right)}^{0.5}}\\
\end{array}
\end{array}
if dX.u < 186Initial program 79.1%
Simplified79.3%
Applied egg-rr79.5%
Taylor expanded in dX.u 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.f3270.1%
Simplified70.1%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
unpow2N/A
swap-sqrN/A
associate-*r*N/A
Applied egg-rr70.1%
if 186 < dX.u Initial program 75.4%
Simplified75.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3270.7%
Simplified70.7%
Applied egg-rr71.0%
unpow-prod-downN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
*-lowering-*.f3271.0%
Applied egg-rr71.0%
Final simplification70.4%
(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 (* (floor w) dY.u) 2.0) (pow t_1 2.0)))
(t_4 (pow t_2 2.0))
(t_5 (pow (fmax (+ t_0 t_4) t_3) 0.5))
(t_6 (/ t_2 t_5))
(t_7 (/ t_1 t_5)))
(if (<= dX.u 186.0) (if (>= t_4 t_3) t_6 t_7) (if (>= t_0 t_3) t_6 t_7))))
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((floorf(w) * dY_46_u), 2.0f) + powf(t_1, 2.0f);
float t_4 = powf(t_2, 2.0f);
float t_5 = powf(fmaxf((t_0 + t_4), t_3), 0.5f);
float t_6 = t_2 / t_5;
float t_7 = t_1 / t_5;
float tmp_1;
if (dX_46_u <= 186.0f) {
float tmp_2;
if (t_4 >= t_3) {
tmp_2 = t_6;
} else {
tmp_2 = t_7;
}
tmp_1 = tmp_2;
} else if (t_0 >= t_3) {
tmp_1 = t_6;
} else {
tmp_1 = t_7;
}
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((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = t_2 ^ Float32(2.0) t_5 = ((Float32(t_0 + t_4) != Float32(t_0 + t_4)) ? t_3 : ((t_3 != t_3) ? Float32(t_0 + t_4) : max(Float32(t_0 + t_4), t_3))) ^ Float32(0.5) t_6 = Float32(t_2 / t_5) t_7 = Float32(t_1 / t_5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(186.0)) tmp_2 = Float32(0.0) if (t_4 >= t_3) tmp_2 = t_6; else tmp_2 = t_7; end tmp_1 = tmp_2; elseif (t_0 >= t_3) tmp_1 = t_6; else tmp_1 = t_7; 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 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = t_2 ^ single(2.0); t_5 = max((t_0 + t_4), t_3) ^ single(0.5); t_6 = t_2 / t_5; t_7 = t_1 / t_5; tmp_2 = single(0.0); if (dX_46_u <= single(186.0)) tmp_3 = single(0.0); if (t_4 >= t_3) tmp_3 = t_6; else tmp_3 = t_7; end tmp_2 = tmp_3; elseif (t_0 >= t_3) tmp_2 = t_6; else tmp_2 = t_7; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_2 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_3 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {t\_1}^{2}\\
t_4 := {t\_2}^{2}\\
t_5 := {\left(\mathsf{max}\left(t\_0 + t\_4, t\_3\right)\right)}^{0.5}\\
t_6 := \frac{t\_2}{t\_5}\\
t_7 := \frac{t\_1}{t\_5}\\
\mathbf{if}\;dX.u \leq 186:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_3:\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_3:\\
\;\;\;\;t\_6\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
if dX.u < 186Initial program 79.1%
Simplified79.3%
Applied egg-rr79.5%
Taylor expanded in dX.u 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.f3270.1%
Simplified70.1%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
unpow2N/A
swap-sqrN/A
associate-*r*N/A
Applied egg-rr70.1%
if 186 < dX.u Initial program 75.4%
Simplified75.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3270.7%
Simplified70.7%
Applied egg-rr71.0%
Final simplification70.4%
(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 (* (floor w) dY.u) 2.0) (pow t_1 2.0)))
(t_4 (pow t_2 2.0))
(t_5 (pow (fmax (+ t_0 t_4) t_3) 0.5))
(t_6 (/ t_1 t_5)))
(if (<= dX.u 186.0)
(if (>= t_4 t_3) (/ t_2 t_5) t_6)
(if (>= t_0 t_3) (* dX.v (/ (floor h) 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((floorf(w) * dY_46_u), 2.0f) + powf(t_1, 2.0f);
float t_4 = powf(t_2, 2.0f);
float t_5 = powf(fmaxf((t_0 + t_4), t_3), 0.5f);
float t_6 = t_1 / t_5;
float tmp_1;
if (dX_46_u <= 186.0f) {
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 = dX_46_v * (floorf(h) / 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((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = t_2 ^ Float32(2.0) t_5 = ((Float32(t_0 + t_4) != Float32(t_0 + t_4)) ? t_3 : ((t_3 != t_3) ? Float32(t_0 + t_4) : max(Float32(t_0 + t_4), t_3))) ^ Float32(0.5) t_6 = Float32(t_1 / t_5) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(186.0)) 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(dX_46_v * Float32(floor(h) / 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 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_1 ^ single(2.0)); t_4 = t_2 ^ single(2.0); t_5 = max((t_0 + t_4), t_3) ^ single(0.5); t_6 = t_1 / t_5; tmp_2 = single(0.0); if (dX_46_u <= single(186.0)) 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 = dX_46_v * (floor(h) / t_5); else tmp_2 = t_6; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2}\\
t_1 := \left\lfloorh\right\rfloor \cdot dY.v\\
t_2 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_3 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {t\_1}^{2}\\
t_4 := {t\_2}^{2}\\
t_5 := {\left(\mathsf{max}\left(t\_0 + t\_4, t\_3\right)\right)}^{0.5}\\
t_6 := \frac{t\_1}{t\_5}\\
\mathbf{if}\;dX.u \leq 186:\\
\;\;\;\;\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:\\
\;\;\;\;dX.v \cdot \frac{\left\lfloorh\right\rfloor}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.u < 186Initial program 79.1%
Simplified79.3%
Applied egg-rr79.5%
Taylor expanded in dX.u 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.f3270.1%
Simplified70.1%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
unpow2N/A
swap-sqrN/A
associate-*r*N/A
Applied egg-rr70.1%
if 186 < dX.u Initial program 75.4%
Simplified75.2%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3270.7%
Simplified70.7%
Applied egg-rr71.0%
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
*-commutativeN/A
*-lowering-*.f32N/A
*-lowering-*.f32N/A
pow-lowering-pow.f32N/A
floor-lowering-floor.f3271.0%
Applied egg-rr71.0%
Applied egg-rr70.9%
Final simplification70.3%
(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 (* (floor w) dY.u) 2.0) (pow t_0 2.0)))
(t_3 (pow t_1 2.0))
(t_4 (pow (fmax (+ (pow (* (floor w) dX.u) 2.0) t_3) 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((floorf(w) * dY_46_u), 2.0f) + powf(t_0, 2.0f);
float t_3 = powf(t_1, 2.0f);
float t_4 = powf(fmaxf((powf((floorf(w) * dX_46_u), 2.0f) + t_3), 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((Float32(floor(w) * dY_46_u) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = t_1 ^ Float32(2.0) t_4 = ((Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + t_3) != Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + t_3)) ? t_2 : ((t_2 != t_2) ? Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + t_3) : max(Float32((Float32(floor(w) * dX_46_u) ^ Float32(2.0)) + t_3), 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 = ((floor(w) * dY_46_u) ^ single(2.0)) + (t_0 ^ single(2.0)); t_3 = t_1 ^ single(2.0); t_4 = max((((floor(w) * dX_46_u) ^ single(2.0)) + t_3), 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\lfloorh\right\rfloor \cdot dY.v\\
t_1 := \left\lfloorh\right\rfloor \cdot dX.v\\
t_2 := {\left(\left\lfloorw\right\rfloor \cdot dY.u\right)}^{2} + {t\_0}^{2}\\
t_3 := {t\_1}^{2}\\
t_4 := {\left(\mathsf{max}\left({\left(\left\lfloorw\right\rfloor \cdot dX.u\right)}^{2} + t\_3, 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 78.2%
Simplified78.3%
Applied egg-rr78.5%
Taylor expanded in dX.u 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.f3266.3%
Simplified66.3%
unpow2N/A
associate-*r*N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
>=-lowering->=.f32N/A
associate-*r*N/A
unpow2N/A
pow-lowering-pow.f32N/A
*-commutativeN/A
*-lowering-*.f32N/A
floor-lowering-floor.f32N/A
pow2N/A
unpow2N/A
swap-sqrN/A
associate-*r*N/A
Applied egg-rr66.3%
Final simplification66.3%
herbie shell --seed 2024155
(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))))