
(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 4 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 (* dX.u (floor w)))
(t_1 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow (* dY.v (floor h)) 2.0) (pow t_2 2.0)))
(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 = dX_46_u * floorf(w);
float t_1 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_2, 2.0f);
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(dX_46_u * floor(w)) t_1 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) 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 = dX_46_u * floor(w); t_1 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_2 ^ single(2.0)); 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 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_2}^{2}\\
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 79.4%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites79.6%
Applied rewrites79.6%
Applied rewrites79.6%
lift-*.f32N/A
pow2N/A
lift-pow.f3279.6
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3279.6
lift-*.f32N/A
pow2N/A
lower-pow.f3279.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3279.6
lift-*.f32N/A
pow2N/A
Applied rewrites79.6%
Final simplification79.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* (floor h) dX.v))
(t_2 (* t_1 t_1))
(t_3 (* dX.u (floor w)))
(t_4 (pow t_3 2.0))
(t_5 (* dY.v (floor h))))
(if (>= (+ t_2 t_4) (* (* (pow (floor h) 2.0) dY.v) dY.v))
(/
t_3
(sqrt (fmax (+ (pow t_1 2.0) t_4) (+ (pow t_5 2.0) (pow t_0 2.0)))))
(*
(/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_2) (+ (* t_5 t_5) (* t_0 t_0)))))
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 = dY_46_u * floorf(w);
float t_1 = floorf(h) * dX_46_v;
float t_2 = t_1 * t_1;
float t_3 = dX_46_u * floorf(w);
float t_4 = powf(t_3, 2.0f);
float t_5 = dY_46_v * floorf(h);
float tmp;
if ((t_2 + t_4) >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp = t_3 / sqrtf(fmaxf((powf(t_1, 2.0f) + t_4), (powf(t_5, 2.0f) + powf(t_0, 2.0f))));
} else {
tmp = (1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_2), ((t_5 * t_5) + (t_0 * t_0))))) * t_0;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(dX_46_u * floor(w)) t_4 = t_3 ^ Float32(2.0) t_5 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (Float32(t_2 + t_4) >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp = Float32(t_3 / sqrt(((Float32((t_1 ^ Float32(2.0)) + t_4) != Float32((t_1 ^ Float32(2.0)) + t_4)) ? Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : ((Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? Float32((t_1 ^ Float32(2.0)) + t_4) : max(Float32((t_1 ^ Float32(2.0)) + t_4), Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))))))); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + t_2) != Float32(Float32(t_3 * t_3) + t_2)) ? Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) : ((Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) != Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0))) ? Float32(Float32(t_3 * t_3) + t_2) : max(Float32(Float32(t_3 * t_3) + t_2), Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0))))))) * 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 = dY_46_u * floor(w); t_1 = floor(h) * dX_46_v; t_2 = t_1 * t_1; t_3 = dX_46_u * floor(w); t_4 = t_3 ^ single(2.0); t_5 = dY_46_v * floor(h); tmp = single(0.0); if ((t_2 + t_4) >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp = t_3 / sqrt(max(((t_1 ^ single(2.0)) + t_4), ((t_5 ^ single(2.0)) + (t_0 ^ single(2.0))))); else tmp = (single(1.0) / sqrt(max(((t_3 * t_3) + t_2), ((t_5 * t_5) + (t_0 * t_0))))) * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_4 := {t\_3}^{2}\\
t_5 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;t\_2 + t\_4 \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left({t\_1}^{2} + t\_4, {t\_5}^{2} + {t\_0}^{2}\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_2, t\_5 \cdot t\_5 + t\_0 \cdot t\_0\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 79.4%
Taylor expanded in dY.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.7
Applied rewrites66.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.7
Applied rewrites66.7%
Applied rewrites66.9%
Final simplification66.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* t_0 t_0))
(t_2 (* dY.v (floor h)))
(t_3 (* dY.u (floor w)))
(t_4 (+ (* t_2 t_2) (* t_3 t_3)))
(t_5 (* (floor h) dX.v)))
(if (>=
(+ (pow t_5 2.0) (pow t_0 2.0))
(* (* (pow (floor h) 2.0) dY.v) dY.v))
(*
(/
1.0
(sqrt (fmax (+ (* (* (* dX.v dX.v) (floor h)) (floor h)) t_1) t_4)))
t_0)
(* (/ 1.0 (sqrt (fmax (+ t_1 (* t_5 t_5)) t_4))) t_3))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_u * floorf(w);
float t_1 = t_0 * t_0;
float t_2 = dY_46_v * floorf(h);
float t_3 = dY_46_u * floorf(w);
float t_4 = (t_2 * t_2) + (t_3 * t_3);
float t_5 = floorf(h) * dX_46_v;
float tmp;
if ((powf(t_5, 2.0f) + powf(t_0, 2.0f)) >= ((powf(floorf(h), 2.0f) * dY_46_v) * dY_46_v)) {
tmp = (1.0f / sqrtf(fmaxf(((((dX_46_v * dX_46_v) * floorf(h)) * floorf(h)) + t_1), t_4))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_1 + (t_5 * t_5)), t_4))) * t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) t_1 = Float32(t_0 * t_0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(dY_46_u * floor(w)) t_4 = Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3)) t_5 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (Float32((t_5 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= Float32(Float32((floor(h) ^ Float32(2.0)) * dY_46_v) * dY_46_v)) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(Float32(dX_46_v * dX_46_v) * floor(h)) * floor(h)) + t_1) != Float32(Float32(Float32(Float32(dX_46_v * dX_46_v) * floor(h)) * floor(h)) + t_1)) ? t_4 : ((t_4 != t_4) ? Float32(Float32(Float32(Float32(dX_46_v * dX_46_v) * floor(h)) * floor(h)) + t_1) : max(Float32(Float32(Float32(Float32(dX_46_v * dX_46_v) * floor(h)) * floor(h)) + t_1), t_4))))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_1 + Float32(t_5 * t_5)) != Float32(t_1 + Float32(t_5 * t_5))) ? t_4 : ((t_4 != t_4) ? Float32(t_1 + Float32(t_5 * t_5)) : max(Float32(t_1 + Float32(t_5 * t_5)), t_4))))) * t_3); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = dX_46_u * floor(w); t_1 = t_0 * t_0; t_2 = dY_46_v * floor(h); t_3 = dY_46_u * floor(w); t_4 = (t_2 * t_2) + (t_3 * t_3); t_5 = floor(h) * dX_46_v; tmp = single(0.0); if (((t_5 ^ single(2.0)) + (t_0 ^ single(2.0))) >= (((floor(h) ^ single(2.0)) * dY_46_v) * dY_46_v)) tmp = (single(1.0) / sqrt(max(((((dX_46_v * dX_46_v) * floor(h)) * floor(h)) + t_1), t_4))) * t_0; else tmp = (single(1.0) / sqrt(max((t_1 + (t_5 * t_5)), t_4))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := t\_0 \cdot t\_0\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := t\_2 \cdot t\_2 + t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;{t\_5}^{2} + {t\_0}^{2} \geq \left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dY.v\right) \cdot dY.v:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor + t\_1, t\_4\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_1 + t\_5 \cdot t\_5, t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 79.4%
Taylor expanded in dY.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.7
Applied rewrites66.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.7
Applied rewrites66.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.7
Applied rewrites66.7%
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-*.f3266.8
Applied rewrites66.8%
Final simplification66.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) dX.v))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_0 t_0) (* t_3 t_3)) (+ (* t_2 t_2) (* t_1 t_1)))))))
(if (>= (+ (pow t_3 2.0) (pow t_0 2.0)) (pow t_2 2.0))
(* t_4 t_0)
(* 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 = dX_46_u * floorf(w);
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) * dX_46_v;
float t_4 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + (t_3 * t_3)), ((t_2 * t_2) + (t_1 * t_1))));
float tmp;
if ((powf(t_3, 2.0f) + powf(t_0, 2.0f)) >= powf(t_2, 2.0f)) {
tmp = t_4 * t_0;
} 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(dX_46_u * floor(w)) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) != Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3))) ? Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) : max(Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)), Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) >= (t_2 ^ Float32(2.0))) tmp = Float32(t_4 * t_0); 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 = dX_46_u * floor(w); t_1 = dY_46_u * floor(w); t_2 = dY_46_v * floor(h); t_3 = floor(h) * dX_46_v; t_4 = single(1.0) / sqrt(max(((t_0 * t_0) + (t_3 * t_3)), ((t_2 * t_2) + (t_1 * t_1)))); tmp = single(0.0); if (((t_3 ^ single(2.0)) + (t_0 ^ single(2.0))) >= (t_2 ^ single(2.0))) tmp = t_4 * t_0; else tmp = t_4 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_3 \cdot t\_3, t\_2 \cdot t\_2 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_3}^{2} + {t\_0}^{2} \geq {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 79.4%
Taylor expanded in dY.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.7
Applied rewrites66.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.7
Applied rewrites66.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.7
Applied rewrites66.7%
Applied rewrites66.7%
Final simplification66.7%
herbie shell --seed 2024295
(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))))