
(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 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_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 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(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 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(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 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(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 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
\end{array}
Herbie found 3 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 h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2))))
(t_5 (sqrt t_4))
(t_6 (fabs (- (* t_3 t_2) (* t_0 t_1)))))
(log2
(if (> (/ t_4 t_6) (floor maxAniso))
(/ t_5 (floor maxAniso))
(/ t_6 t_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 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)));
float t_5 = sqrtf(t_4);
float t_6 = fabsf(((t_3 * t_2) - (t_0 * t_1)));
float tmp;
if ((t_4 / t_6) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = t_6 / t_5;
}
return log2f(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 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) t_5 = sqrt(t_4) t_6 = abs(Float32(Float32(t_3 * t_2) - Float32(t_0 * t_1))) tmp = Float32(0.0) if (Float32(t_4 / t_6) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(t_6 / t_5); end return log2(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 = max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))); t_5 = sqrt(t_4); t_6 = abs(((t_3 * t_2) - (t_0 * t_1))); tmp = single(0.0); if ((t_4 / t_6) > floor(maxAniso)) tmp = t_5 / floor(maxAniso); else tmp = t_6 / t_5; end tmp_2 = log2(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 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)\\
t_5 := \sqrt{t\_4}\\
t_6 := \left|t\_3 \cdot t\_2 - t\_0 \cdot t\_1\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_4}{t\_6} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_5}\\
\end{array}
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2
(fmax
(fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))
(fma (* t_0 dY.v) dY.v (* (* dY.u dY.u) t_1))))
(t_3 (sqrt t_2))
(t_4
(fabs (* (* (floor h) (floor w)) (- (* dY.u dX.v) (* dY.v dX.u))))))
(log2
(if (> (/ t_2 t_4) (floor maxAniso))
(/ t_3 (floor maxAniso))
(/ 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) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaxf(fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_1)));
float t_3 = sqrtf(t_2);
float t_4 = fabsf(((floorf(h) * floorf(w)) * ((dY_46_u * dX_46_v) - (dY_46_v * dX_46_u))));
float tmp;
if ((t_2 / t_4) > floorf(maxAniso)) {
tmp = t_3 / floorf(maxAniso);
} else {
tmp = t_4 / t_3;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fmax(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_1))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(floor(h) * floor(w)) * Float32(Float32(dY_46_u * dX_46_v) - Float32(dY_46_v * dX_46_u)))) tmp = Float32(0.0) if (Float32(t_2 / t_4) > floor(maxAniso)) tmp = Float32(t_3 / floor(maxAniso)); else tmp = Float32(t_4 / t_3); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dY.u \cdot dX.v - dY.v \cdot dX.u\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_4} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_3}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_3}\\
\end{array}
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1
(fmax
(fma (* (* (floor w) (floor w)) dX.u) dX.u (* t_0 (* dX.v dX.v)))
(fma (* t_0 dY.v) dY.v (* (* (* (floor w) dY.u) (floor w)) dY.u))))
(t_2 (sqrt t_1))
(t_3 (fabs (* (* (floor h) (floor w)) (* dY.u dX.v)))))
(log2
(if (> (/ t_1 t_3) (floor maxAniso))
(/ t_2 (floor maxAniso))
(/ t_3 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) * floorf(h);
float t_1 = fmaxf(fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_v), dY_46_v, (((floorf(w) * dY_46_u) * floorf(w)) * dY_46_u)));
float t_2 = sqrtf(t_1);
float t_3 = fabsf(((floorf(h) * floorf(w)) * (dY_46_u * dX_46_v)));
float tmp;
if ((t_1 / t_3) > floorf(maxAniso)) {
tmp = t_2 / floorf(maxAniso);
} else {
tmp = t_3 / t_2;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = fmax(fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_v), dY_46_v, Float32(Float32(Float32(floor(w) * dY_46_u) * floor(w)) * dY_46_u))) t_2 = sqrt(t_1) t_3 = abs(Float32(Float32(floor(h) * floor(w)) * Float32(dY_46_u * dX_46_v))) tmp = Float32(0.0) if (Float32(t_1 / t_3) > floor(maxAniso)) tmp = Float32(t_2 / floor(maxAniso)); else tmp = Float32(t_3 / t_2); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.v, dY.v, \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)\right)\\
t_2 := \sqrt{t\_1}\\
t_3 := \left|\left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dY.u \cdot dX.v\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_1}{t\_3} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_2}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{t\_2}\\
\end{array}
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.0
Applied rewrites75.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.0
Applied rewrites75.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.0
Applied rewrites75.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.0
Applied rewrites75.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.0
Applied rewrites75.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2
(fmax
(fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))
(fma (* t_0 dY.v) dY.v (* (* dY.u dY.u) t_1))))
(t_3 (sqrt t_2))
(t_4 (fabs (* (* (floor h) (floor w)) (* dY.u dX.v)))))
(log2
(if (> (/ t_2 t_4) (floor maxAniso))
(/ t_3 (floor maxAniso))
(/ 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) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaxf(fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_1)));
float t_3 = sqrtf(t_2);
float t_4 = fabsf(((floorf(h) * floorf(w)) * (dY_46_u * dX_46_v)));
float tmp;
if ((t_2 / t_4) > floorf(maxAniso)) {
tmp = t_3 / floorf(maxAniso);
} else {
tmp = t_4 / t_3;
}
return log2f(tmp);
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fmax(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_1))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(floor(h) * floor(w)) * Float32(dY_46_u * dX_46_v))) tmp = Float32(0.0) if (Float32(t_2 / t_4) > floor(maxAniso)) tmp = Float32(t_3 / floor(maxAniso)); else tmp = Float32(t_4 / t_3); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dY.u \cdot dX.v\right)\right|\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_2}{t\_4} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_3}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_3}\\
\end{array}
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.0
Applied rewrites75.0%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.0
Applied rewrites75.0%
herbie shell --seed 2025142
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (LOD)"
: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))
(log2 (if (> (/ (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)))) (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u))))) (floor maxAniso)) (/ (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 maxAniso)) (/ (fabs (- (* (* (floor w) dX.u) (* (floor h) dY.v)) (* (* (floor h) dX.v) (* (floor w) dY.u)))) (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)))))))))