
(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 5 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 h) dX.v))
(t_2
(fabs (* (* (floor h) (floor w)) (- (* dY.u dX.v) (* dY.v dX.u)))))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor w) (floor w)))
(t_6
(fmax
(fma (* t_5 dX.u) dX.u (* t_0 (* dX.v dX.v)))
(fma
(* t_0 dY.v)
dY.v
(exp (fma 2.0 (log (- (floor w))) (* 2.0 (log dY.u)))))))
(t_7 (sqrt t_6))
(t_8 (+ (* t_4 t_4) (* t_1 t_1)))
(t_9 (* (floor h) dY.v))
(t_10 (fmax t_8 (+ (* t_3 t_3) (* t_9 t_9))))
(t_11 (sqrt t_10))
(t_12 (fabs (- (* t_4 t_9) (* t_1 t_3))))
(t_13 (fmax t_8 (fma (* t_5 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(t_14 (sqrt t_13)))
(if (<=
(log2
(if (> (/ t_10 t_12) (floor maxAniso))
(/ t_11 (floor maxAniso))
(/ t_12 t_11)))
100.0)
(log2
(if (> (/ t_13 t_12) (floor maxAniso))
(/ t_14 (floor maxAniso))
(/ t_12 t_14)))
(log2
(if (> (/ t_6 t_2) (floor maxAniso))
(/ t_7 (floor maxAniso))
(/ t_2 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) * floorf(h);
float t_1 = floorf(h) * dX_46_v;
float t_2 = fabsf(((floorf(h) * floorf(w)) * ((dY_46_u * dX_46_v) - (dY_46_v * dX_46_u))));
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(w) * floorf(w);
float t_6 = fmaxf(fmaf((t_5 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_v), dY_46_v, expf(fmaf(2.0f, logf(-floorf(w)), (2.0f * logf(dY_46_u))))));
float t_7 = sqrtf(t_6);
float t_8 = (t_4 * t_4) + (t_1 * t_1);
float t_9 = floorf(h) * dY_46_v;
float t_10 = fmaxf(t_8, ((t_3 * t_3) + (t_9 * t_9)));
float t_11 = sqrtf(t_10);
float t_12 = fabsf(((t_4 * t_9) - (t_1 * t_3)));
float t_13 = fmaxf(t_8, fmaf((t_5 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0)));
float t_14 = sqrtf(t_13);
float tmp;
if ((t_10 / t_12) > floorf(maxAniso)) {
tmp = t_11 / floorf(maxAniso);
} else {
tmp = t_12 / t_11;
}
float tmp_2;
if (log2f(tmp) <= 100.0f) {
float tmp_3;
if ((t_13 / t_12) > floorf(maxAniso)) {
tmp_3 = t_14 / floorf(maxAniso);
} else {
tmp_3 = t_12 / t_14;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_6 / t_2) > floorf(maxAniso)) {
tmp_4 = t_7 / floorf(maxAniso);
} else {
tmp_4 = t_2 / t_7;
}
tmp_2 = log2f(tmp_4);
}
return tmp_2;
}
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(h) * dX_46_v) t_2 = abs(Float32(Float32(floor(h) * floor(w)) * Float32(Float32(dY_46_u * dX_46_v) - Float32(dY_46_v * dX_46_u)))) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(w) * floor(w)) t_6 = fmax(fma(Float32(t_5 * 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, exp(fma(Float32(2.0), log(Float32(-floor(w))), Float32(Float32(2.0) * log(dY_46_u)))))) t_7 = sqrt(t_6) t_8 = Float32(Float32(t_4 * t_4) + Float32(t_1 * t_1)) t_9 = Float32(floor(h) * dY_46_v) t_10 = fmax(t_8, Float32(Float32(t_3 * t_3) + Float32(t_9 * t_9))) t_11 = sqrt(t_10) t_12 = abs(Float32(Float32(t_4 * t_9) - Float32(t_1 * t_3))) t_13 = fmax(t_8, fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0))) t_14 = sqrt(t_13) tmp = Float32(0.0) if (Float32(t_10 / t_12) > floor(maxAniso)) tmp = Float32(t_11 / floor(maxAniso)); else tmp = Float32(t_12 / t_11); end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(100.0)) tmp_3 = Float32(0.0) if (Float32(t_13 / t_12) > floor(maxAniso)) tmp_3 = Float32(t_14 / floor(maxAniso)); else tmp_3 = Float32(t_12 / t_14); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_6 / t_2) > floor(maxAniso)) tmp_4 = Float32(t_7 / floor(maxAniso)); else tmp_4 = Float32(t_2 / t_7); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \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|\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_6 := \mathsf{max}\left(\mathsf{fma}\left(t\_5 \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, e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor w\right\rfloor \right), 2 \cdot \log dY.u\right)}\right)\right)\\
t_7 := \sqrt{t\_6}\\
t_8 := t\_4 \cdot t\_4 + t\_1 \cdot t\_1\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := \mathsf{max}\left(t\_8, t\_3 \cdot t\_3 + t\_9 \cdot t\_9\right)\\
t_11 := \sqrt{t\_10}\\
t_12 := \left|t\_4 \cdot t\_9 - t\_1 \cdot t\_3\right|\\
t_13 := \mathsf{max}\left(t\_8, \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\right)\\
t_14 := \sqrt{t\_13}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_10}{t\_12} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_11}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_12}{t\_11}\\
\end{array} \leq 100:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_13}{t\_12} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_14}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_12}{t\_14}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_6}{t\_2} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_7}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_7}\\
\end{array}\\
\end{array}
\end{array}
if (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) < 100Initial program 76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
if 100 < (log2.f32 (if (>.f32 (/.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u))))) (floor.f32 maxAniso)) (/.f32 (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v))))) (floor.f32 maxAniso)) (/.f32 (fabs.f32 (-.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 h) dY.v)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 w) dY.u)))) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))))) Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3264.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3264.2
Applied rewrites64.2%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3261.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.8
Applied rewrites61.8%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3265.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.8
Applied rewrites65.8%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
sum-logN/A
distribute-lft-inN/A
log-pow-revN/A
pow2N/A
lift-*.f32N/A
log-pow-revN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites53.6%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
sum-logN/A
distribute-lft-inN/A
log-pow-revN/A
pow2N/A
lift-*.f32N/A
log-pow-revN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites51.3%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
sum-logN/A
distribute-lft-inN/A
log-pow-revN/A
pow2N/A
lift-*.f32N/A
log-pow-revN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
log-prodN/A
*-commutativeN/A
log-prodN/A
Applied rewrites55.0%
(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) dX.u))
(t_2
(fmax
(+ (* t_1 t_1) (* t_0 t_0))
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(* (* dY.v dY.v) (* (floor h) (floor h))))))
(t_3 (sqrt t_2))
(t_4 (fabs (- (* t_1 (* (floor h) dY.v)) (* t_0 (* (floor w) dY.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) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = fmaxf(((t_1 * t_1) + (t_0 * t_0)), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))));
float t_3 = sqrtf(t_2);
float t_4 = fabsf(((t_1 * (floorf(h) * dY_46_v)) - (t_0 * (floorf(w) * dY_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) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = fmax(Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(t_1 * Float32(floor(h) * dY_46_v)) - Float32(t_0 * Float32(floor(w) * dY_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 dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \mathsf{max}\left(t\_1 \cdot t\_1 + t\_0 \cdot t\_0, \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|t\_1 \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right) - t\_0 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.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 76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
associate-*r*N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites76.6%
(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 76.6%
Applied rewrites76.6%
(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) dX.u))
(t_2
(fmax
(+ (* t_1 t_1) (* t_0 t_0))
(fma
dY.u
(* (* (floor w) (floor w)) dY.u)
(* (* (* (floor h) dY.v) (floor h)) dY.v))))
(t_3 (sqrt t_2))
(t_4 (fabs (* (* (- dX.v) dY.u) (* (floor h) (floor w))))))
(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) * dX_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = fmaxf(((t_1 * t_1) + (t_0 * t_0)), fmaf(dY_46_u, ((floorf(w) * floorf(w)) * dY_46_u), (((floorf(h) * dY_46_v) * floorf(h)) * dY_46_v)));
float t_3 = sqrtf(t_2);
float t_4 = fabsf(((-dX_46_v * dY_46_u) * (floorf(h) * floorf(w))));
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) * dX_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = fmax(Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)), fma(dY_46_u, Float32(Float32(floor(w) * floor(w)) * dY_46_u), Float32(Float32(Float32(floor(h) * dY_46_v) * floor(h)) * dY_46_v))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(Float32(-dX_46_v) * dY_46_u) * Float32(floor(h) * floor(w)))) 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 dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \mathsf{max}\left(t\_1 \cdot t\_1 + t\_0 \cdot t\_0, \mathsf{fma}\left(dY.u, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left(\left(-dX.v\right) \cdot dY.u\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \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 76.6%
Taylor expanded in dX.u around 0
mul-1-negN/A
distribute-lft-neg-outN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3275.6
Applied rewrites75.6%
Taylor expanded in dX.u around 0
mul-1-negN/A
distribute-lft-neg-outN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3275.6
Applied rewrites75.6%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
unswap-sqrN/A
sqr-neg-revN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-*.f32N/A
lower-neg.f3275.6
Applied rewrites75.6%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
unswap-sqrN/A
sqr-neg-revN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-*.f32N/A
lower-neg.f3275.6
Applied rewrites75.6%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
unswap-sqrN/A
sqr-neg-revN/A
lift-*.f32N/A
associate-*l*N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-*.f32N/A
lower-neg.f3275.6
Applied rewrites75.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-neg.f32N/A
lift-neg.f32N/A
sqr-neg-revN/A
associate-*l*N/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f3275.6
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-*.f32N/A
Applied rewrites75.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-neg.f32N/A
lift-neg.f32N/A
sqr-neg-revN/A
associate-*l*N/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f3275.6
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-*.f32N/A
Applied rewrites75.6%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-neg.f32N/A
lift-neg.f32N/A
sqr-neg-revN/A
associate-*l*N/A
lower-fma.f32N/A
*-commutativeN/A
lower-*.f3275.6
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-*.f32N/A
Applied rewrites75.6%
(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 76.6%
Applied rewrites76.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.6
Applied rewrites75.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3275.6
Applied rewrites75.6%
herbie shell --seed 2025137
(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)))))))))