
(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 7 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 w) (floor w)))
(t_1
(fabs (* (* (floor h) (floor w)) (- (* dY.u dX.v) (* dY.v dX.u)))))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* (floor w) dX.u))
(t_5 (* t_0 dX.u))
(t_6 (* (floor h) (floor h)))
(t_7 (* (floor h) dX.v))
(t_8 (fabs (- (* t_4 t_3) (* t_7 t_2))))
(t_9 (fma (* t_6 dY.v) dY.v (* (* dY.u dY.u) t_0)))
(t_10
(fmax
(fma t_5 dX.u (exp (fma 2.0 (log (- (floor h))) (* 2.0 (log dX.v)))))
t_9))
(t_11 (fmax (+ (* t_4 t_4) (* t_7 t_7)) (+ (* t_2 t_2) (* t_3 t_3))))
(t_12 (sqrt t_11))
(t_13 (/ t_12 (floor maxAniso)))
(t_14 (/ t_8 t_12))
(t_15 (sqrt t_10)))
(if (<= (log2 (if (> (/ t_11 t_8) (floor maxAniso)) t_13 t_14)) 60.0)
(log2
(if (>
(/ (fmax (fma t_5 dX.u (* t_6 (* dX.v dX.v))) t_9) t_1)
(floor maxAniso))
t_13
t_14))
(log2
(if (> (/ t_10 t_1) (floor maxAniso))
(/ t_15 (floor maxAniso))
(/ t_1 t_15))))))
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(w) * floorf(w);
float t_1 = fabsf(((floorf(h) * floorf(w)) * ((dY_46_u * dX_46_v) - (dY_46_v * dX_46_u))));
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_0 * dX_46_u;
float t_6 = floorf(h) * floorf(h);
float t_7 = floorf(h) * dX_46_v;
float t_8 = fabsf(((t_4 * t_3) - (t_7 * t_2)));
float t_9 = fmaf((t_6 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_0));
float t_10 = fmaxf(fmaf(t_5, dX_46_u, expf(fmaf(2.0f, logf(-floorf(h)), (2.0f * logf(dX_46_v))))), t_9);
float t_11 = fmaxf(((t_4 * t_4) + (t_7 * t_7)), ((t_2 * t_2) + (t_3 * t_3)));
float t_12 = sqrtf(t_11);
float t_13 = t_12 / floorf(maxAniso);
float t_14 = t_8 / t_12;
float t_15 = sqrtf(t_10);
float tmp;
if ((t_11 / t_8) > floorf(maxAniso)) {
tmp = t_13;
} else {
tmp = t_14;
}
float tmp_2;
if (log2f(tmp) <= 60.0f) {
float tmp_3;
if ((fmaxf(fmaf(t_5, dX_46_u, (t_6 * (dX_46_v * dX_46_v))), t_9) / t_1) > floorf(maxAniso)) {
tmp_3 = t_13;
} else {
tmp_3 = t_14;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_10 / t_1) > floorf(maxAniso)) {
tmp_4 = t_15 / floorf(maxAniso);
} else {
tmp_4 = t_1 / t_15;
}
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(w) * floor(w)) t_1 = abs(Float32(Float32(floor(h) * floor(w)) * Float32(Float32(dY_46_u * dX_46_v) - Float32(dY_46_v * dX_46_u)))) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_0 * dX_46_u) t_6 = Float32(floor(h) * floor(h)) t_7 = Float32(floor(h) * dX_46_v) t_8 = abs(Float32(Float32(t_4 * t_3) - Float32(t_7 * t_2))) t_9 = fma(Float32(t_6 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0)) t_10 = fmax(fma(t_5, dX_46_u, exp(fma(Float32(2.0), log(Float32(-floor(h))), Float32(Float32(2.0) * log(dX_46_v))))), t_9) t_11 = fmax(Float32(Float32(t_4 * t_4) + Float32(t_7 * t_7)), Float32(Float32(t_2 * t_2) + Float32(t_3 * t_3))) t_12 = sqrt(t_11) t_13 = Float32(t_12 / floor(maxAniso)) t_14 = Float32(t_8 / t_12) t_15 = sqrt(t_10) tmp = Float32(0.0) if (Float32(t_11 / t_8) > floor(maxAniso)) tmp = t_13; else tmp = t_14; end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(60.0)) tmp_3 = Float32(0.0) if (Float32(fmax(fma(t_5, dX_46_u, Float32(t_6 * Float32(dX_46_v * dX_46_v))), t_9) / t_1) > floor(maxAniso)) tmp_3 = t_13; else tmp_3 = t_14; end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_10 / t_1) > floor(maxAniso)) tmp_4 = Float32(t_15 / floor(maxAniso)); else tmp_4 = Float32(t_1 / t_15); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \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_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_0 \cdot dX.u\\
t_6 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_7 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_8 := \left|t\_4 \cdot t\_3 - t\_7 \cdot t\_2\right|\\
t_9 := \mathsf{fma}\left(t\_6 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\\
t_10 := \mathsf{max}\left(\mathsf{fma}\left(t\_5, dX.u, e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor h\right\rfloor \right), 2 \cdot \log dX.v\right)}\right), t\_9\right)\\
t_11 := \mathsf{max}\left(t\_4 \cdot t\_4 + t\_7 \cdot t\_7, t\_2 \cdot t\_2 + t\_3 \cdot t\_3\right)\\
t_12 := \sqrt{t\_11}\\
t_13 := \frac{t\_12}{\left\lfloor maxAniso\right\rfloor }\\
t_14 := \frac{t\_8}{t\_12}\\
t_15 := \sqrt{t\_10}\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_11}{t\_8} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_13\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array} \leq 60:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(\mathsf{fma}\left(t\_5, dX.u, t\_6 \cdot \left(dX.v \cdot dX.v\right)\right), t\_9\right)}{t\_1} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_13\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_10}{t\_1} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_15}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_15}\\
\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)))))))) < 60Initial program 75.9%
Applied rewrites75.9%
if 60 < (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 75.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3263.7
Applied rewrites63.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3261.4
Applied rewrites61.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
exp-to-powN/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3265.1
Applied rewrites65.1%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
log-prodN/A
+-commutativeN/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 rewrites52.8%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
log-prodN/A
+-commutativeN/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 rewrites50.6%
lift-*.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
log-prodN/A
+-commutativeN/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 rewrites54.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) (floor h)))
(t_5
(fmax
(fma
(* dX.v dX.v)
t_4
(exp (fma 2.0 (log (- (floor w))) (* 2.0 (log dX.u)))))
(fma (* dY.u dY.u) t_0 (* (* dY.v dY.v) t_4))))
(t_6 (sqrt t_5))
(t_7 (* (floor w) dY.u))
(t_8 (fmax (+ (* t_3 t_3) (* t_1 t_1)) (+ (* t_7 t_7) (* t_2 t_2))))
(t_9 (sqrt t_8))
(t_10 (/ t_9 (floor maxAniso)))
(t_11 (fabs (- (* t_3 t_2) (* t_1 t_7))))
(t_12 (/ t_11 t_9))
(t_13 (* (floor h) (floor w)))
(t_14 (fabs (* t_13 (* (- dX.v) dY.u)))))
(if (<= (log2 (if (> (/ t_8 t_11) (floor maxAniso)) t_10 t_12)) 60.0)
(log2
(if (>
(/
(fmax
(fma (* t_0 dX.u) dX.u (* t_4 (* dX.v dX.v)))
(fma (* t_4 dY.v) dY.v (* (* dY.u dY.u) t_0)))
(fabs (* t_13 (- (* dY.u dX.v) (* dY.v dX.u)))))
(floor maxAniso))
t_10
t_12))
(log2
(if (> (/ t_5 t_14) (floor maxAniso))
(/ t_6 (floor maxAniso))
(/ t_14 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 = floorf(w) * floorf(w);
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * floorf(h);
float t_5 = fmaxf(fmaf((dX_46_v * dX_46_v), t_4, expf(fmaf(2.0f, logf(-floorf(w)), (2.0f * logf(dX_46_u))))), fmaf((dY_46_u * dY_46_u), t_0, ((dY_46_v * dY_46_v) * t_4)));
float t_6 = sqrtf(t_5);
float t_7 = floorf(w) * dY_46_u;
float t_8 = fmaxf(((t_3 * t_3) + (t_1 * t_1)), ((t_7 * t_7) + (t_2 * t_2)));
float t_9 = sqrtf(t_8);
float t_10 = t_9 / floorf(maxAniso);
float t_11 = fabsf(((t_3 * t_2) - (t_1 * t_7)));
float t_12 = t_11 / t_9;
float t_13 = floorf(h) * floorf(w);
float t_14 = fabsf((t_13 * (-dX_46_v * dY_46_u)));
float tmp;
if ((t_8 / t_11) > floorf(maxAniso)) {
tmp = t_10;
} else {
tmp = t_12;
}
float tmp_2;
if (log2f(tmp) <= 60.0f) {
float tmp_3;
if ((fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v))), fmaf((t_4 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_0))) / fabsf((t_13 * ((dY_46_u * dX_46_v) - (dY_46_v * dX_46_u))))) > floorf(maxAniso)) {
tmp_3 = t_10;
} else {
tmp_3 = t_12;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_5 / t_14) > floorf(maxAniso)) {
tmp_4 = t_6 / floorf(maxAniso);
} else {
tmp_4 = t_14 / t_6;
}
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(w) * floor(w)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * floor(h)) t_5 = fmax(fma(Float32(dX_46_v * dX_46_v), t_4, exp(fma(Float32(2.0), log(Float32(-floor(w))), Float32(Float32(2.0) * log(dX_46_u))))), fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(dY_46_v * dY_46_v) * t_4))) t_6 = sqrt(t_5) t_7 = Float32(floor(w) * dY_46_u) t_8 = fmax(Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)), Float32(Float32(t_7 * t_7) + Float32(t_2 * t_2))) t_9 = sqrt(t_8) t_10 = Float32(t_9 / floor(maxAniso)) t_11 = abs(Float32(Float32(t_3 * t_2) - Float32(t_1 * t_7))) t_12 = Float32(t_11 / t_9) t_13 = Float32(floor(h) * floor(w)) t_14 = abs(Float32(t_13 * Float32(Float32(-dX_46_v) * dY_46_u))) tmp = Float32(0.0) if (Float32(t_8 / t_11) > floor(maxAniso)) tmp = t_10; else tmp = t_12; end tmp_2 = Float32(0.0) if (log2(tmp) <= Float32(60.0)) tmp_3 = Float32(0.0) if (Float32(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_4 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0))) / abs(Float32(t_13 * Float32(Float32(dY_46_u * dX_46_v) - Float32(dY_46_v * dX_46_u))))) > floor(maxAniso)) tmp_3 = t_10; else tmp_3 = t_12; end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_5 / t_14) > floor(maxAniso)) tmp_4 = Float32(t_6 / floor(maxAniso)); else tmp_4 = Float32(t_14 / t_6); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \mathsf{max}\left(\mathsf{fma}\left(dX.v \cdot dX.v, t\_4, e^{\mathsf{fma}\left(2, \log \left(-\left\lfloor w\right\rfloor \right), 2 \cdot \log dX.u\right)}\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(dY.v \cdot dY.v\right) \cdot t\_4\right)\right)\\
t_6 := \sqrt{t\_5}\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \mathsf{max}\left(t\_3 \cdot t\_3 + t\_1 \cdot t\_1, t\_7 \cdot t\_7 + t\_2 \cdot t\_2\right)\\
t_9 := \sqrt{t\_8}\\
t_10 := \frac{t\_9}{\left\lfloor maxAniso\right\rfloor }\\
t_11 := \left|t\_3 \cdot t\_2 - t\_1 \cdot t\_7\right|\\
t_12 := \frac{t\_11}{t\_9}\\
t_13 := \left\lfloor h\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_14 := \left|t\_13 \cdot \left(\left(-dX.v\right) \cdot dY.u\right)\right|\\
\mathbf{if}\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_8}{t\_11} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_12\\
\end{array} \leq 60:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_4 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}{\left|t\_13 \cdot \left(dY.u \cdot dX.v - dY.v \cdot dX.u\right)\right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_12\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_5}{t\_14} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_6}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_14}{t\_6}\\
\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)))))))) < 60Initial program 75.9%
Applied rewrites75.9%
if 60 < (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 75.9%
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.f3274.8
Applied rewrites74.8%
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.f3274.8
Applied rewrites74.8%
Applied rewrites74.8%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3266.6
Applied rewrites66.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3260.0
Applied rewrites60.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
unpow-prod-downN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3262.3
Applied rewrites62.3%
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
log-powN/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
log-prodN/A
Applied rewrites54.1%
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
log-powN/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
log-prodN/A
Applied rewrites48.8%
lift-log.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
log-powN/A
lift-*.f32N/A
lift-floor.f32N/A
unpow-prod-downN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
log-prodN/A
Applied rewrites51.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor h) (floor h)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5
(sqrt (fmax (+ (* t_3 t_3) (* t_4 t_4)) (+ (* t_0 t_0) (* t_1 t_1)))))
(t_6 (* (floor w) (floor w))))
(log2
(if (>
(/
(fmax
(fma (* t_6 dX.u) dX.u (* t_2 (* dX.v dX.v)))
(fma (* t_2 dY.v) dY.v (* (* dY.u dY.u) t_6)))
(fabs (* (* (floor h) (floor w)) (- (* dY.u dX.v) (* dY.v dX.u)))))
(floor maxAniso))
(/ t_5 (floor maxAniso))
(/ (fabs (- (* t_3 t_1) (* t_4 t_0))) 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(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(h) * floorf(h);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = sqrtf(fmaxf(((t_3 * t_3) + (t_4 * t_4)), ((t_0 * t_0) + (t_1 * t_1))));
float t_6 = floorf(w) * floorf(w);
float tmp;
if ((fmaxf(fmaf((t_6 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v))), fmaf((t_2 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_6))) / fabsf(((floorf(h) * floorf(w)) * ((dY_46_u * dX_46_v) - (dY_46_v * dX_46_u))))) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = fabsf(((t_3 * t_1) - (t_4 * t_0))) / 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(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = sqrt(fmax(Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4)), Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)))) t_6 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (Float32(fmax(fma(Float32(t_6 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_2 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_6))) / abs(Float32(Float32(floor(h) * floor(w)) * Float32(Float32(dY_46_u * dX_46_v) - Float32(dY_46_v * dX_46_u))))) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(abs(Float32(Float32(t_3 * t_1) - Float32(t_4 * t_0))) / t_5); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_4 \cdot t\_4, t\_0 \cdot t\_0 + t\_1 \cdot t\_1\right)}\\
t_6 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(\mathsf{fma}\left(t\_6 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_2 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_6\right)\right)}{\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|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|t\_3 \cdot t\_1 - t\_4 \cdot t\_0\right|}{t\_5}\\
\end{array}
\end{array}
\end{array}
Initial program 75.9%
Applied rewrites75.9%
(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 75.9%
Applied rewrites75.9%
(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 (* (* (* (- dX.v) dY.u) (floor h)) (floor w)))))
(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((((-dX_46_v * dY_46_u) * floorf(h)) * floorf(w)));
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(Float32(Float32(-dX_46_v) * dY_46_u) * floor(h)) * floor(w))) 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((((-dX_46_v * dY_46_u) * floor(h)) * floor(w))); 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|\left(\left(\left(-dX.v\right) \cdot dY.u\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \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}
Initial program 75.9%
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.f3274.8
Applied rewrites74.8%
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.f3274.8
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3274.8
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3274.8
Applied rewrites74.8%
(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(floor(h) * Float32(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\lfloor h\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dX.v\right)\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 75.9%
Applied rewrites75.9%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3274.8
Applied rewrites74.8%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3274.8
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lower-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f3274.8
Applied rewrites74.8%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lower-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f3274.8
Applied rewrites74.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor w) dY.u))
(t_2
(fmax
(fma (* (floor h) (* (floor h) dX.v)) dX.v (* t_0 t_0))
(fma (* dY.v dY.v) (* (floor h) (floor h)) (* t_1 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(w) * dX_46_u;
float t_1 = floorf(w) * dY_46_u;
float t_2 = fmaxf(fmaf((floorf(h) * (floorf(h) * dX_46_v)), dX_46_v, (t_0 * t_0)), fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), (t_1 * 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(w) * dX_46_u) t_1 = Float32(floor(w) * dY_46_u) t_2 = fmax(fma(Float32(floor(h) * Float32(floor(h) * dX_46_v)), dX_46_v, Float32(t_0 * t_0)), fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), Float32(t_1 * 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 w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right), dX.v, t\_0 \cdot t\_0\right), \mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , t\_1 \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 75.9%
Applied rewrites75.9%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3274.8
Applied rewrites74.8%
Taylor expanded in dX.u around 0
*-commutativeN/A
lift-*.f3274.8
Applied rewrites74.8%
Applied rewrites74.8%
herbie shell --seed 2025138
(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)))))))))