
(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 (* dY.v (floor h)))
(t_1 (* (floor h) dX.v))
(t_2 (* dY.u (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (* t_0 t_0))
(t_6 (* dX.u (floor w)))
(t_7 (* (floor w) dY.u))
(t_8 (fabs (* (* (- dX.v) dY.u) (* (floor h) (floor w)))))
(t_9 (* (floor h) dY.v))
(t_10 (fmax (+ t_4 (* t_1 t_1)) (+ (* t_7 t_7) (* t_9 t_9))))
(t_11 (sqrt t_10))
(t_12 (fabs (- (* t_3 t_9) (* t_1 t_7))))
(t_13
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_6 dX.u) (floor w))))
(t_14
(sqrt
(fmax t_13 (fma (* t_2 dY.u) (floor w) (* (* t_0 dY.v) (floor h))))))
(t_15
(fmax
t_13
(fma
(* (* (sqrt (* t_2 (- dY.u))) (sqrt (floor w))) dY.u)
(floor w)
t_5)))
(t_16 (sqrt t_15)))
(if (<=
(log2
(if (> (/ t_10 t_12) (floor maxAniso))
(/ t_11 (floor maxAniso))
(/ t_12 t_11)))
59.79999923706055)
(log2
(if (>
(/
(fmax
(fma t_1 t_1 t_4)
(fma (* (* (floor w) (floor w)) dY.u) dY.u t_5))
(fabs
(*
(fma (* dY.v (floor w)) dX.u (* (* (floor w) dX.v) dY.u))
(floor h))))
(floor maxAniso))
(/ t_14 (floor maxAniso))
(/ (fabs (* (floor h) (- (* t_2 dX.v) (* dY.v t_6)))) t_14)))
(log2
(if (> (/ t_15 t_8) (floor maxAniso))
(/ t_16 (floor maxAniso))
(/ t_8 t_16))))))
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_v * floorf(h);
float t_1 = floorf(h) * dX_46_v;
float t_2 = dY_46_u * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = t_3 * t_3;
float t_5 = t_0 * t_0;
float t_6 = dX_46_u * floorf(w);
float t_7 = floorf(w) * dY_46_u;
float t_8 = fabsf(((-dX_46_v * dY_46_u) * (floorf(h) * floorf(w))));
float t_9 = floorf(h) * dY_46_v;
float t_10 = fmaxf((t_4 + (t_1 * t_1)), ((t_7 * t_7) + (t_9 * t_9)));
float t_11 = sqrtf(t_10);
float t_12 = fabsf(((t_3 * t_9) - (t_1 * t_7)));
float t_13 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_6 * dX_46_u) * floorf(w)));
float t_14 = sqrtf(fmaxf(t_13, fmaf((t_2 * dY_46_u), floorf(w), ((t_0 * dY_46_v) * floorf(h)))));
float t_15 = fmaxf(t_13, fmaf(((sqrtf((t_2 * -dY_46_u)) * sqrtf(floorf(w))) * dY_46_u), floorf(w), t_5));
float t_16 = sqrtf(t_15);
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) <= 59.79999923706055f) {
float tmp_3;
if ((fmaxf(fmaf(t_1, t_1, t_4), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_5)) / fabsf((fmaf((dY_46_v * floorf(w)), dX_46_u, ((floorf(w) * dX_46_v) * dY_46_u)) * floorf(h)))) > floorf(maxAniso)) {
tmp_3 = t_14 / floorf(maxAniso);
} else {
tmp_3 = fabsf((floorf(h) * ((t_2 * dX_46_v) - (dY_46_v * t_6)))) / t_14;
}
tmp_2 = log2f(tmp_3);
} else {
float tmp_4;
if ((t_15 / t_8) > floorf(maxAniso)) {
tmp_4 = t_16 / floorf(maxAniso);
} else {
tmp_4 = t_8 / t_16;
}
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(dY_46_v * floor(h)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(t_3 * t_3) t_5 = Float32(t_0 * t_0) t_6 = Float32(dX_46_u * floor(w)) t_7 = Float32(floor(w) * dY_46_u) t_8 = abs(Float32(Float32(Float32(-dX_46_v) * dY_46_u) * Float32(floor(h) * floor(w)))) t_9 = Float32(floor(h) * dY_46_v) t_10 = fmax(Float32(t_4 + Float32(t_1 * t_1)), Float32(Float32(t_7 * t_7) + Float32(t_9 * t_9))) t_11 = sqrt(t_10) t_12 = abs(Float32(Float32(t_3 * t_9) - Float32(t_1 * t_7))) t_13 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_6 * dX_46_u) * floor(w))) t_14 = sqrt(fmax(t_13, fma(Float32(t_2 * dY_46_u), floor(w), Float32(Float32(t_0 * dY_46_v) * floor(h))))) t_15 = fmax(t_13, fma(Float32(Float32(sqrt(Float32(t_2 * Float32(-dY_46_u))) * sqrt(floor(w))) * dY_46_u), floor(w), t_5)) t_16 = sqrt(t_15) 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(59.79999923706055)) tmp_3 = Float32(0.0) if (Float32(fmax(fma(t_1, t_1, t_4), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_5)) / abs(Float32(fma(Float32(dY_46_v * floor(w)), dX_46_u, Float32(Float32(floor(w) * dX_46_v) * dY_46_u)) * floor(h)))) > floor(maxAniso)) tmp_3 = Float32(t_14 / floor(maxAniso)); else tmp_3 = Float32(abs(Float32(floor(h) * Float32(Float32(t_2 * dX_46_v) - Float32(dY_46_v * t_6)))) / t_14); end tmp_2 = log2(tmp_3); else tmp_4 = Float32(0.0) if (Float32(t_15 / t_8) > floor(maxAniso)) tmp_4 = Float32(t_16 / floor(maxAniso)); else tmp_4 = Float32(t_8 / t_16); end tmp_2 = log2(tmp_4); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_0 \cdot t\_0\\
t_6 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \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|\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := \mathsf{max}\left(t\_4 + t\_1 \cdot t\_1, t\_7 \cdot t\_7 + t\_9 \cdot t\_9\right)\\
t_11 := \sqrt{t\_10}\\
t_12 := \left|t\_3 \cdot t\_9 - t\_1 \cdot t\_7\right|\\
t_13 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_6 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_14 := \sqrt{\mathsf{max}\left(t\_13, \mathsf{fma}\left(t\_2 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_0 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\\
t_15 := \mathsf{max}\left(t\_13, \mathsf{fma}\left(\left(\sqrt{t\_2 \cdot \left(-dY.u\right)} \cdot \sqrt{\left\lfloor w\right\rfloor }\right) \cdot dY.u, \left\lfloor w\right\rfloor , t\_5\right)\right)\\
t_16 := \sqrt{t\_15}\\
\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 59.79999923706055:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_4\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_5\right)\right)}{\left|\mathsf{fma}\left(dY.v \cdot \left\lfloor w\right\rfloor , dX.u, \left(\left\lfloor w\right\rfloor \cdot dX.v\right) \cdot dY.u\right) \cdot \left\lfloor h\right\rfloor \right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_14}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor h\right\rfloor \cdot \left(t\_2 \cdot dX.v - dY.v \cdot t\_6\right)\right|}{t\_14}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{t\_15}{t\_8} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_16}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{t\_16}\\
\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)))))))) < 59.7999992Initial program 76.4%
Applied rewrites76.3%
Applied rewrites76.3%
if 59.7999992 < (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 76.4%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.5
Applied rewrites75.5%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.5
Applied rewrites75.5%
Applied rewrites75.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3275.5
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
*-commutativeN/A
sqr-neg-revN/A
lift-neg.f32N/A
lift-neg.f32N/A
swap-sqrN/A
lift-*.f32N/A
Applied rewrites59.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.3
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
*-commutativeN/A
sqr-neg-revN/A
lift-neg.f32N/A
lift-neg.f32N/A
swap-sqrN/A
lift-*.f32N/A
Applied rewrites47.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3247.2
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
*-commutativeN/A
sqr-neg-revN/A
lift-neg.f32N/A
lift-neg.f32N/A
swap-sqrN/A
lift-*.f32N/A
Applied rewrites51.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) dX.v))
(t_4 (* dX.u (floor w)))
(t_5
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_4 dX.u) (floor w)))
(fma (* t_0 dY.u) (floor w) (* (* t_2 dY.v) (floor h)))))))
(log2
(if (>
(/
(fmax
(fma t_3 t_3 (* t_1 t_1))
(fma (* (* (floor w) (floor w)) dY.u) dY.u (* t_2 t_2)))
(fabs
(*
(fma (* dY.v (floor w)) dX.u (* (* (floor w) dX.v) dY.u))
(floor h))))
(floor maxAniso))
(/ t_5 (floor maxAniso))
(/ (fabs (* (floor h) (- (* t_0 dX.v) (* dY.v t_4)))) 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 = dY_46_u * floorf(w);
float t_1 = floorf(w) * dX_46_u;
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) * dX_46_v;
float t_4 = dX_46_u * floorf(w);
float t_5 = sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_4 * dX_46_u) * floorf(w))), fmaf((t_0 * dY_46_u), floorf(w), ((t_2 * dY_46_v) * floorf(h)))));
float tmp;
if ((fmaxf(fmaf(t_3, t_3, (t_1 * t_1)), fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, (t_2 * t_2))) / fabsf((fmaf((dY_46_v * floorf(w)), dX_46_u, ((floorf(w) * dX_46_v) * dY_46_u)) * floorf(h)))) > floorf(maxAniso)) {
tmp = t_5 / floorf(maxAniso);
} else {
tmp = fabsf((floorf(h) * ((t_0 * dX_46_v) - (dY_46_v * t_4)))) / 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(dY_46_u * floor(w)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(dX_46_u * floor(w)) t_5 = sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_4 * dX_46_u) * floor(w))), fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))))) tmp = Float32(0.0) if (Float32(fmax(fma(t_3, t_3, Float32(t_1 * t_1)), fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(t_2 * t_2))) / abs(Float32(fma(Float32(dY_46_v * floor(w)), dX_46_u, Float32(Float32(floor(w) * dX_46_v) * dY_46_u)) * floor(h)))) > floor(maxAniso)) tmp = Float32(t_5 / floor(maxAniso)); else tmp = Float32(abs(Float32(floor(h) * Float32(Float32(t_0 * dX_46_v) - Float32(dY_46_v * t_4)))) / t_5); end return log2(tmp) end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_5 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_4 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(t\_0 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}\\
\log_{2} \begin{array}{l}
\mathbf{if}\;\frac{\mathsf{max}\left(\mathsf{fma}\left(t\_3, t\_3, t\_1 \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_2 \cdot t\_2\right)\right)}{\left|\mathsf{fma}\left(dY.v \cdot \left\lfloor w\right\rfloor , dX.u, \left(\left\lfloor w\right\rfloor \cdot dX.v\right) \cdot dY.u\right) \cdot \left\lfloor h\right\rfloor \right|} > \left\lfloor maxAniso\right\rfloor :\\
\;\;\;\;\frac{t\_5}{\left\lfloor maxAniso\right\rfloor }\\
\mathbf{else}:\\
\;\;\;\;\frac{\left|\left\lfloor h\right\rfloor \cdot \left(t\_0 \cdot dX.v - dY.v \cdot t\_4\right)\right|}{t\_5}\\
\end{array}
\end{array}
\end{array}
Initial program 76.4%
Applied rewrites76.3%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2
(fmax
(fma t_1 t_1 (* t_0 t_0))
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(* (* dY.u dY.u) (* (floor w) (floor w))))))
(t_3 (sqrt t_2))
(t_4 (fabs (* (* t_1 dY.u) (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(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = fmaxf(fmaf(t_1, t_1, (t_0 * t_0)), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))));
float t_3 = sqrtf(t_2);
float t_4 = fabsf(((t_1 * dY_46_u) * 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(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = fmax(fma(t_1, t_1, Float32(t_0 * t_0)), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))))) t_3 = sqrt(t_2) t_4 = abs(Float32(Float32(t_1 * dY_46_u) * 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 w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_0 \cdot t\_0\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\right)\\
t_3 := \sqrt{t\_2}\\
t_4 := \left|\left(t\_1 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \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 76.4%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.5
Applied rewrites75.5%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3275.5
Applied rewrites75.5%
Applied rewrites75.5%
Applied rewrites75.5%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites75.5%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites75.5%
lift-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites75.5%
herbie shell --seed 2025162
(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)))))))))