
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
Herbie found 8 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor w) dX.u) t_4) (/ (* (floor w) dY.u) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(w) * dX_46_u) / t_4;
} else {
tmp = (floorf(w) * dY_46_u) / t_4;
}
return 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 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(w) * dX_46_u) / t_4); else tmp = Float32(Float32(floor(w) * dY_46_u) / t_4); end return 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{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_4}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
(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 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(/ (* (floor w) dX.u) (sqrt (fmax t_3 t_2)))
(*
(/
dY.u
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* (floor h) dX.v) dX.v) (floor h)))
(fma (* (* (floor h) dY.v) dY.v) (floor h) (* (* dY.u dY.u) t_1)))))
(floor w)))))
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 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_3, t_2));
} else {
tmp = (dY_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h))), fmaf(((floorf(h) * dY_46_v) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_1))))) * floorf(w);
}
return 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 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_3, t_2))); else tmp = Float32(Float32(dY_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))), fma(Float32(Float32(floor(h) * dY_46_v) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_1))))) * floor(w)); end return 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{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* dY.v dY.v) t_0))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dY.u) dY.u t_1))
(t_4 (* t_2 dX.u))
(t_5 (fma t_4 dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_5 t_3)
(/
(* (floor w) dX.u)
(sqrt (fmax (fma t_4 dX.u (exp (* (log (* (floor h) dX.v)) 2.0))) t_1)))
(/ (* (floor w) dY.u) (sqrt (fmax t_5 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 = (dY_46_v * dY_46_v) * t_0;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dY_46_u), dY_46_u, t_1);
float t_4 = t_2 * dX_46_u;
float t_5 = fmaf(t_4, dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_5 >= t_3) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(fmaf(t_4, dX_46_u, expf((logf((floorf(h) * dX_46_v)) * 2.0f))), t_1));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_5, t_3));
}
return 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(Float32(dY_46_v * dY_46_v) * t_0) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dY_46_u), dY_46_u, t_1) t_4 = Float32(t_2 * dX_46_u) t_5 = fma(t_4, dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_5 >= t_3) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(fma(t_4, dX_46_u, exp(Float32(log(Float32(floor(h) * dX_46_v)) * Float32(2.0)))), t_1))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_5, t_3))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(dY.v \cdot dY.v\right) \cdot t\_0\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, t\_1\right)\\
t_4 := t\_2 \cdot dX.u\\
t_5 := \mathsf{fma}\left(t\_4, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4, dX.u, e^{\log \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot 2}\right), t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_5, t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
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
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3259.9
Applied rewrites59.9%
Taylor expanded in dY.u around 0
Applied rewrites59.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor h) (floor h)))
(t_4 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_5 (* (floor h) dY.v))
(t_6 (* (floor h) dX.v))
(t_7
(/
1.0
(sqrt
(fmax (+ (* t_2 t_2) (* t_6 t_6)) (+ (* t_1 t_1) (* t_5 t_5))))))
(t_8 (sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_3 (* dX.v dX.v))) t_4))))
(if (<= dX.v 0.029999999329447746)
(if (>=
(* (* dX.u dX.u) t_0)
(fma (* t_5 dY.v) (floor h) (* (* dY.u dY.u) t_0)))
(* t_7 t_2)
(* t_7 t_1))
(if (>= (* (* t_6 dX.v) (floor h)) t_4) (/ t_2 t_8) (/ t_1 t_8)))))
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(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(h) * floorf(h);
float t_4 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_3));
float t_5 = floorf(h) * dY_46_v;
float t_6 = floorf(h) * dX_46_v;
float t_7 = 1.0f / sqrtf(fmaxf(((t_2 * t_2) + (t_6 * t_6)), ((t_1 * t_1) + (t_5 * t_5))));
float t_8 = sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v))), t_4));
float tmp_1;
if (dX_46_v <= 0.029999999329447746f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * t_0) >= fmaf((t_5 * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_0))) {
tmp_2 = t_7 * t_2;
} else {
tmp_2 = t_7 * t_1;
}
tmp_1 = tmp_2;
} else if (((t_6 * dX_46_v) * floorf(h)) >= t_4) {
tmp_1 = t_2 / t_8;
} else {
tmp_1 = t_1 / t_8;
}
return tmp_1;
}
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(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(h) * floor(h)) t_4 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_3)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)), Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5))))) t_8 = sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))), t_4)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(0.029999999329447746)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= fma(Float32(t_5 * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_0))) tmp_2 = Float32(t_7 * t_2); else tmp_2 = Float32(t_7 * t_1); end tmp_1 = tmp_2; elseif (Float32(Float32(t_6 * dX_46_v) * floor(h)) >= t_4) tmp_1 = Float32(t_2 / t_8); else tmp_1 = Float32(t_1 / t_8); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_3\right)\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_6 \cdot t\_6, t\_1 \cdot t\_1 + t\_5 \cdot t\_5\right)}}\\
t_8 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}\\
\mathbf{if}\;dX.v \leq 0.029999999329447746:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq \mathsf{fma}\left(t\_5 \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_0\right):\\
\;\;\;\;t\_7 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;\left(t\_6 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_4:\\
\;\;\;\;\frac{t\_2}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_8}\\
\end{array}
\end{array}
if dX.v < 0.0299999993Initial program 77.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3269.5
Applied rewrites69.5%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f3269.5
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites69.5%
if 0.0299999993 < dX.v Initial program 72.9%
Applied rewrites73.0%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
pow2N/A
associate-*r*N/A
Applied rewrites69.4%
(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) dX.v))
(t_2 (* t_1 (floor h)))
(t_3 (* (* (floor h) dY.v) dY.v))
(t_4 (* dY.u (floor w)))
(t_5 (* (floor h) (floor h)))
(t_6 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_5)))
(t_7 (sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_5 (* dX.v dX.v))) t_6))))
(if (<= dX.v 0.029999999329447746)
(if (>= (* (* dX.u dX.u) t_0) (fma t_3 (floor h) (* (* dY.u dY.u) t_0)))
(*
(/
(floor w)
(sqrt
(fmax
(fma t_1 (floor h) (* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* dY.v dY.v) t_5 (* (* t_4 dY.u) (floor w))))))
dX.u)
(*
(/
(floor w)
(sqrt
(fmax
(fma (* (* dX.u dX.u) (floor w)) (floor w) t_2)
(fma t_3 (floor h) (* (* t_4 (floor w)) dY.u)))))
dY.u))
(if (>= t_2 t_6) (/ (* (floor w) dX.u) t_7) (/ (* (floor w) dY.u) 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(w) * floorf(w);
float t_1 = (floorf(h) * dX_46_v) * dX_46_v;
float t_2 = t_1 * floorf(h);
float t_3 = (floorf(h) * dY_46_v) * dY_46_v;
float t_4 = dY_46_u * floorf(w);
float t_5 = floorf(h) * floorf(h);
float t_6 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_5));
float t_7 = sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_5 * (dX_46_v * dX_46_v))), t_6));
float tmp_1;
if (dX_46_v <= 0.029999999329447746f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * t_0) >= fmaf(t_3, floorf(h), ((dY_46_u * dY_46_u) * t_0))) {
tmp_2 = (floorf(w) / sqrtf(fmaxf(fmaf(t_1, floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((dY_46_v * dY_46_v), t_5, ((t_4 * dY_46_u) * floorf(w)))))) * dX_46_u;
} else {
tmp_2 = (floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), t_2), fmaf(t_3, floorf(h), ((t_4 * floorf(w)) * dY_46_u))))) * dY_46_u;
}
tmp_1 = tmp_2;
} else if (t_2 >= t_6) {
tmp_1 = (floorf(w) * dX_46_u) / t_7;
} else {
tmp_1 = (floorf(w) * dY_46_u) / t_7;
}
return tmp_1;
}
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(Float32(floor(h) * dX_46_v) * dX_46_v) t_2 = Float32(t_1 * floor(h)) t_3 = Float32(Float32(floor(h) * dY_46_v) * dY_46_v) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(floor(h) * floor(h)) t_6 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_5)) t_7 = sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_5 * Float32(dX_46_v * dX_46_v))), t_6)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(0.029999999329447746)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= fma(t_3, floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_0))) tmp_2 = Float32(Float32(floor(w) / sqrt(fmax(fma(t_1, floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(dY_46_v * dY_46_v), t_5, Float32(Float32(t_4 * dY_46_u) * floor(w)))))) * dX_46_u); else tmp_2 = Float32(Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), t_2), fma(t_3, floor(h), Float32(Float32(t_4 * floor(w)) * dY_46_u))))) * dY_46_u); end tmp_1 = tmp_2; elseif (t_2 >= t_6) tmp_1 = Float32(Float32(floor(w) * dX_46_u) / t_7); else tmp_1 = Float32(Float32(floor(w) * dY_46_u) / t_7); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\\
t_2 := t\_1 \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_6 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_5\right)\\
t_7 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_5 \cdot \left(dX.v \cdot dX.v\right)\right), t\_6\right)}\\
\mathbf{if}\;dX.v \leq 0.029999999329447746:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq \mathsf{fma}\left(t\_3, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_0\right):\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_5, \left(t\_4 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_2\right), \mathsf{fma}\left(t\_3, \left\lfloor h\right\rfloor , \left(t\_4 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)\right)}} \cdot dY.u\\
\end{array}\\
\mathbf{elif}\;t\_2 \geq t\_6:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_7}\\
\end{array}
\end{array}
if dX.v < 0.0299999993Initial program 77.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3269.5
Applied rewrites69.5%
Applied rewrites69.3%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lower-*.f3269.3
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3269.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3269.3
Applied rewrites69.3%
Applied rewrites69.4%
if 0.0299999993 < dX.v Initial program 72.9%
Applied rewrites73.0%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
pow2N/A
associate-*r*N/A
Applied rewrites69.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (floor h) dY.v) dY.v))
(t_1 (* dY.u (floor w)))
(t_2 (* (floor w) (floor w)))
(t_3 (* (* (floor h) dX.v) dX.v)))
(if (>= (* (* dX.u dX.u) t_2) (fma t_0 (floor h) (* (* dY.u dY.u) t_2)))
(*
(/
(floor w)
(sqrt
(fmax
(fma t_3 (floor h) (* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* dY.v dY.v)
(* (floor h) (floor h))
(* (* t_1 dY.u) (floor w))))))
dX.u)
(*
(/
(floor w)
(sqrt
(fmax
(fma (* (* dX.u dX.u) (floor w)) (floor w) (* t_3 (floor h)))
(fma t_0 (floor h) (* (* t_1 (floor w)) dY.u)))))
dY.u))))
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) * dY_46_v) * dY_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = floorf(w) * floorf(w);
float t_3 = (floorf(h) * dX_46_v) * dX_46_v;
float tmp;
if (((dX_46_u * dX_46_u) * t_2) >= fmaf(t_0, floorf(h), ((dY_46_u * dY_46_u) * t_2))) {
tmp = (floorf(w) / sqrtf(fmaxf(fmaf(t_3, floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((dY_46_v * dY_46_v), (floorf(h) * floorf(h)), ((t_1 * dY_46_u) * floorf(w)))))) * dX_46_u;
} else {
tmp = (floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (t_3 * floorf(h))), fmaf(t_0, floorf(h), ((t_1 * floorf(w)) * dY_46_u))))) * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(floor(h) * dY_46_v) * dY_46_v) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(Float32(floor(h) * dX_46_v) * dX_46_v) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= fma(t_0, floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_2))) tmp = Float32(Float32(floor(w) / sqrt(fmax(fma(t_3, floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(dY_46_v * dY_46_v), Float32(floor(h) * floor(h)), Float32(Float32(t_1 * dY_46_u) * floor(w)))))) * dX_46_u); else tmp = Float32(Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(t_3 * floor(h))), fma(t_0, floor(h), Float32(Float32(t_1 * floor(w)) * dY_46_u))))) * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq \mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_2\right):\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(dY.v \cdot dY.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , \left(t\_1 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_3 \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , \left(t\_1 \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.3%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.7
Applied rewrites64.7%
Applied rewrites64.6%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lower-*.f3264.6
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3264.6
lift-*.f32N/A
*-commutativeN/A
lower-*.f3264.6
Applied rewrites64.6%
Applied rewrites64.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (floor h) dY.v) dY.v))
(t_1 (* (floor w) (floor w)))
(t_2 (fma t_0 (floor h) (* (* dY.u dY.u) t_1)))
(t_3
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* (floor h) dX.v) dX.v) (floor h)))))
(if (>= (* (* dX.u dX.u) t_1) t_2)
(* (/ (floor w) (sqrt (fmax t_3 t_2))) dX.u)
(*
(/
(floor w)
(sqrt
(fmax
t_3
(fma t_0 (floor h) (* (* (* dY.u (floor w)) (floor w)) dY.u)))))
dY.u))))
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) * dY_46_v) * dY_46_v;
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf(t_0, floorf(h), ((dY_46_u * dY_46_u) * t_1));
float t_3 = fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h)));
float tmp;
if (((dX_46_u * dX_46_u) * t_1) >= t_2) {
tmp = (floorf(w) / sqrtf(fmaxf(t_3, t_2))) * dX_46_u;
} else {
tmp = (floorf(w) / sqrtf(fmaxf(t_3, fmaf(t_0, floorf(h), (((dY_46_u * floorf(w)) * floorf(w)) * dY_46_u))))) * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(floor(h) * dY_46_v) * dY_46_v) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(t_0, floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_1)) t_3 = fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_1) >= t_2) tmp = Float32(Float32(floor(w) / sqrt(fmax(t_3, t_2))) * dX_46_u); else tmp = Float32(Float32(floor(w) / sqrt(fmax(t_3, fma(t_0, floor(h), Float32(Float32(Float32(dY_46_u * floor(w)) * floor(w)) * dY_46_u))))) * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\\
t_3 := \mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_1 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right)\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.3%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.7
Applied rewrites64.7%
Applied rewrites64.6%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lower-*.f3264.6
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3264.6
lift-*.f32N/A
*-commutativeN/A
lower-*.f3264.6
Applied rewrites64.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (fma (* (* (floor h) dY.v) dY.v) (floor h) (* (* dY.u dY.u) t_0)))
(t_2
(/
(floor w)
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* (floor h) dX.v) dX.v) (floor h)))
t_1)))))
(if (>= (* (* dX.u dX.u) t_0) t_1) (* t_2 dX.u) (* t_2 dY.u))))
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 = fmaf(((floorf(h) * dY_46_v) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_0));
float t_2 = floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h))), t_1));
float tmp;
if (((dX_46_u * dX_46_u) * t_0) >= t_1) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
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 = fma(Float32(Float32(floor(h) * dY_46_v) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_0)) t_2 = Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))), t_1))) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= t_1) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_1\right)}}\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq t\_1:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.3%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.7
Applied rewrites64.7%
Applied rewrites64.6%
herbie shell --seed 2025110
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, u)"
:precision binary32
:pre (and (and (and (and (and (and (and (<= 1.0 w) (<= w 16384.0)) (and (<= 1.0 h) (<= h 16384.0))) (and (<= 1e-20 (fabs dX.u)) (<= (fabs dX.u) 1e+20))) (and (<= 1e-20 (fabs dX.v)) (<= (fabs dX.v) 1e+20))) (and (<= 1e-20 (fabs dY.u)) (<= (fabs dY.u) 1e+20))) (and (<= 1e-20 (fabs dY.v)) (<= (fabs dY.v) 1e+20))) (== maxAniso 16.0))
(if (>= (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor w) dX.u)) (* (/ 1.0 (sqrt (fmax (+ (* (* (floor w) dX.u) (* (floor w) dX.u)) (* (* (floor h) dX.v) (* (floor h) dX.v))) (+ (* (* (floor w) dY.u) (* (floor w) dY.u)) (* (* (floor h) dY.v) (* (floor h) dY.v)))))) (* (floor w) dY.u))))