
(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 15 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 (* t_1 dX.u))
(t_4 (fma t_3 dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_4 t_2)
(/
(* (floor w) dX.u)
(sqrt
(fmax (fma t_3 dX.u (* (* (* dX.v dX.v) (floor h)) (floor h))) t_2)))
(/ (* (floor w) dY.u) (sqrt (fmax t_4 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = 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 = t_1 * dX_46_u;
float t_4 = fmaf(t_3, dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_4 >= t_2) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(fmaf(t_3, dX_46_u, (((dX_46_v * dX_46_v) * floorf(h)) * floorf(h))), t_2));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_4, t_2));
}
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 = Float32(t_1 * dX_46_u) t_4 = fma(t_3, dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_4 >= t_2) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(fma(t_3, dX_46_u, Float32(Float32(Float32(dX_46_v * dX_46_v) * floor(h)) * floor(h))), t_2))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_4, t_2))); 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 := t\_1 \cdot dX.u\\
t_4 := \mathsf{fma}\left(t\_3, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.u, \left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right), t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_4, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f3276.2
Applied rewrites76.2%
(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 76.1%
Applied rewrites76.2%
(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 (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_0))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* (floor h) dY.v) dY.v) (floor h))))))
(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 * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_0)), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h)))))) * 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 * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)))))) * 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 \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.1%
(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)
(*
(/
dX.u
(sqrt
(fmax
(fma (* (* dX.u (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_0))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* (floor h) dY.v) dY.v) (floor h))))))
(floor w))
(/ (* (floor w) dY.u) (sqrt (fmax t_3 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = 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 = (dX_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_0)), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h)))))) * floorf(w);
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_3, t_2));
}
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(dX_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)))))) * floor(w)); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_3, t_2))); 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{dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.1%
(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)
(*
dX.u
(/
(floor w)
(sqrt
(fmax
(fma (* (* dX.u (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_0))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* (floor h) dY.v) dY.v) (floor h)))))))
(/ (* (floor w) dY.u) (sqrt (fmax t_3 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = 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 = dX_46_u * (floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_0)), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h))))));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_3, t_2));
}
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(dX_46_u * Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h))))))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_3, t_2))); 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:\\
\;\;\;\;dX.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_1
(fma
(* (* dX.v (floor h)) (floor h))
dX.v
(* (* dX.u dX.u) (* (floor w) (floor w)))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_1 = fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * (floorf(w) * floorf(w))));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
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 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * Float32(floor(w) * floor(w)))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) 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 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f3276.2
Applied rewrites76.2%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) (floor h)))
(t_4 (fma (* t_1 dX.u) dX.u (* t_3 (* dX.v dX.v))))
(t_5 (* (* dY.v dY.v) t_3))
(t_6 (fma (* t_1 dY.u) dY.u t_5))
(t_7 (* dY.u (floor w)))
(t_8 (sqrt (fmax t_4 t_6))))
(if (<= dX.v 2.4000000953674316)
(if (>= (* (* (* dX.u (floor w)) dX.u) (floor w)) t_6)
(/ (* (floor w) dX.u) t_8)
(/ t_2 (sqrt (fmax t_4 (fma (* t_7 (floor w)) dY.u t_5)))))
(if (>= (* (* t_0 dX.v) (floor h)) t_6)
(*
(/
dX.u
(sqrt
(fmax
(fma (* t_0 (floor h)) dX.v (* (* dX.u dX.u) t_1))
(fma (* t_7 dY.u) (floor w) t_5))))
(floor w))
(/ t_2 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 = dX_46_v * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * floorf(h);
float t_4 = fmaf((t_1 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v)));
float t_5 = (dY_46_v * dY_46_v) * t_3;
float t_6 = fmaf((t_1 * dY_46_u), dY_46_u, t_5);
float t_7 = dY_46_u * floorf(w);
float t_8 = sqrtf(fmaxf(t_4, t_6));
float tmp_1;
if (dX_46_v <= 2.4000000953674316f) {
float tmp_2;
if ((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)) >= t_6) {
tmp_2 = (floorf(w) * dX_46_u) / t_8;
} else {
tmp_2 = t_2 / sqrtf(fmaxf(t_4, fmaf((t_7 * floorf(w)), dY_46_u, t_5)));
}
tmp_1 = tmp_2;
} else if (((t_0 * dX_46_v) * floorf(h)) >= t_6) {
tmp_1 = (dX_46_u / sqrtf(fmaxf(fmaf((t_0 * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_1)), fmaf((t_7 * dY_46_u), floorf(w), t_5)))) * floorf(w);
} else {
tmp_1 = t_2 / 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(dX_46_v * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * floor(h)) t_4 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))) t_5 = Float32(Float32(dY_46_v * dY_46_v) * t_3) t_6 = fma(Float32(t_1 * dY_46_u), dY_46_u, t_5) t_7 = Float32(dY_46_u * floor(w)) t_8 = sqrt(fmax(t_4, t_6)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(2.4000000953674316)) tmp_2 = Float32(0.0) if (Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) >= t_6) tmp_2 = Float32(Float32(floor(w) * dX_46_u) / t_8); else tmp_2 = Float32(t_2 / sqrt(fmax(t_4, fma(Float32(t_7 * floor(w)), dY_46_u, t_5)))); end tmp_1 = tmp_2; elseif (Float32(Float32(t_0 * dX_46_v) * floor(h)) >= t_6) tmp_1 = Float32(Float32(dX_46_u / sqrt(fmax(fma(Float32(t_0 * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(t_7 * dY_46_u), floor(w), t_5)))) * floor(w)); else tmp_1 = Float32(t_2 / t_8); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_5 := \left(dY.v \cdot dY.v\right) \cdot t\_3\\
t_6 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, t\_5\right)\\
t_7 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_8 := \sqrt{\mathsf{max}\left(t\_4, t\_6\right)}\\
\mathbf{if}\;dX.v \leq 2.4000000953674316:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_6:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_7 \cdot \left\lfloor w\right\rfloor , dY.u, t\_5\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;\left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_6:\\
\;\;\;\;\frac{dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(t\_7 \cdot dY.u, \left\lfloor w\right\rfloor , t\_5\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_8}\\
\end{array}
\end{array}
if dX.v < 2.4000001Initial program 77.9%
Applied rewrites78.0%
Taylor expanded in dX.u around inf
Applied rewrites69.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3269.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.6
Applied rewrites69.6%
if 2.4000001 < dX.v Initial program 70.6%
Applied rewrites70.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f3270.7
Applied rewrites70.7%
Applied rewrites70.6%
Taylor expanded in dX.u around 0
Applied rewrites67.0%
(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) (floor h)))
(t_2 (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))))
(t_3 (* (* dY.v dY.v) t_1))
(t_4 (fma (* (* dX.v (floor h)) dX.v) (floor h) (* (* dX.u dX.u) t_0)))
(t_5 (fma (* t_0 dY.u) dY.u t_3))
(t_6 (* (* dY.u dY.u) t_0))
(t_7
(/
(floor w)
(sqrt (fmax t_4 (fma (* (* dY.v (floor h)) (floor h)) dY.v t_6))))))
(if (<= dX.v 2.0)
(if (>= (* (* (* dX.u (floor w)) dX.u) (floor w)) t_5)
(/ (* (floor w) dX.u) (sqrt (fmax t_2 t_5)))
(/
(* (floor w) dY.u)
(sqrt (fmax t_2 (fma (* (* dY.u (floor w)) (floor w)) dY.u t_3)))))
(if (>= t_4 t_6) (* t_7 dX.u) (* t_7 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 = floorf(h) * floorf(h);
float t_2 = fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v)));
float t_3 = (dY_46_v * dY_46_v) * t_1;
float t_4 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_0));
float t_5 = fmaf((t_0 * dY_46_u), dY_46_u, t_3);
float t_6 = (dY_46_u * dY_46_u) * t_0;
float t_7 = floorf(w) / sqrtf(fmaxf(t_4, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, t_6)));
float tmp_1;
if (dX_46_v <= 2.0f) {
float tmp_2;
if ((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)) >= t_5) {
tmp_2 = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_2, t_5));
} else {
tmp_2 = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_2, fmaf(((dY_46_u * floorf(w)) * floorf(w)), dY_46_u, t_3)));
}
tmp_1 = tmp_2;
} else if (t_4 >= t_6) {
tmp_1 = t_7 * dX_46_u;
} else {
tmp_1 = t_7 * dY_46_u;
}
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(h) * floor(h)) t_2 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_4 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_0)) t_5 = fma(Float32(t_0 * dY_46_u), dY_46_u, t_3) t_6 = Float32(Float32(dY_46_u * dY_46_u) * t_0) t_7 = Float32(floor(w) / sqrt(fmax(t_4, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, t_6)))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(2.0)) tmp_2 = Float32(0.0) if (Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) >= t_5) tmp_2 = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_2, t_5))); else tmp_2 = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_2, fma(Float32(Float32(dY_46_u * floor(w)) * floor(w)), dY_46_u, t_3)))); end tmp_1 = tmp_2; elseif (t_4 >= t_6) tmp_1 = Float32(t_7 * dX_46_u); else tmp_1 = Float32(t_7 * dY_46_u); 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_4 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_0\right)\\
t_5 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_3\right)\\
t_6 := \left(dY.u \cdot dY.u\right) \cdot t\_0\\
t_7 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_6\right)\right)}}\\
\mathbf{if}\;dX.v \leq 2:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_5:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_2, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor , dY.u, t\_3\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_4 \geq t\_6:\\
\;\;\;\;t\_7 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot dY.u\\
\end{array}
\end{array}
if dX.v < 2Initial program 77.9%
Applied rewrites78.0%
Taylor expanded in dX.u around inf
Applied rewrites69.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3269.5
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.5
Applied rewrites69.5%
if 2 < dX.v Initial program 70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3262.5
Applied rewrites62.5%
Applied rewrites62.7%
Applied rewrites62.5%
(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 (* dX.v (floor h)))
(t_3 (* (floor w) (floor w)))
(t_4 (* (* dX.u dX.u) t_3))
(t_5 (fma (* t_3 dY.u) dY.u t_1))
(t_6 (fma (* t_2 dX.v) (floor h) t_4))
(t_7 (* (* dY.u dY.u) t_3))
(t_8
(/
(floor w)
(sqrt (fmax t_6 (fma (* (* dY.v (floor h)) (floor h)) dY.v t_7))))))
(if (<= dX.v 1.399999976158142)
(if (>= (* (* (* dX.u (floor w)) dX.u) (floor w)) t_5)
(/
(* (floor w) dX.u)
(sqrt (fmax (fma (* t_3 dX.u) dX.u (* t_0 (* dX.v dX.v))) t_5)))
(*
(/
dY.u
(sqrt
(fmax
(fma (* t_2 (floor h)) dX.v t_4)
(fma (* (* dY.u (floor w)) dY.u) (floor w) t_1))))
(floor w)))
(if (>= t_6 t_7) (* t_8 dX.u) (* t_8 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) * floorf(h);
float t_1 = (dY_46_v * dY_46_v) * t_0;
float t_2 = dX_46_v * floorf(h);
float t_3 = floorf(w) * floorf(w);
float t_4 = (dX_46_u * dX_46_u) * t_3;
float t_5 = fmaf((t_3 * dY_46_u), dY_46_u, t_1);
float t_6 = fmaf((t_2 * dX_46_v), floorf(h), t_4);
float t_7 = (dY_46_u * dY_46_u) * t_3;
float t_8 = floorf(w) / sqrtf(fmaxf(t_6, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, t_7)));
float tmp_1;
if (dX_46_v <= 1.399999976158142f) {
float tmp_2;
if ((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)) >= t_5) {
tmp_2 = (floorf(w) * dX_46_u) / sqrtf(fmaxf(fmaf((t_3 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), t_5));
} else {
tmp_2 = (dY_46_u / sqrtf(fmaxf(fmaf((t_2 * floorf(h)), dX_46_v, t_4), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), t_1)))) * floorf(w);
}
tmp_1 = tmp_2;
} else if (t_6 >= t_7) {
tmp_1 = t_8 * dX_46_u;
} else {
tmp_1 = t_8 * dY_46_u;
}
return tmp_1;
}
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(dX_46_v * floor(h)) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(Float32(dX_46_u * dX_46_u) * t_3) t_5 = fma(Float32(t_3 * dY_46_u), dY_46_u, t_1) t_6 = fma(Float32(t_2 * dX_46_v), floor(h), t_4) t_7 = Float32(Float32(dY_46_u * dY_46_u) * t_3) t_8 = Float32(floor(w) / sqrt(fmax(t_6, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, t_7)))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(1.399999976158142)) tmp_2 = Float32(0.0) if (Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) >= t_5) tmp_2 = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_5))); else tmp_2 = Float32(Float32(dY_46_u / sqrt(fmax(fma(Float32(t_2 * floor(h)), dX_46_v, t_4), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), t_1)))) * floor(w)); end tmp_1 = tmp_2; elseif (t_6 >= t_7) tmp_1 = Float32(t_8 * dX_46_u); else tmp_1 = Float32(t_8 * dY_46_u); end return tmp_1 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 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \left(dX.u \cdot dX.u\right) \cdot t\_3\\
t_5 := \mathsf{fma}\left(t\_3 \cdot dY.u, dY.u, t\_1\right)\\
t_6 := \mathsf{fma}\left(t\_2 \cdot dX.v, \left\lfloor h\right\rfloor , t\_4\right)\\
t_7 := \left(dY.u \cdot dY.u\right) \cdot t\_3\\
t_8 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_7\right)\right)}}\\
\mathbf{if}\;dX.v \leq 1.399999976158142:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_5:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot \left\lfloor h\right\rfloor , dX.v, t\_4\right), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , t\_1\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_7:\\
\;\;\;\;t\_8 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot dY.u\\
\end{array}
\end{array}
if dX.v < 1.39999998Initial program 77.9%
Applied rewrites78.0%
Taylor expanded in dX.u around inf
Applied rewrites69.5%
Applied rewrites69.4%
if 1.39999998 < dX.v Initial program 70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3262.6
Applied rewrites62.6%
Applied rewrites62.8%
Applied rewrites62.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* dX.v (floor h)))
(t_2 (* (floor w) (floor w)))
(t_3 (* (* dY.u dY.u) t_2))
(t_4 (* (* dX.u dX.u) t_2))
(t_5 (fma (* t_1 dX.v) (floor h) t_4))
(t_6 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_7
(/
(floor w)
(sqrt (fmax t_5 (fma (* (* dY.v (floor h)) (floor h)) dY.v t_3))))))
(if (<= dX.v 2.0)
(if (>= (* (* (* dX.u (floor w)) dX.u) (floor w)) t_6)
(/
(* (floor w) dX.u)
(sqrt (fmax (fma (* t_1 (floor h)) dX.v t_4) t_6)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.v))) t_6))))
(if (>= t_5 t_3) (* t_7 dX.u) (* t_7 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) * floorf(h);
float t_1 = dX_46_v * floorf(h);
float t_2 = floorf(w) * floorf(w);
float t_3 = (dY_46_u * dY_46_u) * t_2;
float t_4 = (dX_46_u * dX_46_u) * t_2;
float t_5 = fmaf((t_1 * dX_46_v), floorf(h), t_4);
float t_6 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_7 = floorf(w) / sqrtf(fmaxf(t_5, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, t_3)));
float tmp_1;
if (dX_46_v <= 2.0f) {
float tmp_2;
if ((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)) >= t_6) {
tmp_2 = (floorf(w) * dX_46_u) / sqrtf(fmaxf(fmaf((t_1 * floorf(h)), dX_46_v, t_4), t_6));
} else {
tmp_2 = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), t_6));
}
tmp_1 = tmp_2;
} else if (t_5 >= t_3) {
tmp_1 = t_7 * dX_46_u;
} else {
tmp_1 = t_7 * dY_46_u;
}
return tmp_1;
}
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(dX_46_v * floor(h)) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(Float32(dY_46_u * dY_46_u) * t_2) t_4 = Float32(Float32(dX_46_u * dX_46_u) * t_2) t_5 = fma(Float32(t_1 * dX_46_v), floor(h), t_4) t_6 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_7 = Float32(floor(w) / sqrt(fmax(t_5, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, t_3)))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(2.0)) tmp_2 = Float32(0.0) if (Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) >= t_6) tmp_2 = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(fma(Float32(t_1 * floor(h)), dX_46_v, t_4), t_6))); else tmp_2 = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_6))); end tmp_1 = tmp_2; elseif (t_5 >= t_3) tmp_1 = Float32(t_7 * dX_46_u); else tmp_1 = Float32(t_7 * dY_46_u); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left(dY.u \cdot dY.u\right) \cdot t\_2\\
t_4 := \left(dX.u \cdot dX.u\right) \cdot t\_2\\
t_5 := \mathsf{fma}\left(t\_1 \cdot dX.v, \left\lfloor h\right\rfloor , t\_4\right)\\
t_6 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_7 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_3\right)\right)}}\\
\mathbf{if}\;dX.v \leq 2:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_6:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot \left\lfloor h\right\rfloor , dX.v, t\_4\right), t\_6\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_6\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_3:\\
\;\;\;\;t\_7 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot dY.u\\
\end{array}
\end{array}
if dX.v < 2Initial program 77.9%
Applied rewrites78.0%
Taylor expanded in dX.u around inf
Applied rewrites69.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3269.5
lower-fma.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites69.5%
if 2 < dX.v Initial program 70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3262.5
Applied rewrites62.5%
Applied rewrites62.7%
Applied rewrites62.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (* dY.u dY.u) t_0))
(t_2 (* (* dX.u dX.u) t_0))
(t_3
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_4 (* dX.v (floor h)))
(t_5 (fma (* t_4 dX.v) (floor h) t_2))
(t_6
(/
(floor w)
(sqrt (fmax t_5 (fma (* (* dY.v (floor h)) (floor h)) dY.v t_1)))))
(t_7 (/ (floor w) (sqrt (fmax (fma (* t_4 (floor h)) dX.v t_2) t_3)))))
(if (<= dX.v 0.4000000059604645)
(if (>= t_2 t_3) (* t_7 dX.u) (* t_7 dY.u))
(if (>= t_5 t_1) (* t_6 dX.u) (* t_6 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 = (dY_46_u * dY_46_u) * t_0;
float t_2 = (dX_46_u * dX_46_u) * t_0;
float t_3 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_4 = dX_46_v * floorf(h);
float t_5 = fmaf((t_4 * dX_46_v), floorf(h), t_2);
float t_6 = floorf(w) / sqrtf(fmaxf(t_5, fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, t_1)));
float t_7 = floorf(w) / sqrtf(fmaxf(fmaf((t_4 * floorf(h)), dX_46_v, t_2), t_3));
float tmp_1;
if (dX_46_v <= 0.4000000059604645f) {
float tmp_2;
if (t_2 >= t_3) {
tmp_2 = t_7 * dX_46_u;
} else {
tmp_2 = t_7 * dY_46_u;
}
tmp_1 = tmp_2;
} else if (t_5 >= t_1) {
tmp_1 = t_6 * dX_46_u;
} else {
tmp_1 = t_6 * dY_46_u;
}
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(dY_46_u * dY_46_u) * t_0) t_2 = Float32(Float32(dX_46_u * dX_46_u) * t_0) t_3 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_4 = Float32(dX_46_v * floor(h)) t_5 = fma(Float32(t_4 * dX_46_v), floor(h), t_2) t_6 = Float32(floor(w) / sqrt(fmax(t_5, fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, t_1)))) t_7 = Float32(floor(w) / sqrt(fmax(fma(Float32(t_4 * floor(h)), dX_46_v, t_2), t_3))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(0.4000000059604645)) tmp_2 = Float32(0.0) if (t_2 >= t_3) tmp_2 = Float32(t_7 * dX_46_u); else tmp_2 = Float32(t_7 * dY_46_u); end tmp_1 = tmp_2; elseif (t_5 >= t_1) tmp_1 = Float32(t_6 * dX_46_u); else tmp_1 = Float32(t_6 * dY_46_u); 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(dY.u \cdot dY.u\right) \cdot t\_0\\
t_2 := \left(dX.u \cdot dX.u\right) \cdot t\_0\\
t_3 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := \mathsf{fma}\left(t\_4 \cdot dX.v, \left\lfloor h\right\rfloor , t\_2\right)\\
t_6 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, t\_1\right)\right)}}\\
t_7 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot \left\lfloor h\right\rfloor , dX.v, t\_2\right), t\_3\right)}}\\
\mathbf{if}\;dX.v \leq 0.4000000059604645:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;t\_7 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot dY.u\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_1:\\
\;\;\;\;t\_6 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot dY.u\\
\end{array}
\end{array}
if dX.v < 0.400000006Initial program 77.9%
Applied rewrites77.9%
Taylor expanded in dX.u around inf
Applied rewrites69.4%
Applied rewrites69.3%
if 0.400000006 < dX.v Initial program 71.0%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3262.8
Applied rewrites62.8%
Applied rewrites62.9%
Applied rewrites62.8%
(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) (floor h)))
(t_3 (* (* dY.v dY.v) t_2))
(t_4 (fma (* t_0 dY.u) dY.u t_3))
(t_5 (* (floor w) dY.u))
(t_6 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_7 (* (floor h) dY.v))
(t_8 (+ (* t_5 t_5) (* t_7 t_7)))
(t_9 (* (floor w) dX.u))
(t_10 (+ (* t_9 t_9) (* t_1 t_1)))
(t_11 (/ 1.0 (sqrt (fmax t_10 t_8))))
(t_12 (if (>= t_10 t_8) (* t_11 t_9) (* t_11 t_5)))
(t_13 (sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_2 (* dX.v dX.v))) t_4)))
(t_14 (/ t_5 t_13))
(t_15
(if (>= t_6 (* (* (* dY.u (floor w)) dY.u) (floor w)))
(/ t_9 t_13)
t_14)))
(if (<= t_12 -1.0000000116860974e-7)
t_15
(if (<= t_12 0.004999999888241291)
(if (>= t_6 t_3)
(/ t_9 (sqrt (fmax (* (* (* dX.v (floor h)) dX.v) (floor h)) t_4)))
t_14)
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 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * floorf(h);
float t_3 = (dY_46_v * dY_46_v) * t_2;
float t_4 = fmaf((t_0 * dY_46_u), dY_46_u, t_3);
float t_5 = floorf(w) * dY_46_u;
float t_6 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_7 = floorf(h) * dY_46_v;
float t_8 = (t_5 * t_5) + (t_7 * t_7);
float t_9 = floorf(w) * dX_46_u;
float t_10 = (t_9 * t_9) + (t_1 * t_1);
float t_11 = 1.0f / sqrtf(fmaxf(t_10, t_8));
float tmp;
if (t_10 >= t_8) {
tmp = t_11 * t_9;
} else {
tmp = t_11 * t_5;
}
float t_12 = tmp;
float t_13 = sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v))), t_4));
float t_14 = t_5 / t_13;
float tmp_1;
if (t_6 >= (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))) {
tmp_1 = t_9 / t_13;
} else {
tmp_1 = t_14;
}
float t_15 = tmp_1;
float tmp_2;
if (t_12 <= -1.0000000116860974e-7f) {
tmp_2 = t_15;
} else if (t_12 <= 0.004999999888241291f) {
float tmp_3;
if (t_6 >= t_3) {
tmp_3 = t_9 / sqrtf(fmaxf((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)), t_4));
} else {
tmp_3 = t_14;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_15;
}
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) * floor(h)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_2) t_4 = fma(Float32(t_0 * dY_46_u), dY_46_u, t_3) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_7 = Float32(floor(h) * dY_46_v) t_8 = Float32(Float32(t_5 * t_5) + Float32(t_7 * t_7)) t_9 = Float32(floor(w) * dX_46_u) t_10 = Float32(Float32(t_9 * t_9) + Float32(t_1 * t_1)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_10, t_8))) tmp = Float32(0.0) if (t_10 >= t_8) tmp = Float32(t_11 * t_9); else tmp = Float32(t_11 * t_5); end t_12 = tmp t_13 = sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))), t_4)) t_14 = Float32(t_5 / t_13) tmp_1 = Float32(0.0) if (t_6 >= Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))) tmp_1 = Float32(t_9 / t_13); else tmp_1 = t_14; end t_15 = tmp_1 tmp_2 = Float32(0.0) if (t_12 <= Float32(-1.0000000116860974e-7)) tmp_2 = t_15; elseif (t_12 <= Float32(0.004999999888241291)) tmp_3 = Float32(0.0) if (t_6 >= t_3) tmp_3 = Float32(t_9 / sqrt(fmax(Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)), t_4))); else tmp_3 = t_14; end tmp_2 = tmp_3; else tmp_2 = t_15; 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 \left\lfloor h\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_2\\
t_4 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_3\right)\\
t_5 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_7 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_8 := t\_5 \cdot t\_5 + t\_7 \cdot t\_7\\
t_9 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_10 := t\_9 \cdot t\_9 + t\_1 \cdot t\_1\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_10, t\_8\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_10 \geq t\_8:\\
\;\;\;\;t\_11 \cdot t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_5\\
\end{array}\\
t_13 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}\\
t_14 := \frac{t\_5}{t\_13}\\
t_15 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\frac{t\_9}{t\_13}\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}\\
\mathbf{if}\;t\_12 \leq -1.0000000116860974 \cdot 10^{-7}:\\
\;\;\;\;t\_15\\
\mathbf{elif}\;t\_12 \leq 0.004999999888241291:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_3:\\
\;\;\;\;\frac{t\_9}{\sqrt{\mathsf{max}\left(\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array}
\end{array}
if (if (>=.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)))) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dY.u))) < -1.00000001e-7 or 0.00499999989 < (if (>=.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)))) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dY.u))) Initial program 99.3%
Applied rewrites99.3%
Taylor expanded in dX.u around inf
Applied rewrites97.9%
Taylor expanded in dY.u around inf
unpow2N/A
lift-floor.f32N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
associate-*l*N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3297.8
Applied rewrites97.8%
if -1.00000001e-7 < (if (>=.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)))) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (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)))))) (*.f32 (floor.f32 w) dY.u))) < 0.00499999989Initial program 58.8%
Applied rewrites58.9%
Taylor expanded in dX.u around inf
Applied rewrites39.8%
Taylor expanded in dY.u around 0
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f3244.1
Applied rewrites44.1%
Taylor expanded in dX.u around 0
Applied rewrites44.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* dX.u dX.u) (* (floor w) (floor w))))
(t_1
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_2
(/
(floor w)
(sqrt (fmax (fma (* (* dX.v (floor h)) (floor h)) dX.v t_0) t_1)))))
(if (>= 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 = (dX_46_u * dX_46_u) * (floorf(w) * floorf(w));
float t_1 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_2 = floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, t_0), t_1));
float tmp;
if (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(Float32(dX_46_u * dX_46_u) * Float32(floor(w) * floor(w))) t_1 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_2 = Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, t_0), t_1))) tmp = Float32(0.0) if (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(dX.u \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dX.v, t\_0\right), t\_1\right)}}\\
\mathbf{if}\;t\_0 \geq t\_1:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Taylor expanded in dX.u around inf
Applied rewrites64.7%
Applied rewrites64.5%
(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) (floor h)))
(t_2 (* dX.u (floor w)))
(t_3 (* (* dY.v dY.v) t_1))
(t_4 (fma (* t_0 dY.u) dY.u t_3)))
(if (>= (* (* t_2 dX.u) (floor w)) t_3)
(/
(* (floor w) dX.u)
(sqrt
(fmax (fma t_2 t_2 (* (* (* dX.v (floor h)) dX.v) (floor h))) t_4)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))) 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(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = dX_46_u * floorf(w);
float t_3 = (dY_46_v * dY_46_v) * t_1;
float t_4 = fmaf((t_0 * dY_46_u), dY_46_u, t_3);
float tmp;
if (((t_2 * dX_46_u) * floorf(w)) >= t_3) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(fmaf(t_2, t_2, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), t_4));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(dX_46_u * floor(w)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_4 = fma(Float32(t_0 * dY_46_u), dY_46_u, t_3) tmp = Float32(0.0) if (Float32(Float32(t_2 * dX_46_u) * floor(w)) >= t_3) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(fma(t_2, t_2, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), t_4))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_4))); 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 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_4 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_3\right)\\
\mathbf{if}\;\left(t\_2 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, t\_2, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Taylor expanded in dX.u around inf
Applied rewrites64.7%
Taylor expanded in dY.u around 0
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f3259.2
Applied rewrites59.2%
Applied rewrites59.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (* dX.u dX.u) t_0))
(t_2
(/
(floor w)
(sqrt
(fmax
(fma (* (* dX.v (floor h)) dX.v) (floor h) t_1)
(fma
(* (* dY.v (floor h)) (floor h))
dY.v
(* (* dY.u dY.u) t_0)))))))
(if (>= t_1 (* (* dY.v dY.v) (* (floor h) (floor h))))
(* 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 = (dX_46_u * dX_46_u) * t_0;
float t_2 = floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_1), fmaf(((dY_46_v * floorf(h)) * floorf(h)), dY_46_v, ((dY_46_u * dY_46_u) * t_0))));
float tmp;
if (t_1 >= ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h)))) {
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 = Float32(Float32(dX_46_u * dX_46_u) * t_0) t_2 = Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_1), fma(Float32(Float32(dY_46_v * floor(h)) * floor(h)), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0))))) tmp = Float32(0.0) if (t_1 >= Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) 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 := \left(dX.u \cdot dX.u\right) \cdot t\_0\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\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 , t\_1\right), \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor , dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}}\\
\mathbf{if}\;t\_1 \geq \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right):\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Taylor expanded in dX.u around inf
Applied rewrites64.7%
Taylor expanded in dY.u around 0
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f32N/A
lift-*.f3259.2
Applied rewrites59.2%
Applied rewrites59.1%
herbie shell --seed 2025101
(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))))