
(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_0) (* t_6 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) * 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_0;
} else {
tmp = t_6 * 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) * 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_0); else tmp = Float32(t_6 * t_4); 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
Herbie found 13 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_0) (* t_6 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) * 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_0;
} else {
tmp = t_6 * 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) * 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_0); else tmp = Float32(t_6 * t_4); 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_0; else tmp = t_6 * t_4; 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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\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 (* dY.v (floor h)))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_4 (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_4 t_3)
(/ (* (floor h) dX.v) (sqrt (fmax t_4 t_3)))
(/
(* (floor h) dY.v)
(sqrt
(fmax t_4 (fma t_1 t_1 (* (* (* (floor w) dY.u) dY.u) (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 = dY_46_v * floorf(h);
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_4 = fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_4 >= t_3) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(t_4, t_3));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_4, fmaf(t_1, t_1, (((floorf(w) * dY_46_u) * dY_46_u) * 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(dY_46_v * floor(h)) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_4 = fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_4 >= t_3) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(t_4, t_3))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_4, fma(t_1, t_1, Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * 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 := dY.v \cdot \left\lfloor h\right\rfloor \\
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, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_4 := \mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_4 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(t\_4, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Applied rewrites76.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_3 (* dY.v (floor h))))
(if (>= t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(*
dX.v
(/
(floor h)
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) (floor h))
dX.v
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* dY.v dY.v) t_0 (* (* dY.u dY.u) t_1))))))
(/
(* (floor h) dY.v)
(sqrt
(fmax t_2 (fma t_3 t_3 (* (* (* (floor w) dY.u) dY.u) (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 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_3 = dY_46_v * floorf(h);
float tmp;
if (t_2 >= fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0))) {
tmp = dX_46_v * (floorf(h) / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * floorf(h)), dX_46_v, (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((dY_46_v * dY_46_v), t_0, ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_2, fmaf(t_3, t_3, (((floorf(w) * dY_46_u) * dY_46_u) * 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 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_3 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (t_2 >= fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0))) tmp = Float32(dX_46_v * Float32(floor(h) / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * floor(h)), dX_46_v, Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1)))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_2, fma(t_3, t_3, Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * 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 dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;t\_2 \geq \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right):\\
\;\;\;\;dX.v \cdot \frac{\left\lfloor h\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, \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\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(t\_3, t\_3, \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Applied rewrites76.9%
Applied rewrites76.8%
Applied rewrites76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (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.v
(/
(floor h)
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* (floor w) dX.u) dX.u) (floor w)))
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* (* (floor w) dY.u) dY.u) (floor w)))))))
(/ (* (floor h) dY.v) (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_v * (floorf(h) / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((floorf(w) * dX_46_u) * dX_46_u) * floorf(w))), fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), (((floorf(w) * dY_46_u) * dY_46_u) * floorf(w))))));
} else {
tmp = (floorf(h) * dY_46_v) / 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_v * Float32(floor(h) / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w))), fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * floor(w))))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / 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.v \cdot \frac{\left\lfloor h\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 , \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
Applied rewrites76.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) (floor h)))
(t_5 (* (* dY.v dY.v) t_4))
(t_6 (* (floor w) (floor w)))
(t_7 (fma (* t_6 dX.u) dX.u (* t_4 (* dX.v dX.v))))
(t_8 (sqrt (fmax t_7 (fma (* t_6 dY.u) dY.u t_5))))
(t_9 (+ (* t_3 t_3) (* t_1 t_1)))
(t_10 (/ 1.0 (sqrt (fmax t_9 (+ (* t_0 t_0) (* t_2 t_2)))))))
(if (<= dY.u 25.0)
(if (>= t_9 t_5) (* t_10 t_1) (* t_10 t_2))
(if (>= t_7 (* (* t_0 dY.u) (floor w))) (/ t_1 t_8) (/ 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 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * floorf(h);
float t_5 = (dY_46_v * dY_46_v) * t_4;
float t_6 = floorf(w) * floorf(w);
float t_7 = fmaf((t_6 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v)));
float t_8 = sqrtf(fmaxf(t_7, fmaf((t_6 * dY_46_u), dY_46_u, t_5)));
float t_9 = (t_3 * t_3) + (t_1 * t_1);
float t_10 = 1.0f / sqrtf(fmaxf(t_9, ((t_0 * t_0) + (t_2 * t_2))));
float tmp_1;
if (dY_46_u <= 25.0f) {
float tmp_2;
if (t_9 >= t_5) {
tmp_2 = t_10 * t_1;
} else {
tmp_2 = t_10 * t_2;
}
tmp_1 = tmp_2;
} else if (t_7 >= ((t_0 * dY_46_u) * floorf(w))) {
tmp_1 = t_1 / t_8;
} 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(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * floor(h)) t_5 = Float32(Float32(dY_46_v * dY_46_v) * t_4) t_6 = Float32(floor(w) * floor(w)) t_7 = fma(Float32(t_6 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))) t_8 = sqrt(fmax(t_7, fma(Float32(t_6 * dY_46_u), dY_46_u, t_5))) t_9 = Float32(Float32(t_3 * t_3) + Float32(t_1 * t_1)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_9, Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2))))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(25.0)) tmp_2 = Float32(0.0) if (t_9 >= t_5) tmp_2 = Float32(t_10 * t_1); else tmp_2 = Float32(t_10 * t_2); end tmp_1 = tmp_2; elseif (t_7 >= Float32(Float32(t_0 * dY_46_u) * floor(w))) tmp_1 = Float32(t_1 / t_8); else tmp_1 = Float32(t_2 / t_8); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \left(dY.v \cdot dY.v\right) \cdot t\_4\\
t_6 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_7 := \mathsf{fma}\left(t\_6 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_8 := \sqrt{\mathsf{max}\left(t\_7, \mathsf{fma}\left(t\_6 \cdot dY.u, dY.u, t\_5\right)\right)}\\
t_9 := t\_3 \cdot t\_3 + t\_1 \cdot t\_1\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_0 \cdot t\_0 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;dY.u \leq 25:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_9 \geq t\_5:\\
\;\;\;\;t\_10 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_2\\
\end{array}\\
\mathbf{elif}\;t\_7 \geq \left(t\_0 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\frac{t\_1}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_8}\\
\end{array}
\end{array}
if dY.u < 25Initial program 78.4%
Taylor expanded in dY.u around 0
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3270.5
Applied rewrites70.5%
if 25 < dY.u Initial program 71.5%
Applied rewrites71.6%
Taylor expanded in dY.u around inf
unpow2N/A
pow2N/A
swap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3268.9
Applied rewrites68.9%
(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) dY.v))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) (floor h)))
(t_4 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_5 (fma (* t_0 dX.u) dX.u (* t_3 (* dX.v dX.v))))
(t_6 (* dX.v (floor h))))
(if (<= dX.u 10.0)
(if (>= (* (* t_6 dX.v) (floor h)) t_4)
(/
(* (floor h) dX.v)
(sqrt (fmax (fma (* t_6 (floor h)) dX.v (* (* dX.u dX.u) t_0)) t_4)))
(/ t_1 (sqrt (fmax t_5 t_4))))
(if (>= (* (* (* dX.u (floor w)) dX.u) (floor w)) t_4)
(*
t_6
(sqrt
(/
1.0
(fmax
(fma (* dX.u dX.u) t_0 (* (* dX.v dX.v) t_3))
(fma (* t_2 dY.v) (floor h) (* (* dY.u dY.u) t_0))))))
(/
t_1
(sqrt
(fmax
t_5
(fma t_2 t_2 (* (* (* (floor w) dY.u) dY.u) (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(w) * floorf(w);
float t_1 = floorf(h) * dY_46_v;
float t_2 = dY_46_v * floorf(h);
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 = fmaf((t_0 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v)));
float t_6 = dX_46_v * floorf(h);
float tmp_1;
if (dX_46_u <= 10.0f) {
float tmp_2;
if (((t_6 * dX_46_v) * floorf(h)) >= t_4) {
tmp_2 = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf((t_6 * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_0)), t_4));
} else {
tmp_2 = t_1 / sqrtf(fmaxf(t_5, t_4));
}
tmp_1 = tmp_2;
} else if ((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)) >= t_4) {
tmp_1 = t_6 * sqrtf((1.0f / fmaxf(fmaf((dX_46_u * dX_46_u), t_0, ((dX_46_v * dX_46_v) * t_3)), fmaf((t_2 * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_0)))));
} else {
tmp_1 = t_1 / sqrtf(fmaxf(t_5, fmaf(t_2, t_2, (((floorf(w) * dY_46_u) * dY_46_u) * floorf(w)))));
}
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) * dY_46_v) t_2 = Float32(dY_46_v * floor(h)) 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 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))) t_6 = Float32(dX_46_v * floor(h)) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(10.0)) tmp_2 = Float32(0.0) if (Float32(Float32(t_6 * dX_46_v) * floor(h)) >= t_4) tmp_2 = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(Float32(t_6 * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_0)), t_4))); else tmp_2 = Float32(t_1 / sqrt(fmax(t_5, t_4))); end tmp_1 = tmp_2; elseif (Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) >= t_4) tmp_1 = Float32(t_6 * sqrt(Float32(Float32(1.0) / fmax(fma(Float32(dX_46_u * dX_46_u), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_3)), fma(Float32(t_2 * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_0)))))); else tmp_1 = Float32(t_1 / sqrt(fmax(t_5, fma(t_2, t_2, Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * floor(w)))))); 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 dY.v\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
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 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_6 := dX.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.u \leq 10:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(t\_6 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_4:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6 \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_0\right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_5, t\_4\right)}}\\
\end{array}\\
\mathbf{elif}\;\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_4:\\
\;\;\;\;t\_6 \cdot \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_3\right), \mathsf{fma}\left(t\_2 \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_2, t\_2, \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
if dX.u < 10Initial program 78.3%
Applied rewrites78.3%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites70.4%
Applied rewrites70.5%
if 10 < dX.u Initial program 71.7%
Applied rewrites71.8%
Applied rewrites71.8%
Applied rewrites71.8%
Taylor expanded in dX.u around inf
Applied rewrites68.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1
(fma
(* t_0 dY.v)
(floor h)
(* (* (* (floor w) dY.u) dY.u) (floor w))))
(t_2 (* (floor w) (floor w)))
(t_3 (* (floor h) (floor h)))
(t_4 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_5 (* dX.v (floor h)))
(t_6 (* t_5 dX.v))
(t_7 (* (* (* (floor w) dX.u) dX.u) (floor w)))
(t_8 (sqrt (fmax (fma t_6 (floor h) t_7) t_1))))
(if (<= dX.u 7600.0)
(if (>= (* t_6 (floor h)) t_4)
(/
(* (floor h) dX.v)
(sqrt (fmax (fma (* t_5 (floor h)) dX.v (* (* dX.u dX.u) t_2)) t_4)))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_3 (* dX.v dX.v))) t_4))))
(if (>= t_7 t_1) (/ t_5 t_8) (/ t_0 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 = dY_46_v * floorf(h);
float t_1 = fmaf((t_0 * dY_46_v), floorf(h), (((floorf(w) * dY_46_u) * dY_46_u) * floorf(w)));
float t_2 = floorf(w) * floorf(w);
float t_3 = floorf(h) * floorf(h);
float t_4 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_3));
float t_5 = dX_46_v * floorf(h);
float t_6 = t_5 * dX_46_v;
float t_7 = ((floorf(w) * dX_46_u) * dX_46_u) * floorf(w);
float t_8 = sqrtf(fmaxf(fmaf(t_6, floorf(h), t_7), t_1));
float tmp_1;
if (dX_46_u <= 7600.0f) {
float tmp_2;
if ((t_6 * floorf(h)) >= t_4) {
tmp_2 = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf((t_5 * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_2)), t_4));
} else {
tmp_2 = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v))), t_4));
}
tmp_1 = tmp_2;
} else if (t_7 >= t_1) {
tmp_1 = t_5 / t_8;
} else {
tmp_1 = t_0 / 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(dY_46_v * floor(h)) t_1 = fma(Float32(t_0 * dY_46_v), floor(h), Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * floor(w))) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(floor(h) * floor(h)) t_4 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_3)) t_5 = Float32(dX_46_v * floor(h)) t_6 = Float32(t_5 * dX_46_v) t_7 = Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w)) t_8 = sqrt(fmax(fma(t_6, floor(h), t_7), t_1)) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(7600.0)) tmp_2 = Float32(0.0) if (Float32(t_6 * floor(h)) >= t_4) tmp_2 = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(Float32(t_5 * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_2)), t_4))); else tmp_2 = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))), t_4))); end tmp_1 = tmp_2; elseif (t_7 >= t_1) tmp_1 = Float32(t_5 / t_8); else tmp_1 = Float32(t_0 / t_8); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \mathsf{fma}\left(t\_0 \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_3\right)\\
t_5 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_6 := t\_5 \cdot dX.v\\
t_7 := \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_8 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor h\right\rfloor , t\_7\right), t\_1\right)}\\
\mathbf{if}\;dX.u \leq 7600:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \cdot \left\lfloor h\right\rfloor \geq t\_4:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_2\right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_7 \geq t\_1:\\
\;\;\;\;\frac{t\_5}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_8}\\
\end{array}
\end{array}
if dX.u < 7600Initial program 78.6%
Applied rewrites78.6%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites70.6%
Applied rewrites70.6%
if 7600 < dX.u Initial program 69.3%
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-*.f3266.7
Applied rewrites66.7%
Applied rewrites66.9%
(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.v (floor h)))
(t_3 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_1))))
(if (>= (* (* t_2 dX.v) (floor h)) t_3)
(/
(* (floor h) dX.v)
(sqrt (fmax (fma (* t_2 (floor h)) dX.v (* (* dX.u dX.u) t_0)) t_3)))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))) t_3))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float tmp;
if (((t_2 * dX_46_v) * floorf(h)) >= t_3) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf((t_2 * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_0)), t_3));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp = Float32(0.0) if (Float32(Float32(t_2 * dX_46_v) * floor(h)) >= t_3) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(Float32(t_2 * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_0)), t_3))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_3))); 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.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
\mathbf{if}\;\left(t\_2 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_0\right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\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\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Applied rewrites65.9%
(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 dY.u) dY.u (* (* dY.v dY.v) t_1))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) t_2)
(*
dX.v
(/
(floor h)
(sqrt
(fmax
(fma (* dX.u dX.u) t_0 (* (* dX.v dX.v) t_1))
(fma (* (* dY.v (floor h)) dY.v) (floor h) (* (* dY.u dY.u) t_0))))))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))) 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(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= t_2) {
tmp = dX_46_v * (floorf(h) / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_0, ((dX_46_v * dX_46_v) * t_1)), fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_0)))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= t_2) tmp = Float32(dX_46_v * Float32(floor(h) / sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), t_0, Float32(Float32(dX_46_v * dX_46_v) * t_1)), fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_0)))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_2))); 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 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_2:\\
\;\;\;\;dX.v \cdot \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_0, \left(dX.v \cdot dX.v\right) \cdot t\_1\right), \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\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\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Applied rewrites65.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* t_0 dY.u))
(t_2 (* (floor h) (floor h)))
(t_3 (fma (* t_0 dX.u) dX.u (* t_2 (* dX.v dX.v)))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) (* (* dY.u dY.u) t_0))
(/
(* (floor h) dX.v)
(sqrt (fmax t_3 (fma t_1 dY.u (exp (* (log (* dY.v (floor h))) 2.0))))))
(/
(* (floor h) dY.v)
(sqrt (fmax t_3 (fma t_1 dY.u (* (* dY.v dY.v) 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(w) * floorf(w);
float t_1 = t_0 * dY_46_u;
float t_2 = floorf(h) * floorf(h);
float t_3 = fmaf((t_0 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v)));
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(t_3, fmaf(t_1, dY_46_u, expf((logf((dY_46_v * floorf(h))) * 2.0f)))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, fmaf(t_1, dY_46_u, ((dY_46_v * dY_46_v) * 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(w) * floor(w)) t_1 = Float32(t_0 * dY_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(t_3, fma(t_1, dY_46_u, exp(Float32(log(Float32(dY_46_v * floor(h))) * Float32(2.0))))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, fma(t_1, dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_2))))); 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 := t\_0 \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_1, dY.u, e^{\log \left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot 2}\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_1, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.u around inf
Applied rewrites60.6%
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3262.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.1
Applied rewrites62.1%
(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 (* t_0 dX.u))
(t_3 (* dX.v (floor h)))
(t_4 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_1))))
(if (>= (* (* t_3 dX.v) (floor h)) (* (* dY.u dY.u) t_0))
(/
(* (floor h) dX.v)
(sqrt (fmax (fma t_2 dX.u (* (* t_3 (floor h)) dX.v)) t_4)))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma t_2 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 = t_0 * dX_46_u;
float t_3 = dX_46_v * floorf(h);
float t_4 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float tmp;
if (((t_3 * dX_46_v) * floorf(h)) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf(t_2, dX_46_u, ((t_3 * floorf(h)) * dX_46_v)), t_4));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf(t_2, 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(t_0 * dX_46_u) t_3 = Float32(dX_46_v * floor(h)) t_4 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp = Float32(0.0) if (Float32(Float32(t_3 * dX_46_v) * floor(h)) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(t_2, dX_46_u, Float32(Float32(t_3 * floor(h)) * dX_46_v)), t_4))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(t_2, 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 := t\_0 \cdot dX.u\\
t_3 := dX.v \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\_1\right)\\
\mathbf{if}\;\left(t\_3 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, dX.u, \left(t\_3 \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.u around inf
Applied rewrites60.6%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f3260.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.7
Applied rewrites60.7%
(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 (* (* dY.u dY.u) t_0))
(t_3 (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))))
(t_4 (* dY.v (floor h))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) t_2)
(/
(* (floor h) dX.v)
(sqrt (fmax t_3 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_1)))))
(/ (* (floor h) dY.v) (sqrt (fmax t_3 (fma t_4 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(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_u * dY_46_u) * t_0;
float t_3 = fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v)));
float t_4 = dY_46_v * floorf(h);
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= t_2) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(t_3, fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, fmaf(t_4, 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_u * dY_46_u) * t_0) t_3 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))) t_4 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(t_3, fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, fma(t_4, t_4, t_2)))); 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 := \left(dY.u \cdot dY.u\right) \cdot t\_0\\
t_3 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_4, t\_4, t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.u around inf
Applied rewrites60.6%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites60.6%
(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 dY.u) dY.u (* (* dY.v dY.v) t_1))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) (* (* dY.u dY.u) t_0))
(/
(* (floor h) dX.v)
(sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))) t_2)))
(/
(* (floor h) dY.v)
(sqrt (fmax (* (* (* dX.u (floor w)) dX.u) (floor w)) 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(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), t_2));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)), 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_2))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)), t_2))); 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 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\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\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.u around inf
Applied rewrites60.6%
Taylor expanded in dX.u around inf
Applied rewrites60.6%
(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 dY.u) dY.u (* (* dY.v dY.v) t_1))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) (* (* dY.u dY.u) t_0))
(/
(* (floor h) dX.v)
(sqrt (fmax (* (* (* dX.u (floor w)) dX.u) (floor w)) t_2)))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))) 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(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= ((dY_46_u * dY_46_u) * t_0)) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf((((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)), t_2));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), 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(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)), t_2))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_2))); 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 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\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\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.7%
Applied rewrites76.8%
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
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.8%
Taylor expanded in dY.u around inf
Applied rewrites60.6%
Taylor expanded in dX.u around inf
Applied rewrites43.8%
herbie shell --seed 2025113
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:name "Anisotropic x16 LOD (line direction, v)"
: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 h) dX.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 h) dY.v))))