
(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 9 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 (* (* dY.u (floor w)) dY.u))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (* (* dY.v (floor h)) dY.v))
(t_3 (fma t_0 (floor w) (* t_2 (floor h)))))
(if (>= t_1 t_3)
(/ (* (- dX.u) (floor w)) (- (sqrt (fmax t_1 t_3))))
(/
(* (- dY.u) (floor w))
(- (sqrt (fmax t_1 (fma t_2 (floor h) (* t_0 (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 = (dY_46_u * floorf(w)) * dY_46_u;
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = (dY_46_v * floorf(h)) * dY_46_v;
float t_3 = fmaf(t_0, floorf(w), (t_2 * floorf(h)));
float tmp;
if (t_1 >= t_3) {
tmp = (-dX_46_u * floorf(w)) / -sqrtf(fmaxf(t_1, t_3));
} else {
tmp = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(t_1, fmaf(t_2, floorf(h), (t_0 * floorf(w)))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(Float32(dY_46_v * floor(h)) * dY_46_v) t_3 = fma(t_0, floor(w), Float32(t_2 * floor(h))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(Float32(-dX_46_u) * floor(w)) / Float32(-sqrt(fmax(t_1, t_3)))); else tmp = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(t_1, fma(t_2, floor(h), Float32(t_0 * floor(w))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_3 := \mathsf{fma}\left(t\_0, \left\lfloor w\right\rfloor , t\_2 \cdot \left\lfloor h\right\rfloor \right)\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{\left(-dX.u\right) \cdot \left\lfloor w\right\rfloor }{-\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{-\sqrt{\mathsf{max}\left(t\_1, \mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , t\_0 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.4%
Applied rewrites76.5%
Applied rewrites76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (* (floor h) dX.v))
(t_6 (* dY.u (floor w))))
(if (>= (+ t_4 (* t_5 t_5)) (+ (* t_1 t_1) (* t_2 t_2)))
(/ t_3 (sqrt (fmax (fma t_5 t_5 t_4) (fma t_6 t_6 (* t_0 t_0)))))
(*
(/
dY.u
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* t_6 dY.u) (floor w) (* (* t_0 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 = dY_46_v * floorf(h);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = t_3 * t_3;
float t_5 = floorf(h) * dX_46_v;
float t_6 = dY_46_u * floorf(w);
float tmp;
if ((t_4 + (t_5 * t_5)) >= ((t_1 * t_1) + (t_2 * t_2))) {
tmp = t_3 / sqrtf(fmaxf(fmaf(t_5, t_5, t_4), fmaf(t_6, t_6, (t_0 * t_0))));
} else {
tmp = (dY_46_u / sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((t_6 * dY_46_u), floorf(w), ((t_0 * 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(dY_46_v * floor(h)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(t_3 * t_3) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(dY_46_u * floor(w)) tmp = Float32(0.0) if (Float32(t_4 + Float32(t_5 * t_5)) >= Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) tmp = Float32(t_3 / sqrt(fmax(fma(t_5, t_5, t_4), fma(t_6, t_6, Float32(t_0 * t_0))))); else tmp = Float32(Float32(dY_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(t_6 * dY_46_u), floor(w), Float32(Float32(t_0 * dY_46_v) * floor(h)))))) * floor(w)); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := dY.u \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;t\_4 + t\_5 \cdot t\_5 \geq t\_1 \cdot t\_1 + t\_2 \cdot t\_2:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, t\_5, t\_4\right), \mathsf{fma}\left(t\_6, t\_6, t\_0 \cdot t\_0\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\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(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(t\_6 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_0 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.4%
Applied rewrites76.5%
Applied rewrites76.4%
Applied rewrites76.4%
(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 (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (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 * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * 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(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * 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(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
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 76.3%
Applied rewrites76.3%
(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 (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* dX.u (floor w)) dX.u) (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 * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * 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(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * 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(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
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 76.3%
Applied rewrites76.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.2
Applied rewrites76.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.2
Applied rewrites76.2%
(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_4 (* (fabs (floor w)) (floor w)))
(t_5 (* dY.u (* (* (sqrt dY.u) (sqrt dY.u)) t_4)))
(t_6 (sqrt (- dY.u)))
(t_7 (* dY.u (* (* t_6 t_6) t_4)))
(t_8 (+ t_3 (* t_0 t_0)))
(t_9 (* (floor h) dY.v))
(t_10 (+ (* t_1 t_1) (* t_9 t_9)))
(t_11 (/ 1.0 (sqrt (fmax t_8 t_10))))
(t_12 (if (>= t_8 t_10) (* t_11 t_2) (* t_11 t_1)))
(t_13 (* dY.u (floor w)))
(t_14
(/
(* (- dY.u) (floor w))
(-
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* t_13 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))))
(if (<= t_12 -1.999999982195158e-37)
(if (>= t_8 t_5) (* (/ 1.0 (sqrt (fmax t_8 t_5))) t_2) t_14)
(if (<= t_12 0.00039999998989515007)
(if (>= t_8 t_7) (* (/ 1.0 (sqrt (fmax t_8 t_7))) t_2) t_14)
(if (>= t_8 (* dY.u (* (fabs dY.u) t_4)))
(sqrt (/ t_3 (fmax (fma t_0 t_0 t_3) (* (fabs t_13) t_13))))
t_14)))))
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;
float t_4 = fabsf(floorf(w)) * floorf(w);
float t_5 = dY_46_u * ((sqrtf(dY_46_u) * sqrtf(dY_46_u)) * t_4);
float t_6 = sqrtf(-dY_46_u);
float t_7 = dY_46_u * ((t_6 * t_6) * t_4);
float t_8 = t_3 + (t_0 * t_0);
float t_9 = floorf(h) * dY_46_v;
float t_10 = (t_1 * t_1) + (t_9 * t_9);
float t_11 = 1.0f / sqrtf(fmaxf(t_8, t_10));
float tmp;
if (t_8 >= t_10) {
tmp = t_11 * t_2;
} else {
tmp = t_11 * t_1;
}
float t_12 = tmp;
float t_13 = dY_46_u * floorf(w);
float t_14 = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((t_13 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
float tmp_2;
if (t_12 <= -1.999999982195158e-37f) {
float tmp_3;
if (t_8 >= t_5) {
tmp_3 = (1.0f / sqrtf(fmaxf(t_8, t_5))) * t_2;
} else {
tmp_3 = t_14;
}
tmp_2 = tmp_3;
} else if (t_12 <= 0.00039999998989515007f) {
float tmp_4;
if (t_8 >= t_7) {
tmp_4 = (1.0f / sqrtf(fmaxf(t_8, t_7))) * t_2;
} else {
tmp_4 = t_14;
}
tmp_2 = tmp_4;
} else if (t_8 >= (dY_46_u * (fabsf(dY_46_u) * t_4))) {
tmp_2 = sqrtf((t_3 / fmaxf(fmaf(t_0, t_0, t_3), (fabsf(t_13) * t_13))));
} else {
tmp_2 = t_14;
}
return tmp_2;
}
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(t_2 * t_2) t_4 = Float32(abs(floor(w)) * floor(w)) t_5 = Float32(dY_46_u * Float32(Float32(sqrt(dY_46_u) * sqrt(dY_46_u)) * t_4)) t_6 = sqrt(Float32(-dY_46_u)) t_7 = Float32(dY_46_u * Float32(Float32(t_6 * t_6) * t_4)) t_8 = Float32(t_3 + Float32(t_0 * t_0)) t_9 = Float32(floor(h) * dY_46_v) t_10 = Float32(Float32(t_1 * t_1) + Float32(t_9 * t_9)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_10))) tmp = Float32(0.0) if (t_8 >= t_10) tmp = Float32(t_11 * t_2); else tmp = Float32(t_11 * t_1); end t_12 = tmp t_13 = Float32(dY_46_u * floor(w)) t_14 = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(t_13 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))) tmp_2 = Float32(0.0) if (t_12 <= Float32(-1.999999982195158e-37)) tmp_3 = Float32(0.0) if (t_8 >= t_5) tmp_3 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_8, t_5))) * t_2); else tmp_3 = t_14; end tmp_2 = tmp_3; elseif (t_12 <= Float32(0.00039999998989515007)) tmp_4 = Float32(0.0) if (t_8 >= t_7) tmp_4 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_8, t_7))) * t_2); else tmp_4 = t_14; end tmp_2 = tmp_4; elseif (t_8 >= Float32(dY_46_u * Float32(abs(dY_46_u) * t_4))) tmp_2 = sqrt(Float32(t_3 / fmax(fma(t_0, t_0, t_3), Float32(abs(t_13) * t_13)))); else tmp_2 = t_14; end return tmp_2 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_4 := \left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \\
t_5 := dY.u \cdot \left(\left(\sqrt{dY.u} \cdot \sqrt{dY.u}\right) \cdot t\_4\right)\\
t_6 := \sqrt{-dY.u}\\
t_7 := dY.u \cdot \left(\left(t\_6 \cdot t\_6\right) \cdot t\_4\right)\\
t_8 := t\_3 + t\_0 \cdot t\_0\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := t\_1 \cdot t\_1 + t\_9 \cdot t\_9\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_10\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_10:\\
\;\;\;\;t\_11 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_1\\
\end{array}\\
t_13 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_14 := \frac{\left(-dY.u\right) \cdot \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 , \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(t\_13 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\mathbf{if}\;t\_12 \leq -1.999999982195158 \cdot 10^{-37}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_5:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_5\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}\\
\mathbf{elif}\;t\_12 \leq 0.00039999998989515007:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_7:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_7\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}\\
\mathbf{elif}\;t\_8 \geq dY.u \cdot \left(\left|dY.u\right| \cdot t\_4\right):\\
\;\;\;\;\sqrt{\frac{t\_3}{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_3\right), \left|t\_13\right| \cdot t\_13\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\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))) < -1.99999998e-37Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
lift-fabs.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3254.2
Applied rewrites54.2%
lift-fabs.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3254.2
Applied rewrites54.2%
if -1.99999998e-37 < (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))) < 3.9999999e-4Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
lift-fabs.f32N/A
neg-fabsN/A
lift-neg.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3245.4
Applied rewrites45.4%
lift-fabs.f32N/A
neg-fabsN/A
lift-neg.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3245.4
Applied rewrites45.4%
if 3.9999999e-4 < (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 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
Applied rewrites33.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* t_0 t_0))
(t_2 (* dY.v (floor h)))
(t_3 (* (floor h) dX.v))
(t_4 (fma t_3 t_3 t_1))
(t_5 (+ t_1 (* t_3 t_3)))
(t_6 (* dY.u (floor w)))
(t_7 (* (fabs t_6) t_6))
(t_8
(*
dY.u
(* (* (sqrt dY.u) (sqrt dY.u)) (* (fabs (floor w)) (floor w))))))
(if (<= dY.v 6500000.0)
(if (>= t_4 t_7)
(/ t_0 (sqrt (fmax t_4 t_7)))
(* (/ (floor w) (sqrt (fmax t_4 (fma t_6 t_6 (* t_2 t_2))))) dY.u))
(if (>= t_5 t_8)
(* (/ 1.0 (sqrt (fmax t_5 t_8))) t_0)
(/
(* (- dY.u) (floor w))
(-
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* t_6 dY.u) (floor w) (* (* t_2 dY.v) (floor h)))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = t_0 * t_0;
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) * dX_46_v;
float t_4 = fmaf(t_3, t_3, t_1);
float t_5 = t_1 + (t_3 * t_3);
float t_6 = dY_46_u * floorf(w);
float t_7 = fabsf(t_6) * t_6;
float t_8 = dY_46_u * ((sqrtf(dY_46_u) * sqrtf(dY_46_u)) * (fabsf(floorf(w)) * floorf(w)));
float tmp_1;
if (dY_46_v <= 6500000.0f) {
float tmp_2;
if (t_4 >= t_7) {
tmp_2 = t_0 / sqrtf(fmaxf(t_4, t_7));
} else {
tmp_2 = (floorf(w) / sqrtf(fmaxf(t_4, fmaf(t_6, t_6, (t_2 * t_2))))) * dY_46_u;
}
tmp_1 = tmp_2;
} else if (t_5 >= t_8) {
tmp_1 = (1.0f / sqrtf(fmaxf(t_5, t_8))) * t_0;
} else {
tmp_1 = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((t_6 * dY_46_u), floorf(w), ((t_2 * dY_46_v) * floorf(h)))));
}
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) * dX_46_u) t_1 = Float32(t_0 * t_0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) * dX_46_v) t_4 = fma(t_3, t_3, t_1) t_5 = Float32(t_1 + Float32(t_3 * t_3)) t_6 = Float32(dY_46_u * floor(w)) t_7 = Float32(abs(t_6) * t_6) t_8 = Float32(dY_46_u * Float32(Float32(sqrt(dY_46_u) * sqrt(dY_46_u)) * Float32(abs(floor(w)) * floor(w)))) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(6500000.0)) tmp_2 = Float32(0.0) if (t_4 >= t_7) tmp_2 = Float32(t_0 / sqrt(fmax(t_4, t_7))); else tmp_2 = Float32(Float32(floor(w) / sqrt(fmax(t_4, fma(t_6, t_6, Float32(t_2 * t_2))))) * dY_46_u); end tmp_1 = tmp_2; elseif (t_5 >= t_8) tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, t_8))) * t_0); else tmp_1 = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(t_6 * dY_46_u), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))))))); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := t\_0 \cdot t\_0\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \mathsf{fma}\left(t\_3, t\_3, t\_1\right)\\
t_5 := t\_1 + t\_3 \cdot t\_3\\
t_6 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_7 := \left|t\_6\right| \cdot t\_6\\
t_8 := dY.u \cdot \left(\left(\sqrt{dY.u} \cdot \sqrt{dY.u}\right) \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\\
\mathbf{if}\;dY.v \leq 6500000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_7:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_4, t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_6, t\_6, t\_2 \cdot t\_2\right)\right)}} \cdot dY.u\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_8:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_8\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \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 , \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(t\_6 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
if dY.v < 6.5e6Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
Applied rewrites56.5%
if 6.5e6 < dY.v Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
lift-fabs.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3254.2
Applied rewrites54.2%
lift-fabs.f32N/A
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lower-unsound-sqrt.f3254.2
Applied rewrites54.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.u (floor w)))
(t_2
(/
(* (- dY.u) (floor w))
(-
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* t_1 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))))
(t_3 (* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w)))))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4))
(t_6 (+ t_5 (* (- dX.v) (fabs (* t_0 (floor h)))))))
(if (<= dY.v 149999992832.0)
(if (>= (+ t_5 (* t_0 t_0)) t_3)
(/ 1.0 (/ (sqrt (fmax (fma t_0 t_0 t_5) (* (fabs t_1) t_1))) t_4))
t_2)
(if (>= t_6 t_3) (* (/ 1.0 (sqrt (fmax t_6 t_3))) 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) * dX_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((t_1 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
float t_3 = dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w)));
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float t_6 = t_5 + (-dX_46_v * fabsf((t_0 * floorf(h))));
float tmp_1;
if (dY_46_v <= 149999992832.0f) {
float tmp_2;
if ((t_5 + (t_0 * t_0)) >= t_3) {
tmp_2 = 1.0f / (sqrtf(fmaxf(fmaf(t_0, t_0, t_5), (fabsf(t_1) * t_1))) / t_4);
} else {
tmp_2 = t_2;
}
tmp_1 = tmp_2;
} else if (t_6 >= t_3) {
tmp_1 = (1.0f / sqrtf(fmaxf(t_6, t_3))) * t_4;
} else {
tmp_1 = t_2;
}
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) * dX_46_v) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))) t_3 = Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w)))) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) t_6 = Float32(t_5 + Float32(Float32(-dX_46_v) * abs(Float32(t_0 * floor(h))))) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(149999992832.0)) tmp_2 = Float32(0.0) if (Float32(t_5 + Float32(t_0 * t_0)) >= t_3) tmp_2 = Float32(Float32(1.0) / Float32(sqrt(fmax(fma(t_0, t_0, t_5), Float32(abs(t_1) * t_1))) / t_4)); else tmp_2 = t_2; end tmp_1 = tmp_2; elseif (t_6 >= t_3) tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_6, t_3))) * t_4); else tmp_1 = t_2; end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \frac{\left(-dY.u\right) \cdot \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 , \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(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
t_3 := dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
t_6 := t\_5 + \left(-dX.v\right) \cdot \left|t\_0 \cdot \left\lfloor h\right\rfloor \right|\\
\mathbf{if}\;dY.v \leq 149999992832:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 + t\_0 \cdot t\_0 \geq t\_3:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_5\right), \left|t\_1\right| \cdot t\_1\right)}}{t\_4}}\\
\mathbf{else}:\\
\;\;\;\;t\_2\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_3:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_2\\
\end{array}
\end{array}
if dY.v < 149999993000Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
Applied rewrites56.6%
if 149999993000 < dY.v Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
lift-*.f32N/A
fabs-sqrN/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
fabs-mulN/A
neg-fabsN/A
lift-neg.f32N/A
rem-exp-logN/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
rem-exp-logN/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f3250.8
Applied rewrites50.8%
lift-*.f32N/A
fabs-sqrN/A
lift-*.f32N/A
associate-*r*N/A
*-commutativeN/A
fabs-mulN/A
neg-fabsN/A
lift-neg.f32N/A
rem-exp-logN/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
rem-exp-logN/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f3250.1
Applied rewrites50.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.u (floor w)))
(t_2 (* (floor w) dX.u))
(t_3 (* t_2 t_2)))
(if (>=
(+ t_3 (* t_0 t_0))
(* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w)))))
(/ 1.0 (/ (sqrt (fmax (fma t_0 t_0 t_3) (* (fabs t_1) t_1))) t_2))
(/
(* (- dY.u) (floor w))
(-
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* t_1 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))))))
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 = dY_46_u * floorf(w);
float t_2 = floorf(w) * dX_46_u;
float t_3 = t_2 * t_2;
float tmp;
if ((t_3 + (t_0 * t_0)) >= (dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w))))) {
tmp = 1.0f / (sqrtf(fmaxf(fmaf(t_0, t_0, t_3), (fabsf(t_1) * t_1))) / t_2);
} else {
tmp = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))), fmaf((t_1 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
}
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(dY_46_u * floor(w)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(t_2 * t_2) tmp = Float32(0.0) if (Float32(t_3 + Float32(t_0 * t_0)) >= Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w))))) tmp = Float32(Float32(1.0) / Float32(sqrt(fmax(fma(t_0, t_0, t_3), Float32(abs(t_1) * t_1))) / t_2)); else tmp = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))), fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2\\
\mathbf{if}\;t\_3 + t\_0 \cdot t\_0 \geq dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right):\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_3\right), \left|t\_1\right| \cdot t\_1\right)}}{t\_2}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \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 , \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(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
Applied rewrites56.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* dY.u (floor w)))
(t_2 (* (fabs t_1) t_1))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor w) dX.u))
(t_5 (fma t_3 t_3 (* t_4 t_4))))
(if (>= t_5 t_2)
(/ t_4 (sqrt (fmax t_5 t_2)))
(* (/ (floor w) (sqrt (fmax t_5 (fma t_1 t_1 (* t_0 t_0))))) 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 = dY_46_v * floorf(h);
float t_1 = dY_46_u * floorf(w);
float t_2 = fabsf(t_1) * t_1;
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = fmaf(t_3, t_3, (t_4 * t_4));
float tmp;
if (t_5 >= t_2) {
tmp = t_4 / sqrtf(fmaxf(t_5, t_2));
} else {
tmp = (floorf(w) / sqrtf(fmaxf(t_5, fmaf(t_1, t_1, (t_0 * t_0))))) * 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(dY_46_v * floor(h)) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(abs(t_1) * t_1) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = fma(t_3, t_3, Float32(t_4 * t_4)) tmp = Float32(0.0) if (t_5 >= t_2) tmp = Float32(t_4 / sqrt(fmax(t_5, t_2))); else tmp = Float32(Float32(floor(w) / sqrt(fmax(t_5, fma(t_1, t_1, Float32(t_0 * t_0))))) * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left|t\_1\right| \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \mathsf{fma}\left(t\_3, t\_3, t\_4 \cdot t\_4\right)\\
\mathbf{if}\;t\_5 \geq t\_2:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_5, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_1, t\_1, t\_0 \cdot t\_0\right)\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.4%
Applied rewrites68.1%
Applied rewrites70.7%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.9
Applied rewrites56.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.5
Applied rewrites56.5%
Applied rewrites56.5%
herbie shell --seed 2025154
(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))))