
(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 5 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
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_1 (* dY.v (floor h)))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.u (floor w)))
(t_4 (fma (* t_3 dY.u) (floor w) (* (* t_1 dY.v) (floor h))))
(t_5 (* (floor w) dX.u)))
(if (>= t_0 t_4)
(* (/ (floor h) (sqrt (fmax t_0 t_4))) dX.v)
(/
t_1
(sqrt (fmax (fma t_2 t_2 (* t_5 t_5)) (fma t_1 t_1 (* t_3 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 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_1 = dY_46_v * floorf(h);
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_u * floorf(w);
float t_4 = fmaf((t_3 * dY_46_u), floorf(w), ((t_1 * dY_46_v) * floorf(h)));
float t_5 = floorf(w) * dX_46_u;
float tmp;
if (t_0 >= t_4) {
tmp = (floorf(h) / sqrtf(fmaxf(t_0, t_4))) * dX_46_v;
} else {
tmp = t_1 / sqrtf(fmaxf(fmaf(t_2, t_2, (t_5 * t_5)), fmaf(t_1, t_1, (t_3 * t_3))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = 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_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_u * floor(w)) t_4 = fma(Float32(t_3 * dY_46_u), floor(w), Float32(Float32(t_1 * dY_46_v) * floor(h))) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (t_0 >= t_4) tmp = Float32(Float32(floor(h) / sqrt(fmax(t_0, t_4))) * dX_46_v); else tmp = Float32(t_1 / sqrt(fmax(fma(t_2, t_2, Float32(t_5 * t_5)), fma(t_1, t_1, Float32(t_3 * t_3))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \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_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_3 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_1 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;t\_0 \geq t\_4:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_0, t\_4\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, t\_2, t\_5 \cdot t\_5\right), \mathsf{fma}\left(t\_1, t\_1, t\_3 \cdot t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* dY.u (floor w)))
(t_3 (* t_2 dY.u))
(t_4 (* dY.v (floor h))))
(if (>=
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma t_3 (floor w) (* (* t_4 dY.v) (floor h))))
(*
(/
dX.v
(sqrt (fmax (fma t_0 t_0 (* t_1 t_1)) (fma t_3 (floor w) (* t_4 t_4)))))
(floor h))
(/
t_4
(sqrt (fmax (fma t_1 t_1 (* t_0 t_0)) (fma t_4 t_4 (* t_2 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) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = dY_46_u * floorf(w);
float t_3 = t_2 * dY_46_u;
float t_4 = dY_46_v * floorf(h);
float tmp;
if (fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))) >= fmaf(t_3, floorf(w), ((t_4 * dY_46_v) * floorf(h)))) {
tmp = (dX_46_v / sqrtf(fmaxf(fmaf(t_0, t_0, (t_1 * t_1)), fmaf(t_3, floorf(w), (t_4 * t_4))))) * floorf(h);
} else {
tmp = t_4 / sqrtf(fmaxf(fmaf(t_1, t_1, (t_0 * t_0)), fmaf(t_4, t_4, (t_2 * 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) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(t_2 * dY_46_u) t_4 = Float32(dY_46_v * floor(h)) tmp = Float32(0.0) if (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(t_3, floor(w), Float32(Float32(t_4 * dY_46_v) * floor(h)))) tmp = Float32(Float32(dX_46_v / sqrt(fmax(fma(t_0, t_0, Float32(t_1 * t_1)), fma(t_3, floor(w), Float32(t_4 * t_4))))) * floor(h)); else tmp = Float32(t_4 / sqrt(fmax(fma(t_1, t_1, Float32(t_0 * t_0)), fma(t_4, t_4, Float32(t_2 * t_2))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := t\_2 \cdot dY.u\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;\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) \geq \mathsf{fma}\left(t\_3, \left\lfloor w\right\rfloor , \left(t\_4 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right):\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_1 \cdot t\_1\right), \mathsf{fma}\left(t\_3, \left\lfloor w\right\rfloor , t\_4 \cdot t\_4\right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_0 \cdot t\_0\right), \mathsf{fma}\left(t\_4, t\_4, t\_2 \cdot t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Applied rewrites76.1%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(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)))
(t_3 (fma t_0 t_0 (* t_2 t_2)))
(t_4 (/ (floor h) (sqrt (fmax t_1 t_3)))))
(if (>= t_1 t_3) (* t_4 dX.v) (* t_4 dY.v))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_u * floorf(w);
float t_1 = 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);
float t_3 = fmaf(t_0, t_0, (t_2 * t_2));
float t_4 = floorf(h) / sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_4 * dX_46_v;
} else {
tmp = t_4 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = 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(dY_46_v * floor(h)) t_3 = fma(t_0, t_0, Float32(t_2 * t_2)) t_4 = Float32(floor(h) / sqrt(fmax(t_1, t_3))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_4 * dX_46_v); else tmp = Float32(t_4 * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
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 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_0, t\_0, t\_2 \cdot t\_2\right)\\
t_4 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;t\_4 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-fma.f3276.0
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-fma.f3276.0
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-fma.f3276.0
Applied rewrites76.0%
(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 dX.u) (floor w)) (floor w))))
(t_2 (/ (floor h) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.v) (* t_2 dY.v))))
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 * dX_46_u) * floorf(w)) * floorf(w)));
float t_2 = floorf(h) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_v;
} else {
tmp = t_2 * dY_46_v;
}
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 * dX_46_u) * floor(w)) * floor(w))) t_2 = Float32(floor(h) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_v); else tmp = Float32(t_2 * dY_46_v); 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 dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3275.9
Applied rewrites75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (* (floor w) (fabs (* dY.u (floor w))))))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor h) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.v) (* t_2 dY.v))))
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) * fabsf((dY_46_u * floorf(w))));
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(h) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_v;
} else {
tmp = t_2 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * Float32(floor(w) * abs(Float32(dY_46_u * floor(w))))) 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(h) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_v); else tmp = Float32(t_2 * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left(\left\lfloor w\right\rfloor \cdot \left|dY.u \cdot \left\lfloor w\right\rfloor \right|\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 h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.0
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
associate-*r*N/A
sqrt-prodN/A
lower-unsound-*.f32N/A
Applied rewrites59.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.2
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
associate-*r*N/A
sqrt-prodN/A
lower-unsound-*.f32N/A
Applied rewrites59.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3259.2
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
rem-sqrt-square-revN/A
lift-*.f32N/A
associate-*r*N/A
sqrt-prodN/A
lower-unsound-*.f32N/A
Applied rewrites43.7%
Taylor expanded in dY.v around 0
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-sqrt.f32N/A
lower-sqrt.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3238.1
Applied rewrites38.1%
Taylor expanded in dY.v around 0
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-sqrt.f32N/A
lower-sqrt.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3240.0
Applied rewrites40.0%
Taylor expanded in dY.v around 0
lower-*.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-*.f32N/A
lower-sqrt.f32N/A
lower-sqrt.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3236.8
Applied rewrites36.8%
lift-*.f32N/A
lift-sqrt.f32N/A
lift-sqrt.f32N/A
sqrt-unprodN/A
lift-*.f32N/A
associate-*l*N/A
lift-pow.f32N/A
pow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
rem-sqrt-square-revN/A
lower-fabs.f3254.6
Applied rewrites54.6%
lift-*.f32N/A
lift-sqrt.f32N/A
lift-sqrt.f32N/A
sqrt-unprodN/A
lift-*.f32N/A
associate-*l*N/A
lift-pow.f32N/A
pow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
rem-sqrt-square-revN/A
lower-fabs.f3254.6
Applied rewrites54.6%
lift-*.f32N/A
lift-sqrt.f32N/A
lift-sqrt.f32N/A
sqrt-unprodN/A
lift-*.f32N/A
associate-*l*N/A
lift-pow.f32N/A
pow2N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
rem-sqrt-square-revN/A
lower-fabs.f3254.6
Applied rewrites54.6%
herbie shell --seed 2025164
(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))))