
(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 6 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 (* (floor w) dX.u))
(t_2 (* dY.v (floor h)))
(t_3 (* (* t_2 (floor h)) dY.v))
(t_4 (* (floor w) (floor w))))
(if (>=
(fma t_1 t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma (* (* dY.u dY.u) (floor w)) (floor w) t_3))
(/
(* (* (floor h) dX.v) -1.0)
(-
(sqrt
(fmax
(fma (* (* dX.u (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_0))
(fma (* dY.u dY.u) t_4 (* (* t_2 dY.v) (floor h)))))))
(/
(* (* (floor h) dY.v) -1.0)
(-
(sqrt
(fmax
(fma (* t_4 dX.u) dX.u (* t_0 (* dX.v dX.v)))
(fma (* t_4 dY.u) dY.u t_3))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * dX_46_u;
float t_2 = dY_46_v * floorf(h);
float t_3 = (t_2 * floorf(h)) * dY_46_v;
float t_4 = floorf(w) * floorf(w);
float tmp;
if (fmaf(t_1, t_1, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))) >= fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), t_3)) {
tmp = ((floorf(h) * dX_46_v) * -1.0f) / -sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_0)), fmaf((dY_46_u * dY_46_u), t_4, ((t_2 * dY_46_v) * floorf(h)))));
} else {
tmp = ((floorf(h) * dY_46_v) * -1.0f) / -sqrtf(fmaxf(fmaf((t_4 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), fmaf((t_4 * dY_46_u), dY_46_u, t_3)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(t_2 * floor(h)) * dY_46_v) t_4 = Float32(floor(w) * floor(w)) tmp = Float32(0.0) if (fma(t_1, t_1, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) >= fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), t_3)) tmp = Float32(Float32(Float32(floor(h) * dX_46_v) * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(dY_46_u * dY_46_u), t_4, Float32(Float32(t_2 * dY_46_v) * floor(h))))))); else tmp = Float32(Float32(Float32(floor(h) * dY_46_v) * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_4 * dY_46_u), dY_46_u, t_3))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(t\_2 \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;\mathsf{fma}\left(t\_1, t\_1, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \geq \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_3\right):\\
\;\;\;\;\frac{\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_4, \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Applied rewrites76.1%
Applied rewrites76.1%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.1
Applied rewrites76.1%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.2
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) (floor h)))
(t_3 (* (* dY.v dY.v) t_2)))
(if (>=
(fma t_1 t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma (* (* dY.u dY.u) (floor w)) (floor w) t_3))
(/
(* (* (floor h) dX.v) -1.0)
(-
(sqrt
(fmax
(fma (* (* dX.u (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_2))
(fma (* dY.u dY.u) t_0 (* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(/
(* (* (floor h) dY.v) -1.0)
(-
(sqrt
(fmax
(fma (* t_0 dX.u) dX.u (* t_2 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u 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(w) * dX_46_u;
float t_2 = floorf(h) * floorf(h);
float t_3 = (dY_46_v * dY_46_v) * t_2;
float tmp;
if (fmaf(t_1, t_1, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))) >= fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), t_3)) {
tmp = ((floorf(h) * dX_46_v) * -1.0f) / -sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_2)), fmaf((dY_46_u * dY_46_u), t_0, (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
} else {
tmp = ((floorf(h) * dY_46_v) * -1.0f) / -sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, 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(w) * dX_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_2) tmp = Float32(0.0) if (fma(t_1, t_1, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) >= fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), t_3)) tmp = Float32(Float32(Float32(floor(h) * dX_46_v) * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_2)), fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = Float32(Float32(Float32(floor(h) * dY_46_v) * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, 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 w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_2\\
\mathbf{if}\;\mathsf{fma}\left(t\_1, t\_1, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \geq \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_3\right):\\
\;\;\;\;\frac{\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Applied rewrites76.1%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor h) (floor h)))
(t_3 (* (* dY.v dY.v) t_2)))
(if (>=
(fma t_1 t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma (* (* dY.u dY.u) (floor w)) (floor w) t_3))
(*
(floor h)
(/
dX.v
(sqrt
(fmax
(fma (* (* dX.u (floor w)) dX.u) (floor w) (* (* dX.v dX.v) t_2))
(fma (* dY.u dY.u) t_0 (* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(/
(* (* (floor h) dY.v) -1.0)
(-
(sqrt
(fmax
(fma (* t_0 dX.u) dX.u (* t_2 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u 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(w) * dX_46_u;
float t_2 = floorf(h) * floorf(h);
float t_3 = (dY_46_v * dY_46_v) * t_2;
float tmp;
if (fmaf(t_1, t_1, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))) >= fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), t_3)) {
tmp = floorf(h) * (dX_46_v / sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((dX_46_v * dX_46_v) * t_2)), fmaf((dY_46_u * dY_46_u), t_0, (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))))));
} else {
tmp = ((floorf(h) * dY_46_v) * -1.0f) / -sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, 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(w) * dX_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_2) tmp = Float32(0.0) if (fma(t_1, t_1, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) >= fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), t_3)) tmp = Float32(floor(h) * Float32(dX_46_v / sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_2)), fma(Float32(dY_46_u * dY_46_u), t_0, Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = Float32(Float32(Float32(floor(h) * dY_46_v) * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, 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 w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_2\\
\mathbf{if}\;\mathsf{fma}\left(t\_1, t\_1, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \geq \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_3\right):\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.u \cdot dY.u, t\_0, \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.1%
Applied rewrites76.1%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* (floor w) dY.u) (floor w))
dY.u
(* (* (* dY.v (floor h)) (floor h)) dY.v)))
(t_1
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (floor h) (floor h)) (* dX.v dX.v))))
(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(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, (((dY_46_v * floorf(h)) * floorf(h)) * dY_46_v));
float t_1 = fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)));
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(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(Float32(dY_46_v * floor(h)) * floor(h)) * dY_46_v)) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) 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(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right)\\
t_1 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\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 76.1%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* (floor w) (floor w)))
(t_2 (* dY.v (floor h)))
(t_3
(/
(floor h)
(sqrt
(fmax
(fma (* dX.u dX.u) t_1 (* (* (* (floor h) dX.v) dX.v) (floor h)))
(fma (* t_2 dY.v) (floor h) (* t_0 t_0)))))))
(if (>=
(fma (* t_1 dX.u) dX.u (* (* (floor h) (floor h)) (* dX.v dX.v)))
(fma (* (* (floor w) dY.u) (floor w)) dY.u (* (* t_2 (floor h)) dY.v)))
(* t_3 dX.v)
(* t_3 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 = floorf(w) * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = floorf(h) / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_1, (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h))), fmaf((t_2 * dY_46_v), floorf(h), (t_0 * t_0))));
float tmp;
if (fmaf((t_1 * dX_46_u), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v))) >= fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, ((t_2 * floorf(h)) * dY_46_v))) {
tmp = t_3 * dX_46_v;
} else {
tmp = t_3 * 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 = Float32(floor(w) * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(floor(h) / sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), t_1, Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))), fma(Float32(t_2 * dY_46_v), floor(h), Float32(t_0 * t_0))))) tmp = Float32(0.0) if (fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) >= fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(t_2 * floor(h)) * dY_46_v))) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * 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 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(t\_2 \cdot dY.v, \left\lfloor h\right\rfloor , t\_0 \cdot t\_0\right)\right)}}\\
\mathbf{if}\;\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right) \geq \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(t\_2 \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right):\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-*.f3276.0
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.0
Applied rewrites76.0%
Applied rewrites76.0%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (* (floor w) dY.u) (floor w)) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (fma (* (* (floor w) (floor w)) dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_3 (/ (floor h) (sqrt (fmax t_2 t_1)))))
(if (>= t_2 t_1) (* t_3 dX.v) (* t_3 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 = floorf(h) * floorf(h);
float t_1 = fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_3 = floorf(h) / sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_3 * dX_46_v;
} else {
tmp = t_3 * 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(floor(h) * floor(h)) t_1 = fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_3 = Float32(floor(h) / sqrt(fmax(t_2, t_1))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); 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 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
herbie shell --seed 2025143
(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))))