
(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}
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}
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}
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}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.v (floor h)))
(t_2 (* t_1 dY.v))
(t_3 (* dX.v (floor h)))
(t_4
(fma
(* t_3 dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_5 (* dY.u (floor w)))
(t_6 (* t_5 dY.u)))
(if (>= t_4 (fma t_2 (floor h) (* t_6 (floor w))))
(/ 1.0 (/ (sqrt (fmax (fma t_6 (floor w) (* t_2 (floor h))) t_4)) t_3))
(/
(- (- dY.v))
(/
(sqrt
(fmax
(fma t_1 t_1 (* t_5 t_5))
(fma (* (* (floor w) (floor w)) dX.u) dX.u (* t_0 t_0))))
(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_v * floorf(h);
float t_2 = t_1 * dY_46_v;
float t_3 = dX_46_v * floorf(h);
float t_4 = fmaf((t_3 * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_5 = dY_46_u * floorf(w);
float t_6 = t_5 * dY_46_u;
float tmp;
if (t_4 >= fmaf(t_2, floorf(h), (t_6 * floorf(w)))) {
tmp = 1.0f / (sqrtf(fmaxf(fmaf(t_6, floorf(w), (t_2 * floorf(h))), t_4)) / t_3);
} else {
tmp = -(-dY_46_v) / (sqrtf(fmaxf(fmaf(t_1, t_1, (t_5 * t_5)), fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (t_0 * t_0)))) / 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_v * floor(h)) t_2 = Float32(t_1 * dY_46_v) t_3 = Float32(dX_46_v * floor(h)) t_4 = fma(Float32(t_3 * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_5 = Float32(dY_46_u * floor(w)) t_6 = Float32(t_5 * dY_46_u) tmp = Float32(0.0) if (t_4 >= fma(t_2, floor(h), Float32(t_6 * floor(w)))) tmp = Float32(Float32(1.0) / Float32(sqrt(fmax(fma(t_6, floor(w), Float32(t_2 * floor(h))), t_4)) / t_3)); else tmp = Float32(Float32(-Float32(-dY_46_v)) / Float32(sqrt(fmax(fma(t_1, t_1, Float32(t_5 * t_5)), fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(t_0 * t_0)))) / floor(h))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := t\_1 \cdot dY.v\\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_3 \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_5 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_6 := t\_5 \cdot dY.u\\
\mathbf{if}\;t\_4 \geq \mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , t\_6 \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor w\right\rfloor , t\_2 \cdot \left\lfloor h\right\rfloor \right), t\_4\right)}}{t\_3}}\\
\mathbf{else}:\\
\;\;\;\;\frac{-\left(-dY.v\right)}{\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_5 \cdot t\_5\right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, t\_0 \cdot t\_0\right)\right)}}{\left\lfloor h\right\rfloor }}\\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
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 (* (* dY.u (floor w)) dY.u))
(t_2 (* (floor h) dX.v))
(t_3
(/
1.0
(sqrt
(fmax
(+ (* t_0 t_0) (* t_2 t_2))
(fma (* (* (floor h) (floor h)) dY.v) dY.v (* t_1 (floor w))))))))
(if (>=
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma t_1 (floor w) (* (* (* dY.v (floor h)) dY.v) (floor h))))
(* t_3 t_2)
(* t_3 (* (floor h) 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(w) * dX_46_u;
float t_1 = (dY_46_u * floorf(w)) * dY_46_u;
float t_2 = floorf(h) * dX_46_v;
float t_3 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + (t_2 * t_2)), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, (t_1 * floorf(w)))));
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_1, floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))) {
tmp = t_3 * t_2;
} else {
tmp = t_3 * (floorf(h) * 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(w) * dX_46_u) t_1 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(t_1 * floor(w)))))) 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_1, floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))) tmp = Float32(t_3 * t_2); else tmp = Float32(t_3 * Float32(floor(h) * dY_46_v)); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, t\_1 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\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\_1, \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\_3 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)\\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
Applied rewrites76.0%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
Applied rewrites75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* dY.u (floor w)) dY.u))
(t_2 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_3
(/
1.0
(sqrt
(fmax
(fma t_0 (* dX.v dX.v) t_2)
(fma (* t_0 dY.v) dY.v (* t_1 (floor w))))))))
(if (>=
(fma (* (* dX.v (floor h)) dX.v) (floor h) t_2)
(fma t_1 (floor w) (* (* (* dY.v (floor h)) dY.v) (floor h))))
(* t_3 (* (floor h) dX.v))
(* t_3 (* (floor h) 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 = (dY_46_u * floorf(w)) * dY_46_u;
float t_2 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_3 = 1.0f / sqrtf(fmaxf(fmaf(t_0, (dX_46_v * dX_46_v), t_2), fmaf((t_0 * dY_46_v), dY_46_v, (t_1 * floorf(w)))));
float tmp;
if (fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_2) >= fmaf(t_1, floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))) {
tmp = t_3 * (floorf(h) * dX_46_v);
} else {
tmp = t_3 * (floorf(h) * 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 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_2 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_3 = Float32(Float32(1.0) / sqrt(fmax(fma(t_0, Float32(dX_46_v * dX_46_v), t_2), fma(Float32(t_0 * dY_46_v), dY_46_v, Float32(t_1 * floor(w)))))) tmp = Float32(0.0) if (fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_2) >= fma(t_1, floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))) tmp = Float32(t_3 * Float32(floor(h) * dX_46_v)); else tmp = Float32(t_3 * Float32(floor(h) * dY_46_v)); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_2 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dX.v \cdot dX.v, t\_2\right), \mathsf{fma}\left(t\_0 \cdot dY.v, dY.v, t\_1 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\mathbf{if}\;\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_2\right) \geq \mathsf{fma}\left(t\_1, \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\_3 \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right)\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)\\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-+.f32N/A
+-commutativeN/A
*-commutativeN/A
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-+.f32N/A
+-commutativeN/A
*-commutativeN/A
Applied rewrites75.9%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f3275.9%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites75.9%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f3275.8%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (* dY.u (floor w)) dY.u))
(t_2 (* (* dY.v dY.v) (floor h)))
(t_3
(/
(floor h)
(sqrt
(fmax
(fma t_0 t_0 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma t_2 (floor h) (* t_1 (floor w))))))))
(if (>=
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma t_1 (floor w) (* t_2 (floor h))))
(* 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(w) * dX_46_u;
float t_1 = (dY_46_u * floorf(w)) * dY_46_u;
float t_2 = (dY_46_v * dY_46_v) * floorf(h);
float t_3 = floorf(h) / sqrtf(fmaxf(fmaf(t_0, t_0, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), fmaf(t_2, floorf(h), (t_1 * floorf(w)))));
float tmp;
if (fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))) >= fmaf(t_1, floorf(w), (t_2 * floorf(h)))) {
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(w) * dX_46_u) t_1 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_2 = Float32(Float32(dY_46_v * dY_46_v) * floor(h)) t_3 = Float32(floor(h) / sqrt(fmax(fma(t_0, t_0, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), fma(t_2, floor(h), Float32(t_1 * floor(w)))))) tmp = Float32(0.0) if (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))) >= fma(t_1, floor(w), Float32(t_2 * floor(h)))) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_2 := \left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , t\_1 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\mathbf{if}\;\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) \geq \mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , t\_2 \cdot \left\lfloor h\right\rfloor \right):\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
Initial program 76.0%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3275.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3275.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3275.9%
Applied rewrites75.9%
Applied rewrites75.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3275.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3275.9%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/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 (* (floor h) dX.v))
(t_1 (* dY.v (* (fabs dY.v) (* (fabs (floor h)) (floor h)))))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0))))
(if (>= t_3 t_1)
(* (/ 1.0 (sqrt (fmax t_3 t_1))) t_0)
(*
(/
dY.v
(sqrt
(fmax
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))))
(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_v * (fabsf(dY_46_v) * (fabsf(floorf(h)) * floorf(h)));
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float tmp;
if (t_3 >= t_1) {
tmp = (1.0f / sqrtf(fmaxf(t_3, t_1))) * t_0;
} else {
tmp = (dY_46_v / sqrtf(fmaxf(fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)))))) * 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_v * Float32(abs(dY_46_v) * Float32(abs(floor(h)) * floor(h)))) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_3, t_1))) * t_0); else tmp = Float32(Float32(dY_46_v / sqrt(fmax(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))), 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)))))) * floor(h)); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dY.v \cdot \left(\left|dY.v\right| \cdot \left(\left|\left\lfloor h\right\rfloor \right| \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_1\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v}{\sqrt{\mathsf{max}\left(\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), \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)\right)}} \cdot \left\lfloor h\right\rfloor \\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Applied rewrites67.6%
Applied rewrites70.2%
Taylor expanded in dY.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.8%
Applied rewrites56.8%
Taylor expanded in dY.u around 0
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.4%
Applied rewrites56.4%
herbie shell --seed 2025184
(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))))