
(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 8 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 (* (floor w) dX.u))
(t_1 (fma t_0 t_0 (* (* (* (floor h) dX.v) (floor h)) dX.v)))
(t_2 (* (floor h) dY.v))
(t_3 (fma (* (* (floor w) (floor w)) dY.u) dY.u (* t_2 t_2)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ (* (floor w) dY.u) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = fmaf(t_0, t_0, (((floorf(h) * dX_46_v) * floorf(h)) * dX_46_v));
float t_2 = floorf(h) * dY_46_v;
float t_3 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, (t_2 * t_2));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = (floorf(w) * dY_46_u) / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = fma(t_0, t_0, Float32(Float32(Float32(floor(h) * dX_46_v) * floor(h)) * dX_46_v)) t_2 = Float32(floor(h) * dY_46_v) t_3 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(t_2 * t_2)) t_4 = sqrt(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(Float32(floor(w) * dY_46_u) / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \mathsf{fma}\left(t\_0, t\_0, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right)\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_2 \cdot t\_2\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_4}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f3276.0
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_1 (* (floor w) dX.u))
(t_2 (fma t_1 t_1 (* (* (* (floor h) dX.v) (floor h)) dX.v)))
(t_3 (sqrt (fmax t_2 t_0))))
(if (>= t_2 t_0) (/ t_1 t_3) (/ (* (floor w) 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 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_1 = floorf(w) * dX_46_u;
float t_2 = fmaf(t_1, t_1, (((floorf(h) * dX_46_v) * floorf(h)) * dX_46_v));
float t_3 = sqrtf(fmaxf(t_2, t_0));
float tmp;
if (t_2 >= t_0) {
tmp = t_1 / t_3;
} else {
tmp = (floorf(w) * 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 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_1 = Float32(floor(w) * dX_46_u) t_2 = fma(t_1, t_1, Float32(Float32(Float32(floor(h) * dX_46_v) * floor(h)) * dX_46_v)) t_3 = sqrt(fmax(t_2, t_0)) tmp = Float32(0.0) if (t_2 >= t_0) tmp = Float32(t_1 / t_3); else tmp = Float32(Float32(floor(w) * dY_46_u) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_0\right)}\\
\mathbf{if}\;t\_2 \geq t\_0:\\
\;\;\;\;\frac{t\_1}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_3}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3276.0
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (fma t_0 t_0 (* (* dX.v dX.v) (* (floor h) (floor h)))))
(t_2 (* (floor h) dY.v))
(t_3 (fma (* (* (floor w) (floor w)) dY.u) dY.u (* t_2 t_2)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ (* (floor w) dY.u) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = fmaf(t_0, t_0, ((dX_46_v * dX_46_v) * (floorf(h) * floorf(h))));
float t_2 = floorf(h) * dY_46_v;
float t_3 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, (t_2 * t_2));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = (floorf(w) * dY_46_u) / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = fma(t_0, t_0, Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h)))) t_2 = Float32(floor(h) * dY_46_v) t_3 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(t_2 * t_2)) t_4 = sqrt(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(Float32(floor(w) * dY_46_u) / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \mathsf{fma}\left(t\_0, t\_0, \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_2 \cdot t\_2\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_4}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f3276.0
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) (floor w)) dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (* (floor w) dX.u))
(t_3 (fma t_2 t_2 (* (* dX.v dX.v) t_0)))
(t_4 (sqrt (fmax t_3 t_1))))
(if (>= t_3 t_1) (/ t_2 t_4) (/ (* (floor w) dY.u) 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) * floorf(h);
float t_1 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = floorf(w) * dX_46_u;
float t_3 = fmaf(t_2, t_2, ((dX_46_v * dX_46_v) * t_0));
float t_4 = sqrtf(fmaxf(t_3, t_1));
float tmp;
if (t_3 >= t_1) {
tmp = t_2 / t_4;
} else {
tmp = (floorf(w) * dY_46_u) / 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) * floor(h)) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = Float32(floor(w) * dX_46_u) t_3 = fma(t_2, t_2, Float32(Float32(dX_46_v * dX_46_v) * t_0)) t_4 = sqrt(fmax(t_3, t_1)) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(t_2 / t_4); else tmp = Float32(Float32(floor(w) * dY_46_u) / t_4); 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 \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \mathsf{fma}\left(t\_2, t\_2, \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_1\right)}\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_4}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (fma (floor h) (floor h) (/ 0.0 dY.v)) (* dY.v dY.v)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (fma (* t_1 (floor w)) dY.u 0.0))
(t_4 (* (floor h) dX.v))
(t_5 (* (* dY.u dY.u) (* (floor w) (floor w))))
(t_6 (+ (* t_2 t_2) (* t_4 t_4)))
(t_7 (/ 1.0 (sqrt (fmax t_6 t_5))))
(t_8 (/ 1.0 (sqrt (fmax t_6 t_3))))
(t_9 (/ 1.0 (sqrt (fmax t_6 t_0))))
(t_10 (* (floor h) dY.v))
(t_11 (+ (* t_1 t_1) (* t_10 t_10)))
(t_12 (/ 1.0 (sqrt (fmax t_6 t_11))))
(t_13 (if (>= t_6 t_11) (* t_12 t_2) (* t_12 t_1))))
(if (<= t_13 -0.800000011920929)
(if (>= t_6 t_5) (* t_7 t_2) (* t_7 t_1))
(if (<= t_13 0.05000000074505806)
(if (>= t_6 t_0) (* t_9 t_2) (* t_9 t_1))
(if (>= t_6 t_3) (* t_8 t_2) (* t_8 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 = fmaf(floorf(h), floorf(h), (0.0f / dY_46_v)) * (dY_46_v * dY_46_v);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = fmaf((t_1 * floorf(w)), dY_46_u, 0.0f);
float t_4 = floorf(h) * dX_46_v;
float t_5 = (dY_46_u * dY_46_u) * (floorf(w) * floorf(w));
float t_6 = (t_2 * t_2) + (t_4 * t_4);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, t_5));
float t_8 = 1.0f / sqrtf(fmaxf(t_6, t_3));
float t_9 = 1.0f / sqrtf(fmaxf(t_6, t_0));
float t_10 = floorf(h) * dY_46_v;
float t_11 = (t_1 * t_1) + (t_10 * t_10);
float t_12 = 1.0f / sqrtf(fmaxf(t_6, t_11));
float tmp;
if (t_6 >= t_11) {
tmp = t_12 * t_2;
} else {
tmp = t_12 * t_1;
}
float t_13 = tmp;
float tmp_2;
if (t_13 <= -0.800000011920929f) {
float tmp_3;
if (t_6 >= t_5) {
tmp_3 = t_7 * t_2;
} else {
tmp_3 = t_7 * t_1;
}
tmp_2 = tmp_3;
} else if (t_13 <= 0.05000000074505806f) {
float tmp_4;
if (t_6 >= t_0) {
tmp_4 = t_9 * t_2;
} else {
tmp_4 = t_9 * t_1;
}
tmp_2 = tmp_4;
} else if (t_6 >= t_3) {
tmp_2 = t_8 * t_2;
} else {
tmp_2 = t_8 * t_1;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(fma(floor(h), floor(h), Float32(Float32(0.0) / dY_46_v)) * Float32(dY_46_v * dY_46_v)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = fma(Float32(t_1 * floor(w)), dY_46_u, Float32(0.0)) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_5))) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_3))) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_0))) t_10 = Float32(floor(h) * dY_46_v) t_11 = Float32(Float32(t_1 * t_1) + Float32(t_10 * t_10)) t_12 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_11))) tmp = Float32(0.0) if (t_6 >= t_11) tmp = Float32(t_12 * t_2); else tmp = Float32(t_12 * t_1); end t_13 = tmp tmp_2 = Float32(0.0) if (t_13 <= Float32(-0.800000011920929)) tmp_3 = Float32(0.0) if (t_6 >= t_5) tmp_3 = Float32(t_7 * t_2); else tmp_3 = Float32(t_7 * t_1); end tmp_2 = tmp_3; elseif (t_13 <= Float32(0.05000000074505806)) tmp_4 = Float32(0.0) if (t_6 >= t_0) tmp_4 = Float32(t_9 * t_2); else tmp_4 = Float32(t_9 * t_1); end tmp_2 = tmp_4; elseif (t_6 >= t_3) tmp_2 = Float32(t_8 * t_2); else tmp_2 = Float32(t_8 * t_1); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \frac{0}{dY.v}\right) \cdot \left(dY.v \cdot dY.v\right)\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \mathsf{fma}\left(t\_1 \cdot \left\lfloor w\right\rfloor , dY.u, 0\right)\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\\
t_6 := t\_2 \cdot t\_2 + t\_4 \cdot t\_4\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_5\right)}}\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_3\right)}}\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_0\right)}}\\
t_10 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_11 := t\_1 \cdot t\_1 + t\_10 \cdot t\_10\\
t_12 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_11\right)}}\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_11:\\
\;\;\;\;t\_12 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_12 \cdot t\_1\\
\end{array}\\
\mathbf{if}\;t\_13 \leq -0.800000011920929:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_5:\\
\;\;\;\;t\_7 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;t\_13 \leq 0.05000000074505806:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_0:\\
\;\;\;\;t\_9 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_3:\\
\;\;\;\;t\_8 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_1\\
\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))) < -0.800000012Initial program 75.8%
Applied rewrites68.6%
Applied rewrites68.6%
Applied rewrites45.6%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites42.8%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites49.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
if -0.800000012 < (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))) < 0.0500000007Initial program 75.8%
Applied rewrites68.6%
Applied rewrites68.6%
Applied rewrites45.6%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites42.8%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites49.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites58.5%
Taylor expanded in dY.v around inf
Applied rewrites48.6%
Taylor expanded in dY.v around inf
Applied rewrites51.8%
Taylor expanded in dY.v around inf
Applied rewrites62.5%
if 0.0500000007 < (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 75.8%
Applied rewrites68.6%
Applied rewrites68.6%
Applied rewrites45.6%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites42.8%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites49.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* dY.v dY.v) t_0))
(t_2 (fma (* (* (floor w) (floor w)) dY.u) dY.u t_1))
(t_3 (* (floor w) dX.u))
(t_4 (fma t_3 t_3 (* (* dX.v dX.v) t_0))))
(if (>= t_4 t_2)
(/ t_3 (sqrt (fmax t_4 t_2)))
(*
(floor w)
(/
dY.u
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* (floor h) dX.v) (floor h)) dX.v))
(fma (* (* dY.u dY.u) (floor w)) (floor w) 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) * floorf(h);
float t_1 = (dY_46_v * dY_46_v) * t_0;
float t_2 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_1);
float t_3 = floorf(w) * dX_46_u;
float t_4 = fmaf(t_3, t_3, ((dX_46_v * dX_46_v) * t_0));
float tmp;
if (t_4 >= t_2) {
tmp = t_3 / sqrtf(fmaxf(t_4, t_2));
} else {
tmp = floorf(w) * (dY_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((floorf(h) * dX_46_v) * floorf(h)) * dX_46_v)), fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), 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) * floor(h)) t_1 = Float32(Float32(dY_46_v * dY_46_v) * t_0) t_2 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_1) t_3 = Float32(floor(w) * dX_46_u) t_4 = fma(t_3, t_3, Float32(Float32(dX_46_v * dX_46_v) * t_0)) tmp = Float32(0.0) if (t_4 >= t_2) tmp = Float32(t_3 / sqrt(fmax(t_4, t_2))); else tmp = Float32(floor(w) * Float32(dY_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(floor(h) * dX_46_v) * floor(h)) * dX_46_v)), fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), t_1))))); 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(dY.v \cdot dY.v\right) \cdot t\_0\\
t_2 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_1\right)\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \mathsf{fma}\left(t\_3, t\_3, \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(t\_4, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \frac{dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_1\right)\right)}}\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
Applied rewrites76.0%
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 (* (floor w) dY.u))
(t_2 (fma (* t_1 (floor w)) dY.u 0.0))
(t_3 (* (floor w) dX.u))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (/ 1.0 (sqrt (fmax t_4 t_2)))))
(if (>= t_4 t_2) (* t_5 t_3) (* t_5 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 = fmaf((t_1 * floorf(w)), dY_46_u, 0.0f);
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = 1.0f / sqrtf(fmaxf(t_4, t_2));
float tmp;
if (t_4 >= t_2) {
tmp = t_5 * t_3;
} else {
tmp = t_5 * 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 = fma(Float32(t_1 * floor(w)), dY_46_u, Float32(0.0)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(Float32(1.0) / sqrt(fmax(t_4, t_2))) tmp = Float32(0.0) if (t_4 >= t_2) tmp = Float32(t_5 * t_3); else tmp = Float32(t_5 * t_1); end return 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 := \mathsf{fma}\left(t\_1 \cdot \left\lfloor w\right\rfloor , dY.u, 0\right)\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_2\right)}}\\
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;t\_5 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot t\_1\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites68.6%
Applied rewrites68.6%
Applied rewrites45.6%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites42.8%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites49.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
Taylor expanded in w around 0
Applied rewrites58.5%
(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 dY.u) (* (floor w) (floor w))))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (/ 1.0 (sqrt (fmax t_3 t_1)))))
(if (>= t_3 t_1) (* t_4 t_2) (* t_4 (* (floor w) 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 = floorf(h) * dX_46_v;
float t_1 = (dY_46_u * dY_46_u) * (floorf(w) * floorf(w));
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = 1.0f / sqrtf(fmaxf(t_3, t_1));
float tmp;
if (t_3 >= t_1) {
tmp = t_4 * t_2;
} else {
tmp = t_4 * (floorf(w) * 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(floor(h) * dX_46_v) t_1 = Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_1))) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(t_4 * t_2); else tmp = Float32(t_4 * Float32(floor(w) * dY_46_u)); 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 = (dY_46_u * dY_46_u) * (floor(w) * floor(w)); t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = single(1.0) / sqrt(max(t_3, t_1)); tmp = single(0.0); if (t_3 >= t_1) tmp = t_4 * t_2; else tmp = t_4 * (floor(w) * dY_46_u); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_1\right)}}\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;t\_4 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
\end{array}
\end{array}
Initial program 75.8%
Applied rewrites68.6%
Applied rewrites68.6%
Applied rewrites45.6%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites42.8%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites49.1%
Taylor expanded in dY.v around 0
+-commutativeN/A
*-commutativeN/A
pow2N/A
pow2N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
Taylor expanded in w around 0
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow2N/A
lift-*.f3258.5
Applied rewrites58.5%
herbie shell --seed 2025127
(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))))