
(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 7 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 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_0 dX.u) (floor w))))
(t_3 (* dY.v (floor h)))
(t_4 (fma (* t_1 (floor w)) dY.u (* t_3 t_3)))
(t_5 (sqrt (fmax t_2 t_4))))
(if (>= t_2 t_4) (/ t_0 t_5) (/ t_1 t_5))))
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 = dX_46_u * floorf(w);
float t_1 = dY_46_u * floorf(w);
float t_2 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_0 * dX_46_u) * floorf(w)));
float t_3 = dY_46_v * floorf(h);
float t_4 = fmaf((t_1 * floorf(w)), dY_46_u, (t_3 * t_3));
float t_5 = sqrtf(fmaxf(t_2, t_4));
float tmp;
if (t_2 >= t_4) {
tmp = t_0 / t_5;
} else {
tmp = t_1 / t_5;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) t_1 = Float32(dY_46_u * floor(w)) t_2 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_0 * dX_46_u) * floor(w))) t_3 = Float32(dY_46_v * floor(h)) t_4 = fma(Float32(t_1 * floor(w)), dY_46_u, Float32(t_3 * t_3)) t_5 = sqrt(fmax(t_2, t_4)) tmp = Float32(0.0) if (t_2 >= t_4) tmp = Float32(t_0 / t_5); else tmp = Float32(t_1 / t_5); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_0 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_1 \cdot \left\lfloor w\right\rfloor , dY.u, t\_3 \cdot t\_3\right)\\
t_5 := \sqrt{\mathsf{max}\left(t\_2, t\_4\right)}\\
\mathbf{if}\;t\_2 \geq t\_4:\\
\;\;\;\;\frac{t\_0}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_5}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2
(fma
(* t_1 dY.u)
(floor w)
(* (* (floor h) (floor h)) (* dY.v dY.v))))
(t_3
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_0 dX.u) (floor w))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ t_0 t_4) (/ t_1 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 = dX_46_u * floorf(w);
float t_1 = dY_46_u * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), floorf(w), ((floorf(h) * floorf(h)) * (dY_46_v * dY_46_v)));
float t_3 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_0 * dX_46_u) * floorf(w)));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_0 / t_4;
} else {
tmp = t_1 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) t_1 = Float32(dY_46_u * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(floor(h) * floor(h)) * Float32(dY_46_v * dY_46_v))) t_3 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_0 * dX_46_u) * floor(w))) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_0 / t_4); else tmp = Float32(t_1 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dY.v \cdot dY.v\right)\right)\\
t_3 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_0 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lower-*.f32N/A
lower-*.f3276.2
Applied rewrites76.2%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lower-*.f32N/A
lower-*.f3276.2
Applied rewrites76.2%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lower-*.f32N/A
lower-*.f3276.2
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* dY.u dY.u) t_1 (* (* (* (floor h) (floor h)) dY.v) dY.v)))
(t_3 (fma (* dX.u dX.u) t_1 (* t_0 t_0)))
(t_4 (/ (floor w) (sqrt (fmax t_3 t_2)))))
(if (>= t_3 t_2) (* t_4 dX.u) (* t_4 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 = floorf(w) * floorf(w);
float t_2 = fmaf((dY_46_u * dY_46_u), t_1, (((floorf(h) * floorf(h)) * dY_46_v) * dY_46_v));
float t_3 = fmaf((dX_46_u * dX_46_u), t_1, (t_0 * t_0));
float t_4 = floorf(w) / sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_4 * dX_46_u;
} else {
tmp = t_4 * 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(floor(w) * floor(w)) t_2 = fma(Float32(dY_46_u * dY_46_u), t_1, Float32(Float32(Float32(floor(h) * floor(h)) * dY_46_v) * dY_46_v)) t_3 = fma(Float32(dX_46_u * dX_46_u), t_1, Float32(t_0 * t_0)) t_4 = Float32(floor(w) / sqrt(fmax(t_3, t_2))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_4 * dX_46_u); else tmp = Float32(t_4 * dY_46_u); 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 \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(dY.u \cdot dY.u, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot dY.v\right)\\
t_3 := \mathsf{fma}\left(dX.u \cdot dX.u, t\_1, t\_0 \cdot t\_0\right)\\
t_4 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;t\_4 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.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) dX.v))
(t_1 (* dY.v (floor h)))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* dY.u dY.u) t_2 (* t_1 t_1)))
(t_4 (fma (* dX.u dX.u) t_2 (* t_0 t_0)))
(t_5 (/ (floor w) (sqrt (fmax t_4 t_3)))))
(if (>= t_4 t_3) (* t_5 dX.u) (* t_5 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_v * floorf(h);
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((dY_46_u * dY_46_u), t_2, (t_1 * t_1));
float t_4 = fmaf((dX_46_u * dX_46_u), t_2, (t_0 * t_0));
float t_5 = floorf(w) / sqrtf(fmaxf(t_4, t_3));
float tmp;
if (t_4 >= t_3) {
tmp = t_5 * dX_46_u;
} else {
tmp = t_5 * 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(dY_46_v * floor(h)) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(dY_46_u * dY_46_u), t_2, Float32(t_1 * t_1)) t_4 = fma(Float32(dX_46_u * dX_46_u), t_2, Float32(t_0 * t_0)) t_5 = Float32(floor(w) / sqrt(fmax(t_4, t_3))) tmp = Float32(0.0) if (t_4 >= t_3) tmp = Float32(t_5 * dX_46_u); else tmp = Float32(t_5 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\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 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(dY.u \cdot dY.u, t\_2, t\_1 \cdot t\_1\right)\\
t_4 := \mathsf{fma}\left(dX.u \cdot dX.u, t\_2, t\_0 \cdot t\_0\right)\\
t_5 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_4, t\_3\right)}}\\
\mathbf{if}\;t\_4 \geq t\_3:\\
\;\;\;\;t\_5 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.0%
(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 (* (floor h) (floor h)))
(t_3 (* (floor w) (floor w)))
(t_4
(/
(floor w)
(sqrt
(fmax
(fma (* t_3 dX.u) dX.u (* (* dX.v dX.v) t_2))
(* (fma dY.u (floor w) t_1) (- (* dY.u (floor w)) t_1)))))))
(if (>=
(fma (* dX.u dX.u) t_3 (* t_0 t_0))
(fma (* dY.u dY.u) t_3 (* (* t_2 dY.v) dY.v)))
(* t_4 dX.u)
(* t_4 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_v * floorf(h);
float t_2 = floorf(h) * floorf(h);
float t_3 = floorf(w) * floorf(w);
float t_4 = floorf(w) / sqrtf(fmaxf(fmaf((t_3 * dX_46_u), dX_46_u, ((dX_46_v * dX_46_v) * t_2)), (fmaf(dY_46_u, floorf(w), t_1) * ((dY_46_u * floorf(w)) - t_1))));
float tmp;
if (fmaf((dX_46_u * dX_46_u), t_3, (t_0 * t_0)) >= fmaf((dY_46_u * dY_46_u), t_3, ((t_2 * dY_46_v) * dY_46_v))) {
tmp = t_4 * dX_46_u;
} else {
tmp = t_4 * 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(dY_46_v * floor(h)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(floor(w) / sqrt(fmax(fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * t_2)), Float32(fma(dY_46_u, floor(w), t_1) * Float32(Float32(dY_46_u * floor(w)) - t_1))))) tmp = Float32(0.0) if (fma(Float32(dX_46_u * dX_46_u), t_3, Float32(t_0 * t_0)) >= fma(Float32(dY_46_u * dY_46_u), t_3, Float32(Float32(t_2 * dY_46_v) * dY_46_v))) tmp = Float32(t_4 * dX_46_u); else tmp = Float32(t_4 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\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 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, \left(dX.v \cdot dX.v\right) \cdot t\_2\right), \mathsf{fma}\left(dY.u, \left\lfloor w\right\rfloor , t\_1\right) \cdot \left(dY.u \cdot \left\lfloor w\right\rfloor - t\_1\right)\right)}}\\
\mathbf{if}\;\mathsf{fma}\left(dX.u \cdot dX.u, t\_3, t\_0 \cdot t\_0\right) \geq \mathsf{fma}\left(dY.u \cdot dY.u, t\_3, \left(t\_2 \cdot dY.v\right) \cdot dY.v\right):\\
\;\;\;\;t\_4 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.0
Applied rewrites76.0%
Applied rewrites76.0%
Applied rewrites59.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1
(fma
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* dX.v dX.v) (* (floor h) (floor h)))))
(t_2 (* dY.u (floor w)))
(t_3 (* (fma dY.u (floor w) t_0) (- t_2 t_0)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ (* (floor w) dX.u) t_4) (/ t_2 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 = dY_46_v * floorf(h);
float t_1 = fmaf(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, ((dX_46_v * dX_46_v) * (floorf(h) * floorf(h))));
float t_2 = dY_46_u * floorf(w);
float t_3 = fmaf(dY_46_u, floorf(w), t_0) * (t_2 - t_0);
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = (floorf(w) * dX_46_u) / t_4;
} else {
tmp = t_2 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h)))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(fma(dY_46_u, floor(w), t_0) * Float32(t_2 - t_0)) t_4 = sqrt(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(floor(w) * dX_46_u) / t_4); else tmp = Float32(t_2 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \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 dX.u, dX.u, \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(dY.u, \left\lfloor w\right\rfloor , t\_0\right) \cdot \left(t\_2 - t\_0\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.2%
lift-fma.f32N/A
lift-*.f32N/A
*-commutativeN/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
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites76.3%
Applied rewrites76.2%
Applied rewrites58.1%
(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)))
(t_2 (fma (* (* (floor h) (floor h)) dX.v) dX.v (* t_0 t_0)))
(t_3 (/ 1.0 (sqrt (fmax t_2 (* dY.u (* (fabs t_1) (floor w))))))))
(if (>=
t_2
(fma (* t_1 dY.u) (floor w) (* (* (* dY.v (floor h)) dY.v) (floor h))))
(* t_3 t_0)
(* t_3 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(w) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = fmaf(((floorf(h) * floorf(h)) * dX_46_v), dX_46_v, (t_0 * t_0));
float t_3 = 1.0f / sqrtf(fmaxf(t_2, (dY_46_u * (fabsf(t_1) * floorf(w)))));
float tmp;
if (t_2 >= fmaf((t_1 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))) {
tmp = t_3 * t_0;
} else {
tmp = t_3 * 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(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = fma(Float32(Float32(floor(h) * floor(h)) * dX_46_v), dX_46_v, Float32(t_0 * t_0)) t_3 = Float32(Float32(1.0) / sqrt(fmax(t_2, Float32(dY_46_u * Float32(abs(t_1) * floor(w)))))) tmp = Float32(0.0) if (t_2 >= fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))) tmp = Float32(t_3 * t_0); else tmp = Float32(t_3 * t_1); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, dX.v, t\_0 \cdot t\_0\right)\\
t_3 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2, dY.u \cdot \left(\left|t\_1\right| \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\mathbf{if}\;t\_2 \geq \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right):\\
\;\;\;\;t\_3 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot t\_1\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites73.3%
Applied rewrites70.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.5
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
lower-fabs.f3270.5
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.5
Applied rewrites70.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.5
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
lower-fabs.f3262.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.9
Applied rewrites62.9%
Applied rewrites67.6%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3267.6
Applied rewrites67.6%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-floor.f32N/A
lower-floor.f3251.8
Applied rewrites51.8%
herbie shell --seed 2025140
(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))))