
(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 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_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) dX.v))
(t_1 (fma (* (* (floor w) (floor w)) (- dX.u)) (- dX.u) (* t_0 t_0)))
(t_2 (* dY.v (floor h)))
(t_3
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* t_2 dY.v) (floor h))))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ (* dX.v (floor h)) 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 = floorf(h) * dX_46_v;
float t_1 = fmaf(((floorf(w) * floorf(w)) * -dX_46_u), -dX_46_u, (t_0 * t_0));
float t_2 = dY_46_v * floorf(h);
float t_3 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((t_2 * dY_46_v) * floorf(h)));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = (dX_46_v * floorf(h)) / 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(floor(h) * dX_46_v) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * Float32(-dX_46_u)), Float32(-dX_46_u), Float32(t_0 * t_0)) t_2 = Float32(dY_46_v * floor(h)) t_3 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))) t_4 = sqrt(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(dX_46_v * floor(h)) / t_4); else tmp = Float32(t_2 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(-dX.u\right), -dX.u, t\_0 \cdot t\_0\right)\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{dX.v \cdot \left\lfloor h\right\rfloor }{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1
(fma
(* t_0 dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (* dY.v (floor h)))
(t_3
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* t_2 dY.v) (floor h))))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 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 = dX_46_v * floorf(h);
float t_1 = fmaf((t_0 * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = dY_46_v * floorf(h);
float t_3 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((t_2 * dY_46_v) * floorf(h)));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / 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(dX_46_v * floor(h)) t_1 = fma(Float32(t_0 * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(dY_46_v * floor(h)) t_3 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))) 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(t_2 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \mathsf{fma}\left(t\_0 \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \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{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1
(fma
(* t_0 dX.v)
(floor h)
(* (* (* dX.u dX.u) (floor w)) (floor w))))
(t_2 (* dY.v (floor h)))
(t_3
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* t_2 dY.v) (floor h))))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 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 = dX_46_v * floorf(h);
float t_1 = fmaf((t_0 * dX_46_v), floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)));
float t_2 = dY_46_v * floorf(h);
float t_3 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), ((t_2 * dY_46_v) * floorf(h)));
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / 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(dX_46_v * floor(h)) t_1 = fma(Float32(t_0 * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))) t_2 = Float32(dY_46_v * floor(h)) t_3 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(t_2 * dY_46_v) * floor(h))) 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(t_2 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \mathsf{fma}\left(t\_0 \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \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{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.4%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* dY.u dY.u)
(* (floor w) (floor w))
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(t_2 (/ (floor h) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.v) (* t_2 dY.v))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)));
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(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) 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(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_2 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
add-flipN/A
sub-flipN/A
Applied rewrites76.4%
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (fma (* dY.u dY.u) (* (floor w) (floor w)) (* t_0 t_0)))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor h) dX.v))
(t_4 (fma t_2 t_2 (* t_3 t_3)))
(t_5 (/ (floor h) (sqrt (fmax t_4 t_1)))))
(if (>= t_4 t_1) (* t_5 dX.v) (* t_5 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_v * floorf(h);
float t_1 = fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (t_0 * t_0));
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(h) * dX_46_v;
float t_4 = fmaf(t_2, t_2, (t_3 * t_3));
float t_5 = floorf(h) / sqrtf(fmaxf(t_4, t_1));
float tmp;
if (t_4 >= t_1) {
tmp = t_5 * dX_46_v;
} else {
tmp = t_5 * 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_v * floor(h)) t_1 = fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), Float32(t_0 * t_0)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(h) * dX_46_v) t_4 = fma(t_2, t_2, Float32(t_3 * t_3)) t_5 = Float32(floor(h) / sqrt(fmax(t_4, t_1))) tmp = Float32(0.0) if (t_4 >= t_1) tmp = Float32(t_5 * dX_46_v); else tmp = Float32(t_5 * dY_46_v); 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(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_0 \cdot t\_0\right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \mathsf{fma}\left(t\_2, t\_2, t\_3 \cdot t\_3\right)\\
t_5 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_4, t\_1\right)}}\\
\mathbf{if}\;t\_4 \geq t\_1:\\
\;\;\;\;t\_5 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.4%
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (* (* dY.u dY.u) (* (floor w) (floor w))) 1.0))
(t_2
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* t_0 dX.v) (floor h))))
(t_3 (* dY.u (floor w))))
(if (>= t_2 t_1)
(/ t_0 (sqrt (fmax t_2 t_1)))
(*
(floor h)
(/
dY.v
(sqrt
(fmax
(fma
(* (* (floor w) dX.u) (floor w))
dX.u
(* (* dX.v dX.v) (* (floor h) (floor h))))
(* (* t_3 t_3) 1.0))))))))
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_v * floorf(h);
float t_1 = ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))) * 1.0f;
float t_2 = fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), ((t_0 * dX_46_v) * floorf(h)));
float t_3 = dY_46_u * floorf(w);
float tmp;
if (t_2 >= t_1) {
tmp = t_0 / sqrtf(fmaxf(t_2, t_1));
} else {
tmp = floorf(h) * (dY_46_v / sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, ((dX_46_v * dX_46_v) * (floorf(h) * floorf(h)))), ((t_3 * t_3) * 1.0f))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) * Float32(1.0)) t_2 = fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), Float32(Float32(t_0 * dX_46_v) * floor(h))) t_3 = Float32(dY_46_u * floor(w)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_0 / sqrt(fmax(t_2, t_1))); else tmp = Float32(floor(h) * Float32(dY_46_v / sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h)))), Float32(Float32(t_3 * t_3) * Float32(1.0)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right) \cdot 1\\
t_2 := \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , \left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right), \left(t\_3 \cdot t\_3\right) \cdot 1\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites73.7%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites73.7%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites68.1%
Taylor expanded in dY.u around inf
Applied rewrites62.0%
Taylor expanded in dY.u around inf
Applied rewrites67.9%
Taylor expanded in dY.u around inf
Applied rewrites62.6%
Applied rewrites62.7%
Applied rewrites62.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (* (* dY.u dY.u) (* (floor w) (floor w))) 1.0))
(t_2
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* t_0 dX.v) (floor h))))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1) (/ t_0 t_3) (/ (* dY.v (floor h)) 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 = dX_46_v * floorf(h);
float t_1 = ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w))) * 1.0f;
float t_2 = fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), ((t_0 * dX_46_v) * floorf(h)));
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_0 / t_3;
} else {
tmp = (dY_46_v * floorf(h)) / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w))) * Float32(1.0)) t_2 = fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(t_0 * dX_46_v) * floor(h))) t_3 = sqrt(fmax(t_2, t_1)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_0 / t_3); else tmp = Float32(Float32(dY_46_v * floor(h)) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right)\right) \cdot 1\\
t_2 := \mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{t\_0}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v \cdot \left\lfloor h\right\rfloor }{t\_3}\\
\end{array}
\end{array}
Initial program 76.2%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites73.7%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites73.7%
lift-+.f32N/A
sum-to-multN/A
lower-*.f32N/A
Applied rewrites68.1%
Taylor expanded in dY.u around inf
Applied rewrites62.0%
Taylor expanded in dY.u around inf
Applied rewrites67.9%
Taylor expanded in dY.u around inf
Applied rewrites62.6%
Applied rewrites62.7%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3262.7
Applied rewrites62.7%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3262.7
Applied rewrites62.7%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3262.7
Applied rewrites62.7%
herbie shell --seed 2025156
(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))))