
(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 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_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
(fma
(* dY.u dY.u)
(* (floor w) (floor w))
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_2
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_0 dX.u) (floor w))))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1) (/ t_0 t_3) (/ (* dY.u (floor w)) 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_u * floorf(w);
float t_1 = fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_2 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_0 * dX_46_u) * floorf(w)));
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_u * floorf(w)) / 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_u * floor(w)) t_1 = 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_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 = 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_u * floor(w)) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \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_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 := \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.u \cdot \left\lfloor w\right\rfloor }{t\_3}\\
\end{array}
\end{array}
Initial program 76.4%
Applied rewrites76.5%
Applied rewrites76.4%
Applied rewrites76.5%
(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.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = 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_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = 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))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \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.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)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.4%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/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
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u dX.u) (floor w)) (floor w))))
(t_1 (* dY.u (floor w)))
(t_2 (fma (* (floor h) (floor h)) (* dY.v dY.v) (* t_1 t_1)))
(t_3 (/ (floor w) (sqrt (fmax t_0 t_2)))))
(if (>= t_0 t_2) (* t_3 dX.u) (* t_3 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 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)));
float t_1 = dY_46_u * floorf(w);
float t_2 = fmaf((floorf(h) * floorf(h)), (dY_46_v * dY_46_v), (t_1 * t_1));
float t_3 = floorf(w) / sqrtf(fmaxf(t_0, t_2));
float tmp;
if (t_0 >= t_2) {
tmp = t_3 * dX_46_u;
} else {
tmp = t_3 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))) t_1 = Float32(dY_46_u * floor(w)) t_2 = fma(Float32(floor(h) * floor(h)), Float32(dY_46_v * dY_46_v), Float32(t_1 * t_1)) t_3 = Float32(floor(w) / sqrt(fmax(t_0, t_2))) tmp = Float32(0.0) if (t_0 >= t_2) tmp = Float32(t_3 * dX_46_u); else tmp = Float32(t_3 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \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 dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right)\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , dY.v \cdot dY.v, t\_1 \cdot t\_1\right)\\
t_3 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_0, t\_2\right)}}\\
\mathbf{if}\;t\_0 \geq t\_2:\\
\;\;\;\;t\_3 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.4%
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (* dX.v (floor h)) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_4 (* (floor w) dX.u))
(t_5 (* dY.u (floor w)))
(t_6 (* dY.u (* (fabs t_5) (floor w))))
(t_7 (+ (* t_4 t_4) (* t_0 t_0)))
(t_8 (* (floor h) dY.v))
(t_9 (+ (* t_2 t_2) (* t_8 t_8)))
(t_10 (/ 1.0 (sqrt (fmax t_7 t_9))))
(t_11 (* dX.u (floor w)))
(t_12 (fma t_1 (floor h) (* (* t_11 dX.u) (floor w))))
(t_13 (fma t_1 (floor h) (* (* (* dX.u dX.u) (floor w)) (floor w))))
(t_14 (/ (floor w) (sqrt (fmax t_13 t_6)))))
(if (<= (if (>= t_7 t_9) (* t_10 t_4) (* t_10 t_2)) 0.9999979734420776)
(if (>= t_12 t_3)
(/ t_11 (sqrt (fmax t_12 t_3)))
(/
t_5
(sqrt (fmax t_12 (fma (* dY.u dY.u) (* (floor w) (floor w)) t_3)))))
(if (>= t_13 t_6) (* t_14 dX.u) (* t_14 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 = (dX_46_v * floorf(h)) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_4 = floorf(w) * dX_46_u;
float t_5 = dY_46_u * floorf(w);
float t_6 = dY_46_u * (fabsf(t_5) * floorf(w));
float t_7 = (t_4 * t_4) + (t_0 * t_0);
float t_8 = floorf(h) * dY_46_v;
float t_9 = (t_2 * t_2) + (t_8 * t_8);
float t_10 = 1.0f / sqrtf(fmaxf(t_7, t_9));
float t_11 = dX_46_u * floorf(w);
float t_12 = fmaf(t_1, floorf(h), ((t_11 * dX_46_u) * floorf(w)));
float t_13 = fmaf(t_1, floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)));
float t_14 = floorf(w) / sqrtf(fmaxf(t_13, t_6));
float tmp;
if (t_7 >= t_9) {
tmp = t_10 * t_4;
} else {
tmp = t_10 * t_2;
}
float tmp_2;
if (tmp <= 0.9999979734420776f) {
float tmp_3;
if (t_12 >= t_3) {
tmp_3 = t_11 / sqrtf(fmaxf(t_12, t_3));
} else {
tmp_3 = t_5 / sqrtf(fmaxf(t_12, fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), t_3)));
}
tmp_2 = tmp_3;
} else if (t_13 >= t_6) {
tmp_2 = t_14 * dX_46_u;
} else {
tmp_2 = t_14 * dY_46_u;
}
return tmp_2;
}
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(dX_46_v * floor(h)) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(dY_46_u * floor(w)) t_6 = Float32(dY_46_u * Float32(abs(t_5) * floor(w))) t_7 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(Float32(t_2 * t_2) + Float32(t_8 * t_8)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_9))) t_11 = Float32(dX_46_u * floor(w)) t_12 = fma(t_1, floor(h), Float32(Float32(t_11 * dX_46_u) * floor(w))) t_13 = fma(t_1, floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))) t_14 = Float32(floor(w) / sqrt(fmax(t_13, t_6))) tmp = Float32(0.0) if (t_7 >= t_9) tmp = Float32(t_10 * t_4); else tmp = Float32(t_10 * t_2); end tmp_2 = Float32(0.0) if (tmp <= Float32(0.9999979734420776)) tmp_3 = Float32(0.0) if (t_12 >= t_3) tmp_3 = Float32(t_11 / sqrt(fmax(t_12, t_3))); else tmp_3 = Float32(t_5 / sqrt(fmax(t_12, fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), t_3)))); end tmp_2 = tmp_3; elseif (t_13 >= t_6) tmp_2 = Float32(t_14 * dX_46_u); else tmp_2 = Float32(t_14 * dY_46_u); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_6 := dY.u \cdot \left(\left|t\_5\right| \cdot \left\lfloor w\right\rfloor \right)\\
t_7 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_2 \cdot t\_2 + t\_8 \cdot t\_8\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_9\right)}}\\
t_11 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_12 := \mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , \left(t\_11 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_13 := \mathsf{fma}\left(t\_1, \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_14 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_13, t\_6\right)}}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_9:\\
\;\;\;\;t\_10 \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_2\\
\end{array} \leq 0.9999979734420776:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_12 \geq t\_3:\\
\;\;\;\;\frac{t\_11}{\sqrt{\mathsf{max}\left(t\_12, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(t\_12, \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_3\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_13 \geq t\_6:\\
\;\;\;\;t\_14 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_14 \cdot dY.u\\
\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))) < 0.999997973Initial program 76.4%
Applied rewrites76.5%
Applied rewrites67.8%
Applied rewrites67.6%
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.f3254.0
Applied rewrites54.0%
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.f3257.8
Applied rewrites57.8%
Applied rewrites68.8%
if 0.999997973 < (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 76.4%
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.3
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.f3268.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.6
Applied rewrites68.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.6
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.f3270.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.7
Applied rewrites70.7%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.7
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.f3270.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.4
Applied rewrites70.4%
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.f3256.4
Applied rewrites56.4%
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.f3256.1
Applied rewrites56.1%
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.f3253.2
Applied rewrites53.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (* (fabs (* dY.u (floor w))) (floor w))))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u dX.u) (floor w)) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = dY_46_u * (fabsf((dY_46_u * floorf(w))) * floorf(w));
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * 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(dY_46_u * Float32(abs(Float32(dY_46_u * floor(w))) * floor(w))) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\\
t_1 := \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 dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.4%
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.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-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.3
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.f3268.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.6
Applied rewrites68.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3268.6
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.f3270.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.7
Applied rewrites70.7%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.7
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.f3270.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3270.4
Applied rewrites70.4%
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.f3256.4
Applied rewrites56.4%
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.f3256.1
Applied rewrites56.1%
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.f3253.2
Applied rewrites53.2%
herbie shell --seed 2025162
(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))))