
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
Herbie found 3 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (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)) (floor w))
dY.u
(* (* t_2 dY.v) (floor h))))
(t_4 (sqrt (fmax t_3 t_1))))
(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)) * floorf(w)), dY_46_u, ((t_2 * dY_46_v) * floorf(h)));
float t_4 = sqrtf(fmaxf(t_3, t_1));
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(floor(w) * floor(w)), Float32(dX_46_u * 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)) * floor(w)), dY_46_u, Float32(Float32(t_2 * dY_46_v) * floor(h))) t_4 = sqrt(fmax(t_3, t_1)) 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}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , dX.u \cdot 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 \left\lfloor w\right\rfloor , dY.u, \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_1\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}
Initial program 76.6%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
lift-fma.f32N/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites76.8%
lift-fma.f32N/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites76.8%
lift-fma.f32N/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites76.8%
(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_3 t_1))))
(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_3, t_1));
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(floor(w) * floor(w)), Float32(dX_46_u * 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_3, t_1)) 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}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , dX.u \cdot 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\_3, t\_1\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}
Initial program 76.6%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dX.u (floor w)))
(t_2 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_3 (fma (* (* dY.u (floor w)) dY.u) (floor w) t_2))
(t_4 (* (floor w) (floor w)))
(t_5 (fma t_4 (* dX.u dX.u) (* t_0 t_0)))
(t_6 (* dX.v (floor h))))
(if (>= t_5 t_3)
(/ t_6 (sqrt (fmax t_3 t_5)))
(*
(/
dY.v
(sqrt
(fmax
(fma (* dY.u dY.u) t_4 t_2)
(fma t_1 t_1 (* (* t_6 dX.v) (floor h))))))
(floor h)))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = dX_46_u * floorf(w);
float t_2 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_3 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), t_2);
float t_4 = floorf(w) * floorf(w);
float t_5 = fmaf(t_4, (dX_46_u * dX_46_u), (t_0 * t_0));
float t_6 = dX_46_v * floorf(h);
float tmp;
if (t_5 >= t_3) {
tmp = t_6 / sqrtf(fmaxf(t_3, t_5));
} else {
tmp = (dY_46_v / sqrtf(fmaxf(fmaf((dY_46_u * dY_46_u), t_4, t_2), fmaf(t_1, t_1, ((t_6 * dX_46_v) * floorf(h)))))) * floorf(h);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(dX_46_u * floor(w)) t_2 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_3 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), t_2) t_4 = Float32(floor(w) * floor(w)) t_5 = fma(t_4, Float32(dX_46_u * dX_46_u), Float32(t_0 * t_0)) t_6 = Float32(dX_46_v * floor(h)) tmp = Float32(0.0) if (t_5 >= t_3) tmp = Float32(t_6 / sqrt(fmax(t_3, t_5))); else tmp = Float32(Float32(dY_46_v / sqrt(fmax(fma(Float32(dY_46_u * dY_46_u), t_4, t_2), fma(t_1, t_1, Float32(Float32(t_6 * dX_46_v) * floor(h)))))) * floor(h)); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \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 , t\_2\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_5 := \mathsf{fma}\left(t\_4, dX.u \cdot dX.u, t\_0 \cdot t\_0\right)\\
t_6 := dX.v \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;t\_5 \geq t\_3:\\
\;\;\;\;\frac{t\_6}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.u \cdot dY.u, t\_4, t\_2\right), \mathsf{fma}\left(t\_1, t\_1, \left(t\_6 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\end{array}
Initial program 76.6%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.8%
Applied rewrites76.7%
herbie shell --seed 2025212
(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))))