
(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 6 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) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(/
(* (floor h) dX.v)
(sqrt
(fmax
(- (* (* (* dX.v (floor h)) dX.v) (floor h)) (* (* (- dX.u) dX.u) t_1))
t_2)))
(/ (* (floor h) dY.v) (sqrt (fmax t_3 t_2))))))
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 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) - ((-dX_46_u * dX_46_u) * t_1)), t_2));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, t_2));
}
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(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(Float32(Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) - Float32(Float32(Float32(-dX_46_u) * dX_46_u) * t_1)), t_2))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, t_2))); 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\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor - \left(\left(-dX.u\right) \cdot dX.u\right) \cdot t\_1, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 75.7%
Applied rewrites75.7%
lift-fma.f32N/A
+-commutativeN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lower-*.f32N/A
*-commutativeN/A
associate-*r*N/A
pow2N/A
fp-cancel-sign-sub-invN/A
distribute-lft-neg-inN/A
lower--.f32N/A
Applied rewrites75.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(*
(/
dX.v
(sqrt
(fmax
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(floor h))
(/ (* (floor h) dY.v) (sqrt (fmax t_3 t_2))))))
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 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = (dX_46_v / sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))) * floorf(h);
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, t_2));
}
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(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(dX_46_v / sqrt(fmax(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))), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))) * floor(h)); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, t_2))); 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\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\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), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \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)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 75.7%
Applied rewrites75.7%
Applied rewrites75.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(*
dX.v
(/
(floor h)
(sqrt
(fmax
(fma
(* (* dX.u (floor w)) dX.u)
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(/ (* (floor h) dY.v) (sqrt (fmax t_3 t_2))))))
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 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = dX_46_v * (floorf(h) / sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h))))));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_3, t_2));
}
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(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(dX_46_v * Float32(floor(h) / sqrt(fmax(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))), fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_3, t_2))); 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\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;dX.v \cdot \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\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), \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \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)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 75.7%
Applied rewrites75.7%
Applied rewrites75.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (* t_0 dX.v) (floor h)))
(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 (* (floor w) (floor w)))
(t_5 (sqrt (fmax (fma (* (* dX.u (floor w)) dX.u) (floor w) t_1) t_3)))
(t_6 (* (floor h) (floor h)))
(t_7 (fma (* t_4 dY.u) dY.u (* (* dY.v dY.v) t_6))))
(if (<= dX.u 6500.0)
(if (>= t_1 t_3) (/ t_0 t_5) (/ t_2 t_5))
(if (>= (* (* (* (floor w) dX.u) dX.u) (floor w)) t_7)
(*
(/
dX.v
(sqrt
(fmax
(fma (* t_0 (floor h)) dX.v (* (* dX.u dX.u) t_4))
(fma (* dY.v dY.v) t_6 (* (* dY.u dY.u) t_4)))))
(floor h))
(/
(* (floor h) dY.v)
(sqrt (fmax (fma (* t_4 dX.u) dX.u (* t_6 (* dX.v dX.v))) t_7)))))))
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 = (t_0 * dX_46_v) * floorf(h);
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 = floorf(w) * floorf(w);
float t_5 = sqrtf(fmaxf(fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), t_1), t_3));
float t_6 = floorf(h) * floorf(h);
float t_7 = fmaf((t_4 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_6));
float tmp_1;
if (dX_46_u <= 6500.0f) {
float tmp_2;
if (t_1 >= t_3) {
tmp_2 = t_0 / t_5;
} else {
tmp_2 = t_2 / t_5;
}
tmp_1 = tmp_2;
} else if ((((floorf(w) * dX_46_u) * dX_46_u) * floorf(w)) >= t_7) {
tmp_1 = (dX_46_v / sqrtf(fmaxf(fmaf((t_0 * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_4)), fmaf((dY_46_v * dY_46_v), t_6, ((dY_46_u * dY_46_u) * t_4))))) * floorf(h);
} else {
tmp_1 = (floorf(h) * dY_46_v) / sqrtf(fmaxf(fmaf((t_4 * dX_46_u), dX_46_u, (t_6 * (dX_46_v * dX_46_v))), t_7));
}
return tmp_1;
}
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(t_0 * dX_46_v) * floor(h)) 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 = Float32(floor(w) * floor(w)) t_5 = sqrt(fmax(fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), t_1), t_3)) t_6 = Float32(floor(h) * floor(h)) t_7 = fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_6)) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(6500.0)) tmp_2 = Float32(0.0) if (t_1 >= t_3) tmp_2 = Float32(t_0 / t_5); else tmp_2 = Float32(t_2 / t_5); end tmp_1 = tmp_2; elseif (Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w)) >= t_7) tmp_1 = Float32(Float32(dX_46_v / sqrt(fmax(fma(Float32(t_0 * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_4)), fma(Float32(dY_46_v * dY_46_v), t_6, Float32(Float32(dY_46_u * dY_46_u) * t_4))))) * floor(h)); else tmp_1 = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(t_6 * Float32(dX_46_v * dX_46_v))), t_7))); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
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 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_5 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , t\_1\right), t\_3\right)}\\
t_6 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_7 := \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_6\right)\\
\mathbf{if}\;dX.u \leq 6500:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_5}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_5}\\
\end{array}\\
\mathbf{elif}\;\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq t\_7:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_4\right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_6, \left(dY.u \cdot dY.u\right) \cdot t\_4\right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.u, dX.u, t\_6 \cdot \left(dX.v \cdot dX.v\right)\right), t\_7\right)}}\\
\end{array}
\end{array}
if dX.u < 6500Initial program 77.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3269.9
Applied rewrites69.9%
Applied rewrites70.1%
if 6500 < dX.u Initial program 67.4%
Applied rewrites67.5%
lift-fma.f32N/A
+-commutativeN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lower-*.f32N/A
*-commutativeN/A
associate-*r*N/A
pow2N/A
fp-cancel-sign-sub-invN/A
distribute-lft-neg-inN/A
lower--.f32N/A
Applied rewrites67.6%
Applied rewrites67.6%
Taylor expanded in dX.u around inf
Applied rewrites65.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* (* t_0 dY.v) (floor h)))
(t_2 (* dX.v (floor h)))
(t_3 (* (* t_2 dX.v) (floor h)))
(t_4 (fma (* (* dX.u (floor w)) dX.u) (floor w) t_3))
(t_5 (fma (* (* dY.u (floor w)) dY.u) (floor w) t_1)))
(if (>= t_3 t_5)
(/ t_2 (sqrt (fmax t_4 t_5)))
(/
t_0
(sqrt (fmax t_4 (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 = dY_46_v * floorf(h);
float t_1 = (t_0 * dY_46_v) * floorf(h);
float t_2 = dX_46_v * floorf(h);
float t_3 = (t_2 * dX_46_v) * floorf(h);
float t_4 = fmaf(((dX_46_u * floorf(w)) * dX_46_u), floorf(w), t_3);
float t_5 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), t_1);
float tmp;
if (t_3 >= t_5) {
tmp = t_2 / sqrtf(fmaxf(t_4, t_5));
} else {
tmp = t_0 / sqrtf(fmaxf(t_4, 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(dY_46_v * floor(h)) t_1 = Float32(Float32(t_0 * dY_46_v) * floor(h)) t_2 = Float32(dX_46_v * floor(h)) t_3 = Float32(Float32(t_2 * dX_46_v) * floor(h)) t_4 = fma(Float32(Float32(dX_46_u * floor(w)) * dX_46_u), floor(w), t_3) t_5 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), t_1) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_2 / sqrt(fmax(t_4, t_5))); else tmp = Float32(t_0 / sqrt(fmax(t_4, 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 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(t\_0 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(t\_2 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, \left\lfloor w\right\rfloor , t\_3\right)\\
t_5 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , t\_1\right)\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_4, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_4, \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 75.7%
Taylor expanded in dX.u around 0
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.9
Applied rewrites64.9%
Applied rewrites65.0%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f3265.0
Applied rewrites65.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* dX.v dX.v) (* (floor h) (floor h))))
(t_1
(fma
(* (* (floor w) dY.u) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_2
(/
(floor h)
(sqrt (fmax (fma (* (* (floor w) dX.u) dX.u) (floor w) t_0) t_1)))))
(if (>= t_0 t_1) (* 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 = (dX_46_v * dX_46_v) * (floorf(h) * floorf(h));
float t_1 = fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_2 = floorf(h) / sqrtf(fmaxf(fmaf(((floorf(w) * dX_46_u) * dX_46_u), floorf(w), t_0), t_1));
float tmp;
if (t_0 >= t_1) {
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 = Float32(Float32(dX_46_v * dX_46_v) * Float32(floor(h) * floor(h))) t_1 = fma(Float32(Float32(floor(w) * dY_46_u) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_2 = Float32(floor(h) / sqrt(fmax(fma(Float32(Float32(floor(w) * dX_46_u) * dX_46_u), floor(w), t_0), t_1))) tmp = Float32(0.0) if (t_0 >= t_1) 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 := \left(dX.v \cdot dX.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \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_2 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u, \left\lfloor w\right\rfloor , t\_0\right), t\_1\right)}}\\
\mathbf{if}\;t\_0 \geq t\_1:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 75.7%
Taylor expanded in dX.u around 0
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.9
Applied rewrites64.9%
Applied rewrites65.0%
Applied rewrites64.7%
Applied rewrites64.8%
herbie shell --seed 2025112
(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))))