
(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 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_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 (* (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.u
(sqrt
(fmax
(fma (* (* (floor h) dX.v) dX.v) (floor h) (* (* dX.u dX.u) t_1))
(fma (* (* (floor h) dY.v) dY.v) (floor h) (* (* dY.u dY.u) t_1)))))
(floor w))
(/ (* (floor w) dY.u) (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_u / sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_1)), fmaf(((floorf(h) * dY_46_v) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_1))))) * floorf(w);
} else {
tmp = (floorf(w) * dY_46_u) / 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_u / sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(Float32(floor(h) * dY_46_v) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_1))))) * floor(w)); else tmp = Float32(Float32(floor(w) * dY_46_u) / 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.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.5%
Applied rewrites76.6%
Applied rewrites76.5%
(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.u
(/
(floor w)
(sqrt
(fmax
(fma (* (* (floor h) dX.v) dX.v) (floor h) (* (* dX.u dX.u) t_1))
(fma (* (* (floor h) dY.v) dY.v) (floor h) (* (* dY.u dY.u) t_1))))))
(/ (* (floor w) dY.u) (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_u * (floorf(w) / sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_1)), fmaf(((floorf(h) * dY_46_v) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = (floorf(w) * dY_46_u) / 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_u * Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_1)), fma(Float32(Float32(floor(h) * dY_46_v) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_1)))))); else tmp = Float32(Float32(floor(w) * dY_46_u) / 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.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_1\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.5%
Applied rewrites76.6%
Applied rewrites76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dX.v dX.v) t_1))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor w) (floor w)))
(t_5 (fma (* t_4 dY.u) dY.u (* (* dY.v dY.v) t_1)))
(t_6 (fma (* dY.v dY.v) t_1 (* (* dY.u dY.u) t_4)))
(t_7 (sqrt (fmax (fma (* t_3 dX.u) (floor w) t_2) t_6))))
(if (<= dX.v 40.0)
(if (>= (* (* dX.u dX.u) t_4) t_5)
(/ t_3 (sqrt (fmax (fma t_3 t_3 t_2) t_5)))
(/ t_0 (sqrt (fmax (fma (* t_4 dX.u) dX.u (* t_1 (* dX.v dX.v))) t_5))))
(if (>= t_2 t_6) (/ t_3 t_7) (/ t_0 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 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * floorf(h);
float t_2 = (dX_46_v * dX_46_v) * t_1;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(w) * floorf(w);
float t_5 = fmaf((t_4 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float t_6 = fmaf((dY_46_v * dY_46_v), t_1, ((dY_46_u * dY_46_u) * t_4));
float t_7 = sqrtf(fmaxf(fmaf((t_3 * dX_46_u), floorf(w), t_2), t_6));
float tmp_1;
if (dX_46_v <= 40.0f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * t_4) >= t_5) {
tmp_2 = t_3 / sqrtf(fmaxf(fmaf(t_3, t_3, t_2), t_5));
} else {
tmp_2 = t_0 / sqrtf(fmaxf(fmaf((t_4 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), t_5));
}
tmp_1 = tmp_2;
} else if (t_2 >= t_6) {
tmp_1 = t_3 / t_7;
} else {
tmp_1 = t_0 / 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(floor(w) * dY_46_u) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dX_46_v * dX_46_v) * t_1) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(w) * floor(w)) t_5 = fma(Float32(t_4 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) t_6 = fma(Float32(dY_46_v * dY_46_v), t_1, Float32(Float32(dY_46_u * dY_46_u) * t_4)) t_7 = sqrt(fmax(fma(Float32(t_3 * dX_46_u), floor(w), t_2), t_6)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(40.0)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_4) >= t_5) tmp_2 = Float32(t_3 / sqrt(fmax(fma(t_3, t_3, t_2), t_5))); else tmp_2 = Float32(t_0 / sqrt(fmax(fma(Float32(t_4 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), t_5))); end tmp_1 = tmp_2; elseif (t_2 >= t_6) tmp_1 = Float32(t_3 / t_7); else tmp_1 = Float32(t_0 / t_7); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dX.v \cdot dX.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_5 := \mathsf{fma}\left(t\_4 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
t_6 := \mathsf{fma}\left(dY.v \cdot dY.v, t\_1, \left(dY.u \cdot dY.u\right) \cdot t\_4\right)\\
t_7 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3 \cdot dX.u, \left\lfloor w\right\rfloor , t\_2\right), t\_6\right)}\\
\mathbf{if}\;dX.v \leq 40:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_4 \geq t\_5:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, t\_3, t\_2\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), t\_5\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_2 \geq t\_6:\\
\;\;\;\;\frac{t\_3}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_7}\\
\end{array}
\end{array}
if dX.v < 40Initial program 78.2%
Applied rewrites78.3%
Taylor expanded in dX.u around inf
Applied rewrites70.4%
Applied rewrites70.5%
if 40 < dX.v Initial program 70.8%
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-*.f3268.0
Applied rewrites68.0%
Applied rewrites68.0%
Applied rewrites68.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor w) dX.u))
(t_5 (* (* dX.v dX.v) t_3))
(t_6 (fma (* dY.v dY.v) t_3 (* (* dY.u dY.u) t_0)))
(t_7 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_8 (* t_4 dX.u))
(t_9 (sqrt (fmax (fma t_8 (floor w) t_5) t_6))))
(if (<= dX.v 40.0)
(if (>= (* (* dX.u dX.u) t_0) t_7)
(/ t_4 (sqrt (fmax (fma t_1 t_1 (* t_8 (floor w))) t_7)))
(/ t_2 (sqrt (fmax (fma (* t_0 dX.u) dX.u (* t_3 (* dX.v dX.v))) t_7))))
(if (>= t_5 t_6) (/ t_4 t_9) (/ t_2 t_9)))))
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) * floorf(w);
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(w) * dX_46_u;
float t_5 = (dX_46_v * dX_46_v) * t_3;
float t_6 = fmaf((dY_46_v * dY_46_v), t_3, ((dY_46_u * dY_46_u) * t_0));
float t_7 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_3));
float t_8 = t_4 * dX_46_u;
float t_9 = sqrtf(fmaxf(fmaf(t_8, floorf(w), t_5), t_6));
float tmp_1;
if (dX_46_v <= 40.0f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * t_0) >= t_7) {
tmp_2 = t_4 / sqrtf(fmaxf(fmaf(t_1, t_1, (t_8 * floorf(w))), t_7));
} else {
tmp_2 = t_2 / sqrtf(fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v))), t_7));
}
tmp_1 = tmp_2;
} else if (t_5 >= t_6) {
tmp_1 = t_4 / t_9;
} else {
tmp_1 = t_2 / t_9;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(dX_46_v * dX_46_v) * t_3) t_6 = fma(Float32(dY_46_v * dY_46_v), t_3, Float32(Float32(dY_46_u * dY_46_u) * t_0)) t_7 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_3)) t_8 = Float32(t_4 * dX_46_u) t_9 = sqrt(fmax(fma(t_8, floor(w), t_5), t_6)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(40.0)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_0) >= t_7) tmp_2 = Float32(t_4 / sqrt(fmax(fma(t_1, t_1, Float32(t_8 * floor(w))), t_7))); else tmp_2 = Float32(t_2 / sqrt(fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))), t_7))); end tmp_1 = tmp_2; elseif (t_5 >= t_6) tmp_1 = Float32(t_4 / t_9); else tmp_1 = Float32(t_2 / t_9); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left(dX.v \cdot dX.v\right) \cdot t\_3\\
t_6 := \mathsf{fma}\left(dY.v \cdot dY.v, t\_3, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\\
t_7 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_3\right)\\
t_8 := t\_4 \cdot dX.u\\
t_9 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , t\_5\right), t\_6\right)}\\
\mathbf{if}\;dX.v \leq 40:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_0 \geq t\_7:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, t\_8 \cdot \left\lfloor w\right\rfloor \right), t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right), t\_7\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_6:\\
\;\;\;\;\frac{t\_4}{t\_9}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_9}\\
\end{array}
\end{array}
if dX.v < 40Initial program 78.2%
Applied rewrites78.3%
Taylor expanded in dX.u around inf
Applied rewrites70.4%
Applied rewrites70.4%
if 40 < dX.v Initial program 70.8%
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-*.f3268.0
Applied rewrites68.0%
Applied rewrites68.0%
Applied rewrites68.1%
(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 h) (floor h)))
(t_3 (* (* dY.v dY.v) t_2))
(t_4 (* (* dX.v dX.v) t_2))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (* (floor w) dX.u))
(t_8 (+ (* t_7 t_7) (* t_0 t_0)))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_6))))
(t_10 (* (floor w) (floor w)))
(t_11 (fma (* dY.v dY.v) t_2 (* (* dY.u dY.u) t_10)))
(t_12
(sqrt
(fmax
(fma (* t_10 dX.u) dX.u (* t_2 (* dX.v dX.v)))
(fma (* t_10 dY.u) dY.u t_3))))
(t_13 (sqrt (fmax (fma (* t_7 dX.u) (floor w) t_4) t_11))))
(if (<= (if (>= t_8 t_6) (* t_9 t_7) (* t_9 t_1)) 5.00000006675716e-11)
(if (>= t_4 t_11) (/ t_7 t_13) (/ t_1 t_13))
(if (>= (exp (* (log t_7) 2.0)) t_3) (/ t_7 t_12) (/ t_1 t_12)))))
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(h) * floorf(h);
float t_3 = (dY_46_v * dY_46_v) * t_2;
float t_4 = (dX_46_v * dX_46_v) * t_2;
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
float t_7 = floorf(w) * dX_46_u;
float t_8 = (t_7 * t_7) + (t_0 * t_0);
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_6));
float t_10 = floorf(w) * floorf(w);
float t_11 = fmaf((dY_46_v * dY_46_v), t_2, ((dY_46_u * dY_46_u) * t_10));
float t_12 = sqrtf(fmaxf(fmaf((t_10 * dX_46_u), dX_46_u, (t_2 * (dX_46_v * dX_46_v))), fmaf((t_10 * dY_46_u), dY_46_u, t_3)));
float t_13 = sqrtf(fmaxf(fmaf((t_7 * dX_46_u), floorf(w), t_4), t_11));
float tmp;
if (t_8 >= t_6) {
tmp = t_9 * t_7;
} else {
tmp = t_9 * t_1;
}
float tmp_2;
if (tmp <= 5.00000006675716e-11f) {
float tmp_3;
if (t_4 >= t_11) {
tmp_3 = t_7 / t_13;
} else {
tmp_3 = t_1 / t_13;
}
tmp_2 = tmp_3;
} else if (expf((logf(t_7) * 2.0f)) >= t_3) {
tmp_2 = t_7 / t_12;
} else {
tmp_2 = t_1 / t_12;
}
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(floor(w) * dY_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_2) t_4 = Float32(Float32(dX_46_v * dX_46_v) * t_2) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = Float32(floor(w) * dX_46_u) t_8 = Float32(Float32(t_7 * t_7) + Float32(t_0 * t_0)) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_6))) t_10 = Float32(floor(w) * floor(w)) t_11 = fma(Float32(dY_46_v * dY_46_v), t_2, Float32(Float32(dY_46_u * dY_46_u) * t_10)) t_12 = sqrt(fmax(fma(Float32(t_10 * dX_46_u), dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_10 * dY_46_u), dY_46_u, t_3))) t_13 = sqrt(fmax(fma(Float32(t_7 * dX_46_u), floor(w), t_4), t_11)) tmp = Float32(0.0) if (t_8 >= t_6) tmp = Float32(t_9 * t_7); else tmp = Float32(t_9 * t_1); end tmp_2 = Float32(0.0) if (tmp <= Float32(5.00000006675716e-11)) tmp_3 = Float32(0.0) if (t_4 >= t_11) tmp_3 = Float32(t_7 / t_13); else tmp_3 = Float32(t_1 / t_13); end tmp_2 = tmp_3; elseif (exp(Float32(log(t_7) * Float32(2.0))) >= t_3) tmp_2 = Float32(t_7 / t_12); else tmp_2 = Float32(t_1 / t_12); end return tmp_2 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_2\\
t_4 := \left(dX.v \cdot dX.v\right) \cdot t\_2\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_8 := t\_7 \cdot t\_7 + t\_0 \cdot t\_0\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_6\right)}}\\
t_10 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_11 := \mathsf{fma}\left(dY.v \cdot dY.v, t\_2, \left(dY.u \cdot dY.u\right) \cdot t\_10\right)\\
t_12 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_10 \cdot dX.u, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_10 \cdot dY.u, dY.u, t\_3\right)\right)}\\
t_13 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_7 \cdot dX.u, \left\lfloor w\right\rfloor , t\_4\right), t\_11\right)}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_6:\\
\;\;\;\;t\_9 \cdot t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_1\\
\end{array} \leq 5.00000006675716 \cdot 10^{-11}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_11:\\
\;\;\;\;\frac{t\_7}{t\_13}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_13}\\
\end{array}\\
\mathbf{elif}\;e^{\log t\_7 \cdot 2} \geq t\_3:\\
\;\;\;\;\frac{t\_7}{t\_12}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_12}\\
\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))) < 5.00000007e-11Initial program 68.8%
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-*.f3261.1
Applied rewrites61.1%
Applied rewrites61.0%
Applied rewrites61.2%
if 5.00000007e-11 < (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 99.2%
Applied rewrites99.2%
Taylor expanded in dX.u around inf
Applied rewrites94.7%
Taylor expanded in dY.u around 0
Applied rewrites81.0%
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow-prod-downN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3289.7
Applied rewrites89.7%
(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) dX.u))
(t_2 (* (floor w) (floor w)))
(t_3 (* (* dY.v dY.v) t_0))
(t_4 (fma (* t_2 dY.u) dY.u t_3)))
(if (>= (* (* dX.u dX.u) t_2) t_3)
(/
t_1
(sqrt
(fmax (fma t_1 t_1 (* (* (* (floor h) dX.v) dX.v) (floor h))) t_4)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.v))) 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) * floorf(h);
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = (dY_46_v * dY_46_v) * t_0;
float t_4 = fmaf((t_2 * dY_46_u), dY_46_u, t_3);
float tmp;
if (((dX_46_u * dX_46_u) * t_2) >= t_3) {
tmp = t_1 / sqrtf(fmaxf(fmaf(t_1, t_1, (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h))), t_4));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), 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) * floor(h)) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(Float32(dY_46_v * dY_46_v) * t_0) t_4 = fma(Float32(t_2 * dY_46_u), dY_46_u, t_3) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= t_3) tmp = Float32(t_1 / sqrt(fmax(fma(t_1, t_1, Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))), t_4))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_4))); 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 dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left(dY.v \cdot dY.v\right) \cdot t\_0\\
t_4 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, t\_3\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq t\_3:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}}\\
\end{array}
\end{array}
Initial program 76.5%
Applied rewrites76.6%
Taylor expanded in dX.u around inf
Applied rewrites65.7%
Taylor expanded in dY.u around 0
Applied rewrites60.4%
Applied rewrites60.5%
herbie shell --seed 2025114
(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))))