
(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 10 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.v (floor h)) dX.v))
(t_1 (* dY.u (floor w)))
(t_2 (* (* dX.u (floor w)) dX.u))
(t_3 (* dY.v (floor h)))
(t_4 (fma (* t_1 dY.u) (floor w) (* (* t_3 dY.v) (floor h))))
(t_5 (* (floor w) dX.u))
(t_6 (* (floor h) dX.v)))
(if (>= (fma t_2 (floor w) (* t_0 (floor h))) t_4)
(*
(/ (* dX.u 1.0) (sqrt (fmax (fma t_0 (floor h) (* t_2 (floor w))) t_4)))
(floor w))
(/
(* (- dY.u) (floor w))
(- (sqrt (fmax (fma t_5 t_5 (* t_6 t_6)) (fma t_1 t_1 (* t_3 t_3)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = (dX_46_v * floorf(h)) * dX_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = (dX_46_u * floorf(w)) * dX_46_u;
float t_3 = dY_46_v * floorf(h);
float t_4 = fmaf((t_1 * dY_46_u), floorf(w), ((t_3 * dY_46_v) * floorf(h)));
float t_5 = floorf(w) * dX_46_u;
float t_6 = floorf(h) * dX_46_v;
float tmp;
if (fmaf(t_2, floorf(w), (t_0 * floorf(h))) >= t_4) {
tmp = ((dX_46_u * 1.0f) / sqrtf(fmaxf(fmaf(t_0, floorf(h), (t_2 * floorf(w))), t_4))) * floorf(w);
} else {
tmp = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(fmaf(t_5, t_5, (t_6 * t_6)), fmaf(t_1, t_1, (t_3 * t_3))));
}
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 * floor(h)) * dX_46_v) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(Float32(dX_46_u * floor(w)) * dX_46_u) t_3 = Float32(dY_46_v * floor(h)) t_4 = fma(Float32(t_1 * dY_46_u), floor(w), Float32(Float32(t_3 * dY_46_v) * floor(h))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(floor(h) * dX_46_v) tmp = Float32(0.0) if (fma(t_2, floor(w), Float32(t_0 * floor(h))) >= t_4) tmp = Float32(Float32(Float32(dX_46_u * Float32(1.0)) / sqrt(fmax(fma(t_0, floor(h), Float32(t_2 * floor(w))), t_4))) * floor(w)); else tmp = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(fma(t_5, t_5, Float32(t_6 * t_6)), fma(t_1, t_1, Float32(t_3 * t_3)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , \left(t\_3 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
\mathbf{if}\;\mathsf{fma}\left(t\_2, \left\lfloor w\right\rfloor , t\_0 \cdot \left\lfloor h\right\rfloor \right) \geq t\_4:\\
\;\;\;\;\frac{dX.u \cdot 1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , t\_2 \cdot \left\lfloor w\right\rfloor \right), t\_4\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, t\_5, t\_6 \cdot t\_6\right), \mathsf{fma}\left(t\_1, t\_1, t\_3 \cdot t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 75.6%
Applied rewrites75.5%
Applied rewrites75.6%
Applied rewrites75.6%
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))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor w) dX.u) t_4) (/ (* (floor w) dY.u) 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) * 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 t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(w) * dX_46_u) / t_4;
} else {
tmp = (floorf(w) * dY_46_u) / 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) * 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))) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(w) * dX_46_u) / t_4); else tmp = Float32(Float32(floor(w) * dY_46_u) / 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 \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)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{t\_4}\\
\end{array}
\end{array}
Initial program 75.6%
Applied rewrites75.7%
(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 (* (* dY.v dY.v) t_1))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor w) (floor w)))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_3 t_3) (* t_4 t_4)))
(t_8 (/ 1.0 (sqrt (fmax t_7 (+ (* t_0 t_0) (* t_6 t_6))))))
(t_9 (* t_8 t_0)))
(if (<= dY.v 0.0008200000156648457)
(if (>= t_7 (* (* (* dY.u (floor w)) dY.u) (floor w)))
(*
(*
(floor w)
(sqrt
(/
1.0
(fmax
(fma (* t_5 dX.u) dX.u (* t_1 (* dX.v dX.v)))
(fma (* t_5 dY.u) dY.u t_2)))))
dX.u)
t_9)
(if (>= t_7 t_2) (* t_8 t_3) 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) * dY_46_u;
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(w) * floorf(w);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_3 * t_3) + (t_4 * t_4);
float t_8 = 1.0f / sqrtf(fmaxf(t_7, ((t_0 * t_0) + (t_6 * t_6))));
float t_9 = t_8 * t_0;
float tmp_1;
if (dY_46_v <= 0.0008200000156648457f) {
float tmp_2;
if (t_7 >= (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))) {
tmp_2 = (floorf(w) * sqrtf((1.0f / fmaxf(fmaf((t_5 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v))), fmaf((t_5 * dY_46_u), dY_46_u, t_2))))) * dX_46_u;
} else {
tmp_2 = t_9;
}
tmp_1 = tmp_2;
} else if (t_7 >= t_2) {
tmp_1 = t_8 * t_3;
} else {
tmp_1 = 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) * dY_46_u) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(w) * floor(w)) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4)) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_7, Float32(Float32(t_0 * t_0) + Float32(t_6 * t_6))))) t_9 = Float32(t_8 * t_0) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(0.0008200000156648457)) tmp_2 = Float32(0.0) if (t_7 >= Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))) tmp_2 = Float32(Float32(floor(w) * sqrt(Float32(Float32(1.0) / fmax(fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_5 * dY_46_u), dY_46_u, t_2))))) * dX_46_u); else tmp_2 = t_9; end tmp_1 = tmp_2; elseif (t_7 >= t_2) tmp_1 = Float32(t_8 * t_3); else tmp_1 = t_9; 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(dY.v \cdot dY.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_3 \cdot t\_3 + t\_4 \cdot t\_4\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_0 \cdot t\_0 + t\_6 \cdot t\_6\right)}}\\
t_9 := t\_8 \cdot t\_0\\
\mathbf{if}\;dY.v \leq 0.0008200000156648457:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\left(\left\lfloor w\right\rfloor \cdot \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, t\_2\right)\right)}}\right) \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{elif}\;t\_7 \geq t\_2:\\
\;\;\;\;t\_8 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}
\end{array}
if dY.v < 8.20000016e-4Initial program 77.1%
Applied rewrites76.9%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3268.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3268.3
Applied rewrites68.3%
if 8.20000016e-4 < dY.v Initial program 72.0%
Taylor expanded in dY.u around 0
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3267.0
Applied rewrites67.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor h) dX.v))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_3 t_3) (* t_5 t_5)))
(t_7 (* (* dX.v (floor h)) dX.v))
(t_8 (* (floor w) (floor w)))
(t_9 (* dX.u (floor w)))
(t_10 (* (* t_9 dX.u) (floor w)))
(t_11 (sqrt (fmax (fma t_7 (floor h) t_10) t_1)))
(t_12 (* (floor h) (floor h))))
(if (<= dX.v 10.0)
(if (>= t_10 t_6)
(* (/ 1.0 (sqrt (fmax (+ (* t_2 t_2) (* t_4 t_4)) t_6))) t_2)
(*
(*
(floor w)
(sqrt
(/
1.0
(fmax
(fma (* t_8 dX.u) dX.u (* t_12 (* dX.v dX.v)))
(fma (* t_8 dY.u) dY.u (* (* dY.v dY.v) t_12))))))
dY.u))
(if (>= (* t_7 (floor h)) t_1) (/ t_9 t_11) (/ t_0 t_11)))))
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 * floorf(w);
float t_1 = fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_3 * t_3) + (t_5 * t_5);
float t_7 = (dX_46_v * floorf(h)) * dX_46_v;
float t_8 = floorf(w) * floorf(w);
float t_9 = dX_46_u * floorf(w);
float t_10 = (t_9 * dX_46_u) * floorf(w);
float t_11 = sqrtf(fmaxf(fmaf(t_7, floorf(h), t_10), t_1));
float t_12 = floorf(h) * floorf(h);
float tmp_1;
if (dX_46_v <= 10.0f) {
float tmp_2;
if (t_10 >= t_6) {
tmp_2 = (1.0f / sqrtf(fmaxf(((t_2 * t_2) + (t_4 * t_4)), t_6))) * t_2;
} else {
tmp_2 = (floorf(w) * sqrtf((1.0f / fmaxf(fmaf((t_8 * dX_46_u), dX_46_u, (t_12 * (dX_46_v * dX_46_v))), fmaf((t_8 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_12)))))) * dY_46_u;
}
tmp_1 = tmp_2;
} else if ((t_7 * floorf(h)) >= t_1) {
tmp_1 = t_9 / t_11;
} else {
tmp_1 = t_0 / t_11;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_3 * t_3) + Float32(t_5 * t_5)) t_7 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_8 = Float32(floor(w) * floor(w)) t_9 = Float32(dX_46_u * floor(w)) t_10 = Float32(Float32(t_9 * dX_46_u) * floor(w)) t_11 = sqrt(fmax(fma(t_7, floor(h), t_10), t_1)) t_12 = Float32(floor(h) * floor(h)) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(10.0)) tmp_2 = Float32(0.0) if (t_10 >= t_6) tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4)), t_6))) * t_2); else tmp_2 = Float32(Float32(floor(w) * sqrt(Float32(Float32(1.0) / fmax(fma(Float32(t_8 * dX_46_u), dX_46_u, Float32(t_12 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_8 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_12)))))) * dY_46_u); end tmp_1 = tmp_2; elseif (Float32(t_7 * floor(h)) >= t_1) tmp_1 = Float32(t_9 / t_11); else tmp_1 = Float32(t_0 / t_11); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \mathsf{fma}\left(t\_0 \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 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_3 \cdot t\_3 + t\_5 \cdot t\_5\\
t_7 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_8 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_9 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_10 := \left(t\_9 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_11 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_7, \left\lfloor h\right\rfloor , t\_10\right), t\_1\right)}\\
t_12 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;dX.v \leq 10:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_10 \geq t\_6:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_2 \cdot t\_2 + t\_4 \cdot t\_4, t\_6\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;\left(\left\lfloor w\right\rfloor \cdot \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_8 \cdot dX.u, dX.u, t\_12 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_8 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_12\right)\right)}}\right) \cdot dY.u\\
\end{array}\\
\mathbf{elif}\;t\_7 \cdot \left\lfloor h\right\rfloor \geq t\_1:\\
\;\;\;\;\frac{t\_9}{t\_11}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_11}\\
\end{array}
\end{array}
if dX.v < 10Initial program 77.4%
Applied rewrites77.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3269.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3269.3
Applied rewrites69.3%
if 10 < dX.v Initial program 70.0%
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-*.f3266.5
Applied rewrites66.5%
Applied rewrites66.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_2 (* (* dX.v (floor h)) dX.v))
(t_3 (* dX.u (floor w)))
(t_4 (sqrt (fmax (fma t_2 (floor h) (* (* t_3 dX.u) (floor w))) t_1))))
(if (>= (* t_2 (floor h)) t_1) (/ t_3 t_4) (/ t_0 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 = dY_46_u * floorf(w);
float t_1 = fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_2 = (dX_46_v * floorf(h)) * dX_46_v;
float t_3 = dX_46_u * floorf(w);
float t_4 = sqrtf(fmaxf(fmaf(t_2, floorf(h), ((t_3 * dX_46_u) * floorf(w))), t_1));
float tmp;
if ((t_2 * floorf(h)) >= t_1) {
tmp = t_3 / t_4;
} else {
tmp = t_0 / t_4;
}
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 * floor(w)) t_1 = fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_2 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_3 = Float32(dX_46_u * floor(w)) t_4 = sqrt(fmax(fma(t_2, floor(h), Float32(Float32(t_3 * dX_46_u) * floor(w))), t_1)) tmp = Float32(0.0) if (Float32(t_2 * floor(h)) >= t_1) tmp = Float32(t_3 / t_4); else tmp = Float32(t_0 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \mathsf{fma}\left(t\_0 \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 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_3 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_4 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , \left(t\_3 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), t\_1\right)}\\
\mathbf{if}\;t\_2 \cdot \left\lfloor h\right\rfloor \geq t\_1:\\
\;\;\;\;\frac{t\_3}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Applied rewrites64.7%
(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) dY.v))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5 (* t_4 t_4)))
(if (>=
(* (* (floor h) (floor h)) (* dX.v dX.v))
(* (* (* dY.u (floor w)) dY.u) (floor w)))
(* (/ 1.0 (sqrt (fmax (+ (* t_3 t_3) t_5) t_2))) t_3)
(*
(/ 1.0 (sqrt (fmax (+ (exp (* (log (* dX.u (floor w))) 2.0)) t_5) t_2)))
t_0))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = t_4 * t_4;
float tmp;
if (((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)) >= (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w))) {
tmp = (1.0f / sqrtf(fmaxf(((t_3 * t_3) + t_5), t_2))) * t_3;
} else {
tmp = (1.0f / sqrtf(fmaxf((expf((logf((dX_46_u * floorf(w))) * 2.0f)) + t_5), t_2))) * t_0;
}
return tmp;
}
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) * dY_46_v) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if (Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v)) >= Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_3 * t_3) + t_5), t_2))) * t_3); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(exp(Float32(log(Float32(dX_46_u * floor(w))) * Float32(2.0))) + t_5), t_2))) * t_0); 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(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = floor(w) * dX_46_u; t_4 = floor(h) * dX_46_v; t_5 = t_4 * t_4; tmp = single(0.0); if (((floor(h) * floor(h)) * (dX_46_v * dX_46_v)) >= (((dY_46_u * floor(w)) * dY_46_u) * floor(w))) tmp = (single(1.0) / sqrt(max(((t_3 * t_3) + t_5), t_2))) * t_3; else tmp = (single(1.0) / sqrt(max((exp((log((dX_46_u * floor(w))) * single(2.0))) + t_5), t_2))) * t_0; end tmp_2 = tmp; 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 dY.v\\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right) \geq \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_5, t\_2\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(e^{\log \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot 2} + t\_5, t\_2\right)}} \cdot t\_0\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
Applied rewrites58.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow-to-expN/A
lift-log.f32N/A
pow-to-expN/A
exp-sumN/A
lower-fma.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3260.5
lift-fma.f32N/A
lift-*.f32N/A
lift-log.f32N/A
distribute-rgt-outN/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites60.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* t_2 t_2))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor h) dX.v))
(t_6 (* t_5 t_5)))
(if (>=
(* (* (floor h) (floor h)) (* dX.v dX.v))
(* (* t_0 dY.u) (floor w)))
(*
(/ 1.0 (sqrt (fmax (+ (* t_4 t_4) t_6) (+ (exp (* (log t_0) 2.0)) t_3))))
t_4)
(*
(/
1.0
(sqrt
(fmax
(+ (exp (* (log (* dX.u (floor w))) 2.0)) t_6)
(+ (* t_1 t_1) t_3))))
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_u * floorf(w);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = t_2 * t_2;
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(h) * dX_46_v;
float t_6 = t_5 * t_5;
float tmp;
if (((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)) >= ((t_0 * dY_46_u) * floorf(w))) {
tmp = (1.0f / sqrtf(fmaxf(((t_4 * t_4) + t_6), (expf((logf(t_0) * 2.0f)) + t_3)))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf((expf((logf((dX_46_u * floorf(w))) * 2.0f)) + t_6), ((t_1 * t_1) + t_3)))) * 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_u * floor(w)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(t_2 * t_2) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(t_5 * t_5) tmp = Float32(0.0) if (Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v)) >= Float32(Float32(t_0 * dY_46_u) * floor(w))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_4 * t_4) + t_6), Float32(exp(Float32(log(t_0) * Float32(2.0))) + t_3)))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(exp(Float32(log(Float32(dX_46_u * floor(w))) * Float32(2.0))) + t_6), Float32(Float32(t_1 * t_1) + t_3)))) * 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 = dY_46_u * floor(w); t_1 = floor(w) * dY_46_u; t_2 = floor(h) * dY_46_v; t_3 = t_2 * t_2; t_4 = floor(w) * dX_46_u; t_5 = floor(h) * dX_46_v; t_6 = t_5 * t_5; tmp = single(0.0); if (((floor(h) * floor(h)) * (dX_46_v * dX_46_v)) >= ((t_0 * dY_46_u) * floor(w))) tmp = (single(1.0) / sqrt(max(((t_4 * t_4) + t_6), (exp((log(t_0) * single(2.0))) + t_3)))) * t_4; else tmp = (single(1.0) / sqrt(max((exp((log((dX_46_u * floor(w))) * single(2.0))) + t_6), ((t_1 * t_1) + t_3)))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_2 \cdot t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_5 \cdot t\_5\\
\mathbf{if}\;\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right) \geq \left(t\_0 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor :\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4 \cdot t\_4 + t\_6, e^{\log t\_0 \cdot 2} + t\_3\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(e^{\log \left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot 2} + t\_6, t\_1 \cdot t\_1 + t\_3\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
Applied rewrites58.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow-to-expN/A
lift-log.f32N/A
pow-to-expN/A
exp-sumN/A
lower-fma.f32N/A
lift-log.f32N/A
lift-*.f32N/A
lift-exp.f3260.5
lift-fma.f32N/A
lift-*.f32N/A
lift-log.f32N/A
distribute-rgt-outN/A
*-commutativeN/A
lower-*.f32N/A
Applied rewrites60.5%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3260.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.4
Applied rewrites60.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(t_2 (fma (* t_0 dX.u) (floor w) t_1))
(t_3 (* dY.u (floor w))))
(if (>= t_1 (* (* dY.u dY.u) (* (floor w) (floor w))))
(/
t_0
(sqrt
(fmax
t_2
(fma
(* t_3 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(/
t_3
(sqrt
(fmax t_2 (fma (* (* (floor h) (floor h)) dY.v) dY.v (* t_3 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 = ((dX_46_v * floorf(h)) * dX_46_v) * floorf(h);
float t_2 = fmaf((t_0 * dX_46_u), floorf(w), t_1);
float t_3 = dY_46_u * floorf(w);
float tmp;
if (t_1 >= ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))) {
tmp = t_0 / sqrtf(fmaxf(t_2, fmaf((t_3 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
} else {
tmp = t_3 / sqrtf(fmaxf(t_2, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, (t_3 * 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 = Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) t_2 = fma(Float32(t_0 * dX_46_u), floor(w), t_1) t_3 = Float32(dY_46_u * floor(w)) tmp = Float32(0.0) if (t_1 >= Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))) tmp = Float32(t_0 / sqrt(fmax(t_2, fma(Float32(t_3 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))); else tmp = Float32(t_3 / sqrt(fmax(t_2, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(t_3 * t_3))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0 \cdot dX.u, \left\lfloor w\right\rfloor , t\_1\right)\\
t_3 := dY.u \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;t\_1 \geq \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(t\_3 \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{t\_3}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, t\_3 \cdot t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
Applied rewrites58.9%
Applied rewrites59.1%
lift-fma.f32N/A
lift-*.f32N/A
+-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lift-*.f32N/A
*-commutativeN/A
lower-fma.f32N/A
Applied rewrites59.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dX.v))
(t_5
(fma
(* t_0 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))
(if (>= t_2 (* (* dY.u dY.u) (* (floor w) (floor w))))
(/ t_1 (sqrt (fmax (fma t_3 t_3 (* t_4 t_4)) t_5)))
(/ t_0 (sqrt (fmax (fma (* t_1 dX.u) (floor w) t_2) t_5))))))
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 * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = ((dX_46_v * floorf(h)) * dX_46_v) * floorf(h);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dX_46_v;
float t_5 = fmaf((t_0 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float tmp;
if (t_2 >= ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))) {
tmp = t_1 / sqrtf(fmaxf(fmaf(t_3, t_3, (t_4 * t_4)), t_5));
} else {
tmp = t_0 / sqrtf(fmaxf(fmaf((t_1 * dX_46_u), floorf(w), t_2), t_5));
}
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 * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dX_46_v) t_5 = fma(Float32(t_0 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) tmp = Float32(0.0) if (t_2 >= Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))) tmp = Float32(t_1 / sqrt(fmax(fma(t_3, t_3, Float32(t_4 * t_4)), t_5))); else tmp = Float32(t_0 / sqrt(fmax(fma(Float32(t_1 * dX_46_u), floor(w), t_2), t_5))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_5 := \mathsf{fma}\left(t\_0 \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)\\
\mathbf{if}\;t\_2 \geq \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, t\_3, t\_4 \cdot t\_4\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, \left\lfloor w\right\rfloor , t\_2\right), t\_5\right)}}\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
Applied rewrites58.9%
Applied rewrites59.1%
lift-fma.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
Applied rewrites59.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* t_1 t_1))
(t_3 (* dY.v (floor h)))
(t_4 (* dY.u (floor w)))
(t_5
(/
(floor w)
(sqrt (fmax (fma t_0 t_0 t_2) (fma t_4 t_4 (* t_3 t_3)))))))
(if (>= t_2 (* t_4 t_4)) (* t_5 dX.u) (* t_5 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(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = t_1 * t_1;
float t_3 = dY_46_v * floorf(h);
float t_4 = dY_46_u * floorf(w);
float t_5 = floorf(w) / sqrtf(fmaxf(fmaf(t_0, t_0, t_2), fmaf(t_4, t_4, (t_3 * t_3))));
float tmp;
if (t_2 >= (t_4 * t_4)) {
tmp = t_5 * dX_46_u;
} else {
tmp = t_5 * 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(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(floor(w) / sqrt(fmax(fma(t_0, t_0, t_2), fma(t_4, t_4, Float32(t_3 * t_3))))) tmp = Float32(0.0) if (t_2 >= Float32(t_4 * t_4)) tmp = Float32(t_5 * dX_46_u); else tmp = Float32(t_5 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_2\right), \mathsf{fma}\left(t\_4, t\_4, t\_3 \cdot t\_3\right)\right)}}\\
\mathbf{if}\;t\_2 \geq t\_4 \cdot t\_4:\\
\;\;\;\;t\_5 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot dY.u\\
\end{array}
\end{array}
Initial program 75.6%
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-*.f3264.5
Applied rewrites64.5%
Taylor expanded in dY.u around inf
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3258.9
Applied rewrites58.9%
Applied rewrites59.1%
Applied rewrites58.9%
herbie shell --seed 2025112
(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))))