
(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 9 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) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* dX.v (floor h)))
(t_4 (* t_3 dX.v))
(t_5 (+ (* t_1 t_1) (* t_0 t_0)))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_2 t_2) (* t_6 t_6)))
(t_8 (* (floor w) (floor w)))
(t_9 (* t_8 dX.u))
(t_10 (/ 1.0 (sqrt (fmax t_5 t_7))))
(t_11 (if (>= t_5 t_7) (* t_10 t_0) (* t_10 t_6)))
(t_12 (* (floor h) (floor h)))
(t_13 (fma (* t_8 dY.u) dY.u (* (* dY.v dY.v) t_12)))
(t_14 (fma t_9 dX.u (* t_12 (* dX.v dX.v))))
(t_15 (/ t_6 (sqrt (fmax t_14 t_13))))
(t_16
(if (>= (* t_4 (floor h)) t_13)
(/ t_0 (sqrt (fmax (fma t_9 dX.u (* (* t_3 (floor h)) dX.v)) t_13)))
t_15)))
(if (<= t_11 -0.9999949932098389)
t_16
(if (<= t_11 0.9900000095367432)
(if (>= t_14 (* (* dY.u dY.u) t_8))
(*
(/
dX.v
(sqrt
(fmax
(fma t_4 (floor h) (* (* t_1 dX.u) (floor w)))
(fma
(* t_2 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))))
(floor h))
t_15)
t_16))))
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) * dX_46_u;
float t_2 = floorf(w) * dY_46_u;
float t_3 = dX_46_v * floorf(h);
float t_4 = t_3 * dX_46_v;
float t_5 = (t_1 * t_1) + (t_0 * t_0);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_2 * t_2) + (t_6 * t_6);
float t_8 = floorf(w) * floorf(w);
float t_9 = t_8 * dX_46_u;
float t_10 = 1.0f / sqrtf(fmaxf(t_5, t_7));
float tmp;
if (t_5 >= t_7) {
tmp = t_10 * t_0;
} else {
tmp = t_10 * t_6;
}
float t_11 = tmp;
float t_12 = floorf(h) * floorf(h);
float t_13 = fmaf((t_8 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_12));
float t_14 = fmaf(t_9, dX_46_u, (t_12 * (dX_46_v * dX_46_v)));
float t_15 = t_6 / sqrtf(fmaxf(t_14, t_13));
float tmp_1;
if ((t_4 * floorf(h)) >= t_13) {
tmp_1 = t_0 / sqrtf(fmaxf(fmaf(t_9, dX_46_u, ((t_3 * floorf(h)) * dX_46_v)), t_13));
} else {
tmp_1 = t_15;
}
float t_16 = tmp_1;
float tmp_2;
if (t_11 <= -0.9999949932098389f) {
tmp_2 = t_16;
} else if (t_11 <= 0.9900000095367432f) {
float tmp_3;
if (t_14 >= ((dY_46_u * dY_46_u) * t_8)) {
tmp_3 = (dX_46_v / sqrtf(fmaxf(fmaf(t_4, floorf(h), ((t_1 * dX_46_u) * floorf(w))), fmaf((t_2 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))))) * floorf(h);
} else {
tmp_3 = t_15;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_16;
}
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) * dX_46_u) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(dX_46_v * floor(h)) t_4 = Float32(t_3 * dX_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)) t_8 = Float32(floor(w) * floor(w)) t_9 = Float32(t_8 * dX_46_u) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_7))) tmp = Float32(0.0) if (t_5 >= t_7) tmp = Float32(t_10 * t_0); else tmp = Float32(t_10 * t_6); end t_11 = tmp t_12 = Float32(floor(h) * floor(h)) t_13 = fma(Float32(t_8 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_12)) t_14 = fma(t_9, dX_46_u, Float32(t_12 * Float32(dX_46_v * dX_46_v))) t_15 = Float32(t_6 / sqrt(fmax(t_14, t_13))) tmp_1 = Float32(0.0) if (Float32(t_4 * floor(h)) >= t_13) tmp_1 = Float32(t_0 / sqrt(fmax(fma(t_9, dX_46_u, Float32(Float32(t_3 * floor(h)) * dX_46_v)), t_13))); else tmp_1 = t_15; end t_16 = tmp_1 tmp_2 = Float32(0.0) if (t_11 <= Float32(-0.9999949932098389)) tmp_2 = t_16; elseif (t_11 <= Float32(0.9900000095367432)) tmp_3 = Float32(0.0) if (t_14 >= Float32(Float32(dY_46_u * dY_46_u) * t_8)) tmp_3 = Float32(Float32(dX_46_v / sqrt(fmax(fma(t_4, floor(h), Float32(Float32(t_1 * dX_46_u) * floor(w))), fma(Float32(t_2 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))))) * floor(h)); else tmp_3 = t_15; end tmp_2 = tmp_3; else tmp_2 = t_16; 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 dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := t\_3 \cdot dX.v\\
t_5 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_2 \cdot t\_2 + t\_6 \cdot t\_6\\
t_8 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_9 := t\_8 \cdot dX.u\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}}\\
t_11 := \begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_7:\\
\;\;\;\;t\_10 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_6\\
\end{array}\\
t_12 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_13 := \mathsf{fma}\left(t\_8 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_12\right)\\
t_14 := \mathsf{fma}\left(t\_9, dX.u, t\_12 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_15 := \frac{t\_6}{\sqrt{\mathsf{max}\left(t\_14, t\_13\right)}}\\
t_16 := \begin{array}{l}
\mathbf{if}\;t\_4 \cdot \left\lfloor h\right\rfloor \geq t\_13:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_9, dX.u, \left(t\_3 \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), t\_13\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array}\\
\mathbf{if}\;t\_11 \leq -0.9999949932098389:\\
\;\;\;\;t\_16\\
\mathbf{elif}\;t\_11 \leq 0.9900000095367432:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_14 \geq \left(dY.u \cdot dY.u\right) \cdot t\_8:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4, \left\lfloor h\right\rfloor , \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(t\_2 \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}:\\
\;\;\;\;t\_15\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_16\\
\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 h) dX.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 h) dY.v))) < -0.999994993 or 0.99000001 < (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 h) dX.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 h) dY.v))) Initial program 99.5%
Applied rewrites99.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3299.6
lift-*.f32N/A
*-commutativeN/A
lower-*.f3299.6
Applied rewrites99.6%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
lift-floor.f32N/A
exp-to-powN/A
pow2N/A
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
Applied rewrites99.6%
if -0.999994993 < (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 h) dX.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 h) dY.v))) < 0.99000001Initial program 63.5%
Applied rewrites63.7%
Applied rewrites63.7%
Taylor expanded in dY.u around inf
Applied rewrites63.6%
(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) dX.u))
(t_2 (* (* t_1 dX.u) (floor w)))
(t_3 (* (floor w) dY.u))
(t_4
(fma
(* t_3 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_5 (* dX.v (floor h)))
(t_6 (* t_5 dX.v))
(t_7 (+ (* t_1 t_1) (* t_0 t_0)))
(t_8 (* (floor h) dY.v))
(t_9 (+ (* t_3 t_3) (* t_8 t_8)))
(t_10 (* (floor w) (floor w)))
(t_11 (* t_10 dX.u))
(t_12 (/ 1.0 (sqrt (fmax t_7 t_9))))
(t_13 (if (>= t_7 t_9) (* t_12 t_0) (* t_12 t_8)))
(t_14 (* (floor h) (floor h)))
(t_15 (fma (* t_10 dY.u) dY.u (* (* dY.v dY.v) t_14)))
(t_16
(if (>= (* t_6 (floor h)) t_15)
(/
t_0
(sqrt (fmax (fma t_11 dX.u (* (* t_5 (floor h)) dX.v)) t_15)))
(/ t_8 (sqrt (fmax (fma t_11 dX.u (* t_14 (* dX.v dX.v))) t_15))))))
(if (<= t_13 -0.9999979734420776)
t_16
(if (<= t_13 4.999999873689376e-6)
(if (>= t_2 t_4)
(*
(/
(floor h)
(sqrt (fmax (fma t_6 (floor h) (* (* t_1 (floor w)) dX.u)) t_4)))
dX.v)
(* (/ (floor h) (sqrt (fmax (fma t_6 (floor h) t_2) t_4))) dY.v))
t_16))))
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) * dX_46_u;
float t_2 = (t_1 * dX_46_u) * floorf(w);
float t_3 = floorf(w) * dY_46_u;
float t_4 = fmaf((t_3 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_5 = dX_46_v * floorf(h);
float t_6 = t_5 * dX_46_v;
float t_7 = (t_1 * t_1) + (t_0 * t_0);
float t_8 = floorf(h) * dY_46_v;
float t_9 = (t_3 * t_3) + (t_8 * t_8);
float t_10 = floorf(w) * floorf(w);
float t_11 = t_10 * dX_46_u;
float t_12 = 1.0f / sqrtf(fmaxf(t_7, t_9));
float tmp;
if (t_7 >= t_9) {
tmp = t_12 * t_0;
} else {
tmp = t_12 * t_8;
}
float t_13 = tmp;
float t_14 = floorf(h) * floorf(h);
float t_15 = fmaf((t_10 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_14));
float tmp_1;
if ((t_6 * floorf(h)) >= t_15) {
tmp_1 = t_0 / sqrtf(fmaxf(fmaf(t_11, dX_46_u, ((t_5 * floorf(h)) * dX_46_v)), t_15));
} else {
tmp_1 = t_8 / sqrtf(fmaxf(fmaf(t_11, dX_46_u, (t_14 * (dX_46_v * dX_46_v))), t_15));
}
float t_16 = tmp_1;
float tmp_2;
if (t_13 <= -0.9999979734420776f) {
tmp_2 = t_16;
} else if (t_13 <= 4.999999873689376e-6f) {
float tmp_3;
if (t_2 >= t_4) {
tmp_3 = (floorf(h) / sqrtf(fmaxf(fmaf(t_6, floorf(h), ((t_1 * floorf(w)) * dX_46_u)), t_4))) * dX_46_v;
} else {
tmp_3 = (floorf(h) / sqrtf(fmaxf(fmaf(t_6, floorf(h), t_2), t_4))) * dY_46_v;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_16;
}
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) * dX_46_u) t_2 = Float32(Float32(t_1 * dX_46_u) * floor(w)) t_3 = Float32(floor(w) * dY_46_u) t_4 = fma(Float32(t_3 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_5 = Float32(dX_46_v * floor(h)) t_6 = Float32(t_5 * dX_46_v) t_7 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(Float32(t_3 * t_3) + Float32(t_8 * t_8)) t_10 = Float32(floor(w) * floor(w)) t_11 = Float32(t_10 * dX_46_u) t_12 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_9))) tmp = Float32(0.0) if (t_7 >= t_9) tmp = Float32(t_12 * t_0); else tmp = Float32(t_12 * t_8); end t_13 = tmp t_14 = Float32(floor(h) * floor(h)) t_15 = fma(Float32(t_10 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_14)) tmp_1 = Float32(0.0) if (Float32(t_6 * floor(h)) >= t_15) tmp_1 = Float32(t_0 / sqrt(fmax(fma(t_11, dX_46_u, Float32(Float32(t_5 * floor(h)) * dX_46_v)), t_15))); else tmp_1 = Float32(t_8 / sqrt(fmax(fma(t_11, dX_46_u, Float32(t_14 * Float32(dX_46_v * dX_46_v))), t_15))); end t_16 = tmp_1 tmp_2 = Float32(0.0) if (t_13 <= Float32(-0.9999979734420776)) tmp_2 = t_16; elseif (t_13 <= Float32(4.999999873689376e-6)) tmp_3 = Float32(0.0) if (t_2 >= t_4) tmp_3 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_6, floor(h), Float32(Float32(t_1 * floor(w)) * dX_46_u)), t_4))) * dX_46_v); else tmp_3 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_6, floor(h), t_2), t_4))) * dY_46_v); end tmp_2 = tmp_3; else tmp_2 = t_16; 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 dX.u\\
t_2 := \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \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)\\
t_5 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_6 := t\_5 \cdot dX.v\\
t_7 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_3 \cdot t\_3 + t\_8 \cdot t\_8\\
t_10 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_11 := t\_10 \cdot dX.u\\
t_12 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_9\right)}}\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_9:\\
\;\;\;\;t\_12 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_12 \cdot t\_8\\
\end{array}\\
t_14 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_15 := \mathsf{fma}\left(t\_10 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_14\right)\\
t_16 := \begin{array}{l}
\mathbf{if}\;t\_6 \cdot \left\lfloor h\right\rfloor \geq t\_15:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_11, dX.u, \left(t\_5 \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), t\_15\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_11, dX.u, t\_14 \cdot \left(dX.v \cdot dX.v\right)\right), t\_15\right)}}\\
\end{array}\\
\mathbf{if}\;t\_13 \leq -0.9999979734420776:\\
\;\;\;\;t\_16\\
\mathbf{elif}\;t\_13 \leq 4.999999873689376 \cdot 10^{-6}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_2 \geq t\_4:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor h\right\rfloor , \left(t\_1 \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right), t\_4\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_6, \left\lfloor h\right\rfloor , t\_2\right), t\_4\right)}} \cdot dY.v\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_16\\
\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 h) dX.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 h) dY.v))) < -0.999997973 or 4.99999987e-6 < (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 h) dX.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 h) dY.v))) Initial program 99.4%
Applied rewrites99.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3299.5
lift-*.f32N/A
*-commutativeN/A
lower-*.f3299.5
Applied rewrites99.5%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
lift-floor.f32N/A
exp-to-powN/A
pow2N/A
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
Applied rewrites98.7%
if -0.999997973 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.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 h) dY.v))) < 4.99999987e-6Initial program 61.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3261.0
Applied rewrites61.0%
Applied rewrites61.1%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3261.1
Applied rewrites61.1%
(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 (* t_1 dX.u))
(t_4 (fma t_3 dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_4 t_2)
(/
(* (floor h) dX.v)
(sqrt
(fmax (fma t_3 dX.u (* (* (* dX.v (floor h)) (floor h)) dX.v)) t_2)))
(/ (* (floor h) dY.v) (sqrt (fmax t_4 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 = t_1 * dX_46_u;
float t_4 = fmaf(t_3, dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_4 >= t_2) {
tmp = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf(t_3, dX_46_u, (((dX_46_v * floorf(h)) * floorf(h)) * dX_46_v)), t_2));
} else {
tmp = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_4, 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 = Float32(t_1 * dX_46_u) t_4 = fma(t_3, dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_4 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(t_3, dX_46_u, Float32(Float32(Float32(dX_46_v * floor(h)) * floor(h)) * dX_46_v)), t_2))); else tmp = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_4, 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 := t\_1 \cdot dX.u\\
t_4 := \mathsf{fma}\left(t\_3, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_4 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.u, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_4, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
(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 (* (floor w) dY.u))
(t_2 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_3 (fma (* t_1 dY.u) (floor w) t_2))
(t_4 (* (floor w) dX.u))
(t_5 (* (floor h) dX.v))
(t_6 (* (* t_4 dX.u) (floor w)))
(t_7 (+ (* t_4 t_4) (* t_5 t_5)))
(t_8 (* (floor h) dY.v))
(t_9 (+ (* t_1 t_1) (* t_8 t_8)))
(t_10 (* (floor w) (floor w)))
(t_11 (/ 1.0 (sqrt (fmax t_7 t_9))))
(t_12 (if (>= t_7 t_9) (* t_11 t_5) (* t_11 t_8)))
(t_13 (* (floor h) (floor h)))
(t_14 (fma (* t_10 dX.u) dX.u (* t_13 (* dX.v dX.v))))
(t_15 (fma (* t_10 dY.u) dY.u (* (* dY.v dY.v) t_13)))
(t_16 (sqrt (fmax t_14 t_15)))
(t_17 (>= (* t_0 (floor h)) t_15))
(t_18 (/ t_5 t_16)))
(if (<= t_12 -0.9999979734420776)
(if t_17 t_18 (/ t_8 (sqrt (fmax t_14 (fma t_1 t_1 t_2)))))
(if (<= t_12 4.999999873689376e-6)
(if (>= t_6 t_3)
(*
(/
(floor h)
(sqrt (fmax (fma t_0 (floor h) (* (* t_4 (floor w)) dX.u)) t_3)))
dX.v)
(* (/ (floor h) (sqrt (fmax (fma t_0 (floor h) t_6) t_3))) dY.v))
(if t_17 t_18 (/ t_8 t_16))))))
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 = floorf(w) * dY_46_u;
float t_2 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_3 = fmaf((t_1 * dY_46_u), floorf(w), t_2);
float t_4 = floorf(w) * dX_46_u;
float t_5 = floorf(h) * dX_46_v;
float t_6 = (t_4 * dX_46_u) * floorf(w);
float t_7 = (t_4 * t_4) + (t_5 * t_5);
float t_8 = floorf(h) * dY_46_v;
float t_9 = (t_1 * t_1) + (t_8 * t_8);
float t_10 = floorf(w) * floorf(w);
float t_11 = 1.0f / sqrtf(fmaxf(t_7, t_9));
float tmp;
if (t_7 >= t_9) {
tmp = t_11 * t_5;
} else {
tmp = t_11 * t_8;
}
float t_12 = tmp;
float t_13 = floorf(h) * floorf(h);
float t_14 = fmaf((t_10 * dX_46_u), dX_46_u, (t_13 * (dX_46_v * dX_46_v)));
float t_15 = fmaf((t_10 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_13));
float t_16 = sqrtf(fmaxf(t_14, t_15));
int t_17 = (t_0 * floorf(h)) >= t_15;
float t_18 = t_5 / t_16;
float tmp_2;
if (t_12 <= -0.9999979734420776f) {
float tmp_3;
if (t_17) {
tmp_3 = t_18;
} else {
tmp_3 = t_8 / sqrtf(fmaxf(t_14, fmaf(t_1, t_1, t_2)));
}
tmp_2 = tmp_3;
} else if (t_12 <= 4.999999873689376e-6f) {
float tmp_4;
if (t_6 >= t_3) {
tmp_4 = (floorf(h) / sqrtf(fmaxf(fmaf(t_0, floorf(h), ((t_4 * floorf(w)) * dX_46_u)), t_3))) * dX_46_v;
} else {
tmp_4 = (floorf(h) / sqrtf(fmaxf(fmaf(t_0, floorf(h), t_6), t_3))) * dY_46_v;
}
tmp_2 = tmp_4;
} else if (t_17) {
tmp_2 = t_18;
} else {
tmp_2 = t_8 / t_16;
}
return tmp_2;
}
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(floor(w) * dY_46_u) t_2 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_3 = fma(Float32(t_1 * dY_46_u), floor(w), t_2) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(t_4 * dX_46_u) * floor(w)) t_7 = Float32(Float32(t_4 * t_4) + Float32(t_5 * t_5)) t_8 = Float32(floor(h) * dY_46_v) t_9 = Float32(Float32(t_1 * t_1) + Float32(t_8 * t_8)) t_10 = Float32(floor(w) * floor(w)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_9))) tmp = Float32(0.0) if (t_7 >= t_9) tmp = Float32(t_11 * t_5); else tmp = Float32(t_11 * t_8); end t_12 = tmp t_13 = Float32(floor(h) * floor(h)) t_14 = fma(Float32(t_10 * dX_46_u), dX_46_u, Float32(t_13 * Float32(dX_46_v * dX_46_v))) t_15 = fma(Float32(t_10 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_13)) t_16 = sqrt(fmax(t_14, t_15)) t_17 = Float32(t_0 * floor(h)) >= t_15 t_18 = Float32(t_5 / t_16) tmp_2 = Float32(0.0) if (t_12 <= Float32(-0.9999979734420776)) tmp_3 = Float32(0.0) if (t_17) tmp_3 = t_18; else tmp_3 = Float32(t_8 / sqrt(fmax(t_14, fma(t_1, t_1, t_2)))); end tmp_2 = tmp_3; elseif (t_12 <= Float32(4.999999873689376e-6)) tmp_4 = Float32(0.0) if (t_6 >= t_3) tmp_4 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_0, floor(h), Float32(Float32(t_4 * floor(w)) * dX_46_u)), t_3))) * dX_46_v); else tmp_4 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_0, floor(h), t_6), t_3))) * dY_46_v); end tmp_2 = tmp_4; elseif (t_17) tmp_2 = t_18; else tmp_2 = Float32(t_8 / t_16); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_1 \cdot dY.u, \left\lfloor w\right\rfloor , t\_2\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \left(t\_4 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_7 := t\_4 \cdot t\_4 + t\_5 \cdot t\_5\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_9 := t\_1 \cdot t\_1 + t\_8 \cdot t\_8\\
t_10 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_9\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_9:\\
\;\;\;\;t\_11 \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_8\\
\end{array}\\
t_13 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_14 := \mathsf{fma}\left(t\_10 \cdot dX.u, dX.u, t\_13 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_15 := \mathsf{fma}\left(t\_10 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_13\right)\\
t_16 := \sqrt{\mathsf{max}\left(t\_14, t\_15\right)}\\
t_17 := t\_0 \cdot \left\lfloor h\right\rfloor \geq t\_15\\
t_18 := \frac{t\_5}{t\_16}\\
\mathbf{if}\;t\_12 \leq -0.9999979734420776:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_17:\\
\;\;\;\;t\_18\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{\sqrt{\mathsf{max}\left(t\_14, \mathsf{fma}\left(t\_1, t\_1, t\_2\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_12 \leq 4.999999873689376 \cdot 10^{-6}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , \left(t\_4 \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right), t\_3\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , t\_6\right), t\_3\right)}} \cdot dY.v\\
\end{array}\\
\mathbf{elif}\;t\_17:\\
\;\;\;\;t\_18\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_8}{t\_16}\\
\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 h) dX.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 h) dY.v))) < -0.999997973Initial program 99.5%
Applied rewrites99.4%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f3299.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
Applied rewrites99.6%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
pow2N/A
lift-*.f32N/A
pow2N/A
pow-prod-downN/A
lift-floor.f32N/A
exp-to-powN/A
pow2N/A
lift-floor.f32N/A
pow2N/A
associate-*r*N/A
Applied rewrites99.6%
if -0.999997973 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.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 h) dY.v))) < 4.99999987e-6Initial program 61.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3261.0
Applied rewrites61.0%
Applied rewrites61.1%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3261.1
Applied rewrites61.1%
if 4.99999987e-6 < (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 h) dX.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 h) dY.v))) Initial program 99.3%
Applied rewrites99.3%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites97.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (* dX.v (floor h)) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) dX.u))
(t_4 (* (* t_3 dX.u) (floor w)))
(t_5 (+ (* t_3 t_3) (* t_0 t_0)))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_2 t_2) (* t_6 t_6)))
(t_8 (* (floor w) (floor w)))
(t_9 (/ 1.0 (sqrt (fmax t_5 t_7))))
(t_10 (if (>= t_5 t_7) (* t_9 t_0) (* t_9 t_6)))
(t_11 (* (floor h) (floor h)))
(t_12 (fma (* t_8 dY.u) dY.u (* (* dY.v dY.v) t_11)))
(t_13
(sqrt (fmax (fma (* t_8 dX.u) dX.u (* t_11 (* dX.v dX.v))) t_12)))
(t_14 (if (>= (* t_1 (floor h)) t_12) (/ t_0 t_13) (/ t_6 t_13)))
(t_15
(fma
(* t_2 dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))
(if (<= t_10 -0.9999979734420776)
t_14
(if (<= t_10 4.999999873689376e-6)
(if (>= t_4 t_15)
(*
(/
(floor h)
(sqrt (fmax (fma t_1 (floor h) (* (* t_3 (floor w)) dX.u)) t_15)))
dX.v)
(* (/ (floor h) (sqrt (fmax (fma t_1 (floor h) t_4) t_15))) dY.v))
t_14))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = (dX_46_v * floorf(h)) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_3 * dX_46_u) * floorf(w);
float t_5 = (t_3 * t_3) + (t_0 * t_0);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_2 * t_2) + (t_6 * t_6);
float t_8 = floorf(w) * floorf(w);
float t_9 = 1.0f / sqrtf(fmaxf(t_5, t_7));
float tmp;
if (t_5 >= t_7) {
tmp = t_9 * t_0;
} else {
tmp = t_9 * t_6;
}
float t_10 = tmp;
float t_11 = floorf(h) * floorf(h);
float t_12 = fmaf((t_8 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_11));
float t_13 = sqrtf(fmaxf(fmaf((t_8 * dX_46_u), dX_46_u, (t_11 * (dX_46_v * dX_46_v))), t_12));
float tmp_1;
if ((t_1 * floorf(h)) >= t_12) {
tmp_1 = t_0 / t_13;
} else {
tmp_1 = t_6 / t_13;
}
float t_14 = tmp_1;
float t_15 = fmaf((t_2 * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float tmp_2;
if (t_10 <= -0.9999979734420776f) {
tmp_2 = t_14;
} else if (t_10 <= 4.999999873689376e-6f) {
float tmp_3;
if (t_4 >= t_15) {
tmp_3 = (floorf(h) / sqrtf(fmaxf(fmaf(t_1, floorf(h), ((t_3 * floorf(w)) * dX_46_u)), t_15))) * dX_46_v;
} else {
tmp_3 = (floorf(h) / sqrtf(fmaxf(fmaf(t_1, floorf(h), t_4), t_15))) * dY_46_v;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_14;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_3 * dX_46_u) * floor(w)) t_5 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)) t_8 = Float32(floor(w) * floor(w)) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_7))) tmp = Float32(0.0) if (t_5 >= t_7) tmp = Float32(t_9 * t_0); else tmp = Float32(t_9 * t_6); end t_10 = tmp t_11 = Float32(floor(h) * floor(h)) t_12 = fma(Float32(t_8 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_11)) t_13 = sqrt(fmax(fma(Float32(t_8 * dX_46_u), dX_46_u, Float32(t_11 * Float32(dX_46_v * dX_46_v))), t_12)) tmp_1 = Float32(0.0) if (Float32(t_1 * floor(h)) >= t_12) tmp_1 = Float32(t_0 / t_13); else tmp_1 = Float32(t_6 / t_13); end t_14 = tmp_1 t_15 = fma(Float32(t_2 * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) tmp_2 = Float32(0.0) if (t_10 <= Float32(-0.9999979734420776)) tmp_2 = t_14; elseif (t_10 <= Float32(4.999999873689376e-6)) tmp_3 = Float32(0.0) if (t_4 >= t_15) tmp_3 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_1, floor(h), Float32(Float32(t_3 * floor(w)) * dX_46_u)), t_15))) * dX_46_v); else tmp_3 = Float32(Float32(floor(h) / sqrt(fmax(fma(t_1, floor(h), t_4), t_15))) * dY_46_v); end tmp_2 = tmp_3; else tmp_2 = t_14; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \left(t\_3 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_5 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_2 \cdot t\_2 + t\_6 \cdot t\_6\\
t_8 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_7:\\
\;\;\;\;t\_9 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_6\\
\end{array}\\
t_11 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_12 := \mathsf{fma}\left(t\_8 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_11\right)\\
t_13 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8 \cdot dX.u, dX.u, t\_11 \cdot \left(dX.v \cdot dX.v\right)\right), t\_12\right)}\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_1 \cdot \left\lfloor h\right\rfloor \geq t\_12:\\
\;\;\;\;\frac{t\_0}{t\_13}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_6}{t\_13}\\
\end{array}\\
t_15 := \mathsf{fma}\left(t\_2 \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\_10 \leq -0.9999979734420776:\\
\;\;\;\;t\_14\\
\mathbf{elif}\;t\_10 \leq 4.999999873689376 \cdot 10^{-6}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_15:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , \left(t\_3 \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right), t\_15\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , t\_4\right), t\_15\right)}} \cdot dY.v\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\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 h) dX.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 h) dY.v))) < -0.999997973 or 4.99999987e-6 < (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 h) dX.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 h) dY.v))) Initial program 99.4%
Applied rewrites99.4%
Taylor expanded in dX.u around 0
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites98.5%
if -0.999997973 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.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 h) dY.v))) < 4.99999987e-6Initial program 61.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3261.0
Applied rewrites61.0%
Applied rewrites61.1%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3261.1
Applied rewrites61.1%
(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.v (floor h)) dX.v)
(floor h)
(* (* (* (floor w) dX.u) dX.u) (floor w)))
(fma
(* (* (floor w) dY.u) 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_v * floorf(h)) * dX_46_v), floorf(h), (((floorf(w) * dX_46_u) * dX_46_u) * floorf(w))), fmaf(((floorf(w) * dY_46_u) * 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_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w))), 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)))))) * 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.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \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)\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 76.2%
Applied rewrites76.3%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* t_0 dX.u))
(t_2 (* (floor h) (floor h)))
(t_3 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_2)))
(t_4 (fma t_1 dX.u (* t_2 (* dX.v dX.v))))
(t_5 (* dX.v (floor h)))
(t_6 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_7 (/ (* (floor h) dY.v) (sqrt (fmax t_4 t_3)))))
(if (<= dY.u 4.0)
(if (>= t_4 t_6)
(/
(* (floor h) dX.v)
(sqrt (fmax (fma t_1 dX.u (* (* t_5 (floor h)) dX.v)) t_3)))
t_7)
(if (>= t_4 (* (* dY.u dY.u) t_0))
(*
(/
dX.v
(sqrt
(fmax
(fma
(* t_5 dX.v)
(floor h)
(* (* (* (floor w) dX.u) dX.u) (floor w)))
(fma (* (* (floor w) dY.u) dY.u) (floor w) t_6))))
(floor h))
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) * floorf(w);
float t_1 = t_0 * dX_46_u;
float t_2 = floorf(h) * floorf(h);
float t_3 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_2));
float t_4 = fmaf(t_1, dX_46_u, (t_2 * (dX_46_v * dX_46_v)));
float t_5 = dX_46_v * floorf(h);
float t_6 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_7 = (floorf(h) * dY_46_v) / sqrtf(fmaxf(t_4, t_3));
float tmp_1;
if (dY_46_u <= 4.0f) {
float tmp_2;
if (t_4 >= t_6) {
tmp_2 = (floorf(h) * dX_46_v) / sqrtf(fmaxf(fmaf(t_1, dX_46_u, ((t_5 * floorf(h)) * dX_46_v)), t_3));
} else {
tmp_2 = t_7;
}
tmp_1 = tmp_2;
} else if (t_4 >= ((dY_46_u * dY_46_u) * t_0)) {
tmp_1 = (dX_46_v / sqrtf(fmaxf(fmaf((t_5 * dX_46_v), floorf(h), (((floorf(w) * dX_46_u) * dX_46_u) * floorf(w))), fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), t_6)))) * floorf(h);
} else {
tmp_1 = 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) * floor(w)) t_1 = Float32(t_0 * dX_46_u) t_2 = Float32(floor(h) * floor(h)) t_3 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_2)) t_4 = fma(t_1, dX_46_u, Float32(t_2 * Float32(dX_46_v * dX_46_v))) t_5 = Float32(dX_46_v * floor(h)) t_6 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_7 = Float32(Float32(floor(h) * dY_46_v) / sqrt(fmax(t_4, t_3))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(4.0)) tmp_2 = Float32(0.0) if (t_4 >= t_6) tmp_2 = Float32(Float32(floor(h) * dX_46_v) / sqrt(fmax(fma(t_1, dX_46_u, Float32(Float32(t_5 * floor(h)) * dX_46_v)), t_3))); else tmp_2 = t_7; end tmp_1 = tmp_2; elseif (t_4 >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp_1 = Float32(Float32(dX_46_v / sqrt(fmax(fma(Float32(t_5 * dX_46_v), floor(h), Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w))), fma(Float32(Float32(floor(w) * dY_46_u) * dY_46_u), floor(w), t_6)))) * floor(h)); else tmp_1 = t_7; 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 := t\_0 \cdot dX.u\\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_2\right)\\
t_4 := \mathsf{fma}\left(t\_1, dX.u, t\_2 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_5 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_6 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_7 := \frac{\left\lfloor h\right\rfloor \cdot dY.v}{\sqrt{\mathsf{max}\left(t\_4, t\_3\right)}}\\
\mathbf{if}\;dY.u \leq 4:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, dX.u, \left(t\_5 \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}\\
\mathbf{elif}\;t\_4 \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;\frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u, \left\lfloor w\right\rfloor , t\_6\right)\right)}} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
if dY.u < 4Initial program 78.0%
Applied rewrites78.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3278.1
lift-*.f32N/A
*-commutativeN/A
lower-*.f3278.1
Applied rewrites78.1%
Taylor expanded in dY.u around 0
pow2N/A
lift-floor.f32N/A
pow2N/A
unswap-sqrN/A
lift-floor.f32N/A
associate-*l*N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3270.0
Applied rewrites70.0%
if 4 < dY.u Initial program 70.8%
Applied rewrites71.0%
Applied rewrites70.9%
Taylor expanded in dY.u around inf
Applied rewrites67.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (* t_0 dX.u) (floor w)))
(t_2 (* (* dX.v (floor h)) dX.v))
(t_3
(fma
(* (* (floor w) dY.u) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))
(if (>= t_1 t_3)
(*
(/
(floor h)
(sqrt (fmax (fma t_2 (floor h) (* (* t_0 (floor w)) dX.u)) t_3)))
dX.v)
(* (/ (floor h) (sqrt (fmax (fma t_2 (floor h) t_1) t_3))) 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 = floorf(w) * dX_46_u;
float t_1 = (t_0 * dX_46_u) * floorf(w);
float t_2 = (dX_46_v * floorf(h)) * dX_46_v;
float t_3 = fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float tmp;
if (t_1 >= t_3) {
tmp = (floorf(h) / sqrtf(fmaxf(fmaf(t_2, floorf(h), ((t_0 * floorf(w)) * dX_46_u)), t_3))) * dX_46_v;
} else {
tmp = (floorf(h) / sqrtf(fmaxf(fmaf(t_2, floorf(h), t_1), t_3))) * 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(floor(w) * dX_46_u) t_1 = Float32(Float32(t_0 * dX_46_u) * floor(w)) t_2 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_3 = 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))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(floor(h) / sqrt(fmax(fma(t_2, floor(h), Float32(Float32(t_0 * floor(w)) * dX_46_u)), t_3))) * dX_46_v); else tmp = Float32(Float32(floor(h) / sqrt(fmax(fma(t_2, floor(h), t_1), t_3))) * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left(t\_0 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_3 := \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)\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , \left(t\_0 \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right), t\_3\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , t\_1\right), t\_3\right)}} \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.2%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.3
Applied rewrites64.3%
Applied rewrites64.3%
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-floor.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f3264.3
Applied rewrites64.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (* (floor w) dX.u) dX.u) (floor w)))
(t_1 (* (* dX.v (floor h)) dX.v))
(t_2
(fma
(* (* (floor w) dY.u) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))
(if (>= t_0 t_2)
(*
(/
(floor h)
(sqrt
(fmax
(fma t_1 (floor h) (* (* (* dX.u dX.u) (floor w)) (floor w)))
t_2)))
dX.v)
(* (/ (floor h) (sqrt (fmax (fma t_1 (floor h) t_0) 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 = ((floorf(w) * dX_46_u) * dX_46_u) * floorf(w);
float t_1 = (dX_46_v * floorf(h)) * dX_46_v;
float t_2 = fmaf(((floorf(w) * dY_46_u) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float tmp;
if (t_0 >= t_2) {
tmp = (floorf(h) / sqrtf(fmaxf(fmaf(t_1, floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w))), t_2))) * dX_46_v;
} else {
tmp = (floorf(h) / sqrtf(fmaxf(fmaf(t_1, floorf(h), t_0), 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(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w)) t_1 = Float32(Float32(dX_46_v * floor(h)) * dX_46_v) t_2 = 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))) tmp = Float32(0.0) if (t_0 >= t_2) tmp = Float32(Float32(floor(h) / sqrt(fmax(fma(t_1, floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))), t_2))) * dX_46_v); else tmp = Float32(Float32(floor(h) / sqrt(fmax(fma(t_1, floor(h), t_0), t_2))) * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_1 := \left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\\
t_2 := \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)\\
\mathbf{if}\;t\_0 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right), t\_2\right)}} \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , t\_0\right), t\_2\right)}} \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.2%
Taylor expanded in dX.u around inf
*-commutativeN/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unpow2N/A
lower-*.f3264.3
Applied rewrites64.3%
Applied rewrites64.3%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f3264.3
Applied rewrites64.3%
herbie shell --seed 2025110
(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))))