
(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 8 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 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (pow (floor h) 2.0))
(t_3 (+ (pow (* dX.v (floor h)) 2.0) (pow t_1 2.0)))
(t_4 (pow t_0 2.0))
(t_5 (+ (pow (* dY.v (floor h)) 2.0) t_4)))
(if (>= t_3 t_5)
(/
t_1
(sqrt (fmax (fma (* t_2 dX.v) dX.v (pow (* (floor w) dX.u) 2.0)) t_5)))
(/ t_0 (sqrt (fmax t_3 (fma t_2 (* dY.v dY.v) t_4)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = powf(floorf(h), 2.0f);
float t_3 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_1, 2.0f);
float t_4 = powf(t_0, 2.0f);
float t_5 = powf((dY_46_v * floorf(h)), 2.0f) + t_4;
float tmp;
if (t_3 >= t_5) {
tmp = t_1 / sqrtf(fmaxf(fmaf((t_2 * dX_46_v), dX_46_v, powf((floorf(w) * dX_46_u), 2.0f)), t_5));
} else {
tmp = t_0 / sqrtf(fmaxf(t_3, fmaf(t_2, (dY_46_v * dY_46_v), t_4)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = floor(h) ^ Float32(2.0) t_3 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) t_4 = t_0 ^ Float32(2.0) t_5 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + t_4) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_1 / sqrt(fmax(fma(Float32(t_2 * dX_46_v), dX_46_v, (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_5))); else tmp = Float32(t_0 / sqrt(fmax(t_3, fma(t_2, Float32(dY_46_v * dY_46_v), t_4)))); 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\lfloor h\right\rfloor \right)}^{2}\\
t_3 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_1}^{2}\\
t_4 := {t\_0}^{2}\\
t_5 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + t\_4\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.v, dX.v, {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_2, dY.v \cdot dY.v, t\_4\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
lift-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.3
lower-pow.f32N/A
pow2N/A
lift-*.f3276.3
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites76.3%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-fma.f32N/A
pow2N/A
lower-*.f3276.3
Applied rewrites76.3%
(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 (* dY.u (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (pow t_3 2.0))
(t_5 (pow (floor h) 2.0))
(t_6 (pow t_2 2.0))
(t_7 (+ (pow (* dY.v (floor h)) 2.0) t_6))
(t_8 (+ (* t_3 t_3) (* t_0 t_0)))
(t_9 (* (floor h) dY.v))
(t_10 (+ (* t_1 t_1) (* t_9 t_9)))
(t_11 (/ 1.0 (sqrt (fmax t_8 t_10))))
(t_12 (if (>= t_8 t_10) (* t_11 t_3) (* t_11 t_1)))
(t_13 (* dX.u (floor w)))
(t_14 (+ (pow (* dX.v (floor h)) 2.0) (pow t_13 2.0)))
(t_15
(if (>= t_14 (* (pow dY.u 2.0) (pow (floor w) 2.0)))
(* (/ (floor w) (sqrt (fmax (+ (pow t_0 2.0) t_4) t_7))) dX.u)
(/ t_2 (sqrt (fmax t_14 t_7))))))
(if (<= t_12 -0.9980000257492065)
t_15
(if (<= t_12 0.004999999888241291)
(if (>= t_14 (* (pow dY.v 2.0) t_5))
(/ t_13 (sqrt (fmax (fma (* t_5 dX.v) dX.v t_4) t_7)))
(/ t_2 (sqrt (fmax t_14 (fma t_5 (* dY.v dY.v) t_6)))))
t_15))))
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 = dY_46_u * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = powf(t_3, 2.0f);
float t_5 = powf(floorf(h), 2.0f);
float t_6 = powf(t_2, 2.0f);
float t_7 = powf((dY_46_v * floorf(h)), 2.0f) + t_6;
float t_8 = (t_3 * t_3) + (t_0 * t_0);
float t_9 = floorf(h) * dY_46_v;
float t_10 = (t_1 * t_1) + (t_9 * t_9);
float t_11 = 1.0f / sqrtf(fmaxf(t_8, t_10));
float tmp;
if (t_8 >= t_10) {
tmp = t_11 * t_3;
} else {
tmp = t_11 * t_1;
}
float t_12 = tmp;
float t_13 = dX_46_u * floorf(w);
float t_14 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_13, 2.0f);
float tmp_1;
if (t_14 >= (powf(dY_46_u, 2.0f) * powf(floorf(w), 2.0f))) {
tmp_1 = (floorf(w) / sqrtf(fmaxf((powf(t_0, 2.0f) + t_4), t_7))) * dX_46_u;
} else {
tmp_1 = t_2 / sqrtf(fmaxf(t_14, t_7));
}
float t_15 = tmp_1;
float tmp_2;
if (t_12 <= -0.9980000257492065f) {
tmp_2 = t_15;
} else if (t_12 <= 0.004999999888241291f) {
float tmp_3;
if (t_14 >= (powf(dY_46_v, 2.0f) * t_5)) {
tmp_3 = t_13 / sqrtf(fmaxf(fmaf((t_5 * dX_46_v), dX_46_v, t_4), t_7));
} else {
tmp_3 = t_2 / sqrtf(fmaxf(t_14, fmaf(t_5, (dY_46_v * dY_46_v), t_6)));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_15;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = t_3 ^ Float32(2.0) t_5 = floor(h) ^ Float32(2.0) t_6 = t_2 ^ Float32(2.0) t_7 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + t_6) t_8 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_9 = Float32(floor(h) * dY_46_v) t_10 = Float32(Float32(t_1 * t_1) + Float32(t_9 * t_9)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_10))) tmp = Float32(0.0) if (t_8 >= t_10) tmp = Float32(t_11 * t_3); else tmp = Float32(t_11 * t_1); end t_12 = tmp t_13 = Float32(dX_46_u * floor(w)) t_14 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_13 ^ Float32(2.0))) tmp_1 = Float32(0.0) if (t_14 >= Float32((dY_46_u ^ Float32(2.0)) * (floor(w) ^ Float32(2.0)))) tmp_1 = Float32(Float32(floor(w) / sqrt(fmax(Float32((t_0 ^ Float32(2.0)) + t_4), t_7))) * dX_46_u); else tmp_1 = Float32(t_2 / sqrt(fmax(t_14, t_7))); end t_15 = tmp_1 tmp_2 = Float32(0.0) if (t_12 <= Float32(-0.9980000257492065)) tmp_2 = t_15; elseif (t_12 <= Float32(0.004999999888241291)) tmp_3 = Float32(0.0) if (t_14 >= Float32((dY_46_v ^ Float32(2.0)) * t_5)) tmp_3 = Float32(t_13 / sqrt(fmax(fma(Float32(t_5 * dX_46_v), dX_46_v, t_4), t_7))); else tmp_3 = Float32(t_2 / sqrt(fmax(t_14, fma(t_5, Float32(dY_46_v * dY_46_v), t_6)))); end tmp_2 = tmp_3; else tmp_2 = t_15; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := {t\_3}^{2}\\
t_5 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_6 := {t\_2}^{2}\\
t_7 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + t\_6\\
t_8 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := t\_1 \cdot t\_1 + t\_9 \cdot t\_9\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_10\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_10:\\
\;\;\;\;t\_11 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_1\\
\end{array}\\
t_13 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_14 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_13}^{2}\\
t_15 := \begin{array}{l}
\mathbf{if}\;t\_14 \geq {dY.u}^{2} \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left({t\_0}^{2} + t\_4, t\_7\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_14, t\_7\right)}}\\
\end{array}\\
\mathbf{if}\;t\_12 \leq -0.9980000257492065:\\
\;\;\;\;t\_15\\
\mathbf{elif}\;t\_12 \leq 0.004999999888241291:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_14 \geq {dY.v}^{2} \cdot t\_5:\\
\;\;\;\;\frac{t\_13}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot dX.v, dX.v, t\_4\right), t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_14, \mathsf{fma}\left(t\_5, dY.v \cdot dY.v, t\_6\right)\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) < -0.998000026 or 0.00499999989 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) Initial program 99.4%
Applied rewrites99.8%
Applied rewrites99.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3299.3
Applied rewrites99.3%
if -0.998000026 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) < 0.00499999989Initial program 61.8%
Applied rewrites61.9%
lift-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f3261.9
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.9
lower-pow.f32N/A
pow2N/A
lift-*.f3261.9
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites61.9%
lift-+.f32N/A
lift-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-fma.f32N/A
pow2N/A
lower-*.f3261.9
Applied rewrites61.9%
Taylor expanded in dY.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3261.9
Applied rewrites61.9%
(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 (+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_3 (* (floor w) dX.u))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (/ 1.0 (sqrt (fmax t_4 t_6))))
(t_8 (* dY.u (floor w)))
(t_9 (+ (pow (* dY.v (floor h)) 2.0) (pow t_8 2.0)))
(t_10
(*
(/ (floor w) (sqrt (fmax (+ (pow t_0 2.0) (pow t_3 2.0)) t_9)))
dX.u))
(t_11 (/ t_8 (sqrt (fmax t_2 t_9)))))
(if (<= (if (>= t_4 t_6) (* t_7 t_3) (* t_7 t_1)) 0.0020000000949949026)
(if (>= (* (pow dX.v 2.0) (pow (floor h) 2.0)) t_9) t_10 t_11)
(if (>= t_2 (* (pow dY.u 2.0) (pow (floor w) 2.0))) t_10 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 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
float t_7 = 1.0f / sqrtf(fmaxf(t_4, t_6));
float t_8 = dY_46_u * floorf(w);
float t_9 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_8, 2.0f);
float t_10 = (floorf(w) / sqrtf(fmaxf((powf(t_0, 2.0f) + powf(t_3, 2.0f)), t_9))) * dX_46_u;
float t_11 = t_8 / sqrtf(fmaxf(t_2, t_9));
float tmp;
if (t_4 >= t_6) {
tmp = t_7 * t_3;
} else {
tmp = t_7 * t_1;
}
float tmp_2;
if (tmp <= 0.0020000000949949026f) {
float tmp_3;
if ((powf(dX_46_v, 2.0f) * powf(floorf(h), 2.0f)) >= t_9) {
tmp_3 = t_10;
} else {
tmp_3 = t_11;
}
tmp_2 = tmp_3;
} else if (t_2 >= (powf(dY_46_u, 2.0f) * powf(floorf(w), 2.0f))) {
tmp_2 = t_10;
} else {
tmp_2 = t_11;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_4, t_6))) t_8 = Float32(dY_46_u * floor(w)) t_9 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_8 ^ Float32(2.0))) t_10 = Float32(Float32(floor(w) / sqrt(fmax(Float32((t_0 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))), t_9))) * dX_46_u) t_11 = Float32(t_8 / sqrt(fmax(t_2, t_9))) tmp = Float32(0.0) if (t_4 >= t_6) tmp = Float32(t_7 * t_3); else tmp = Float32(t_7 * t_1); end tmp_2 = Float32(0.0) if (tmp <= Float32(0.0020000000949949026)) tmp_3 = Float32(0.0) if (Float32((dX_46_v ^ Float32(2.0)) * (floor(h) ^ Float32(2.0))) >= t_9) tmp_3 = t_10; else tmp_3 = t_11; end tmp_2 = tmp_3; elseif (t_2 >= Float32((dY_46_u ^ Float32(2.0)) * (floor(w) ^ Float32(2.0)))) tmp_2 = t_10; else tmp_2 = t_11; end return tmp_2 end
function tmp_5 = 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 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_3 = floor(w) * dX_46_u; t_4 = (t_3 * t_3) + (t_0 * t_0); t_5 = floor(h) * dY_46_v; t_6 = (t_1 * t_1) + (t_5 * t_5); t_7 = single(1.0) / sqrt(max(t_4, t_6)); t_8 = dY_46_u * floor(w); t_9 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_8 ^ single(2.0)); t_10 = (floor(w) / sqrt(max(((t_0 ^ single(2.0)) + (t_3 ^ single(2.0))), t_9))) * dX_46_u; t_11 = t_8 / sqrt(max(t_2, t_9)); tmp = single(0.0); if (t_4 >= t_6) tmp = t_7 * t_3; else tmp = t_7 * t_1; end tmp_3 = single(0.0); if (tmp <= single(0.0020000000949949026)) tmp_4 = single(0.0); if (((dX_46_v ^ single(2.0)) * (floor(h) ^ single(2.0))) >= t_9) tmp_4 = t_10; else tmp_4 = t_11; end tmp_3 = tmp_4; elseif (t_2 >= ((dY_46_u ^ single(2.0)) * (floor(w) ^ single(2.0)))) tmp_3 = t_10; else tmp_3 = t_11; end tmp_5 = tmp_3; 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(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_6\right)}}\\
t_8 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_9 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_8}^{2}\\
t_10 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left({t\_0}^{2} + {t\_3}^{2}, t\_9\right)}} \cdot dX.u\\
t_11 := \frac{t\_8}{\sqrt{\mathsf{max}\left(t\_2, t\_9\right)}}\\
\mathbf{if}\;\begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;t\_7 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_1\\
\end{array} \leq 0.0020000000949949026:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{dX.v}^{2} \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2} \geq t\_9:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
\mathbf{elif}\;t\_2 \geq {dY.u}^{2} \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) < 0.00200000009Initial program 70.0%
Applied rewrites70.2%
Applied rewrites70.1%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3263.3
Applied rewrites63.3%
if 0.00200000009 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) Initial program 99.3%
Applied rewrites99.7%
Applied rewrites99.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3298.9
Applied rewrites98.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow t_0 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow (* dY.v (floor h)) 2.0) (pow t_2 2.0))))
(if (>= t_1 t_3)
(/
t_0
(sqrt
(fmax
(fma (* (pow (floor h) 2.0) dX.v) dX.v (pow (* (floor w) dX.u) 2.0))
t_3)))
(/ t_2 (sqrt (fmax t_1 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 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_0, 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_2, 2.0f);
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / sqrtf(fmaxf(fmaf((powf(floorf(h), 2.0f) * dX_46_v), dX_46_v, powf((floorf(w) * dX_46_u), 2.0f)), t_3));
} else {
tmp = t_2 / sqrtf(fmaxf(t_1, 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(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / sqrt(fmax(fma(Float32((floor(h) ^ Float32(2.0)) * dX_46_v), dX_46_v, (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_3))); else tmp = Float32(t_2 / sqrt(fmax(t_1, 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(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_0}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_2}^{2}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v, dX.v, {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
lift-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.3
lower-pow.f32N/A
pow2N/A
lift-*.f3276.3
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (* dY.u (floor w)))
(t_2 (+ (pow (* dX.v (floor h)) 2.0) (pow t_0 2.0)))
(t_3 (pow t_1 2.0))
(t_4 (+ (pow (* dY.v (floor h)) 2.0) t_3)))
(if (>= t_2 t_4)
(/ t_0 (sqrt (fmax t_2 t_4)))
(/ t_1 (sqrt (fmax t_2 (fma (pow (floor h) 2.0) (* dY.v dY.v) 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 = dY_46_u * floorf(w);
float t_2 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_0, 2.0f);
float t_3 = powf(t_1, 2.0f);
float t_4 = powf((dY_46_v * floorf(h)), 2.0f) + t_3;
float tmp;
if (t_2 >= t_4) {
tmp = t_0 / sqrtf(fmaxf(t_2, t_4));
} else {
tmp = t_1 / sqrtf(fmaxf(t_2, fmaf(powf(floorf(h), 2.0f), (dY_46_v * dY_46_v), 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(dY_46_u * floor(w)) t_2 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = t_1 ^ Float32(2.0) t_4 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + t_3) tmp = Float32(0.0) if (t_2 >= t_4) tmp = Float32(t_0 / sqrt(fmax(t_2, t_4))); else tmp = Float32(t_1 / sqrt(fmax(t_2, fma((floor(h) ^ Float32(2.0)), Float32(dY_46_v * dY_46_v), t_3)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_0}^{2}\\
t_3 := {t\_1}^{2}\\
t_4 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + t\_3\\
\mathbf{if}\;t\_2 \geq t\_4:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_2, t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left({\left(\left\lfloor h\right\rfloor \right)}^{2}, dY.v \cdot dY.v, t\_3\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.3
lower-pow.f32N/A
pow2N/A
lift-*.f3276.3
lower-+.f32N/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lower-fma.f32N/A
lower-pow.f32N/A
pow2N/A
lower-*.f3276.3
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.u (floor w)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow t_0 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (+ (pow (* dY.v (floor h)) 2.0) (pow t_2 2.0)))
(t_4 (sqrt (fmax t_1 t_3))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ t_2 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_u * floorf(w);
float t_1 = powf((dX_46_v * floorf(h)), 2.0f) + powf(t_0, 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_2, 2.0f);
float t_4 = sqrtf(fmaxf(t_1, t_3));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = t_2 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_u * floor(w)) t_1 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_4 = sqrt(fmax(t_1, t_3)) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(t_2 / 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 = dX_46_u * floor(w); t_1 = ((dX_46_v * floor(h)) ^ single(2.0)) + (t_0 ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_2 ^ single(2.0)); t_4 = sqrt(max(t_1, t_3)); tmp = single(0.0); if (t_1 >= t_3) tmp = t_0 / t_4; else tmp = t_2 / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_1 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_0}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_2}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_1, t\_3\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_1 (* dY.u (floor w)))
(t_2 (+ (pow (* dY.v (floor h)) 2.0) (pow t_1 2.0))))
(if (>= t_0 t_2)
(*
(floor w)
(/
dX.u
(sqrt
(fmax
(+ (pow (* (floor h) dX.v) 2.0) (pow (* (floor w) dX.u) 2.0))
t_2))))
(/ t_1 (sqrt (fmax t_0 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 = powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_1 = dY_46_u * floorf(w);
float t_2 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_1, 2.0f);
float tmp;
if (t_0 >= t_2) {
tmp = floorf(w) * (dX_46_u / sqrtf(fmaxf((powf((floorf(h) * dX_46_v), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f)), t_2)));
} else {
tmp = t_1 / sqrtf(fmaxf(t_0, t_2));
}
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)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_1 ^ Float32(2.0))) tmp = Float32(0.0) if (t_0 >= t_2) tmp = Float32(floor(w) * Float32(dX_46_u / sqrt(fmax(Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_2)))); else tmp = Float32(t_1 / sqrt(fmax(t_0, t_2))); 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 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_1 = dY_46_u * floor(w); t_2 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_1 ^ single(2.0)); tmp = single(0.0); if (t_0 >= t_2) tmp = floor(w) * (dX_46_u / sqrt(max((((floor(h) * dX_46_v) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0))), t_2))); else tmp = t_1 / sqrt(max(t_0, t_2)); end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_1}^{2}\\
\mathbf{if}\;t\_0 \geq t\_2:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \frac{dX.u}{\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_0, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (+ (pow (* dY.v (floor h)) 2.0) (pow t_0 2.0))))
(if (>= (* (pow dX.v 2.0) (pow (floor h) 2.0)) t_1)
(*
(/
(floor w)
(sqrt
(fmax
(+ (pow (* (floor h) dX.v) 2.0) (pow (* (floor w) dX.u) 2.0))
t_1)))
dX.u)
(/
t_0
(sqrt
(fmax
(+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0))
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 = powf((dY_46_v * floorf(h)), 2.0f) + powf(t_0, 2.0f);
float tmp;
if ((powf(dX_46_v, 2.0f) * powf(floorf(h), 2.0f)) >= t_1) {
tmp = (floorf(w) / sqrtf(fmaxf((powf((floorf(h) * dX_46_v), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f)), t_1))) * dX_46_u;
} else {
tmp = t_0 / sqrtf(fmaxf((powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), 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((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) tmp = Float32(0.0) if (Float32((dX_46_v ^ Float32(2.0)) * (floor(h) ^ Float32(2.0))) >= t_1) tmp = Float32(Float32(floor(w) / sqrt(fmax(Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))), t_1))) * dX_46_u); else tmp = Float32(t_0 / sqrt(fmax(Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), 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 = ((dY_46_v * floor(h)) ^ single(2.0)) + (t_0 ^ single(2.0)); tmp = single(0.0); if (((dX_46_v ^ single(2.0)) * (floor(h) ^ single(2.0))) >= t_1) tmp = (floor(w) / sqrt(max((((floor(h) * dX_46_v) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0))), t_1))) * dX_46_u; else tmp = t_0 / sqrt(max((((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0))), 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(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {t\_0}^{2}\\
\mathbf{if}\;{dX.v}^{2} \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2} \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left({\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}, t\_1\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left({\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, t\_1\right)}}\\
\end{array}
\end{array}
Initial program 76.1%
Applied rewrites76.3%
Applied rewrites76.2%
Taylor expanded in dX.u around 0
lower-*.f32N/A
lower-pow.f32N/A
lower-pow.f32N/A
lower-floor.f3265.0
Applied rewrites65.0%
herbie shell --seed 2025111
(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))))