
(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 (* (floor h) (floor h)))
(t_4 (* (* dY.v dY.v) t_3))
(t_5
(/
1.0
(sqrt
(fmax
(fma t_3 (* dX.v dX.v) (* (* (floor w) (floor w)) (* dX.u dX.u)))
(fma (* t_2 (floor w)) dY.u t_4)))))
(t_6 (* (fma (floor w) (floor w) (/ 0.0 dY.u)) (* dY.u dY.u)))
(t_7 (+ (* t_1 t_1) (* t_0 t_0)))
(t_8 (/ 1.0 (sqrt (fmax t_7 t_6))))
(t_9 (* (floor h) dY.v))
(t_10 (if (>= t_7 t_4) (* t_5 t_0) (* t_5 t_9)))
(t_11 (+ (* t_2 t_2) (* t_9 t_9)))
(t_12 (/ 1.0 (sqrt (fmax t_7 t_11))))
(t_13 (if (>= t_7 t_11) (* t_12 t_0) (* t_12 t_9))))
(if (<= t_13 -0.007000000216066837)
t_10
(if (<= t_13 1.0000000116860974e-7)
(if (>= t_7 t_6) (* t_8 t_0) (* t_8 t_9))
t_10))))
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 = floorf(h) * floorf(h);
float t_4 = (dY_46_v * dY_46_v) * t_3;
float t_5 = 1.0f / sqrtf(fmaxf(fmaf(t_3, (dX_46_v * dX_46_v), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), fmaf((t_2 * floorf(w)), dY_46_u, t_4)));
float t_6 = fmaf(floorf(w), floorf(w), (0.0f / dY_46_u)) * (dY_46_u * dY_46_u);
float t_7 = (t_1 * t_1) + (t_0 * t_0);
float t_8 = 1.0f / sqrtf(fmaxf(t_7, t_6));
float t_9 = floorf(h) * dY_46_v;
float tmp;
if (t_7 >= t_4) {
tmp = t_5 * t_0;
} else {
tmp = t_5 * t_9;
}
float t_10 = tmp;
float t_11 = (t_2 * t_2) + (t_9 * t_9);
float t_12 = 1.0f / sqrtf(fmaxf(t_7, t_11));
float tmp_1;
if (t_7 >= t_11) {
tmp_1 = t_12 * t_0;
} else {
tmp_1 = t_12 * t_9;
}
float t_13 = tmp_1;
float tmp_2;
if (t_13 <= -0.007000000216066837f) {
tmp_2 = t_10;
} else if (t_13 <= 1.0000000116860974e-7f) {
float tmp_3;
if (t_7 >= t_6) {
tmp_3 = t_8 * t_0;
} else {
tmp_3 = t_8 * t_9;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_10;
}
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(floor(h) * floor(h)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * t_3) t_5 = Float32(Float32(1.0) / sqrt(fmax(fma(t_3, Float32(dX_46_v * dX_46_v), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), fma(Float32(t_2 * floor(w)), dY_46_u, t_4)))) t_6 = Float32(fma(floor(w), floor(w), Float32(Float32(0.0) / dY_46_u)) * Float32(dY_46_u * dY_46_u)) t_7 = Float32(Float32(t_1 * t_1) + Float32(t_0 * t_0)) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_6))) t_9 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (t_7 >= t_4) tmp = Float32(t_5 * t_0); else tmp = Float32(t_5 * t_9); end t_10 = tmp t_11 = Float32(Float32(t_2 * t_2) + Float32(t_9 * t_9)) t_12 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_11))) tmp_1 = Float32(0.0) if (t_7 >= t_11) tmp_1 = Float32(t_12 * t_0); else tmp_1 = Float32(t_12 * t_9); end t_13 = tmp_1 tmp_2 = Float32(0.0) if (t_13 <= Float32(-0.007000000216066837)) tmp_2 = t_10; elseif (t_13 <= Float32(1.0000000116860974e-7)) tmp_3 = Float32(0.0) if (t_7 >= t_6) tmp_3 = Float32(t_8 * t_0); else tmp_3 = Float32(t_8 * t_9); end tmp_2 = tmp_3; else tmp_2 = t_10; 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 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(dY.v \cdot dY.v\right) \cdot t\_3\\
t_5 := \frac{1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.v \cdot dX.v, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), \mathsf{fma}\left(t\_2 \cdot \left\lfloor w\right\rfloor , dY.u, t\_4\right)\right)}}\\
t_6 := \mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \frac{0}{dY.u}\right) \cdot \left(dY.u \cdot dY.u\right)\\
t_7 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_6\right)}}\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_4:\\
\;\;\;\;t\_5 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot t\_9\\
\end{array}\\
t_11 := t\_2 \cdot t\_2 + t\_9 \cdot t\_9\\
t_12 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_11\right)}}\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_11:\\
\;\;\;\;t\_12 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_12 \cdot t\_9\\
\end{array}\\
\mathbf{if}\;t\_13 \leq -0.007000000216066837:\\
\;\;\;\;t\_10\\
\mathbf{elif}\;t\_13 \leq 1.0000000116860974 \cdot 10^{-7}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_6:\\
\;\;\;\;t\_8 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_9\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\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))) < -0.00700000022 or 1.00000001e-7 < (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 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
Applied rewrites59.0%
Taylor expanded in w around 0
Applied rewrites64.8%
if -0.00700000022 < (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))) < 1.00000001e-7Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in dY.u around inf
Applied rewrites48.6%
Taylor expanded in dY.u around inf
Applied rewrites51.8%
Taylor expanded in dY.u around inf
Applied rewrites62.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) (floor h)))
(t_4 (* (* dY.v dY.v) t_3))
(t_5 (+ (* t_1 t_1) (* t_0 t_0)))
(t_6 (* (floor h) dY.v))
(t_7 (* (* (floor h) t_6) dY.v))
(t_8 (/ 1.0 (sqrt (fmax t_5 t_7))))
(t_9 (+ (* t_2 t_2) (* t_6 t_6)))
(t_10 (/ 1.0 (sqrt (fmax t_5 t_9))))
(t_11 (if (>= t_5 t_9) (* t_10 t_0) (* t_10 t_6)))
(t_12 (* (fma (floor w) (floor w) (/ 0.0 dY.u)) (* dY.u dY.u)))
(t_13 (/ 1.0 (sqrt (fmax t_5 t_12)))))
(if (<= t_11 -0.5)
(if (>= t_5 t_4)
(* (/ 1.0 (sqrt (fmax t_5 t_4))) t_0)
(/
(* 1.0 t_6)
(sqrt
(fmax
(fma t_3 (* dX.v dX.v) (* (* (floor w) (floor w)) (* dX.u dX.u)))
t_4))))
(if (<= t_11 0.5)
(if (>= t_5 t_12) (* t_13 t_0) (* t_13 t_6))
(if (>= t_5 t_7) (* t_8 t_0) (* t_8 t_6))))))
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 = floorf(h) * floorf(h);
float t_4 = (dY_46_v * dY_46_v) * t_3;
float t_5 = (t_1 * t_1) + (t_0 * t_0);
float t_6 = floorf(h) * dY_46_v;
float t_7 = (floorf(h) * t_6) * dY_46_v;
float t_8 = 1.0f / sqrtf(fmaxf(t_5, t_7));
float t_9 = (t_2 * t_2) + (t_6 * t_6);
float t_10 = 1.0f / sqrtf(fmaxf(t_5, t_9));
float tmp;
if (t_5 >= t_9) {
tmp = t_10 * t_0;
} else {
tmp = t_10 * t_6;
}
float t_11 = tmp;
float t_12 = fmaf(floorf(w), floorf(w), (0.0f / dY_46_u)) * (dY_46_u * dY_46_u);
float t_13 = 1.0f / sqrtf(fmaxf(t_5, t_12));
float tmp_2;
if (t_11 <= -0.5f) {
float tmp_3;
if (t_5 >= t_4) {
tmp_3 = (1.0f / sqrtf(fmaxf(t_5, t_4))) * t_0;
} else {
tmp_3 = (1.0f * t_6) / sqrtf(fmaxf(fmaf(t_3, (dX_46_v * dX_46_v), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), t_4));
}
tmp_2 = tmp_3;
} else if (t_11 <= 0.5f) {
float tmp_4;
if (t_5 >= t_12) {
tmp_4 = t_13 * t_0;
} else {
tmp_4 = t_13 * t_6;
}
tmp_2 = tmp_4;
} else if (t_5 >= t_7) {
tmp_2 = t_8 * t_0;
} else {
tmp_2 = t_8 * t_6;
}
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(floor(h) * floor(h)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * t_3) 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(floor(h) * t_6) * dY_46_v) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_7))) t_9 = Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_9))) tmp = Float32(0.0) if (t_5 >= t_9) tmp = Float32(t_10 * t_0); else tmp = Float32(t_10 * t_6); end t_11 = tmp t_12 = Float32(fma(floor(w), floor(w), Float32(Float32(0.0) / dY_46_u)) * Float32(dY_46_u * dY_46_u)) t_13 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_12))) tmp_2 = Float32(0.0) if (t_11 <= Float32(-0.5)) tmp_3 = Float32(0.0) if (t_5 >= t_4) tmp_3 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, t_4))) * t_0); else tmp_3 = Float32(Float32(Float32(1.0) * t_6) / sqrt(fmax(fma(t_3, Float32(dX_46_v * dX_46_v), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), t_4))); end tmp_2 = tmp_3; elseif (t_11 <= Float32(0.5)) tmp_4 = Float32(0.0) if (t_5 >= t_12) tmp_4 = Float32(t_13 * t_0); else tmp_4 = Float32(t_13 * t_6); end tmp_2 = tmp_4; elseif (t_5 >= t_7) tmp_2 = Float32(t_8 * t_0); else tmp_2 = Float32(t_8 * t_6); 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 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(dY.v \cdot dY.v\right) \cdot t\_3\\
t_5 := t\_1 \cdot t\_1 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := \left(\left\lfloor h\right\rfloor \cdot t\_6\right) \cdot dY.v\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}}\\
t_9 := t\_2 \cdot t\_2 + t\_6 \cdot t\_6\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_9\right)}}\\
t_11 := \begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_9:\\
\;\;\;\;t\_10 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_6\\
\end{array}\\
t_12 := \mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \frac{0}{dY.u}\right) \cdot \left(dY.u \cdot dY.u\right)\\
t_13 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_12\right)}}\\
\mathbf{if}\;t\_11 \leq -0.5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_4\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1 \cdot t\_6}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.v \cdot dX.v, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), t\_4\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_11 \leq 0.5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_12:\\
\;\;\;\;t\_13 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_13 \cdot t\_6\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_7:\\
\;\;\;\;t\_8 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_6\\
\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))) < -0.5Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Applied rewrites58.7%
if -0.5 < (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))) < 0.5Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in dY.u around inf
Applied rewrites48.6%
Taylor expanded in dY.u around inf
Applied rewrites51.8%
Taylor expanded in dY.u around inf
Applied rewrites62.1%
if 0.5 < (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 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.8
Applied rewrites58.8%
(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 (* (floor h) (floor h)))
(t_4 (* (* dY.v dY.v) t_3))
(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 (/ 1.0 (sqrt (fmax t_5 t_7))))
(t_9 (* (* (floor h) t_6) dY.v))
(t_10 (if (>= t_5 t_7) (* t_8 t_0) (* t_8 t_6)))
(t_11 (/ 1.0 (sqrt (fmax t_5 t_9))))
(t_12 (* (* (fma (floor w) (floor w) 0.0) dY.u) dY.u))
(t_13 (/ 1.0 (sqrt (fmax t_5 t_12)))))
(if (<= t_10 -0.5)
(if (>= t_5 t_4)
(* (/ 1.0 (sqrt (fmax t_5 t_4))) t_0)
(/
(* 1.0 t_6)
(sqrt
(fmax
(fma t_3 (* dX.v dX.v) (* (* (floor w) (floor w)) (* dX.u dX.u)))
t_4))))
(if (<= t_10 0.5)
(if (>= t_5 t_12) (* t_13 t_0) (* t_13 t_6))
(if (>= t_5 t_9) (* t_11 t_0) (* t_11 t_6))))))
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 = floorf(h) * floorf(h);
float t_4 = (dY_46_v * dY_46_v) * t_3;
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 = 1.0f / sqrtf(fmaxf(t_5, t_7));
float t_9 = (floorf(h) * t_6) * dY_46_v;
float tmp;
if (t_5 >= t_7) {
tmp = t_8 * t_0;
} else {
tmp = t_8 * t_6;
}
float t_10 = tmp;
float t_11 = 1.0f / sqrtf(fmaxf(t_5, t_9));
float t_12 = (fmaf(floorf(w), floorf(w), 0.0f) * dY_46_u) * dY_46_u;
float t_13 = 1.0f / sqrtf(fmaxf(t_5, t_12));
float tmp_2;
if (t_10 <= -0.5f) {
float tmp_3;
if (t_5 >= t_4) {
tmp_3 = (1.0f / sqrtf(fmaxf(t_5, t_4))) * t_0;
} else {
tmp_3 = (1.0f * t_6) / sqrtf(fmaxf(fmaf(t_3, (dX_46_v * dX_46_v), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u))), t_4));
}
tmp_2 = tmp_3;
} else if (t_10 <= 0.5f) {
float tmp_4;
if (t_5 >= t_12) {
tmp_4 = t_13 * t_0;
} else {
tmp_4 = t_13 * t_6;
}
tmp_2 = tmp_4;
} else if (t_5 >= t_9) {
tmp_2 = t_11 * t_0;
} else {
tmp_2 = t_11 * t_6;
}
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(floor(h) * floor(h)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * t_3) 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(Float32(1.0) / sqrt(fmax(t_5, t_7))) t_9 = Float32(Float32(floor(h) * t_6) * dY_46_v) tmp = Float32(0.0) if (t_5 >= t_7) tmp = Float32(t_8 * t_0); else tmp = Float32(t_8 * t_6); end t_10 = tmp t_11 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_9))) t_12 = Float32(Float32(fma(floor(w), floor(w), Float32(0.0)) * dY_46_u) * dY_46_u) t_13 = Float32(Float32(1.0) / sqrt(fmax(t_5, t_12))) tmp_2 = Float32(0.0) if (t_10 <= Float32(-0.5)) tmp_3 = Float32(0.0) if (t_5 >= t_4) tmp_3 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, t_4))) * t_0); else tmp_3 = Float32(Float32(Float32(1.0) * t_6) / sqrt(fmax(fma(t_3, Float32(dX_46_v * dX_46_v), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))), t_4))); end tmp_2 = tmp_3; elseif (t_10 <= Float32(0.5)) tmp_4 = Float32(0.0) if (t_5 >= t_12) tmp_4 = Float32(t_13 * t_0); else tmp_4 = Float32(t_13 * t_6); end tmp_2 = tmp_4; elseif (t_5 >= t_9) tmp_2 = Float32(t_11 * t_0); else tmp_2 = Float32(t_11 * t_6); 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 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(dY.v \cdot dY.v\right) \cdot t\_3\\
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 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_7\right)}}\\
t_9 := \left(\left\lfloor h\right\rfloor \cdot t\_6\right) \cdot dY.v\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_7:\\
\;\;\;\;t\_8 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_6\\
\end{array}\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_9\right)}}\\
t_12 := \left(\mathsf{fma}\left(\left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , 0\right) \cdot dY.u\right) \cdot dY.u\\
t_13 := \frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_12\right)}}\\
\mathbf{if}\;t\_10 \leq -0.5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_4\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1 \cdot t\_6}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.v \cdot dX.v, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right), t\_4\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_10 \leq 0.5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_12:\\
\;\;\;\;t\_13 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_13 \cdot t\_6\\
\end{array}\\
\mathbf{elif}\;t\_5 \geq t\_9:\\
\;\;\;\;t\_11 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_6\\
\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))) < -0.5Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Applied rewrites58.7%
if -0.5 < (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))) < 0.5Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in dY.u around inf
*-commutativeN/A
Applied rewrites48.6%
Taylor expanded in dY.u around inf
*-commutativeN/A
Applied rewrites51.8%
Taylor expanded in dY.u around inf
*-commutativeN/A
Applied rewrites62.1%
if 0.5 < (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 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.7
Applied rewrites58.7%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3258.8
Applied rewrites58.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dY.v))
(t_2 (* (floor w) (floor w)))
(t_3 (* (floor h) (floor h)))
(t_4 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_5 (* (floor h) dX.v))
(t_6 (* t_0 t_0)))
(if (>= (fma (* t_3 dX.v) dX.v t_6) t_4)
(/
(* 1.0 t_5)
(sqrt
(fmax
(fma (* t_5 (floor h)) dX.v t_6)
(fma (* t_1 (floor h)) dY.v (* (* dY.u dY.u) t_2)))))
(/
(* t_1 -1.0)
(- (sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_3 (* dX.v dX.v))) t_4)))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = floorf(w) * floorf(w);
float t_3 = floorf(h) * floorf(h);
float t_4 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_3));
float t_5 = floorf(h) * dX_46_v;
float t_6 = t_0 * t_0;
float tmp;
if (fmaf((t_3 * dX_46_v), dX_46_v, t_6) >= t_4) {
tmp = (1.0f * t_5) / sqrtf(fmaxf(fmaf((t_5 * floorf(h)), dX_46_v, t_6), fmaf((t_1 * floorf(h)), dY_46_v, ((dY_46_u * dY_46_u) * t_2))));
} else {
tmp = (t_1 * -1.0f) / -sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v))), t_4));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(floor(h) * floor(h)) t_4 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_3)) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(t_0 * t_0) tmp = Float32(0.0) if (fma(Float32(t_3 * dX_46_v), dX_46_v, t_6) >= t_4) tmp = Float32(Float32(Float32(1.0) * t_5) / sqrt(fmax(fma(Float32(t_5 * floor(h)), dX_46_v, t_6), fma(Float32(t_1 * floor(h)), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_2))))); else tmp = Float32(Float32(t_1 * Float32(-1.0)) / Float32(-sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))), t_4)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_3\right)\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_0 \cdot t\_0\\
\mathbf{if}\;\mathsf{fma}\left(t\_3 \cdot dX.v, dX.v, t\_6\right) \geq t\_4:\\
\;\;\;\;\frac{1 \cdot t\_5}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5 \cdot \left\lfloor h\right\rfloor , dX.v, t\_6\right), \mathsf{fma}\left(t\_1 \cdot \left\lfloor h\right\rfloor , dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1 \cdot -1}{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right), t\_4\right)}}\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Applied rewrites76.0%
Applied rewrites76.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3276.1
Applied rewrites76.1%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3276.1
Applied rewrites76.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (* (floor w) dY.u) (floor w)) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (fma t_0 (* dX.v dX.v) (* (* (floor w) (floor w)) (* dX.u dX.u))))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1) (/ (* (floor h) dX.v) t_3) (/ (* (floor h) 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 = floorf(h) * floorf(h);
float t_1 = fmaf(((floorf(w) * dY_46_u) * floorf(w)), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = fmaf(t_0, (dX_46_v * dX_46_v), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u)));
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(h) * dX_46_v) / t_3;
} else {
tmp = (floorf(h) * 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(floor(h) * floor(h)) t_1 = fma(Float32(Float32(floor(w) * dY_46_u) * floor(w)), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = fma(t_0, Float32(dX_46_v * dX_46_v), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))) t_3 = sqrt(fmax(t_2, t_1)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(h) * dX_46_v) / t_3); else tmp = Float32(Float32(floor(h) * dY_46_v) / t_3); 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 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \mathsf{fma}\left(t\_0, dX.v \cdot dX.v, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_3}\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites76.0%
Applied rewrites76.0%
Applied rewrites76.1%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* (floor w) (floor w)) dY.u)
dY.u
(* (* (* (floor h) dY.v) (floor h)) dY.v)))
(t_1
(fma
(* (floor w) (* (floor w) dX.u))
dX.u
(* (* (floor h) (floor h)) (* dX.v dX.v))))
(t_2 (/ (floor h) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.v) (* t_2 dY.v))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, (((floorf(h) * dY_46_v) * floorf(h)) * dY_46_v));
float t_1 = fmaf((floorf(w) * (floorf(w) * dX_46_u)), dX_46_u, ((floorf(h) * floorf(h)) * (dX_46_v * dX_46_v)));
float t_2 = floorf(h) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_v;
} else {
tmp = t_2 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(Float32(floor(h) * dY_46_v) * floor(h)) * dY_46_v)) t_1 = fma(Float32(floor(w) * Float32(floor(w) * dX_46_u)), dX_46_u, Float32(Float32(floor(h) * floor(h)) * Float32(dX_46_v * dX_46_v))) t_2 = Float32(floor(h) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_v); else tmp = Float32(t_2 * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right)\\
t_1 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right), dX.u, \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_2 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
unswap-sqrN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (* (floor w) (floor w)) dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (fma (* (floor w) (* (floor w) dX.u)) dX.u (* t_0 (* dX.v dX.v))))
(t_3 (/ (floor h) (sqrt (fmax t_2 t_1)))))
(if (>= t_2 t_1) (* t_3 dX.v) (* 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(h) * floorf(h);
float t_1 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = fmaf((floorf(w) * (floorf(w) * dX_46_u)), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_3 = floorf(h) / sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_3 * dX_46_v;
} else {
tmp = 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(h) * floor(h)) t_1 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = fma(Float32(floor(w) * Float32(floor(w) * dX_46_u)), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_3 = Float32(floor(h) / sqrt(fmax(t_2, t_1))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); 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 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right), dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (floor w) (floor w)) (* dY.u dY.u) (* (* dY.v dY.v) t_0)))
(t_2 (fma (* (floor w) (* (floor w) dX.u)) dX.u (* t_0 (* dX.v dX.v))))
(t_3 (/ (floor h) (sqrt (fmax t_2 t_1)))))
(if (>= t_2 t_1) (* t_3 dX.v) (* 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(h) * floorf(h);
float t_1 = fmaf((floorf(w) * floorf(w)), (dY_46_u * dY_46_u), ((dY_46_v * dY_46_v) * t_0));
float t_2 = fmaf((floorf(w) * (floorf(w) * dX_46_u)), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_3 = floorf(h) / sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_3 * dX_46_v;
} else {
tmp = 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(h) * floor(h)) t_1 = fma(Float32(floor(w) * floor(w)), Float32(dY_46_u * dY_46_u), Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = fma(Float32(floor(w) * Float32(floor(w) * dX_46_u)), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_3 = Float32(floor(h) / sqrt(fmax(t_2, t_1))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); 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 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , dY.u \cdot dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \mathsf{fma}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right), dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-*.f32N/A
lift-floor.f3275.9
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
associate-*l*N/A
unpow2N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
lower-fma.f32N/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
associate-*l*N/A
unpow2N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
lower-fma.f32N/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
associate-*l*N/A
unpow2N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
pow2N/A
pow2N/A
lower-fma.f32N/A
Applied rewrites75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* dY.v dY.v) t_0))
(t_2 (fma t_0 (* dX.v dX.v) (* (* (floor w) (floor w)) (* dX.u dX.u))))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1)
(/ (* (floor h) dX.v) t_3)
(* (/ 1.0 t_3) (* (floor h) 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(h) * floorf(h);
float t_1 = (dY_46_v * dY_46_v) * t_0;
float t_2 = fmaf(t_0, (dX_46_v * dX_46_v), ((floorf(w) * floorf(w)) * (dX_46_u * dX_46_u)));
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(h) * dX_46_v) / t_3;
} else {
tmp = (1.0f / t_3) * (floorf(h) * 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(h) * floor(h)) t_1 = Float32(Float32(dY_46_v * dY_46_v) * t_0) t_2 = fma(t_0, Float32(dX_46_v * dX_46_v), Float32(Float32(floor(w) * floor(w)) * Float32(dX_46_u * dX_46_u))) t_3 = sqrt(fmax(t_2, t_1)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(h) * dX_46_v) / t_3); else tmp = Float32(Float32(Float32(1.0) / t_3) * Float32(floor(h) * dY_46_v)); 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(dY.v \cdot dY.v\right) \cdot t\_0\\
t_2 := \mathsf{fma}\left(t\_0, dX.v \cdot dX.v, \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left(dX.u \cdot dX.u\right)\right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_3} \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)\\
\end{array}
\end{array}
Initial program 76.0%
Applied rewrites69.3%
Applied rewrites69.3%
Applied rewrites46.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites43.3%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites49.1%
Taylor expanded in dY.u around 0
+-commutativeN/A
pow2N/A
associate-*r*N/A
lower-fma.f32N/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
*-commutativeN/A
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Taylor expanded in w around 0
pow2N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3258.7
Applied rewrites58.7%
Applied rewrites58.7%
herbie shell --seed 2025139
(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))))