
(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 15 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor w) dY.u))
(t_5 (* (floor w) (floor w)))
(t_6 (fma (* t_5 dX.u) dX.u (* t_3 (* dX.v dX.v))))
(t_7 (fma (* t_5 dY.u) dY.u (* (* dY.v dY.v) t_3)))
(t_8 (sqrt (fmax t_6 t_7)))
(t_9 (* (* dY.u dY.u) t_5))
(t_10 (+ (* t_0 t_0) (* t_2 t_2)))
(t_11 (* (floor h) dY.v))
(t_12 (+ (* t_4 t_4) (* t_11 t_11)))
(t_13 (/ 1.0 (sqrt (fmax t_10 t_12))))
(t_14 (if (>= t_10 t_12) (* t_13 t_0) (* t_13 t_4)))
(t_15 (* (* t_11 dY.v) (floor h))))
(if (<= t_14 -0.00019999999494757503)
(if (>= (* (* dX.u dX.u) t_5) t_9)
(/ t_0 (sqrt (fmax (fma t_0 t_0 t_1) t_7)))
(/ t_4 t_8))
(if (<= t_14 4.999999858590343e-10)
(if (>= t_6 t_7)
(/ t_0 (sqrt (fmax t_1 t_7)))
(/ t_4 (sqrt (fmax t_6 t_15))))
(if (>= t_6 t_9)
(/ t_0 t_8)
(/ t_4 (sqrt (fmax t_6 (fma t_4 t_4 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(w) * dX_46_u;
float t_1 = ((dX_46_v * floorf(h)) * dX_46_v) * floorf(h);
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(w) * dY_46_u;
float t_5 = floorf(w) * floorf(w);
float t_6 = fmaf((t_5 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v)));
float t_7 = fmaf((t_5 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_3));
float t_8 = sqrtf(fmaxf(t_6, t_7));
float t_9 = (dY_46_u * dY_46_u) * t_5;
float t_10 = (t_0 * t_0) + (t_2 * t_2);
float t_11 = floorf(h) * dY_46_v;
float t_12 = (t_4 * t_4) + (t_11 * t_11);
float t_13 = 1.0f / sqrtf(fmaxf(t_10, t_12));
float tmp;
if (t_10 >= t_12) {
tmp = t_13 * t_0;
} else {
tmp = t_13 * t_4;
}
float t_14 = tmp;
float t_15 = (t_11 * dY_46_v) * floorf(h);
float tmp_2;
if (t_14 <= -0.00019999999494757503f) {
float tmp_3;
if (((dX_46_u * dX_46_u) * t_5) >= t_9) {
tmp_3 = t_0 / sqrtf(fmaxf(fmaf(t_0, t_0, t_1), t_7));
} else {
tmp_3 = t_4 / t_8;
}
tmp_2 = tmp_3;
} else if (t_14 <= 4.999999858590343e-10f) {
float tmp_4;
if (t_6 >= t_7) {
tmp_4 = t_0 / sqrtf(fmaxf(t_1, t_7));
} else {
tmp_4 = t_4 / sqrtf(fmaxf(t_6, t_15));
}
tmp_2 = tmp_4;
} else if (t_6 >= t_9) {
tmp_2 = t_0 / t_8;
} else {
tmp_2 = t_4 / sqrtf(fmaxf(t_6, fmaf(t_4, t_4, 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(w) * dX_46_u) t_1 = Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(w) * dY_46_u) t_5 = Float32(floor(w) * floor(w)) t_6 = fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))) t_7 = fma(Float32(t_5 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_3)) t_8 = sqrt(fmax(t_6, t_7)) t_9 = Float32(Float32(dY_46_u * dY_46_u) * t_5) t_10 = Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) t_11 = Float32(floor(h) * dY_46_v) t_12 = Float32(Float32(t_4 * t_4) + Float32(t_11 * t_11)) t_13 = Float32(Float32(1.0) / sqrt(fmax(t_10, t_12))) tmp = Float32(0.0) if (t_10 >= t_12) tmp = Float32(t_13 * t_0); else tmp = Float32(t_13 * t_4); end t_14 = tmp t_15 = Float32(Float32(t_11 * dY_46_v) * floor(h)) tmp_2 = Float32(0.0) if (t_14 <= Float32(-0.00019999999494757503)) tmp_3 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_5) >= t_9) tmp_3 = Float32(t_0 / sqrt(fmax(fma(t_0, t_0, t_1), t_7))); else tmp_3 = Float32(t_4 / t_8); end tmp_2 = tmp_3; elseif (t_14 <= Float32(4.999999858590343e-10)) tmp_4 = Float32(0.0) if (t_6 >= t_7) tmp_4 = Float32(t_0 / sqrt(fmax(t_1, t_7))); else tmp_4 = Float32(t_4 / sqrt(fmax(t_6, t_15))); end tmp_2 = tmp_4; elseif (t_6 >= t_9) tmp_2 = Float32(t_0 / t_8); else tmp_2 = Float32(t_4 / sqrt(fmax(t_6, fma(t_4, t_4, t_15)))); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_5 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_6 := \mathsf{fma}\left(t\_5 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_7 := \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_3\right)\\
t_8 := \sqrt{\mathsf{max}\left(t\_6, t\_7\right)}\\
t_9 := \left(dY.u \cdot dY.u\right) \cdot t\_5\\
t_10 := t\_0 \cdot t\_0 + t\_2 \cdot t\_2\\
t_11 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_12 := t\_4 \cdot t\_4 + t\_11 \cdot t\_11\\
t_13 := \frac{1}{\sqrt{\mathsf{max}\left(t\_10, t\_12\right)}}\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_10 \geq t\_12:\\
\;\;\;\;t\_13 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_13 \cdot t\_4\\
\end{array}\\
t_15 := \left(t\_11 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;t\_14 \leq -0.00019999999494757503:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_5 \geq t\_9:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_1\right), t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{t\_8}\\
\end{array}\\
\mathbf{elif}\;t\_14 \leq 4.999999858590343 \cdot 10^{-10}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_7:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_1, t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_6, t\_15\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_9:\\
\;\;\;\;\frac{t\_0}{t\_8}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_4, t\_4, t\_15\right)\right)}}\\
\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))) < -1.99999995e-4Initial program 99.3%
Applied rewrites99.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites98.7%
Taylor expanded in dY.u around inf
Applied rewrites98.6%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites98.8%
if -1.99999995e-4 < (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))) < 4.99999986e-10Initial program 56.7%
Applied rewrites56.8%
Taylor expanded in dY.u around 0
Applied rewrites57.5%
Taylor expanded in dX.u around 0
Applied rewrites58.3%
if 4.99999986e-10 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dX.u)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 w) dY.u))) Initial program 99.2%
Applied rewrites99.2%
Applied rewrites99.4%
Taylor expanded in dY.u around inf
Applied rewrites96.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (* (* dX.v (floor h)) dX.v) (floor h)))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor h) (floor h)))
(t_4 (* (* dY.v dY.v) t_3))
(t_5 (* (floor w) dY.u))
(t_6 (* (floor w) (floor w)))
(t_7 (>= (* (* dX.u dX.u) t_6) (* (* dY.u dY.u) t_6)))
(t_8 (fma (* t_6 dX.u) dX.u (* t_3 (* dX.v dX.v))))
(t_9 (fma (* t_6 dY.u) dY.u t_4))
(t_10 (sqrt (fmax t_8 t_9)))
(t_11 (/ t_0 t_10))
(t_12 (+ (* t_0 t_0) (* t_2 t_2)))
(t_13 (* (floor h) dY.v))
(t_14 (+ (* t_5 t_5) (* t_13 t_13)))
(t_15 (/ 1.0 (sqrt (fmax t_12 t_14))))
(t_16 (if (>= t_12 t_14) (* t_15 t_0) (* t_15 t_5))))
(if (<= t_16 -0.019999999552965164)
(if t_7 (/ t_0 (sqrt (fmax (fma t_0 t_0 t_1) t_9))) (/ t_5 t_10))
(if (<= t_16 0.25)
(if (>= t_1 t_9)
t_11
(/ t_5 (sqrt (fmax t_8 (* (* t_13 dY.v) (floor h))))))
(if t_7 t_11 (/ t_5 (sqrt (fmax t_8 (fma t_5 t_5 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 = ((dX_46_v * floorf(h)) * dX_46_v) * floorf(h);
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(h) * floorf(h);
float t_4 = (dY_46_v * dY_46_v) * t_3;
float t_5 = floorf(w) * dY_46_u;
float t_6 = floorf(w) * floorf(w);
int t_7 = ((dX_46_u * dX_46_u) * t_6) >= ((dY_46_u * dY_46_u) * t_6);
float t_8 = fmaf((t_6 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v)));
float t_9 = fmaf((t_6 * dY_46_u), dY_46_u, t_4);
float t_10 = sqrtf(fmaxf(t_8, t_9));
float t_11 = t_0 / t_10;
float t_12 = (t_0 * t_0) + (t_2 * t_2);
float t_13 = floorf(h) * dY_46_v;
float t_14 = (t_5 * t_5) + (t_13 * t_13);
float t_15 = 1.0f / sqrtf(fmaxf(t_12, t_14));
float tmp;
if (t_12 >= t_14) {
tmp = t_15 * t_0;
} else {
tmp = t_15 * t_5;
}
float t_16 = tmp;
float tmp_2;
if (t_16 <= -0.019999999552965164f) {
float tmp_3;
if (t_7) {
tmp_3 = t_0 / sqrtf(fmaxf(fmaf(t_0, t_0, t_1), t_9));
} else {
tmp_3 = t_5 / t_10;
}
tmp_2 = tmp_3;
} else if (t_16 <= 0.25f) {
float tmp_4;
if (t_1 >= t_9) {
tmp_4 = t_11;
} else {
tmp_4 = t_5 / sqrtf(fmaxf(t_8, ((t_13 * dY_46_v) * floorf(h))));
}
tmp_2 = tmp_4;
} else if (t_7) {
tmp_2 = t_11;
} else {
tmp_2 = t_5 / sqrtf(fmaxf(t_8, fmaf(t_5, t_5, t_4)));
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * t_3) t_5 = Float32(floor(w) * dY_46_u) t_6 = Float32(floor(w) * floor(w)) t_7 = Float32(Float32(dX_46_u * dX_46_u) * t_6) >= Float32(Float32(dY_46_u * dY_46_u) * t_6) t_8 = fma(Float32(t_6 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))) t_9 = fma(Float32(t_6 * dY_46_u), dY_46_u, t_4) t_10 = sqrt(fmax(t_8, t_9)) t_11 = Float32(t_0 / t_10) t_12 = Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) t_13 = Float32(floor(h) * dY_46_v) t_14 = Float32(Float32(t_5 * t_5) + Float32(t_13 * t_13)) t_15 = Float32(Float32(1.0) / sqrt(fmax(t_12, t_14))) tmp = Float32(0.0) if (t_12 >= t_14) tmp = Float32(t_15 * t_0); else tmp = Float32(t_15 * t_5); end t_16 = tmp tmp_2 = Float32(0.0) if (t_16 <= Float32(-0.019999999552965164)) tmp_3 = Float32(0.0) if (t_7) tmp_3 = Float32(t_0 / sqrt(fmax(fma(t_0, t_0, t_1), t_9))); else tmp_3 = Float32(t_5 / t_10); end tmp_2 = tmp_3; elseif (t_16 <= Float32(0.25)) tmp_4 = Float32(0.0) if (t_1 >= t_9) tmp_4 = t_11; else tmp_4 = Float32(t_5 / sqrt(fmax(t_8, Float32(Float32(t_13 * dY_46_v) * floor(h))))); end tmp_2 = tmp_4; elseif (t_7) tmp_2 = t_11; else tmp_2 = Float32(t_5 / sqrt(fmax(t_8, fma(t_5, t_5, t_4)))); end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
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 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_6 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_7 := \left(dX.u \cdot dX.u\right) \cdot t\_6 \geq \left(dY.u \cdot dY.u\right) \cdot t\_6\\
t_8 := \mathsf{fma}\left(t\_6 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_9 := \mathsf{fma}\left(t\_6 \cdot dY.u, dY.u, t\_4\right)\\
t_10 := \sqrt{\mathsf{max}\left(t\_8, t\_9\right)}\\
t_11 := \frac{t\_0}{t\_10}\\
t_12 := t\_0 \cdot t\_0 + t\_2 \cdot t\_2\\
t_13 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_14 := t\_5 \cdot t\_5 + t\_13 \cdot t\_13\\
t_15 := \frac{1}{\sqrt{\mathsf{max}\left(t\_12, t\_14\right)}}\\
t_16 := \begin{array}{l}
\mathbf{if}\;t\_12 \geq t\_14:\\
\;\;\;\;t\_15 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_15 \cdot t\_5\\
\end{array}\\
\mathbf{if}\;t\_16 \leq -0.019999999552965164:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, t\_0, t\_1\right), t\_9\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{t\_10}\\
\end{array}\\
\mathbf{elif}\;t\_16 \leq 0.25:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_1 \geq t\_9:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(t\_8, \left(t\_13 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)}}\\
\end{array}\\
\mathbf{elif}\;t\_7:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(t\_8, \mathsf{fma}\left(t\_5, t\_5, t\_4\right)\right)}}\\
\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))) < -0.0199999996Initial program 99.4%
Applied rewrites99.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites99.2%
Taylor expanded in dY.u around inf
Applied rewrites99.2%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites99.4%
if -0.0199999996 < (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))) < 0.25Initial program 61.7%
Applied rewrites61.8%
Taylor expanded in dY.u around 0
Applied rewrites62.0%
Taylor expanded in dX.u around 0
Applied rewrites62.0%
if 0.25 < (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 rewrites99.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites99.3%
Taylor expanded in dY.u around inf
Applied rewrites99.2%
Applied rewrites99.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) (floor h)))
(t_2 (fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_1)))
(t_3 (fma (* t_0 dX.u) dX.u (* t_1 (* dX.v dX.v))))
(t_4 (* (floor w) dY.u)))
(if (>= t_3 t_2)
(/ (* (floor w) dX.u) (sqrt (fmax t_3 t_2)))
(/
t_4
(sqrt
(fmax t_3 (fma t_4 t_4 (* (* (* (floor h) dY.v) dY.v) (floor h)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_1));
float t_3 = fmaf((t_0 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v)));
float t_4 = floorf(w) * dY_46_u;
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_3, t_2));
} else {
tmp = t_4 / sqrtf(fmaxf(t_3, fmaf(t_4, t_4, (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h)))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_1)) t_3 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))) t_4 = Float32(floor(w) * dY_46_u) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_3, t_2))); else tmp = Float32(t_4 / sqrt(fmax(t_3, fma(t_4, t_4, Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_1\right)\\
t_3 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot dY.u\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_4, t\_4, \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_4 (* (floor w) dX.u)))
(if (>= t_3 t_2)
(/ t_4 (sqrt (fmax (fma t_4 t_4 (* (* dX.v dX.v) t_0)) t_2)))
(/ (* (floor w) dY.u) (sqrt (fmax t_3 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_4 = floorf(w) * dX_46_u;
float tmp;
if (t_3 >= t_2) {
tmp = t_4 / sqrtf(fmaxf(fmaf(t_4, t_4, ((dX_46_v * dX_46_v) * t_0)), t_2));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_3, t_2));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_4 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_4 / sqrt(fmax(fma(t_4, t_4, Float32(Float32(dX_46_v * dX_46_v) * t_0)), t_2))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_3, t_2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_4, t\_4, \left(dX.v \cdot dX.v\right) \cdot t\_0\right), t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f3276.4
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-*.f3276.4
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(*
(/
dX.u
(sqrt
(fmax
(fma (* (* (floor h) dX.v) dX.v) (floor h) (* (* dX.u dX.u) t_2))
(fma (* dY.v dY.v) t_0 (* (* dY.u dY.u) t_2)))))
(floor w))
(/
t_1
(sqrt
(fmax t_3 (fma t_1 t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0))) {
tmp = (dX_46_u / sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_2)), fmaf((dY_46_v * dY_46_v), t_0, ((dY_46_u * dY_46_u) * t_2))))) * floorf(w);
} else {
tmp = t_1 / sqrtf(fmaxf(t_3, fmaf(t_1, t_1, (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h)))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0))) tmp = Float32(Float32(dX_46_u / sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_2)), fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_2))))) * floor(w)); else tmp = Float32(t_1 / sqrt(fmax(t_3, fma(t_1, t_1, Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right):\\
\;\;\;\;\frac{dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_2\right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Applied rewrites76.4%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(*
dX.u
(/
(floor w)
(sqrt
(fmax
(fma (* (* (floor h) dX.v) dX.v) (floor h) (* (* dX.u dX.u) t_2))
(fma (* dY.v dY.v) t_0 (* (* dY.u dY.u) t_2))))))
(/
t_1
(sqrt
(fmax t_3 (fma t_1 t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0))) {
tmp = dX_46_u * (floorf(w) / sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * dX_46_v), floorf(h), ((dX_46_u * dX_46_u) * t_2)), fmaf((dY_46_v * dY_46_v), t_0, ((dY_46_u * dY_46_u) * t_2)))));
} else {
tmp = t_1 / sqrtf(fmaxf(t_3, fmaf(t_1, t_1, (((floorf(h) * dY_46_v) * dY_46_v) * floorf(h)))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0))) tmp = Float32(dX_46_u * Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * dX_46_v), floor(h), Float32(Float32(dX_46_u * dX_46_u) * t_2)), fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_2)))))); else tmp = Float32(t_1 / sqrt(fmax(t_3, fma(t_1, t_1, Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right):\\
\;\;\;\;dX.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(dX.u \cdot dX.u\right) \cdot t\_2\right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_3, \mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Applied rewrites76.4%
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v)))))
(if (>= t_3 t_2)
(*
dX.u
(/
(floor w)
(sqrt
(fmax
(fma (* dX.u dX.u) t_1 (* (* dX.v dX.v) t_0))
(fma (* (* (floor h) dY.v) dY.v) (floor h) (* (* dY.u dY.u) t_1))))))
(/ (* (floor w) dY.u) (sqrt (fmax t_3 t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float tmp;
if (t_3 >= t_2) {
tmp = dX_46_u * (floorf(w) / sqrtf(fmaxf(fmaf((dX_46_u * dX_46_u), t_1, ((dX_46_v * dX_46_v) * t_0)), fmaf(((floorf(h) * dY_46_v) * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_1)))));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_3, t_2));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(dX_46_u * Float32(floor(w) / sqrt(fmax(fma(Float32(dX_46_u * dX_46_u), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_0)), fma(Float32(Float32(floor(h) * dY_46_v) * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_1)))))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_3, t_2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;dX.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_0\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor w) (floor w)))
(t_6 (fma (* t_5 dX.u) dX.u (* t_1 (* dX.v dX.v))))
(t_7 (* (floor w) dX.u))
(t_8 (+ (* t_7 t_7) (* t_0 t_0)))
(t_9 (/ 1.0 (sqrt (fmax t_8 (+ (* t_3 t_3) (* t_4 t_4)))))))
(if (<= dY.v 7000.0)
(if (>= t_6 (* (* dY.u dY.u) t_5))
(/ t_7 (sqrt (fmax t_6 (fma (* t_5 dY.u) dY.u t_2))))
(/ t_3 (sqrt (fmax t_6 (fma t_3 t_3 (* (* t_4 dY.v) (floor h)))))))
(if (>= t_8 t_2) (* t_9 t_7) (* t_9 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) * dX_46_v;
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(w) * floorf(w);
float t_6 = fmaf((t_5 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v)));
float t_7 = floorf(w) * dX_46_u;
float t_8 = (t_7 * t_7) + (t_0 * t_0);
float t_9 = 1.0f / sqrtf(fmaxf(t_8, ((t_3 * t_3) + (t_4 * t_4))));
float tmp_1;
if (dY_46_v <= 7000.0f) {
float tmp_2;
if (t_6 >= ((dY_46_u * dY_46_u) * t_5)) {
tmp_2 = t_7 / sqrtf(fmaxf(t_6, fmaf((t_5 * dY_46_u), dY_46_u, t_2)));
} else {
tmp_2 = t_3 / sqrtf(fmaxf(t_6, fmaf(t_3, t_3, ((t_4 * dY_46_v) * floorf(h)))));
}
tmp_1 = tmp_2;
} else if (t_8 >= t_2) {
tmp_1 = t_9 * t_7;
} else {
tmp_1 = t_9 * t_3;
}
return tmp_1;
}
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(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(w) * floor(w)) t_6 = fma(Float32(t_5 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))) t_7 = Float32(floor(w) * dX_46_u) t_8 = Float32(Float32(t_7 * t_7) + Float32(t_0 * t_0)) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_8, Float32(Float32(t_3 * t_3) + Float32(t_4 * t_4))))) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(7000.0)) tmp_2 = Float32(0.0) if (t_6 >= Float32(Float32(dY_46_u * dY_46_u) * t_5)) tmp_2 = Float32(t_7 / sqrt(fmax(t_6, fma(Float32(t_5 * dY_46_u), dY_46_u, t_2)))); else tmp_2 = Float32(t_3 / sqrt(fmax(t_6, fma(t_3, t_3, Float32(Float32(t_4 * dY_46_v) * floor(h)))))); end tmp_1 = tmp_2; elseif (t_8 >= t_2) tmp_1 = Float32(t_9 * t_7); else tmp_1 = Float32(t_9 * t_3); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_6 := \mathsf{fma}\left(t\_5 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_7 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_8 := t\_7 \cdot t\_7 + t\_0 \cdot t\_0\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_3 \cdot t\_3 + t\_4 \cdot t\_4\right)}}\\
\mathbf{if}\;dY.v \leq 7000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6 \geq \left(dY.u \cdot dY.u\right) \cdot t\_5:\\
\;\;\;\;\frac{t\_7}{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_5 \cdot dY.u, dY.u, t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_3, t\_3, \left(t\_4 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_8 \geq t\_2:\\
\;\;\;\;t\_9 \cdot t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_3\\
\end{array}
\end{array}
if dY.v < 7e3Initial program 77.9%
Applied rewrites77.9%
Applied rewrites78.0%
Taylor expanded in dY.u around inf
Applied rewrites70.3%
if 7e3 < dY.v Initial program 69.8%
Taylor expanded in dY.u around 0
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3266.5
Applied rewrites66.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) (floor w)))
(t_3 (* t_2 dX.u))
(t_4 (* (floor h) (floor h)))
(t_5 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_4)))
(t_6 (fma t_3 dX.u (* t_4 (* dX.v dX.v))))
(t_7 (sqrt (fmax t_6 t_5)))
(t_8 (* (floor h) dY.v)))
(if (<= dX.v 0.0005000000237487257)
(if (>= (* (* dX.u dX.u) t_2) t_5)
(/ t_1 t_7)
(/ t_0 (sqrt (fmax t_6 (fma t_8 t_8 (* (* dY.u dY.u) t_2))))))
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) t_5)
(/ t_1 (sqrt (fmax (fma t_3 dX.u (* (* t_4 (- dX.v)) (- dX.v))) t_5)))
(/ t_0 t_7)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = t_2 * dX_46_u;
float t_4 = floorf(h) * floorf(h);
float t_5 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_4));
float t_6 = fmaf(t_3, dX_46_u, (t_4 * (dX_46_v * dX_46_v)));
float t_7 = sqrtf(fmaxf(t_6, t_5));
float t_8 = floorf(h) * dY_46_v;
float tmp_1;
if (dX_46_v <= 0.0005000000237487257f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * t_2) >= t_5) {
tmp_2 = t_1 / t_7;
} else {
tmp_2 = t_0 / sqrtf(fmaxf(t_6, fmaf(t_8, t_8, ((dY_46_u * dY_46_u) * t_2))));
}
tmp_1 = tmp_2;
} else if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= t_5) {
tmp_1 = t_1 / sqrtf(fmaxf(fmaf(t_3, dX_46_u, ((t_4 * -dX_46_v) * -dX_46_v)), t_5));
} else {
tmp_1 = t_0 / t_7;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(t_2 * dX_46_u) t_4 = Float32(floor(h) * floor(h)) t_5 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_4)) t_6 = fma(t_3, dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))) t_7 = sqrt(fmax(t_6, t_5)) t_8 = Float32(floor(h) * dY_46_v) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(0.0005000000237487257)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= t_5) tmp_2 = Float32(t_1 / t_7); else tmp_2 = Float32(t_0 / sqrt(fmax(t_6, fma(t_8, t_8, Float32(Float32(dY_46_u * dY_46_u) * t_2))))); end tmp_1 = tmp_2; elseif (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= t_5) tmp_1 = Float32(t_1 / sqrt(fmax(fma(t_3, dX_46_u, Float32(Float32(t_4 * Float32(-dX_46_v)) * Float32(-dX_46_v))), t_5))); else tmp_1 = Float32(t_0 / t_7); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := t\_2 \cdot dX.u\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_4\right)\\
t_6 := \mathsf{fma}\left(t\_3, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_7 := \sqrt{\mathsf{max}\left(t\_6, t\_5\right)}\\
t_8 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;dX.v \leq 0.0005000000237487257:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq t\_5:\\
\;\;\;\;\frac{t\_1}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_8, t\_8, \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_5:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_3, dX.u, \left(t\_4 \cdot \left(-dX.v\right)\right) \cdot \left(-dX.v\right)\right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_7}\\
\end{array}
\end{array}
if dX.v < 5.00000024e-4Initial program 77.0%
Applied rewrites77.0%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites68.7%
Applied rewrites68.7%
if 5.00000024e-4 < dX.v Initial program 74.4%
Applied rewrites74.5%
lift-*.f32N/A
lift-*.f32N/A
sqr-neg-revN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-neg.f32N/A
lower-neg.f3274.5
Applied rewrites74.5%
Taylor expanded in dX.u around 0
Applied rewrites70.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dY.u))
(t_2 (* (* dX.u dX.u) t_0))
(t_3 (* (floor h) (floor h)))
(t_4 (* (* dY.v dY.v) t_3))
(t_5 (fma (* t_0 dX.u) dX.u (* t_3 (* dX.v dX.v))))
(t_6 (sqrt (fmax t_5 (fma (* t_0 dY.u) dY.u t_4))))
(t_7 (/ (* (floor w) dX.u) t_6)))
(if (<= dY.v 6000.0)
(if (>= t_2 (* (* dY.u dY.u) t_0))
t_7
(/ t_1 (sqrt (fmax t_5 (fma t_1 t_1 t_4)))))
(if (>= t_2 t_4) t_7 (/ t_1 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(w) * floorf(w);
float t_1 = floorf(w) * dY_46_u;
float t_2 = (dX_46_u * dX_46_u) * t_0;
float t_3 = floorf(h) * floorf(h);
float t_4 = (dY_46_v * dY_46_v) * t_3;
float t_5 = fmaf((t_0 * dX_46_u), dX_46_u, (t_3 * (dX_46_v * dX_46_v)));
float t_6 = sqrtf(fmaxf(t_5, fmaf((t_0 * dY_46_u), dY_46_u, t_4)));
float t_7 = (floorf(w) * dX_46_u) / t_6;
float tmp_1;
if (dY_46_v <= 6000.0f) {
float tmp_2;
if (t_2 >= ((dY_46_u * dY_46_u) * t_0)) {
tmp_2 = t_7;
} else {
tmp_2 = t_1 / sqrtf(fmaxf(t_5, fmaf(t_1, t_1, t_4)));
}
tmp_1 = tmp_2;
} else if (t_2 >= t_4) {
tmp_1 = t_7;
} else {
tmp_1 = t_1 / t_6;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(Float32(dX_46_u * dX_46_u) * t_0) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * t_3) t_5 = fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_3 * Float32(dX_46_v * dX_46_v))) t_6 = sqrt(fmax(t_5, fma(Float32(t_0 * dY_46_u), dY_46_u, t_4))) t_7 = Float32(Float32(floor(w) * dX_46_u) / t_6) tmp_1 = Float32(0.0) if (dY_46_v <= Float32(6000.0)) tmp_2 = Float32(0.0) if (t_2 >= Float32(Float32(dY_46_u * dY_46_u) * t_0)) tmp_2 = t_7; else tmp_2 = Float32(t_1 / sqrt(fmax(t_5, fma(t_1, t_1, t_4)))); end tmp_1 = tmp_2; elseif (t_2 >= t_4) tmp_1 = t_7; else tmp_1 = Float32(t_1 / t_6); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left(dX.u \cdot dX.u\right) \cdot t\_0\\
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 := \mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_3 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_6 := \sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, t\_4\right)\right)}\\
t_7 := \frac{\left\lfloor w\right\rfloor \cdot dX.u}{t\_6}\\
\mathbf{if}\;dY.v \leq 6000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_2 \geq \left(dY.u \cdot dY.u\right) \cdot t\_0:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_1, t\_1, t\_4\right)\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_2 \geq t\_4:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_6}\\
\end{array}
\end{array}
if dY.v < 6e3Initial program 77.8%
Applied rewrites77.8%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.5%
Taylor expanded in dY.u around inf
Applied rewrites62.2%
Applied rewrites62.4%
if 6e3 < dY.v Initial program 70.0%
Applied rewrites70.1%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites63.1%
Taylor expanded in dY.u around 0
Applied rewrites61.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(if (>= (* (* dX.u dX.u) t_2) t_3)
(/
t_1
(sqrt
(fmax (fma t_1 t_1 (* (* (* (floor h) dX.v) dX.v) (floor h))) t_3)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.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 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float tmp;
if (((dX_46_u * dX_46_u) * t_2) >= t_3) {
tmp = t_1 / sqrtf(fmaxf(fmaf(t_1, t_1, (((floorf(h) * dX_46_v) * dX_46_v) * floorf(h))), t_3));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), t_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 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= t_3) tmp = Float32(t_1 / sqrt(fmax(fma(t_1, t_1, Float32(Float32(Float32(floor(h) * dX_46_v) * dX_46_v) * floor(h))), t_3))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_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 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq t\_3:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-fma.f3265.1
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
Applied rewrites65.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (* (* dX.u dX.u) t_1))
(t_3 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(if (>= t_2 t_3)
(*
(/
dX.u
(sqrt
(fmax
(fma (* (* (floor h) dX.v) dX.v) (floor h) t_2)
(fma (* dY.v dY.v) t_0 (* (* dY.u dY.u) t_1)))))
(floor w))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.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 = floorf(w) * floorf(w);
float t_2 = (dX_46_u * dX_46_u) * t_1;
float t_3 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float tmp;
if (t_2 >= t_3) {
tmp = (dX_46_u / sqrtf(fmaxf(fmaf(((floorf(h) * dX_46_v) * dX_46_v), floorf(h), t_2), fmaf((dY_46_v * dY_46_v), t_0, ((dY_46_u * dY_46_u) * t_1))))) * floorf(w);
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_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 = Float32(floor(w) * floor(w)) t_2 = Float32(Float32(dX_46_u * dX_46_u) * t_1) t_3 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= t_3) tmp = Float32(Float32(dX_46_u / sqrt(fmax(fma(Float32(Float32(floor(h) * dX_46_v) * dX_46_v), floor(h), t_2), fma(Float32(dY_46_v * dY_46_v), t_0, Float32(Float32(dY_46_u * dY_46_u) * t_1))))) * floor(w)); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_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 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(dX.u \cdot dX.u\right) \cdot t\_1\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;\frac{dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_2\right), \mathsf{fma}\left(dY.v \cdot dY.v, t\_0, \left(dY.u \cdot dY.u\right) \cdot t\_1\right)\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
Applied rewrites65.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) (floor h)))
(t_2 (* (* dY.v dY.v) t_1))
(t_3 (* (floor w) (floor w)))
(t_4 (fma (* t_3 dX.u) dX.u (* t_1 (* dX.v dX.v)))))
(if (>= (* (* dX.u dX.u) t_3) (* (* dY.u dY.u) t_3))
(/ (* (floor w) dX.u) (sqrt (fmax t_4 (fma (* t_3 dY.u) dY.u t_2))))
(/ t_0 (sqrt (fmax t_4 (fma t_0 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 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * floorf(h);
float t_2 = (dY_46_v * dY_46_v) * t_1;
float t_3 = floorf(w) * floorf(w);
float t_4 = fmaf((t_3 * dX_46_u), dX_46_u, (t_1 * (dX_46_v * dX_46_v)));
float tmp;
if (((dX_46_u * dX_46_u) * t_3) >= ((dY_46_u * dY_46_u) * t_3)) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_4, fmaf((t_3 * dY_46_u), dY_46_u, t_2)));
} else {
tmp = t_0 / sqrtf(fmaxf(t_4, fmaf(t_0, 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(floor(w) * dY_46_u) t_1 = Float32(floor(h) * floor(h)) t_2 = Float32(Float32(dY_46_v * dY_46_v) * t_1) t_3 = Float32(floor(w) * floor(w)) t_4 = fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(t_1 * Float32(dX_46_v * dX_46_v))) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_3) >= Float32(Float32(dY_46_u * dY_46_u) * t_3)) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_4, fma(Float32(t_3 * dY_46_u), dY_46_u, t_2)))); else tmp = Float32(t_0 / sqrt(fmax(t_4, fma(t_0, t_0, t_2)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(dY.v \cdot dY.v\right) \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, t\_1 \cdot \left(dX.v \cdot dX.v\right)\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_3 \geq \left(dY.u \cdot dY.u\right) \cdot t\_3:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_3 \cdot dY.u, dY.u, t\_2\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_4, \mathsf{fma}\left(t\_0, t\_0, t\_2\right)\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
Taylor expanded in dY.u around inf
Applied rewrites59.7%
Applied rewrites59.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) dX.u))
(t_2 (* (floor w) (floor w)))
(t_3 (fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(if (>= (* (* dX.u dX.u) t_2) (* (* dY.u dY.u) t_2))
(/
t_1
(sqrt
(fmax (fma t_1 t_1 (* (* (* dX.v (floor h)) dX.v) (floor h))) t_3)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_2 dX.u) dX.u (* t_0 (* dX.v dX.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 = floorf(w) * dX_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float tmp;
if (((dX_46_u * dX_46_u) * t_2) >= ((dY_46_u * dY_46_u) * t_2)) {
tmp = t_1 / sqrtf(fmaxf(fmaf(t_1, t_1, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), t_3));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), t_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 = Float32(floor(w) * dX_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_2) >= Float32(Float32(dY_46_u * dY_46_u) * t_2)) tmp = Float32(t_1 / sqrt(fmax(fma(t_1, t_1, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), t_3))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_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 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_2 \geq \left(dY.u \cdot dY.u\right) \cdot t\_2:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
Taylor expanded in dY.u around inf
Applied rewrites59.7%
lift-fma.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied rewrites59.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0))))
(if (>= (* (* dX.u dX.u) t_1) (* (* dY.u dY.u) t_1))
(/
(* (floor w) dX.u)
(sqrt (fmax (* (* (* dX.v (floor h)) dX.v) (floor h)) t_2)))
(/
(* (floor w) dY.u)
(sqrt (fmax (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v))) t_2))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float tmp;
if (((dX_46_u * dX_46_u) * t_1) >= ((dY_46_u * dY_46_u) * t_1)) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)), t_2));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v))), t_2));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) tmp = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * t_1) >= Float32(Float32(dY_46_u * dY_46_u) * t_1)) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)), t_2))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))), t_2))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot t\_1 \geq \left(dY.u \cdot dY.u\right) \cdot t\_1:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right), t\_2\right)}}\\
\end{array}
\end{array}
Initial program 76.2%
Applied rewrites76.3%
Taylor expanded in dX.u around inf
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
swap-sqrN/A
swap-sqrN/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
Taylor expanded in dY.u around inf
Applied rewrites59.7%
Taylor expanded in dX.u around 0
Applied rewrites45.0%
herbie shell --seed 2025106
(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))))