
(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 11 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) dY.u))
(t_2 (* (* t_1 dY.u) (floor w)))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (fma (* (floor h) t_4) dY.v t_2))
(t_7 (* (floor w) dX.u))
(t_8 (* t_7 dX.u))
(t_9 (+ (* t_7 t_7) (* t_0 t_0)))
(t_10 (/ 1.0 (sqrt (fmax t_9 t_5))))
(t_11 (* t_10 t_4))
(t_12 (* t_10 t_0))
(t_13 (if (>= t_9 (* (* dY.v dY.v) t_3)) t_12 t_11))
(t_14 (if (>= t_9 t_5) t_12 t_11)))
(if (<= t_14 -0.9999799728393555)
t_13
(if (<= t_14 4.999999969612645e-9)
(if (>= (* t_8 (floor w)) t_2)
(/ t_0 (sqrt (fmax (fma t_8 (floor w) (* (* dX.v dX.v) t_3)) t_6)))
(/
t_4
(sqrt
(fmax (fma t_8 (floor w) (* (exp (* (log dX.v) 2.0)) t_3)) t_6))))
t_13))))
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 = (t_1 * dY_46_u) * floorf(w);
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = fmaf((floorf(h) * t_4), dY_46_v, t_2);
float t_7 = floorf(w) * dX_46_u;
float t_8 = t_7 * dX_46_u;
float t_9 = (t_7 * t_7) + (t_0 * t_0);
float t_10 = 1.0f / sqrtf(fmaxf(t_9, t_5));
float t_11 = t_10 * t_4;
float t_12 = t_10 * t_0;
float tmp;
if (t_9 >= ((dY_46_v * dY_46_v) * t_3)) {
tmp = t_12;
} else {
tmp = t_11;
}
float t_13 = tmp;
float tmp_1;
if (t_9 >= t_5) {
tmp_1 = t_12;
} else {
tmp_1 = t_11;
}
float t_14 = tmp_1;
float tmp_2;
if (t_14 <= -0.9999799728393555f) {
tmp_2 = t_13;
} else if (t_14 <= 4.999999969612645e-9f) {
float tmp_3;
if ((t_8 * floorf(w)) >= t_2) {
tmp_3 = t_0 / sqrtf(fmaxf(fmaf(t_8, floorf(w), ((dX_46_v * dX_46_v) * t_3)), t_6));
} else {
tmp_3 = t_4 / sqrtf(fmaxf(fmaf(t_8, floorf(w), (expf((logf(dX_46_v) * 2.0f)) * t_3)), t_6));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_13;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(Float32(t_1 * dY_46_u) * floor(w)) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = fma(Float32(floor(h) * t_4), dY_46_v, t_2) t_7 = Float32(floor(w) * dX_46_u) t_8 = Float32(t_7 * dX_46_u) t_9 = Float32(Float32(t_7 * t_7) + Float32(t_0 * t_0)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_9, t_5))) t_11 = Float32(t_10 * t_4) t_12 = Float32(t_10 * t_0) tmp = Float32(0.0) if (t_9 >= Float32(Float32(dY_46_v * dY_46_v) * t_3)) tmp = t_12; else tmp = t_11; end t_13 = tmp tmp_1 = Float32(0.0) if (t_9 >= t_5) tmp_1 = t_12; else tmp_1 = t_11; end t_14 = tmp_1 tmp_2 = Float32(0.0) if (t_14 <= Float32(-0.9999799728393555)) tmp_2 = t_13; elseif (t_14 <= Float32(4.999999969612645e-9)) tmp_3 = Float32(0.0) if (Float32(t_8 * floor(w)) >= t_2) tmp_3 = Float32(t_0 / sqrt(fmax(fma(t_8, floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_3)), t_6))); else tmp_3 = Float32(t_4 / sqrt(fmax(fma(t_8, floor(w), Float32(exp(Float32(log(dX_46_v) * Float32(2.0))) * t_3)), t_6))); end tmp_2 = tmp_3; else tmp_2 = t_13; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left(t\_1 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot t\_4, dY.v, t\_2\right)\\
t_7 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_8 := t\_7 \cdot dX.u\\
t_9 := t\_7 \cdot t\_7 + t\_0 \cdot t\_0\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_5\right)}}\\
t_11 := t\_10 \cdot t\_4\\
t_12 := t\_10 \cdot t\_0\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_9 \geq \left(dY.v \cdot dY.v\right) \cdot t\_3:\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_9 \geq t\_5:\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
\mathbf{if}\;t\_14 \leq -0.9999799728393555:\\
\;\;\;\;t\_13\\
\mathbf{elif}\;t\_14 \leq 4.999999969612645 \cdot 10^{-9}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \cdot \left\lfloor w\right\rfloor \geq t\_2:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_3\right), t\_6\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , e^{\log dX.v \cdot 2} \cdot t\_3\right), t\_6\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_13\\
\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))) < -0.99997997 or 4.99999997e-9 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) Initial program 99.3%
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.f3297.2
Applied rewrites97.2%
if -0.99997997 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999997e-9Initial program 59.5%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3259.5
Applied rewrites59.5%
Taylor expanded in dX.u around inf
Applied rewrites59.5%
Applied rewrites59.6%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3260.0
Applied rewrites60.0%
(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 (* (* t_1 dY.u) (floor w)))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor h) (floor h)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) t_6))
(t_8 (fma (* (floor h) t_5) dY.v t_2))
(t_9
(sqrt
(/
1.0
(fmax
(fma (* t_0 dX.u) dX.u (* t_4 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_4))))))
(t_10 (* (floor w) dX.u))
(t_11 (+ (* t_10 t_10) (* t_3 t_3)))
(t_12 (/ 1.0 (sqrt (fmax t_11 t_6))))
(t_13 (* t_12 t_3))
(t_14 (if (>= t_11 t_6) t_13 (* t_12 t_5)))
(t_15 (* t_10 dX.u)))
(if (<= t_14 -0.9999799728393555)
(if t_7 t_13 (* (* (floor h) t_9) dY.v))
(if (<= t_14 4.999999873689376e-5)
(if (>= (* t_15 (floor w)) t_2)
(/ t_3 (sqrt (fmax (fma t_15 (floor w) (* (* dX.v dX.v) t_4)) t_8)))
(/
t_5
(sqrt
(fmax (fma t_15 (floor w) (* (exp (* (log dX.v) 2.0)) t_4)) t_8))))
(if t_7 t_13 (* (* t_9 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(w) * dY_46_u;
float t_2 = (t_1 * dY_46_u) * floorf(w);
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(h) * floorf(h);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
int t_7 = (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= t_6;
float t_8 = fmaf((floorf(h) * t_5), dY_46_v, t_2);
float t_9 = sqrtf((1.0f / fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_4)))));
float t_10 = floorf(w) * dX_46_u;
float t_11 = (t_10 * t_10) + (t_3 * t_3);
float t_12 = 1.0f / sqrtf(fmaxf(t_11, t_6));
float t_13 = t_12 * t_3;
float tmp;
if (t_11 >= t_6) {
tmp = t_13;
} else {
tmp = t_12 * t_5;
}
float t_14 = tmp;
float t_15 = t_10 * dX_46_u;
float tmp_2;
if (t_14 <= -0.9999799728393555f) {
float tmp_3;
if (t_7) {
tmp_3 = t_13;
} else {
tmp_3 = (floorf(h) * t_9) * dY_46_v;
}
tmp_2 = tmp_3;
} else if (t_14 <= 4.999999873689376e-5f) {
float tmp_4;
if ((t_15 * floorf(w)) >= t_2) {
tmp_4 = t_3 / sqrtf(fmaxf(fmaf(t_15, floorf(w), ((dX_46_v * dX_46_v) * t_4)), t_8));
} else {
tmp_4 = t_5 / sqrtf(fmaxf(fmaf(t_15, floorf(w), (expf((logf(dX_46_v) * 2.0f)) * t_4)), t_8));
}
tmp_2 = tmp_4;
} else if (t_7) {
tmp_2 = t_13;
} else {
tmp_2 = (t_9 * dY_46_v) * floorf(h);
}
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) * floor(w)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(Float32(t_1 * dY_46_u) * floor(w)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(h) * floor(h)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= t_6 t_8 = fma(Float32(floor(h) * t_5), dY_46_v, t_2) t_9 = sqrt(Float32(Float32(1.0) / fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_4))))) t_10 = Float32(floor(w) * dX_46_u) t_11 = Float32(Float32(t_10 * t_10) + Float32(t_3 * t_3)) t_12 = Float32(Float32(1.0) / sqrt(fmax(t_11, t_6))) t_13 = Float32(t_12 * t_3) tmp = Float32(0.0) if (t_11 >= t_6) tmp = t_13; else tmp = Float32(t_12 * t_5); end t_14 = tmp t_15 = Float32(t_10 * dX_46_u) tmp_2 = Float32(0.0) if (t_14 <= Float32(-0.9999799728393555)) tmp_3 = Float32(0.0) if (t_7) tmp_3 = t_13; else tmp_3 = Float32(Float32(floor(h) * t_9) * dY_46_v); end tmp_2 = tmp_3; elseif (t_14 <= Float32(4.999999873689376e-5)) tmp_4 = Float32(0.0) if (Float32(t_15 * floor(w)) >= t_2) tmp_4 = Float32(t_3 / sqrt(fmax(fma(t_15, floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_4)), t_8))); else tmp_4 = Float32(t_5 / sqrt(fmax(fma(t_15, floor(w), Float32(exp(Float32(log(dX_46_v) * Float32(2.0))) * t_4)), t_8))); end tmp_2 = tmp_4; elseif (t_7) tmp_2 = t_13; else tmp_2 = Float32(Float32(t_9 * dY_46_v) * floor(h)); end return tmp_2 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(t\_1 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_6\\
t_8 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot t\_5, dY.v, t\_2\right)\\
t_9 := \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_4\right)\right)}}\\
t_10 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_11 := t\_10 \cdot t\_10 + t\_3 \cdot t\_3\\
t_12 := \frac{1}{\sqrt{\mathsf{max}\left(t\_11, t\_6\right)}}\\
t_13 := t\_12 \cdot t\_3\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_11 \geq t\_6:\\
\;\;\;\;t\_13\\
\mathbf{else}:\\
\;\;\;\;t\_12 \cdot t\_5\\
\end{array}\\
t_15 := t\_10 \cdot dX.u\\
\mathbf{if}\;t\_14 \leq -0.9999799728393555:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7:\\
\;\;\;\;t\_13\\
\mathbf{else}:\\
\;\;\;\;\left(\left\lfloor h\right\rfloor \cdot t\_9\right) \cdot dY.v\\
\end{array}\\
\mathbf{elif}\;t\_14 \leq 4.999999873689376 \cdot 10^{-5}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_15 \cdot \left\lfloor w\right\rfloor \geq t\_2:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_15, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_4\right), t\_8\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_15, \left\lfloor w\right\rfloor , e^{\log dX.v \cdot 2} \cdot t\_4\right), t\_8\right)}}\\
\end{array}\\
\mathbf{elif}\;t\_7:\\
\;\;\;\;t\_13\\
\mathbf{else}:\\
\;\;\;\;\left(t\_9 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
\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))) < -0.99997997Initial program 99.5%
Applied rewrites99.2%
Taylor expanded in dX.u around 0
Applied rewrites99.2%
if -0.99997997 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999987e-5Initial program 61.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3261.3
Applied rewrites61.3%
Taylor expanded in dX.u around inf
Applied rewrites61.3%
Applied rewrites61.5%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3261.8
Applied rewrites61.8%
if 4.99999987e-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 99.2%
Applied rewrites99.1%
Taylor expanded in dX.u around 0
Applied rewrites98.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor w) dY.u))
(t_2 (* (* t_1 dY.u) (floor w)))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor h) (floor h)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (fma (* (floor h) t_5) dY.v t_2))
(t_8 (* (floor w) dX.u))
(t_9 (+ (* t_8 t_8) (* t_3 t_3)))
(t_10 (/ 1.0 (sqrt (fmax t_9 t_6))))
(t_11 (* t_10 t_3))
(t_12
(if (>= (* (* (* dX.v (floor h)) dX.v) (floor h)) t_6)
t_11
(*
(*
(floor h)
(sqrt
(/
1.0
(fmax
(fma (* t_0 dX.u) dX.u (* t_4 (* dX.v dX.v)))
(fma (* t_0 dY.u) dY.u (* (* dY.v dY.v) t_4))))))
dY.v)))
(t_13 (if (>= t_9 t_6) t_11 (* t_10 t_5)))
(t_14 (* t_8 dX.u)))
(if (<= t_13 -0.9999799728393555)
t_12
(if (<= t_13 4.999999873689376e-5)
(if (>= (* t_14 (floor w)) t_2)
(/ t_3 (sqrt (fmax (fma t_14 (floor w) (* (* dX.v dX.v) t_4)) t_7)))
(/
t_5
(sqrt
(fmax (fma t_14 (floor w) (* (exp (* (log dX.v) 2.0)) t_4)) t_7))))
t_12))))
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 = (t_1 * dY_46_u) * floorf(w);
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(h) * floorf(h);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
float t_7 = fmaf((floorf(h) * t_5), dY_46_v, t_2);
float t_8 = floorf(w) * dX_46_u;
float t_9 = (t_8 * t_8) + (t_3 * t_3);
float t_10 = 1.0f / sqrtf(fmaxf(t_9, t_6));
float t_11 = t_10 * t_3;
float tmp;
if ((((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)) >= t_6) {
tmp = t_11;
} else {
tmp = (floorf(h) * sqrtf((1.0f / fmaxf(fmaf((t_0 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v))), fmaf((t_0 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_4)))))) * dY_46_v;
}
float t_12 = tmp;
float tmp_1;
if (t_9 >= t_6) {
tmp_1 = t_11;
} else {
tmp_1 = t_10 * t_5;
}
float t_13 = tmp_1;
float t_14 = t_8 * dX_46_u;
float tmp_2;
if (t_13 <= -0.9999799728393555f) {
tmp_2 = t_12;
} else if (t_13 <= 4.999999873689376e-5f) {
float tmp_3;
if ((t_14 * floorf(w)) >= t_2) {
tmp_3 = t_3 / sqrtf(fmaxf(fmaf(t_14, floorf(w), ((dX_46_v * dX_46_v) * t_4)), t_7));
} else {
tmp_3 = t_5 / sqrtf(fmaxf(fmaf(t_14, floorf(w), (expf((logf(dX_46_v) * 2.0f)) * t_4)), t_7));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_12;
}
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) * floor(w)) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(Float32(t_1 * dY_46_u) * floor(w)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(h) * floor(h)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = fma(Float32(floor(h) * t_5), dY_46_v, t_2) t_8 = Float32(floor(w) * dX_46_u) t_9 = Float32(Float32(t_8 * t_8) + Float32(t_3 * t_3)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_9, t_6))) t_11 = Float32(t_10 * t_3) tmp = Float32(0.0) if (Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h)) >= t_6) tmp = t_11; else tmp = Float32(Float32(floor(h) * sqrt(Float32(Float32(1.0) / fmax(fma(Float32(t_0 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_0 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_4)))))) * dY_46_v); end t_12 = tmp tmp_1 = Float32(0.0) if (t_9 >= t_6) tmp_1 = t_11; else tmp_1 = Float32(t_10 * t_5); end t_13 = tmp_1 t_14 = Float32(t_8 * dX_46_u) tmp_2 = Float32(0.0) if (t_13 <= Float32(-0.9999799728393555)) tmp_2 = t_12; elseif (t_13 <= Float32(4.999999873689376e-5)) tmp_3 = Float32(0.0) if (Float32(t_14 * floor(w)) >= t_2) tmp_3 = Float32(t_3 / sqrt(fmax(fma(t_14, floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_4)), t_7))); else tmp_3 = Float32(t_5 / sqrt(fmax(fma(t_14, floor(w), Float32(exp(Float32(log(dX_46_v) * Float32(2.0))) * t_4)), t_7))); end tmp_2 = tmp_3; else tmp_2 = t_12; end return tmp_2 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(t\_1 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot t\_5, dY.v, t\_2\right)\\
t_8 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_9 := t\_8 \cdot t\_8 + t\_3 \cdot t\_3\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_6\right)}}\\
t_11 := t\_10 \cdot t\_3\\
t_12 := \begin{array}{l}
\mathbf{if}\;\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \geq t\_6:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\left(\left\lfloor h\right\rfloor \cdot \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_0 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_4\right)\right)}}\right) \cdot dY.v\\
\end{array}\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_9 \geq t\_6:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;t\_10 \cdot t\_5\\
\end{array}\\
t_14 := t\_8 \cdot dX.u\\
\mathbf{if}\;t\_13 \leq -0.9999799728393555:\\
\;\;\;\;t\_12\\
\mathbf{elif}\;t\_13 \leq 4.999999873689376 \cdot 10^{-5}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_14 \cdot \left\lfloor w\right\rfloor \geq t\_2:\\
\;\;\;\;\frac{t\_3}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_14, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_4\right), t\_7\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_5}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_14, \left\lfloor w\right\rfloor , e^{\log dX.v \cdot 2} \cdot t\_4\right), t\_7\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_12\\
\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))) < -0.99997997 or 4.99999987e-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 99.4%
Applied rewrites99.1%
Taylor expanded in dX.u around 0
Applied rewrites98.8%
if -0.99997997 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999987e-5Initial program 61.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3261.3
Applied rewrites61.3%
Taylor expanded in dX.u around inf
Applied rewrites61.3%
Applied rewrites61.5%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3261.8
Applied rewrites61.8%
(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) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) (floor w)))
(t_4 (* (floor h) (floor h)))
(t_5 (* (* dY.v dY.v) t_4))
(t_6 (fma (* t_3 dX.u) dX.u (* t_4 (* dX.v dX.v))))
(t_7 (sqrt (fmax t_6 (fma (* t_3 dY.u) dY.u t_5))))
(t_8 (* t_0 dX.u))
(t_9 (+ (* t_0 t_0) (* t_1 t_1)))
(t_10 (* (floor h) dY.v))
(t_11 (if (>= t_6 t_5) (/ t_1 t_7) (/ t_10 t_7)))
(t_12 (+ (* t_2 t_2) (* t_10 t_10)))
(t_13 (/ 1.0 (sqrt (fmax t_9 t_12))))
(t_14 (if (>= t_9 t_12) (* t_13 t_1) (* t_13 t_10)))
(t_15 (* (* t_2 dY.u) (floor w)))
(t_16 (fma (* (floor h) t_10) dY.v t_15)))
(if (<= t_14 -0.9999995231628418)
t_11
(if (<= t_14 4.999999969612645e-9)
(if (>= (* t_8 (floor w)) t_15)
(/ t_1 (sqrt (fmax (fma t_8 (floor w) (* (* dX.v dX.v) t_4)) t_16)))
(/
t_10
(sqrt
(fmax (fma t_8 (floor w) (* (exp (* (log dX.v) 2.0)) t_4)) t_16))))
t_11))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * floorf(w);
float t_4 = floorf(h) * floorf(h);
float t_5 = (dY_46_v * dY_46_v) * t_4;
float t_6 = fmaf((t_3 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v)));
float t_7 = sqrtf(fmaxf(t_6, fmaf((t_3 * dY_46_u), dY_46_u, t_5)));
float t_8 = t_0 * dX_46_u;
float t_9 = (t_0 * t_0) + (t_1 * t_1);
float t_10 = floorf(h) * dY_46_v;
float tmp;
if (t_6 >= t_5) {
tmp = t_1 / t_7;
} else {
tmp = t_10 / t_7;
}
float t_11 = tmp;
float t_12 = (t_2 * t_2) + (t_10 * t_10);
float t_13 = 1.0f / sqrtf(fmaxf(t_9, t_12));
float tmp_1;
if (t_9 >= t_12) {
tmp_1 = t_13 * t_1;
} else {
tmp_1 = t_13 * t_10;
}
float t_14 = tmp_1;
float t_15 = (t_2 * dY_46_u) * floorf(w);
float t_16 = fmaf((floorf(h) * t_10), dY_46_v, t_15);
float tmp_2;
if (t_14 <= -0.9999995231628418f) {
tmp_2 = t_11;
} else if (t_14 <= 4.999999969612645e-9f) {
float tmp_3;
if ((t_8 * floorf(w)) >= t_15) {
tmp_3 = t_1 / sqrtf(fmaxf(fmaf(t_8, floorf(w), ((dX_46_v * dX_46_v) * t_4)), t_16));
} else {
tmp_3 = t_10 / sqrtf(fmaxf(fmaf(t_8, floorf(w), (expf((logf(dX_46_v) * 2.0f)) * t_4)), t_16));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_11;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(floor(h) * floor(h)) t_5 = Float32(Float32(dY_46_v * dY_46_v) * t_4) t_6 = fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))) t_7 = sqrt(fmax(t_6, fma(Float32(t_3 * dY_46_u), dY_46_u, t_5))) t_8 = Float32(t_0 * dX_46_u) t_9 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_10 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (t_6 >= t_5) tmp = Float32(t_1 / t_7); else tmp = Float32(t_10 / t_7); end t_11 = tmp t_12 = Float32(Float32(t_2 * t_2) + Float32(t_10 * t_10)) t_13 = Float32(Float32(1.0) / sqrt(fmax(t_9, t_12))) tmp_1 = Float32(0.0) if (t_9 >= t_12) tmp_1 = Float32(t_13 * t_1); else tmp_1 = Float32(t_13 * t_10); end t_14 = tmp_1 t_15 = Float32(Float32(t_2 * dY_46_u) * floor(w)) t_16 = fma(Float32(floor(h) * t_10), dY_46_v, t_15) tmp_2 = Float32(0.0) if (t_14 <= Float32(-0.9999995231628418)) tmp_2 = t_11; elseif (t_14 <= Float32(4.999999969612645e-9)) tmp_3 = Float32(0.0) if (Float32(t_8 * floor(w)) >= t_15) tmp_3 = Float32(t_1 / sqrt(fmax(fma(t_8, floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_4)), t_16))); else tmp_3 = Float32(t_10 / sqrt(fmax(fma(t_8, floor(w), Float32(exp(Float32(log(dX_46_v) * Float32(2.0))) * t_4)), t_16))); end tmp_2 = tmp_3; else tmp_2 = t_11; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \left(dY.v \cdot dY.v\right) \cdot t\_4\\
t_6 := \mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_7 := \sqrt{\mathsf{max}\left(t\_6, \mathsf{fma}\left(t\_3 \cdot dY.u, dY.u, t\_5\right)\right)}\\
t_8 := t\_0 \cdot dX.u\\
t_9 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_10 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_11 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_5:\\
\;\;\;\;\frac{t\_1}{t\_7}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_10}{t\_7}\\
\end{array}\\
t_12 := t\_2 \cdot t\_2 + t\_10 \cdot t\_10\\
t_13 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_12\right)}}\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_9 \geq t\_12:\\
\;\;\;\;t\_13 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_13 \cdot t\_10\\
\end{array}\\
t_15 := \left(t\_2 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_16 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot t\_10, dY.v, t\_15\right)\\
\mathbf{if}\;t\_14 \leq -0.9999995231628418:\\
\;\;\;\;t\_11\\
\mathbf{elif}\;t\_14 \leq 4.999999969612645 \cdot 10^{-9}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \cdot \left\lfloor w\right\rfloor \geq t\_15:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_4\right), t\_16\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_10}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_8, \left\lfloor w\right\rfloor , e^{\log dX.v \cdot 2} \cdot t\_4\right), t\_16\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 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))) < -0.999999523 or 4.99999997e-9 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) Initial program 99.3%
Applied rewrites99.3%
Taylor expanded in dY.u around 0
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3297.2
Applied rewrites97.2%
if -0.999999523 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999997e-9Initial program 60.0%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3259.9
Applied rewrites59.9%
Taylor expanded in dX.u around inf
Applied rewrites59.9%
Applied rewrites60.1%
lift-*.f32N/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f3260.4
Applied rewrites60.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor w) (floor w)))
(t_4 (* (floor h) (floor h)))
(t_5 (fma (* t_3 dX.u) dX.u (* t_4 (* dX.v dX.v))))
(t_6 (* t_0 dX.u))
(t_7 (fma t_6 (floor w) (* (* dX.v dX.v) t_4)))
(t_8 (+ (* t_0 t_0) (* t_1 t_1)))
(t_9 (* (floor h) dY.v))
(t_10 (+ (* t_2 t_2) (* t_9 t_9)))
(t_11 (/ 1.0 (sqrt (fmax t_8 t_10))))
(t_12 (if (>= t_8 t_10) (* t_11 t_1) (* t_11 t_9)))
(t_13 (* (* dY.v dY.v) t_4))
(t_14 (sqrt (fmax t_5 (fma (* t_3 dY.u) dY.u t_13))))
(t_15 (if (>= t_5 t_13) (/ t_1 t_14) (/ t_9 t_14)))
(t_16 (* (* t_2 dY.u) (floor w)))
(t_17 (fma (* (floor h) t_9) dY.v t_16)))
(if (<= t_12 -0.4000000059604645)
t_15
(if (<= t_12 4.999999969612645e-9)
(if (>= t_7 t_16)
(/ t_1 (sqrt (fmax (* t_6 (floor w)) t_17)))
(/ t_9 (sqrt (fmax t_7 t_17))))
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 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(w) * floorf(w);
float t_4 = floorf(h) * floorf(h);
float t_5 = fmaf((t_3 * dX_46_u), dX_46_u, (t_4 * (dX_46_v * dX_46_v)));
float t_6 = t_0 * dX_46_u;
float t_7 = fmaf(t_6, floorf(w), ((dX_46_v * dX_46_v) * t_4));
float t_8 = (t_0 * t_0) + (t_1 * t_1);
float t_9 = floorf(h) * dY_46_v;
float t_10 = (t_2 * t_2) + (t_9 * t_9);
float t_11 = 1.0f / sqrtf(fmaxf(t_8, t_10));
float tmp;
if (t_8 >= t_10) {
tmp = t_11 * t_1;
} else {
tmp = t_11 * t_9;
}
float t_12 = tmp;
float t_13 = (dY_46_v * dY_46_v) * t_4;
float t_14 = sqrtf(fmaxf(t_5, fmaf((t_3 * dY_46_u), dY_46_u, t_13)));
float tmp_1;
if (t_5 >= t_13) {
tmp_1 = t_1 / t_14;
} else {
tmp_1 = t_9 / t_14;
}
float t_15 = tmp_1;
float t_16 = (t_2 * dY_46_u) * floorf(w);
float t_17 = fmaf((floorf(h) * t_9), dY_46_v, t_16);
float tmp_2;
if (t_12 <= -0.4000000059604645f) {
tmp_2 = t_15;
} else if (t_12 <= 4.999999969612645e-9f) {
float tmp_3;
if (t_7 >= t_16) {
tmp_3 = t_1 / sqrtf(fmaxf((t_6 * floorf(w)), t_17));
} else {
tmp_3 = t_9 / sqrtf(fmaxf(t_7, t_17));
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_15;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(w) * floor(w)) t_4 = Float32(floor(h) * floor(h)) t_5 = fma(Float32(t_3 * dX_46_u), dX_46_u, Float32(t_4 * Float32(dX_46_v * dX_46_v))) t_6 = Float32(t_0 * dX_46_u) t_7 = fma(t_6, floor(w), Float32(Float32(dX_46_v * dX_46_v) * t_4)) t_8 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_9 = Float32(floor(h) * dY_46_v) t_10 = Float32(Float32(t_2 * t_2) + Float32(t_9 * t_9)) t_11 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_10))) tmp = Float32(0.0) if (t_8 >= t_10) tmp = Float32(t_11 * t_1); else tmp = Float32(t_11 * t_9); end t_12 = tmp t_13 = Float32(Float32(dY_46_v * dY_46_v) * t_4) t_14 = sqrt(fmax(t_5, fma(Float32(t_3 * dY_46_u), dY_46_u, t_13))) tmp_1 = Float32(0.0) if (t_5 >= t_13) tmp_1 = Float32(t_1 / t_14); else tmp_1 = Float32(t_9 / t_14); end t_15 = tmp_1 t_16 = Float32(Float32(t_2 * dY_46_u) * floor(w)) t_17 = fma(Float32(floor(h) * t_9), dY_46_v, t_16) tmp_2 = Float32(0.0) if (t_12 <= Float32(-0.4000000059604645)) tmp_2 = t_15; elseif (t_12 <= Float32(4.999999969612645e-9)) tmp_3 = Float32(0.0) if (t_7 >= t_16) tmp_3 = Float32(t_1 / sqrt(fmax(Float32(t_6 * floor(w)), t_17))); else tmp_3 = Float32(t_9 / sqrt(fmax(t_7, t_17))); end tmp_2 = tmp_3; else tmp_2 = t_15; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_4 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_5 := \mathsf{fma}\left(t\_3 \cdot dX.u, dX.u, t\_4 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_6 := t\_0 \cdot dX.u\\
t_7 := \mathsf{fma}\left(t\_6, \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_4\right)\\
t_8 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_9 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_10 := t\_2 \cdot t\_2 + t\_9 \cdot t\_9\\
t_11 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_10\right)}}\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_10:\\
\;\;\;\;t\_11 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_11 \cdot t\_9\\
\end{array}\\
t_13 := \left(dY.v \cdot dY.v\right) \cdot t\_4\\
t_14 := \sqrt{\mathsf{max}\left(t\_5, \mathsf{fma}\left(t\_3 \cdot dY.u, dY.u, t\_13\right)\right)}\\
t_15 := \begin{array}{l}
\mathbf{if}\;t\_5 \geq t\_13:\\
\;\;\;\;\frac{t\_1}{t\_14}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_9}{t\_14}\\
\end{array}\\
t_16 := \left(t\_2 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_17 := \mathsf{fma}\left(\left\lfloor h\right\rfloor \cdot t\_9, dY.v, t\_16\right)\\
\mathbf{if}\;t\_12 \leq -0.4000000059604645:\\
\;\;\;\;t\_15\\
\mathbf{elif}\;t\_12 \leq 4.999999969612645 \cdot 10^{-9}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_16:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_6 \cdot \left\lfloor w\right\rfloor , t\_17\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_9}{\sqrt{\mathsf{max}\left(t\_7, t\_17\right)}}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_15\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 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))) < -0.400000006 or 4.99999997e-9 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) Initial program 99.3%
Applied rewrites99.3%
Taylor expanded in dY.u around 0
pow2N/A
lift-*.f32N/A
pow2N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f3297.2
Applied rewrites97.2%
if -0.400000006 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999997e-9Initial program 58.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3258.4
Applied rewrites58.4%
Taylor expanded in dX.u around inf
Applied rewrites58.5%
Applied rewrites58.7%
(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}
Initial program 75.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (* (floor h) dY.v))
(t_6 (+ (* t_1 t_1) (* t_5 t_5)))
(t_7 (* (floor h) (floor h))))
(if (>= t_4 t_6)
(* (/ 1.0 (sqrt (fmax t_4 t_6))) t_0)
(*
(*
(floor h)
(sqrt
(/
1.0
(fmax
(fma (* t_2 dX.u) dX.u (* t_7 (* dX.v dX.v)))
(fma (* t_2 dY.u) dY.u (* (* dY.v dY.v) t_7))))))
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) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = floorf(h) * dY_46_v;
float t_6 = (t_1 * t_1) + (t_5 * t_5);
float t_7 = floorf(h) * floorf(h);
float tmp;
if (t_4 >= t_6) {
tmp = (1.0f / sqrtf(fmaxf(t_4, t_6))) * t_0;
} else {
tmp = (floorf(h) * sqrtf((1.0f / fmaxf(fmaf((t_2 * dX_46_u), dX_46_u, (t_7 * (dX_46_v * dX_46_v))), fmaf((t_2 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_7)))))) * 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) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(t_1 * t_1) + Float32(t_5 * t_5)) t_7 = Float32(floor(h) * floor(h)) tmp = Float32(0.0) if (t_4 >= t_6) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_4, t_6))) * t_0); else tmp = Float32(Float32(floor(h) * sqrt(Float32(Float32(1.0) / fmax(fma(Float32(t_2 * dX_46_u), dX_46_u, Float32(t_7 * Float32(dX_46_v * dX_46_v))), fma(Float32(t_2 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_7)))))) * dY_46_v); end return 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 \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := t\_1 \cdot t\_1 + t\_5 \cdot t\_5\\
t_7 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
\mathbf{if}\;t\_4 \geq t\_6:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_6\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\left(\left\lfloor h\right\rfloor \cdot \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_2 \cdot dX.u, dX.u, t\_7 \cdot \left(dX.v \cdot dX.v\right)\right), \mathsf{fma}\left(t\_2 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_7\right)\right)}}\right) \cdot dY.v\\
\end{array}
\end{array}
Initial program 75.9%
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 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor h) dX.v) t_4) (/ (* (floor h) dY.v) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(h) * dX_46_v) / t_4;
} else {
tmp = (floorf(h) * dY_46_v) / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / t_4); else tmp = Float32(Float32(floor(h) * dY_46_v) / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_4}\\
\end{array}
\end{array}
Initial program 75.9%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_1 (* (floor h) dX.v))
(t_2 (* (* (* dY.u (floor w)) dY.u) (floor w)))
(t_3 (* (floor w) dY.u))
(t_4 (* dX.v (floor h)))
(t_5 (* t_4 dX.v))
(t_6 (* (floor h) dY.v))
(t_7 (+ (* t_3 t_3) (* t_6 t_6)))
(t_8 (* (floor w) dX.u))
(t_9 (+ (* t_8 t_8) (* t_1 t_1)))
(t_10 (/ 1.0 (sqrt (fmax t_9 t_7))))
(t_11 (* t_10 t_6))
(t_12 (* t_10 t_1))
(t_13 (if (>= t_9 t_7) t_12 t_11))
(t_14
(if (>= (* t_5 (floor h)) (* (* dY.u dY.u) (* (floor w) (floor w))))
t_12
t_11))
(t_15 (* dY.v (floor h)))
(t_16
(sqrt
(fmax (fma t_5 (floor h) t_0) (fma (* t_15 dY.v) (floor h) t_2)))))
(if (<= t_13 -0.9999995231628418)
t_14
(if (<= t_13 4.999999873689376e-5)
(if (>= t_0 t_2) (/ t_4 t_16) (/ t_15 t_16))
t_14))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_1 = floorf(h) * dX_46_v;
float t_2 = ((dY_46_u * floorf(w)) * dY_46_u) * floorf(w);
float t_3 = floorf(w) * dY_46_u;
float t_4 = dX_46_v * floorf(h);
float t_5 = t_4 * dX_46_v;
float t_6 = floorf(h) * dY_46_v;
float t_7 = (t_3 * t_3) + (t_6 * t_6);
float t_8 = floorf(w) * dX_46_u;
float t_9 = (t_8 * t_8) + (t_1 * t_1);
float t_10 = 1.0f / sqrtf(fmaxf(t_9, t_7));
float t_11 = t_10 * t_6;
float t_12 = t_10 * t_1;
float tmp;
if (t_9 >= t_7) {
tmp = t_12;
} else {
tmp = t_11;
}
float t_13 = tmp;
float tmp_1;
if ((t_5 * floorf(h)) >= ((dY_46_u * dY_46_u) * (floorf(w) * floorf(w)))) {
tmp_1 = t_12;
} else {
tmp_1 = t_11;
}
float t_14 = tmp_1;
float t_15 = dY_46_v * floorf(h);
float t_16 = sqrtf(fmaxf(fmaf(t_5, floorf(h), t_0), fmaf((t_15 * dY_46_v), floorf(h), t_2)));
float tmp_2;
if (t_13 <= -0.9999995231628418f) {
tmp_2 = t_14;
} else if (t_13 <= 4.999999873689376e-5f) {
float tmp_3;
if (t_0 >= t_2) {
tmp_3 = t_4 / t_16;
} else {
tmp_3 = t_15 / t_16;
}
tmp_2 = tmp_3;
} else {
tmp_2 = t_14;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(dX_46_v * floor(h)) t_5 = Float32(t_4 * dX_46_v) t_6 = Float32(floor(h) * dY_46_v) t_7 = Float32(Float32(t_3 * t_3) + Float32(t_6 * t_6)) t_8 = Float32(floor(w) * dX_46_u) t_9 = Float32(Float32(t_8 * t_8) + Float32(t_1 * t_1)) t_10 = Float32(Float32(1.0) / sqrt(fmax(t_9, t_7))) t_11 = Float32(t_10 * t_6) t_12 = Float32(t_10 * t_1) tmp = Float32(0.0) if (t_9 >= t_7) tmp = t_12; else tmp = t_11; end t_13 = tmp tmp_1 = Float32(0.0) if (Float32(t_5 * floor(h)) >= Float32(Float32(dY_46_u * dY_46_u) * Float32(floor(w) * floor(w)))) tmp_1 = t_12; else tmp_1 = t_11; end t_14 = tmp_1 t_15 = Float32(dY_46_v * floor(h)) t_16 = sqrt(fmax(fma(t_5, floor(h), t_0), fma(Float32(t_15 * dY_46_v), floor(h), t_2))) tmp_2 = Float32(0.0) if (t_13 <= Float32(-0.9999995231628418)) tmp_2 = t_14; elseif (t_13 <= Float32(4.999999873689376e-5)) tmp_3 = Float32(0.0) if (t_0 >= t_2) tmp_3 = Float32(t_4 / t_16); else tmp_3 = Float32(t_15 / t_16); end tmp_2 = tmp_3; else tmp_2 = t_14; end return tmp_2 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_5 := t\_4 \cdot dX.v\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_7 := t\_3 \cdot t\_3 + t\_6 \cdot t\_6\\
t_8 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_9 := t\_8 \cdot t\_8 + t\_1 \cdot t\_1\\
t_10 := \frac{1}{\sqrt{\mathsf{max}\left(t\_9, t\_7\right)}}\\
t_11 := t\_10 \cdot t\_6\\
t_12 := t\_10 \cdot t\_1\\
t_13 := \begin{array}{l}
\mathbf{if}\;t\_9 \geq t\_7:\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
t_14 := \begin{array}{l}
\mathbf{if}\;t\_5 \cdot \left\lfloor h\right\rfloor \geq \left(dY.u \cdot dY.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;t\_12\\
\mathbf{else}:\\
\;\;\;\;t\_11\\
\end{array}\\
t_15 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_16 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_5, \left\lfloor h\right\rfloor , t\_0\right), \mathsf{fma}\left(t\_15 \cdot dY.v, \left\lfloor h\right\rfloor , t\_2\right)\right)}\\
\mathbf{if}\;t\_13 \leq -0.9999995231628418:\\
\;\;\;\;t\_14\\
\mathbf{elif}\;t\_13 \leq 4.999999873689376 \cdot 10^{-5}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_0 \geq t\_2:\\
\;\;\;\;\frac{t\_4}{t\_16}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_15}{t\_16}\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_14\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < -0.999999523 or 4.99999987e-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 99.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3269.9
Applied rewrites69.9%
Taylor expanded in dX.u around 0
Applied rewrites76.8%
if -0.999999523 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999987e-5Initial program 61.8%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3261.6
Applied rewrites61.6%
Taylor expanded in dX.u around inf
Applied rewrites61.7%
Applied rewrites61.9%
Applied rewrites61.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (* (* dY.u dY.u) (floor w)) (floor w)))
(t_2
(fma
(* dX.u dX.u)
(* (floor w) (floor w))
(* (* t_0 dX.v) (floor h))))
(t_3
(sqrt (fmax t_2 (fma (* (* (floor h) (floor h)) dY.v) dY.v t_1)))))
(if (>= t_2 t_1) (/ t_0 t_3) (/ (* dY.v (floor h)) t_3))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_v * floorf(h);
float t_1 = ((dY_46_u * dY_46_u) * floorf(w)) * floorf(w);
float t_2 = fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((t_0 * dX_46_v) * floorf(h)));
float t_3 = sqrtf(fmaxf(t_2, fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, t_1)));
float tmp;
if (t_2 >= t_1) {
tmp = t_0 / t_3;
} else {
tmp = (dY_46_v * floorf(h)) / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(Float32(Float32(dY_46_u * dY_46_u) * floor(w)) * floor(w)) t_2 = fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(t_0 * dX_46_v) * floor(h))) t_3 = sqrt(fmax(t_2, fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, t_1))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_0 / t_3); else tmp = Float32(Float32(dY_46_v * floor(h)) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(t\_0 \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, t\_1\right)\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{t\_0}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.v \cdot \left\lfloor h\right\rfloor }{t\_3}\\
\end{array}
\end{array}
Initial program 75.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3264.7
Applied rewrites64.7%
Applied rewrites64.9%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (* (* (* dY.u (floor w)) dY.u) (floor w)))
(t_4
(sqrt
(fmax
(fma (* t_0 dX.v) (floor h) t_1)
(fma (* t_2 dY.v) (floor h) t_3)))))
(if (>= t_1 t_3) (/ t_0 t_4) (/ t_2 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_v * floorf(h);
float t_1 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = ((dY_46_u * floorf(w)) * dY_46_u) * floorf(w);
float t_4 = sqrtf(fmaxf(fmaf((t_0 * dX_46_v), floorf(h), t_1), fmaf((t_2 * dY_46_v), floorf(h), t_3)));
float tmp;
if (t_1 >= t_3) {
tmp = t_0 / t_4;
} else {
tmp = t_2 / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w)) t_4 = sqrt(fmax(fma(Float32(t_0 * dX_46_v), floor(h), t_1), fma(Float32(t_2 * dY_46_v), floor(h), t_3))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(t_0 / t_4); else tmp = Float32(t_2 / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_4 := \sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dX.v, \left\lfloor h\right\rfloor , t\_1\right), \mathsf{fma}\left(t\_2 \cdot dY.v, \left\lfloor h\right\rfloor , t\_3\right)\right)}\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{t\_0}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_2}{t\_4}\\
\end{array}
\end{array}
Initial program 75.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
unpow2N/A
lower-*.f32N/A
lift-floor.f32N/A
lift-floor.f3264.7
Applied rewrites64.7%
Taylor expanded in dX.u around inf
Applied rewrites59.7%
Applied rewrites59.8%
Applied rewrites59.8%
herbie shell --seed 2025119
(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))))