
(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 10 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 h) (floor h)))
(t_1 (* (floor w) dY.u))
(t_2 (fma t_1 t_1 (* (* dY.v dY.v) t_0)))
(t_3 (fma (* (* (floor w) (floor w)) dX.u) dX.u (* t_0 (* dX.v dX.v))))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor w) dX.u) t_4) (/ t_1 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) * dY_46_u;
float t_2 = fmaf(t_1, t_1, ((dY_46_v * dY_46_v) * t_0));
float t_3 = fmaf(((floorf(w) * floorf(w)) * 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(w) * dX_46_u) / t_4;
} else {
tmp = t_1 / 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) * dY_46_u) t_2 = fma(t_1, t_1, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_3 = fma(Float32(Float32(floor(w) * floor(w)) * 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(w) * dX_46_u) / t_4); else tmp = Float32(t_1 / 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 dY.u\\
t_2 := \mathsf{fma}\left(t\_1, t\_1, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \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 w\right\rfloor \cdot dX.u}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{t\_4}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
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
Applied rewrites76.3%
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
Applied rewrites76.3%
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
Applied rewrites76.4%
(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) dY.v))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor h) (floor h)))
(t_4 (* (floor w) (floor w)))
(t_5 (* (floor w) dX.u))
(t_6
(pow
(fmax
(fma (* dX.u dX.u) t_4 (* (* dX.v dX.v) t_3))
(* (* t_1 dY.v) (floor h)))
-0.5))
(t_7 (* (floor h) dX.v))
(t_8 (* (* (- dX.u) dX.u) t_4))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_2))))
(t_10 (if (>= t_8 t_2) (* t_9 t_5) (* t_9 t_0))))
(if (<= dY.u -260.0)
t_10
(if (<= dY.u 20000000000.0)
(if (>= (+ (* t_5 t_5) (* t_7 t_7)) (* (* dY.v dY.v) t_3))
(* t_6 t_5)
(* t_6 t_0))
t_10))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(h) * floorf(h);
float t_4 = floorf(w) * floorf(w);
float t_5 = floorf(w) * dX_46_u;
float t_6 = powf(fmaxf(fmaf((dX_46_u * dX_46_u), t_4, ((dX_46_v * dX_46_v) * t_3)), ((t_1 * dY_46_v) * floorf(h))), -0.5f);
float t_7 = floorf(h) * dX_46_v;
float t_8 = (-dX_46_u * dX_46_u) * t_4;
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_2));
float tmp;
if (t_8 >= t_2) {
tmp = t_9 * t_5;
} else {
tmp = t_9 * t_0;
}
float t_10 = tmp;
float tmp_1;
if (dY_46_u <= -260.0f) {
tmp_1 = t_10;
} else if (dY_46_u <= 20000000000.0f) {
float tmp_2;
if (((t_5 * t_5) + (t_7 * t_7)) >= ((dY_46_v * dY_46_v) * t_3)) {
tmp_2 = t_6 * t_5;
} else {
tmp_2 = t_6 * t_0;
}
tmp_1 = tmp_2;
} else {
tmp_1 = t_10;
}
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(h) * dY_46_v) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(floor(w) * floor(w)) t_5 = Float32(floor(w) * dX_46_u) t_6 = fmax(fma(Float32(dX_46_u * dX_46_u), t_4, Float32(Float32(dX_46_v * dX_46_v) * t_3)), Float32(Float32(t_1 * dY_46_v) * floor(h))) ^ Float32(-0.5) t_7 = Float32(floor(h) * dX_46_v) t_8 = Float32(Float32(Float32(-dX_46_u) * dX_46_u) * t_4) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_2))) tmp = Float32(0.0) if (t_8 >= t_2) tmp = Float32(t_9 * t_5); else tmp = Float32(t_9 * t_0); end t_10 = tmp tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-260.0)) tmp_1 = t_10; elseif (dY_46_u <= Float32(20000000000.0)) tmp_2 = Float32(0.0) if (Float32(Float32(t_5 * t_5) + Float32(t_7 * t_7)) >= Float32(Float32(dY_46_v * dY_46_v) * t_3)) tmp_2 = Float32(t_6 * t_5); else tmp_2 = Float32(t_6 * t_0); end tmp_1 = tmp_2; else tmp_1 = t_10; 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 h\right\rfloor \cdot dY.v\\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := {\left(\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_4, \left(dX.v \cdot dX.v\right) \cdot t\_3\right), \left(t\_1 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\right)}^{-0.5}\\
t_7 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_8 := \left(\left(-dX.u\right) \cdot dX.u\right) \cdot t\_4\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_2\right)}}\\
t_10 := \begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_2:\\
\;\;\;\;t\_9 \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_0\\
\end{array}\\
\mathbf{if}\;dY.u \leq -260:\\
\;\;\;\;t\_10\\
\mathbf{elif}\;dY.u \leq 20000000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_5 \cdot t\_5 + t\_7 \cdot t\_7 \geq \left(dY.v \cdot dY.v\right) \cdot t\_3:\\
\;\;\;\;t\_6 \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_0\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}
\end{array}
if dY.u < -260 or 2e10 < dY.u Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.9%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites69.9%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites46.9%
if -260 < dY.u < 2e10Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.0%
Applied rewrites63.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor w) (floor w)))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor h) dY.v))
(t_4 (+ (* t_0 t_0) (* t_3 t_3)))
(t_5 (* (* t_3 dY.v) (floor h)))
(t_6 (* (* (- dX.u) dX.u) t_1))
(t_7 (/ 1.0 (sqrt (fmax t_6 t_4))))
(t_8 (if (>= t_6 t_4) (* t_7 t_2) (* t_7 t_0)))
(t_9 (* (floor h) (floor h)))
(t_10 (fma (* dX.u dX.u) t_1 (* (* dX.v dX.v) t_9))))
(if (<= dY.u -260.0)
t_8
(if (<= dY.u 20000000000.0)
(if (>= t_10 t_5)
(/ t_2 (sqrt (fmax t_10 t_5)))
(*
(/
dY.u
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(* (* dY.v dY.v) t_9))))
(floor w)))
t_8))))
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) * floorf(w);
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = (t_0 * t_0) + (t_3 * t_3);
float t_5 = (t_3 * dY_46_v) * floorf(h);
float t_6 = (-dX_46_u * dX_46_u) * t_1;
float t_7 = 1.0f / sqrtf(fmaxf(t_6, t_4));
float tmp;
if (t_6 >= t_4) {
tmp = t_7 * t_2;
} else {
tmp = t_7 * t_0;
}
float t_8 = tmp;
float t_9 = floorf(h) * floorf(h);
float t_10 = fmaf((dX_46_u * dX_46_u), t_1, ((dX_46_v * dX_46_v) * t_9));
float tmp_1;
if (dY_46_u <= -260.0f) {
tmp_1 = t_8;
} else if (dY_46_u <= 20000000000.0f) {
float tmp_2;
if (t_10 >= t_5) {
tmp_2 = t_2 / sqrtf(fmaxf(t_10, t_5));
} else {
tmp_2 = (dY_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((dY_46_v * dY_46_v) * t_9)))) * floorf(w);
}
tmp_1 = tmp_2;
} else {
tmp_1 = t_8;
}
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) * floor(w)) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(Float32(t_0 * t_0) + Float32(t_3 * t_3)) t_5 = Float32(Float32(t_3 * dY_46_v) * floor(h)) t_6 = Float32(Float32(Float32(-dX_46_u) * dX_46_u) * t_1) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_4))) tmp = Float32(0.0) if (t_6 >= t_4) tmp = Float32(t_7 * t_2); else tmp = Float32(t_7 * t_0); end t_8 = tmp t_9 = Float32(floor(h) * floor(h)) t_10 = fma(Float32(dX_46_u * dX_46_u), t_1, Float32(Float32(dX_46_v * dX_46_v) * t_9)) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-260.0)) tmp_1 = t_8; elseif (dY_46_u <= Float32(20000000000.0)) tmp_2 = Float32(0.0) if (t_10 >= t_5) tmp_2 = Float32(t_2 / sqrt(fmax(t_10, t_5))); else tmp_2 = Float32(Float32(dY_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(dY_46_v * dY_46_v) * t_9)))) * floor(w)); end tmp_1 = tmp_2; else tmp_1 = t_8; 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 \left\lfloor w\right\rfloor \\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_0 \cdot t\_0 + t\_3 \cdot t\_3\\
t_5 := \left(t\_3 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_6 := \left(\left(-dX.u\right) \cdot dX.u\right) \cdot t\_1\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_4\right)}}\\
t_8 := \begin{array}{l}
\mathbf{if}\;t\_6 \geq t\_4:\\
\;\;\;\;t\_7 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_0\\
\end{array}\\
t_9 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_10 := \mathsf{fma}\left(dX.u \cdot dX.u, t\_1, \left(dX.v \cdot dX.v\right) \cdot t\_9\right)\\
\mathbf{if}\;dY.u \leq -260:\\
\;\;\;\;t\_8\\
\mathbf{elif}\;dY.u \leq 20000000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_10 \geq t\_5:\\
\;\;\;\;\frac{t\_2}{\sqrt{\mathsf{max}\left(t\_10, t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(dY.v \cdot dY.v\right) \cdot t\_9\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_8\\
\end{array}
\end{array}
if dY.u < -260 or 2e10 < dY.u Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.9%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites69.9%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites46.9%
if -260 < dY.u < 2e10Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.1%
Applied rewrites63.1%
(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 (* (floor w) (floor w)))
(t_3 (* (floor w) dX.u))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_0 t_0) (* t_4 t_4)))
(t_6
(sqrt
(/
1.0
(fmax
(fma (* dX.u dX.u) t_2 (* (* dX.v dX.v) t_1))
(fma (* t_4 dY.v) (floor h) (* (* dY.u dY.u) t_2))))))
(t_7 (* (floor h) dX.v))
(t_8 (* (* (- dX.u) dX.u) t_2))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_5)))))
(if (<= dY.u -2000.0)
(if (>= t_8 t_5) (* t_9 t_3) (* t_9 t_0))
(if (>= (+ (* t_3 t_3) (* t_7 t_7)) (* (* dY.v dY.v) t_1))
(* t_6 t_3)
(* t_6 t_0)))))
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 = floorf(w) * floorf(w);
float t_3 = floorf(w) * dX_46_u;
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_0 * t_0) + (t_4 * t_4);
float t_6 = sqrtf((1.0f / fmaxf(fmaf((dX_46_u * dX_46_u), t_2, ((dX_46_v * dX_46_v) * t_1)), fmaf((t_4 * dY_46_v), floorf(h), ((dY_46_u * dY_46_u) * t_2)))));
float t_7 = floorf(h) * dX_46_v;
float t_8 = (-dX_46_u * dX_46_u) * t_2;
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_5));
float tmp_1;
if (dY_46_u <= -2000.0f) {
float tmp_2;
if (t_8 >= t_5) {
tmp_2 = t_9 * t_3;
} else {
tmp_2 = t_9 * t_0;
}
tmp_1 = tmp_2;
} else if (((t_3 * t_3) + (t_7 * t_7)) >= ((dY_46_v * dY_46_v) * t_1)) {
tmp_1 = t_6 * t_3;
} else {
tmp_1 = t_6 * t_0;
}
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(h) * floor(h)) t_2 = Float32(floor(w) * floor(w)) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) t_6 = sqrt(Float32(Float32(1.0) / fmax(fma(Float32(dX_46_u * dX_46_u), t_2, Float32(Float32(dX_46_v * dX_46_v) * t_1)), fma(Float32(t_4 * dY_46_v), floor(h), Float32(Float32(dY_46_u * dY_46_u) * t_2))))) t_7 = Float32(floor(h) * dX_46_v) t_8 = Float32(Float32(Float32(-dX_46_u) * dX_46_u) * t_2) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_5))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-2000.0)) tmp_2 = Float32(0.0) if (t_8 >= t_5) tmp_2 = Float32(t_9 * t_3); else tmp_2 = Float32(t_9 * t_0); end tmp_1 = tmp_2; elseif (Float32(Float32(t_3 * t_3) + Float32(t_7 * t_7)) >= Float32(Float32(dY_46_v * dY_46_v) * t_1)) tmp_1 = Float32(t_6 * t_3); else tmp_1 = Float32(t_6 * t_0); 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 h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
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 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_0 \cdot t\_0 + t\_4 \cdot t\_4\\
t_6 := \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(dX.u \cdot dX.u, t\_2, \left(dX.v \cdot dX.v\right) \cdot t\_1\right), \mathsf{fma}\left(t\_4 \cdot dY.v, \left\lfloor h\right\rfloor , \left(dY.u \cdot dY.u\right) \cdot t\_2\right)\right)}}\\
t_7 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_8 := \left(\left(-dX.u\right) \cdot dX.u\right) \cdot t\_2\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_5\right)}}\\
\mathbf{if}\;dY.u \leq -2000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_5:\\
\;\;\;\;t\_9 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_0\\
\end{array}\\
\mathbf{elif}\;t\_3 \cdot t\_3 + t\_7 \cdot t\_7 \geq \left(dY.v \cdot dY.v\right) \cdot t\_1:\\
\;\;\;\;t\_6 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_0\\
\end{array}
\end{array}
if dY.u < -2e3Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.9%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites69.9%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites46.9%
if -2e3 < dY.u Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Taylor expanded in w around 0
Applied rewrites59.6%
Taylor expanded in w around 0
Applied rewrites65.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_1 t_1) (* t_2 t_2)))
(t_4 (* (* dY.v dY.v) (* (floor h) (floor h))))
(t_5 (* (floor w) (floor w)))
(t_6 (sqrt (/ 1.0 (fmax t_0 (fma (* dY.u dY.u) t_5 t_4)))))
(t_7 (* (floor w) dX.u))
(t_8 (* (* (- dX.u) dX.u) t_5))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_3)))))
(if (<= dY.u -2000.0)
(if (>= t_8 t_3) (* t_9 t_7) (* t_9 t_1))
(if (>= t_0 t_4) (* t_6 t_7) (* 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 = fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)));
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = (t_1 * t_1) + (t_2 * t_2);
float t_4 = (dY_46_v * dY_46_v) * (floorf(h) * floorf(h));
float t_5 = floorf(w) * floorf(w);
float t_6 = sqrtf((1.0f / fmaxf(t_0, fmaf((dY_46_u * dY_46_u), t_5, t_4))));
float t_7 = floorf(w) * dX_46_u;
float t_8 = (-dX_46_u * dX_46_u) * t_5;
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_3));
float tmp_1;
if (dY_46_u <= -2000.0f) {
float tmp_2;
if (t_8 >= t_3) {
tmp_2 = t_9 * t_7;
} else {
tmp_2 = t_9 * t_1;
}
tmp_1 = tmp_2;
} else if (t_0 >= t_4) {
tmp_1 = t_6 * t_7;
} else {
tmp_1 = t_6 * t_1;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) t_4 = Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))) t_5 = Float32(floor(w) * floor(w)) t_6 = sqrt(Float32(Float32(1.0) / fmax(t_0, fma(Float32(dY_46_u * dY_46_u), t_5, t_4)))) t_7 = Float32(floor(w) * dX_46_u) t_8 = Float32(Float32(Float32(-dX_46_u) * dX_46_u) * t_5) t_9 = Float32(Float32(1.0) / sqrt(fmax(t_8, t_3))) tmp_1 = Float32(0.0) if (dY_46_u <= Float32(-2000.0)) tmp_2 = Float32(0.0) if (t_8 >= t_3) tmp_2 = Float32(t_9 * t_7); else tmp_2 = Float32(t_9 * t_1); end tmp_1 = tmp_2; elseif (t_0 >= t_4) tmp_1 = Float32(t_6 * t_7); else tmp_1 = Float32(t_6 * t_1); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_1 \cdot t\_1 + t\_2 \cdot t\_2\\
t_4 := \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\\
t_5 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_6 := \sqrt{\frac{1}{\mathsf{max}\left(t\_0, \mathsf{fma}\left(dY.u \cdot dY.u, t\_5, t\_4\right)\right)}}\\
t_7 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_8 := \left(\left(-dX.u\right) \cdot dX.u\right) \cdot t\_5\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_3\right)}}\\
\mathbf{if}\;dY.u \leq -2000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_3:\\
\;\;\;\;t\_9 \cdot t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_9 \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;t\_0 \geq t\_4:\\
\;\;\;\;t\_6 \cdot t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
if dY.u < -2e3Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.9%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
Applied rewrites69.9%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites44.8%
Taylor expanded in dX.u around inf
mul-1-negN/A
lower-neg.f64N/A
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
lower-special-*N/A
Applied rewrites46.9%
if -2e3 < dY.u Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Taylor expanded in w around 0
Applied rewrites59.6%
Taylor expanded in w around 0
Applied rewrites65.1%
Applied rewrites65.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (* dY.v dY.v) (* (floor h) (floor h))))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0))))
(if (>= t_3 t_1)
(* (/ 1.0 (sqrt (fmax t_3 t_1))) t_2)
(/
1.0
(/
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(fma (* dY.u dY.u) (* (floor w) (floor w)) t_1)))
(* (floor w) dY.u))))))
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 = (dY_46_v * dY_46_v) * (floorf(h) * floorf(h));
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float tmp;
if (t_3 >= t_1) {
tmp = (1.0f / sqrtf(fmaxf(t_3, t_1))) * t_2;
} else {
tmp = 1.0f / (sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), fmaf((dY_46_u * dY_46_u), (floorf(w) * floorf(w)), t_1))) / (floorf(w) * dY_46_u));
}
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(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h))) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) tmp = Float32(0.0) if (t_3 >= t_1) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_3, t_1))) * t_2); else tmp = Float32(Float32(1.0) / Float32(sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), fma(Float32(dY_46_u * dY_46_u), Float32(floor(w) * floor(w)), t_1))) / Float32(floor(w) * dY_46_u))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
\mathbf{if}\;t\_3 \geq t\_1:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_1\right)}} \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \mathsf{fma}\left(dY.u \cdot dY.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , t\_1\right)\right)}}{\left\lfloor w\right\rfloor \cdot dY.u}}\\
\end{array}
\end{array}
Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.0%
Taylor expanded in w around 0
Applied rewrites68.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))
(t_2
(fma (* dX.u dX.u) (* (floor w) (floor w)) (* (* dX.v dX.v) t_0))))
(if (>= t_2 t_1)
(/ (* (floor w) dX.u) (sqrt (fmax t_2 t_1)))
(*
(/
dY.u
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(* (* dY.v dY.v) t_0))))
(floor w)))))
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(h) * dY_46_v) * dY_46_v) * floorf(h);
float t_2 = fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((dX_46_v * dX_46_v) * t_0));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_2, t_1));
} else {
tmp = (dY_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((dY_46_v * dY_46_v) * t_0)))) * floorf(w);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)) t_2 = fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dX_46_v * dX_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_2, t_1))); else tmp = Float32(Float32(dY_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(dY_46_v * dY_46_v) * t_0)))) * floor(w)); 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(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{dY.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(dY.v \cdot dY.v\right) \cdot t\_0\right)}} \cdot \left\lfloor w\right\rfloor \\
\end{array}
\end{array}
Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.1%
Applied rewrites63.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))
(t_2
(fma (* dX.u dX.u) (* (floor w) (floor w)) (* (* dX.v dX.v) t_0))))
(if (>= t_2 t_1)
(/ (* (floor w) dX.u) (sqrt (fmax t_2 t_1)))
(*
dY.u
(/
(floor w)
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(* (* dY.v dY.v) t_0))))))))
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(h) * dY_46_v) * dY_46_v) * floorf(h);
float t_2 = fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((dX_46_v * dX_46_v) * t_0));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(w) * dX_46_u) / sqrtf(fmaxf(t_2, t_1));
} else {
tmp = dY_46_u * (floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((dY_46_v * dY_46_v) * t_0))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)) t_2 = fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dX_46_v * dX_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(w) * dX_46_u) / sqrt(fmax(t_2, t_1))); else tmp = Float32(dY_46_u * Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(dY_46_v * dY_46_v) * t_0))))); 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(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dX.u}{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\mathbf{else}:\\
\;\;\;\;dY.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(dY.v \cdot dY.v\right) \cdot t\_0\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.1%
Applied rewrites63.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))
(t_2
(fma (* dX.u dX.u) (* (floor w) (floor w)) (* (* dX.v dX.v) t_0))))
(if (>= t_2 t_1)
(*
(/
dX.u
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(* (* dY.v dY.v) t_0))))
(floor w))
(/ (* (floor w) dY.u) (sqrt (fmax t_2 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) * floorf(h);
float t_1 = ((floorf(h) * dY_46_v) * dY_46_v) * floorf(h);
float t_2 = fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((dX_46_v * dX_46_v) * t_0));
float tmp;
if (t_2 >= t_1) {
tmp = (dX_46_u / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((dY_46_v * dY_46_v) * t_0)))) * floorf(w);
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_2, 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) * floor(h)) t_1 = Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)) t_2 = fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dX_46_v * dX_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(dX_46_u / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(dY_46_v * dY_46_v) * t_0)))) * floor(w)); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_2, t_1))); 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(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{dX.u}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(dY.v \cdot dY.v\right) \cdot t\_0\right)}} \cdot \left\lfloor w\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.1%
Applied rewrites63.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* (* (* (floor h) dY.v) dY.v) (floor h)))
(t_2
(fma (* dX.u dX.u) (* (floor w) (floor w)) (* (* dX.v dX.v) t_0))))
(if (>= t_2 t_1)
(*
dX.u
(/
(floor w)
(sqrt
(fmax
(fma
(* (* dX.u dX.u) (floor w))
(floor w)
(* (* (* dX.v (floor h)) dX.v) (floor h)))
(* (* dY.v dY.v) t_0)))))
(/ (* (floor w) dY.u) (sqrt (fmax t_2 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) * floorf(h);
float t_1 = ((floorf(h) * dY_46_v) * dY_46_v) * floorf(h);
float t_2 = fmaf((dX_46_u * dX_46_u), (floorf(w) * floorf(w)), ((dX_46_v * dX_46_v) * t_0));
float tmp;
if (t_2 >= t_1) {
tmp = dX_46_u * (floorf(w) / sqrtf(fmaxf(fmaf(((dX_46_u * dX_46_u) * floorf(w)), floorf(w), (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h))), ((dY_46_v * dY_46_v) * t_0))));
} else {
tmp = (floorf(w) * dY_46_u) / sqrtf(fmaxf(t_2, 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) * floor(h)) t_1 = Float32(Float32(Float32(floor(h) * dY_46_v) * dY_46_v) * floor(h)) t_2 = fma(Float32(dX_46_u * dX_46_u), Float32(floor(w) * floor(w)), Float32(Float32(dX_46_v * dX_46_v) * t_0)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(dX_46_u * Float32(floor(w) / sqrt(fmax(fma(Float32(Float32(dX_46_u * dX_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))), Float32(Float32(dY_46_v * dY_46_v) * t_0))))); else tmp = Float32(Float32(floor(w) * dY_46_u) / sqrt(fmax(t_2, t_1))); 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(\left(\left\lfloor h\right\rfloor \cdot dY.v\right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(dX.u \cdot dX.u, \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor , \left(dX.v \cdot dX.v\right) \cdot t\_0\right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;dX.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right), \left(dY.v \cdot dY.v\right) \cdot t\_0\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor \cdot dY.u}{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.7%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.5%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
swap-sqrN/A
unpow2N/A
unpow2N/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f32N/A
pow-prod-downN/A
*-commutativeN/A
+-commutativeN/A
Applied rewrites67.0%
Taylor expanded in dY.u around 0
Applied rewrites56.6%
Taylor expanded in dY.u around 0
Applied rewrites60.0%
Taylor expanded in dY.u around 0
Applied rewrites63.0%
Applied rewrites63.1%
Applied rewrites63.0%
herbie shell --seed 2025132
(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))))