
(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 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_2) (* t_6 t_1))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_2;
} else {
tmp = t_6 * t_1;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_2); else tmp = Float32(t_6 * t_1); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_2; else tmp = t_6 * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_2\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_1\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* dY.v (floor h)) dY.v))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (* (* dY.u (floor w)) dY.u))
(t_3 (- (sqrt (fmax t_1 (fma t_2 (floor w) (* t_0 (floor h))))))))
(if (>= t_1 (fma t_0 (floor h) (* t_2 (floor w))))
(/ (* (- dX.u) (floor w)) t_3)
(/ (* (- dY.u) (floor w)) 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 = (dY_46_v * floorf(h)) * dY_46_v;
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = (dY_46_u * floorf(w)) * dY_46_u;
float t_3 = -sqrtf(fmaxf(t_1, fmaf(t_2, floorf(w), (t_0 * floorf(h)))));
float tmp;
if (t_1 >= fmaf(t_0, floorf(h), (t_2 * floorf(w)))) {
tmp = (-dX_46_u * floorf(w)) / t_3;
} else {
tmp = (-dY_46_u * floorf(w)) / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(dY_46_v * floor(h)) * dY_46_v) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_3 = Float32(-sqrt(fmax(t_1, fma(t_2, floor(w), Float32(t_0 * floor(h)))))) tmp = Float32(0.0) if (t_1 >= fma(t_0, floor(h), Float32(t_2 * floor(w)))) tmp = Float32(Float32(Float32(-dX_46_u) * floor(w)) / t_3); else tmp = Float32(Float32(Float32(-dY_46_u) * floor(w)) / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_3 := -\sqrt{\mathsf{max}\left(t\_1, \mathsf{fma}\left(t\_2, \left\lfloor w\right\rfloor , t\_0 \cdot \left\lfloor h\right\rfloor \right)\right)}\\
\mathbf{if}\;t\_1 \geq \mathsf{fma}\left(t\_0, \left\lfloor h\right\rfloor , t\_2 \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;\frac{\left(-dX.u\right) \cdot \left\lfloor w\right\rfloor }{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{t\_3}\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.7%
Applied rewrites76.8%
lift-neg.f32N/A
lift-sqrt.f32N/A
pow1/2N/A
Applied rewrites76.8%
Applied rewrites76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* dY.u (floor w)) dY.u))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (* (* dY.v (floor h)) dY.v))
(t_3 (fma t_0 (floor w) (* t_2 (floor h)))))
(if (>= t_1 t_3)
(/ (* (- dX.u) (floor w)) (- (sqrt (fmax t_1 t_3))))
(/
(* (- dY.u) (floor w))
(- (sqrt (fmax t_1 (fma t_2 (floor h) (* 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 = (dY_46_u * floorf(w)) * dY_46_u;
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = (dY_46_v * floorf(h)) * dY_46_v;
float t_3 = fmaf(t_0, floorf(w), (t_2 * floorf(h)));
float tmp;
if (t_1 >= t_3) {
tmp = (-dX_46_u * floorf(w)) / -sqrtf(fmaxf(t_1, t_3));
} else {
tmp = (-dY_46_u * floorf(w)) / -sqrtf(fmaxf(t_1, fmaf(t_2, floorf(h), (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(Float32(dY_46_u * floor(w)) * dY_46_u) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(Float32(dY_46_v * floor(h)) * dY_46_v) t_3 = fma(t_0, floor(w), Float32(t_2 * floor(h))) tmp = Float32(0.0) if (t_1 >= t_3) tmp = Float32(Float32(Float32(-dX_46_u) * floor(w)) / Float32(-sqrt(fmax(t_1, t_3)))); else tmp = Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-sqrt(fmax(t_1, fma(t_2, floor(h), Float32(t_0 * floor(w))))))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_3 := \mathsf{fma}\left(t\_0, \left\lfloor w\right\rfloor , t\_2 \cdot \left\lfloor h\right\rfloor \right)\\
\mathbf{if}\;t\_1 \geq t\_3:\\
\;\;\;\;\frac{\left(-dX.u\right) \cdot \left\lfloor w\right\rfloor }{-\sqrt{\mathsf{max}\left(t\_1, t\_3\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{-\sqrt{\mathsf{max}\left(t\_1, \mathsf{fma}\left(t\_2, \left\lfloor h\right\rfloor , t\_0 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.7%
Applied rewrites76.8%
Applied rewrites76.8%
(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 (floor h)) dY.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (* dY.u (floor w)) dY.u))
(t_4 (* (floor h) dY.v))
(t_5 (* (floor w) dX.u))
(t_6 (* dX.u (floor w)))
(t_7
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_6 dX.u) (floor w)))))
(if (>= (+ (* t_5 t_5) (* t_0 t_0)) (+ (* t_2 t_2) (* t_4 t_4)))
(/ t_6 (sqrt (fmax t_7 (fma t_1 (floor h) (* t_3 (floor w))))))
(*
(floor w)
(/ dY.u (sqrt (fmax t_7 (fma t_3 (floor w) (* t_1 (floor h))))))))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = (dY_46_v * floorf(h)) * dY_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = (dY_46_u * floorf(w)) * dY_46_u;
float t_4 = floorf(h) * dY_46_v;
float t_5 = floorf(w) * dX_46_u;
float t_6 = dX_46_u * floorf(w);
float t_7 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_6 * dX_46_u) * floorf(w)));
float tmp;
if (((t_5 * t_5) + (t_0 * t_0)) >= ((t_2 * t_2) + (t_4 * t_4))) {
tmp = t_6 / sqrtf(fmaxf(t_7, fmaf(t_1, floorf(h), (t_3 * floorf(w)))));
} else {
tmp = floorf(w) * (dY_46_u / sqrtf(fmaxf(t_7, fmaf(t_3, floorf(w), (t_1 * floorf(h))))));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(Float32(dY_46_v * floor(h)) * dY_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(Float32(dY_46_u * floor(w)) * dY_46_u) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(dX_46_u * floor(w)) t_7 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_6 * dX_46_u) * floor(w))) tmp = Float32(0.0) if (Float32(Float32(t_5 * t_5) + Float32(t_0 * t_0)) >= Float32(Float32(t_2 * t_2) + Float32(t_4 * t_4))) tmp = Float32(t_6 / sqrt(fmax(t_7, fma(t_1, floor(h), Float32(t_3 * floor(w)))))); else tmp = Float32(floor(w) * Float32(dY_46_u / sqrt(fmax(t_7, fma(t_3, floor(w), Float32(t_1 * floor(h))))))); 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 \left\lfloor h\right\rfloor \right) \cdot dY.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_7 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_6 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;t\_5 \cdot t\_5 + t\_0 \cdot t\_0 \geq t\_2 \cdot t\_2 + t\_4 \cdot t\_4:\\
\;\;\;\;\frac{t\_6}{\sqrt{\mathsf{max}\left(t\_7, \mathsf{fma}\left(t\_1, \left\lfloor h\right\rfloor , t\_3 \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \frac{dY.u}{\sqrt{\mathsf{max}\left(t\_7, \mathsf{fma}\left(t\_3, \left\lfloor w\right\rfloor , t\_1 \cdot \left\lfloor h\right\rfloor \right)\right)}}\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.7%
Applied rewrites76.8%
lift-/.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-/l*N/A
lift-neg.f32N/A
Applied rewrites76.7%
Applied rewrites76.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v dY.v) (floor h)) (floor h))))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)));
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.5
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.5
Applied rewrites76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v dY.v) (floor h)) (floor h))))
(t_1
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)));
float t_1 = fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.5
Applied rewrites76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_1
(fma
(* (* dX.v dX.v) (floor h))
(floor h)
(* (* (* dX.u dX.u) (floor w)) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_1 = fmaf(((dX_46_v * dX_46_v) * floorf(h)), floorf(h), (((dX_46_u * dX_46_u) * floorf(w)) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * dY_46_u;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = fma(Float32(Float32(dY_46_u * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))) t_1 = fma(Float32(Float32(dX_46_v * dX_46_v) * floor(h)), floor(h), Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, \left\lfloor w\right\rfloor , \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f3276.6
lift-pow.f32N/A
unpow2N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.6
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.5
Applied rewrites76.5%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3276.5
Applied rewrites76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.u (floor w)))
(t_2 (* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w)))))
(t_3
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_1 dX.u) (floor w)))))
(if (>= t_3 t_2)
(/ t_1 (sqrt (fmax t_3 (* (fabs t_0) t_0))))
(* (/ (floor w) (sqrt (fmax t_3 t_2))) 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 = dY_46_u * floorf(w);
float t_1 = dX_46_u * floorf(w);
float t_2 = dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w)));
float t_3 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_1 * dX_46_u) * floorf(w)));
float tmp;
if (t_3 >= t_2) {
tmp = t_1 / sqrtf(fmaxf(t_3, (fabsf(t_0) * t_0)));
} else {
tmp = (floorf(w) / sqrtf(fmaxf(t_3, t_2))) * 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(dY_46_u * floor(w)) t_1 = Float32(dX_46_u * floor(w)) t_2 = Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w)))) t_3 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_1 * dX_46_u) * floor(w))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_1 / sqrt(fmax(t_3, Float32(abs(t_0) * t_0)))); else tmp = Float32(Float32(floor(w) / sqrt(fmax(t_3, t_2))) * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_3 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_1 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_3, \left|t\_0\right| \cdot t\_0\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}} \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3268.1
Applied rewrites68.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.4
Applied rewrites70.4%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.0
Applied rewrites70.0%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.4
Applied rewrites56.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3255.9
Applied rewrites55.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3253.1
Applied rewrites53.1%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
Applied rewrites53.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dY.u (* (fabs dY.u) (* (fabs (floor w)) (floor w)))))
(t_2
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))))
(if (>= t_2 t_1)
(* (/ (floor w) (sqrt (fmax t_2 t_1))) dX.u)
(/ t_0 (sqrt (fmax t_2 (* (fabs t_0) 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 = dY_46_u * floorf(w);
float t_1 = dY_46_u * (fabsf(dY_46_u) * (fabsf(floorf(w)) * floorf(w)));
float t_2 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(w) / sqrtf(fmaxf(t_2, t_1))) * dX_46_u;
} else {
tmp = t_0 / sqrtf(fmaxf(t_2, (fabsf(t_0) * t_0)));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_u * floor(w)) t_1 = Float32(dY_46_u * Float32(abs(dY_46_u) * Float32(abs(floor(w)) * floor(w)))) t_2 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(w) / sqrt(fmax(t_2, t_1))) * dX_46_u); else tmp = Float32(t_0 / sqrt(fmax(t_2, Float32(abs(t_0) * t_0)))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.u \cdot \left(\left|dY.u\right| \cdot \left(\left|\left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right)\right)\\
t_2 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_2, t\_1\right)}} \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{\sqrt{\mathsf{max}\left(t\_2, \left|t\_0\right| \cdot t\_0\right)}}\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3268.1
Applied rewrites68.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.4
Applied rewrites70.4%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.0
Applied rewrites70.0%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.4
Applied rewrites56.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3255.9
Applied rewrites55.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3253.1
Applied rewrites53.1%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
Applied rewrites53.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (fabs (* dY.u (floor w))) (floor w)) dY.u))
(t_1
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_1 t_0)))))
(if (>= t_1 t_0) (* t_2 dX.u) (* t_2 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 = (fabsf((dY_46_u * floorf(w))) * floorf(w)) * dY_46_u;
float t_1 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_1, t_0));
float tmp;
if (t_1 >= t_0) {
tmp = t_2 * dX_46_u;
} else {
tmp = t_2 * 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(Float32(abs(Float32(dY_46_u * floor(w))) * floor(w)) * dY_46_u) t_1 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_1, t_0))) tmp = Float32(0.0) if (t_1 >= t_0) tmp = Float32(t_2 * dX_46_u); else tmp = Float32(t_2 * dY_46_u); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left|dY.u \cdot \left\lfloor w\right\rfloor \right| \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_1 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_1, t\_0\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
\end{array}
Initial program 76.6%
Applied rewrites76.6%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3268.1
Applied rewrites68.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.4
Applied rewrites70.4%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-log.f32N/A
exp-fabsN/A
lift-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
rem-exp-logN/A
lift-*.f32N/A
fabs-mulN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-fabs.f3270.0
Applied rewrites70.0%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3256.4
Applied rewrites56.4%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3255.9
Applied rewrites55.9%
Taylor expanded in dY.u around inf
lower-*.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-*.f32N/A
lower-fabs.f32N/A
lower-floor.f32N/A
lower-floor.f3253.1
Applied rewrites53.1%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3253.1
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-fabs.f32N/A
lift-fabs.f32N/A
mul-fabsN/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3253.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3253.1
Applied rewrites53.1%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3253.1
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-fabs.f32N/A
lift-fabs.f32N/A
mul-fabsN/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3253.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3253.1
Applied rewrites53.1%
lift-*.f32N/A
*-commutativeN/A
lower-*.f3253.1
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-fabs.f32N/A
lift-fabs.f32N/A
mul-fabsN/A
*-commutativeN/A
lift-*.f32N/A
lower-*.f32N/A
lower-fabs.f3253.1
lift-*.f32N/A
*-commutativeN/A
lift-*.f3253.1
Applied rewrites53.1%
herbie shell --seed 2025158
(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))))