
(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}
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}
Herbie found 6 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}
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}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (sqrt (* (fabs dY.u) (floor w))))
(t_1 (* (* t_0 t_0) (fabs dY.u)))
(t_2 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_3 (* (floor h) (floor h)))
(t_4 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_5 (fma (* (* dX.v (floor h)) dX.v) (floor h) t_4)))
(if (>=
(fma (* dX.v dX.v) t_3 t_4)
(fma (* (* (floor w) (floor w)) (fabs dY.u)) (fabs dY.u) t_2))
(/
(* (- dX.v) (floor h))
(- (sqrt (fmax (fma (* dY.v dY.v) t_3 (* t_1 (floor w))) t_5))))
(/
(* (- dY.v) (floor h))
(- (sqrt (fmax (fma t_1 (floor w) t_2) t_5)))))))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 = sqrtf((fabsf(dY_46_u) * floorf(w)));
float t_1 = (t_0 * t_0) * fabsf(dY_46_u);
float t_2 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_3 = floorf(h) * floorf(h);
float t_4 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_5 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_4);
float tmp;
if (fmaf((dX_46_v * dX_46_v), t_3, t_4) >= fmaf(((floorf(w) * floorf(w)) * fabsf(dY_46_u)), fabsf(dY_46_u), t_2)) {
tmp = (-dX_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf((dY_46_v * dY_46_v), t_3, (t_1 * floorf(w))), t_5));
} else {
tmp = (-dY_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf(t_1, floorf(w), t_2), t_5));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = sqrt(Float32(abs(dY_46_u) * floor(w))) t_1 = Float32(Float32(t_0 * t_0) * abs(dY_46_u)) t_2 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_3 = Float32(floor(h) * floor(h)) t_4 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_5 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_4) tmp = Float32(0.0) if (fma(Float32(dX_46_v * dX_46_v), t_3, t_4) >= fma(Float32(Float32(floor(w) * floor(w)) * abs(dY_46_u)), abs(dY_46_u), t_2)) tmp = Float32(Float32(Float32(-dX_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(Float32(dY_46_v * dY_46_v), t_3, Float32(t_1 * floor(w))), t_5)))); else tmp = Float32(Float32(Float32(-dY_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(t_1, floor(w), t_2), t_5)))); end return tmp end
\begin{array}{l}
t_0 := \sqrt{\left|dY.u\right| \cdot \left\lfloor w\right\rfloor }\\
t_1 := \left(t\_0 \cdot t\_0\right) \cdot \left|dY.u\right|\\
t_2 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_5 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_4\right)\\
\mathbf{if}\;\mathsf{fma}\left(dX.v \cdot dX.v, t\_3, t\_4\right) \geq \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot \left|dY.u\right|, \left|dY.u\right|, t\_2\right):\\
\;\;\;\;\frac{\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.v \cdot dY.v, t\_3, t\_1 \cdot \left\lfloor w\right\rfloor \right), t\_5\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_1, \left\lfloor w\right\rfloor , t\_2\right), t\_5\right)}}\\
\end{array}
Initial program 76.0%
Applied rewrites76.1%
Applied rewrites76.2%
Applied rewrites76.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.2%
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
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-unsound-sqrt.f3276.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.2%
Applied rewrites76.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3276.2%
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
rem-sqrt-square-revN/A
sqrt-prodN/A
lower-unsound-*.f32N/A
lower-unsound-sqrt.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-unsound-sqrt.f3260.5%
lift-*.f32N/A
*-commutativeN/A
lift-*.f3260.5%
Applied rewrites60.5%
(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 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_2 (* (floor h) (floor h)))
(t_3 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_4 (fma (* (* dX.v (floor h)) dX.v) (floor h) t_3)))
(if (>=
(fma (* dX.v dX.v) t_2 t_3)
(fma (* (* (floor w) (floor w)) dY.u) dY.u t_1))
(/
(* (- dX.v) (floor h))
(- (sqrt (fmax (fma (* dY.v dY.v) t_2 (* t_0 (floor w))) t_4))))
(/
(* (- dY.v) (floor h))
(- (sqrt (fmax (fma t_0 (floor w) 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 = (dY_46_u * floorf(w)) * dY_46_u;
float t_1 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_2 = floorf(h) * floorf(h);
float t_3 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_4 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_3);
float tmp;
if (fmaf((dX_46_v * dX_46_v), t_2, t_3) >= fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_1)) {
tmp = (-dX_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf((dY_46_v * dY_46_v), t_2, (t_0 * floorf(w))), t_4));
} else {
tmp = (-dY_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf(t_0, floorf(w), 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(Float32(dY_46_u * floor(w)) * dY_46_u) t_1 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_4 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_3) tmp = Float32(0.0) if (fma(Float32(dX_46_v * dX_46_v), t_2, t_3) >= fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_1)) tmp = Float32(Float32(Float32(-dX_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(Float32(dY_46_v * dY_46_v), t_2, Float32(t_0 * floor(w))), t_4)))); else tmp = Float32(Float32(Float32(-dY_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(t_0, floor(w), t_1), t_4)))); end return tmp end
\begin{array}{l}
t_0 := \left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\\
t_1 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_4 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_3\right)\\
\mathbf{if}\;\mathsf{fma}\left(dX.v \cdot dX.v, t\_2, t\_3\right) \geq \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_1\right):\\
\;\;\;\;\frac{\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.v \cdot dY.v, t\_2, t\_0 \cdot \left\lfloor w\right\rfloor \right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor w\right\rfloor , t\_1\right), t\_4\right)}}\\
\end{array}
Initial program 76.0%
Applied rewrites76.1%
Applied rewrites76.2%
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* dY.u dY.u) (floor w)))
(t_1 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_2 (* (floor h) (floor h)))
(t_3 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_4 (fma (* (* dX.v (floor h)) dX.v) (floor h) t_3)))
(if (>=
(fma (* dX.v dX.v) t_2 t_3)
(fma (* (* (floor w) (floor w)) dY.u) dY.u t_1))
(/
(* (- dX.v) (floor h))
(- (sqrt (fmax (fma (* dY.v dY.v) t_2 (* t_0 (floor w))) t_4))))
(/
(* (- dY.v) (floor h))
(- (sqrt (fmax (fma t_0 (floor w) 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 = (dY_46_u * dY_46_u) * floorf(w);
float t_1 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_2 = floorf(h) * floorf(h);
float t_3 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_4 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_3);
float tmp;
if (fmaf((dX_46_v * dX_46_v), t_2, t_3) >= fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_1)) {
tmp = (-dX_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf((dY_46_v * dY_46_v), t_2, (t_0 * floorf(w))), t_4));
} else {
tmp = (-dY_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf(t_0, floorf(w), 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(Float32(dY_46_u * dY_46_u) * floor(w)) t_1 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_2 = Float32(floor(h) * floor(h)) t_3 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_4 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_3) tmp = Float32(0.0) if (fma(Float32(dX_46_v * dX_46_v), t_2, t_3) >= fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_1)) tmp = Float32(Float32(Float32(-dX_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(Float32(dY_46_v * dY_46_v), t_2, Float32(t_0 * floor(w))), t_4)))); else tmp = Float32(Float32(Float32(-dY_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(t_0, floor(w), t_1), t_4)))); end return tmp end
\begin{array}{l}
t_0 := \left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \\
t_1 := \left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_2 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_4 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_3\right)\\
\mathbf{if}\;\mathsf{fma}\left(dX.v \cdot dX.v, t\_2, t\_3\right) \geq \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_1\right):\\
\;\;\;\;\frac{\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(dY.v \cdot dY.v, t\_2, t\_0 \cdot \left\lfloor w\right\rfloor \right), t\_4\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, \left\lfloor w\right\rfloor , t\_1\right), t\_4\right)}}\\
\end{array}
Initial program 76.0%
Applied rewrites76.1%
Applied rewrites76.2%
Applied rewrites76.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3276.2%
Applied rewrites76.2%
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3276.2%
Applied rewrites76.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* (floor w) (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (* (* t_2 dY.v) (floor h)))
(t_4 (* (* (* dX.u (floor w)) dX.u) (floor w))))
(if (>=
(fma (* dX.v dX.v) (* (floor h) (floor h)) t_4)
(fma (* t_1 dY.u) dY.u t_3))
(*
(floor h)
(/
dX.v
(sqrt
(fmax
(fma t_2 t_2 (* t_0 t_0))
(fma (* (* (floor h) dX.v) (floor h)) dX.v (* (* dX.u dX.u) t_1))))))
(/
(* (- dY.v) (floor h))
(-
(sqrt
(fmax
(fma (* t_0 dY.u) (floor w) t_3)
(fma (* (* dX.v (floor h)) dX.v) (floor h) 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 = dY_46_u * floorf(w);
float t_1 = floorf(w) * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = (t_2 * dY_46_v) * floorf(h);
float t_4 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float tmp;
if (fmaf((dX_46_v * dX_46_v), (floorf(h) * floorf(h)), t_4) >= fmaf((t_1 * dY_46_u), dY_46_u, t_3)) {
tmp = floorf(h) * (dX_46_v / sqrtf(fmaxf(fmaf(t_2, t_2, (t_0 * t_0)), fmaf(((floorf(h) * dX_46_v) * floorf(h)), dX_46_v, ((dX_46_u * dX_46_u) * t_1)))));
} else {
tmp = (-dY_46_v * floorf(h)) / -sqrtf(fmaxf(fmaf((t_0 * dY_46_u), floorf(w), t_3), fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_4)));
}
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(floor(w) * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(Float32(t_2 * dY_46_v) * floor(h)) t_4 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) tmp = Float32(0.0) if (fma(Float32(dX_46_v * dX_46_v), Float32(floor(h) * floor(h)), t_4) >= fma(Float32(t_1 * dY_46_u), dY_46_u, t_3)) tmp = Float32(floor(h) * Float32(dX_46_v / sqrt(fmax(fma(t_2, t_2, Float32(t_0 * t_0)), fma(Float32(Float32(floor(h) * dX_46_v) * floor(h)), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_1)))))); else tmp = Float32(Float32(Float32(-dY_46_v) * floor(h)) / Float32(-sqrt(fmax(fma(Float32(t_0 * dY_46_u), floor(w), t_3), fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_4))))); end return tmp end
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \left(t\_2 \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_4 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
\mathbf{if}\;\mathsf{fma}\left(dX.v \cdot dX.v, \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor , t\_4\right) \geq \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, t\_3\right):\\
\;\;\;\;\left\lfloor h\right\rfloor \cdot \frac{dX.v}{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_2, t\_2, t\_0 \cdot t\_0\right), \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor , dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_1\right)\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left(-dY.v\right) \cdot \left\lfloor h\right\rfloor }{-\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0 \cdot dY.u, \left\lfloor w\right\rfloor , t\_3\right), \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_4\right)\right)}}\\
\end{array}
Initial program 76.0%
Applied rewrites76.1%
Applied rewrites76.2%
Applied rewrites76.2%
Applied rewrites76.1%
(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 (* (* (* dY.v (floor h)) dY.v) (floor h)))
(t_3
(/
(floor h)
(sqrt
(fmax
(fma (* (* (floor w) (floor w)) dY.u) dY.u t_2)
(fma t_0 t_0 (* t_1 t_1)))))))
(if (>=
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma (* (* dY.u dY.u) (floor w)) (floor w) t_2))
(* t_3 dX.v)
(* t_3 dY.v))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dX_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = ((dY_46_v * floorf(h)) * dY_46_v) * floorf(h);
float t_3 = floorf(h) / sqrtf(fmaxf(fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, t_2), fmaf(t_0, t_0, (t_1 * t_1))));
float tmp;
if (fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), (((dX_46_u * floorf(w)) * dX_46_u) * floorf(w))) >= fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), t_2)) {
tmp = t_3 * dX_46_v;
} else {
tmp = t_3 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)) t_3 = Float32(floor(h) / sqrt(fmax(fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, t_2), fma(t_0, t_0, Float32(t_1 * t_1))))) tmp = Float32(0.0) if (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))) >= fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), t_2)) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); end return tmp end
\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(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, t\_2\right), \mathsf{fma}\left(t\_0, t\_0, t\_1 \cdot t\_1\right)\right)}}\\
\mathbf{if}\;\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) \geq \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \left\lfloor w\right\rfloor , t\_2\right):\\
\;\;\;\;t\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
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-*.f3276.0%
Applied rewrites76.0%
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-*.f3276.0%
Applied rewrites76.0%
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-*.f3275.9%
Applied rewrites75.9%
Applied rewrites75.9%
Applied rewrites75.9%
Applied rewrites76.0%
Applied rewrites76.0%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (* dY.u (floor w)))
(t_2 (* (* (* dX.u (floor w)) dX.u) (floor w)))
(t_3
(/
(floor h)
(sqrt
(fmax
(fma t_0 (* dY.v dY.v) (* t_1 t_1))
(fma (* dX.v dX.v) t_0 t_2))))))
(if (>=
(fma (* (* dX.v (floor h)) dX.v) (floor h) t_2)
(fma
(* (* dY.u dY.u) (floor w))
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(* t_3 dX.v)
(* t_3 dY.v))))float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * floorf(h);
float t_1 = dY_46_u * floorf(w);
float t_2 = ((dX_46_u * floorf(w)) * dX_46_u) * floorf(w);
float t_3 = floorf(h) / sqrtf(fmaxf(fmaf(t_0, (dY_46_v * dY_46_v), (t_1 * t_1)), fmaf((dX_46_v * dX_46_v), t_0, t_2)));
float tmp;
if (fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), t_2) >= fmaf(((dY_46_u * dY_46_u) * floorf(w)), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))) {
tmp = t_3 * dX_46_v;
} else {
tmp = t_3 * dY_46_v;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * floor(h)) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(Float32(Float32(dX_46_u * floor(w)) * dX_46_u) * floor(w)) t_3 = Float32(floor(h) / sqrt(fmax(fma(t_0, Float32(dY_46_v * dY_46_v), Float32(t_1 * t_1)), fma(Float32(dX_46_v * dX_46_v), t_0, t_2)))) tmp = Float32(0.0) if (fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), t_2) >= fma(Float32(Float32(dY_46_u * dY_46_u) * floor(w)), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h)))) tmp = Float32(t_3 * dX_46_v); else tmp = Float32(t_3 * dY_46_v); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \left(\left(dX.u \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(\mathsf{fma}\left(t\_0, dY.v \cdot dY.v, t\_1 \cdot t\_1\right), \mathsf{fma}\left(dX.v \cdot dX.v, t\_0, t\_2\right)\right)}}\\
\mathbf{if}\;\mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , t\_2\right) \geq \mathsf{fma}\left(\left(dY.u \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor , \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\_3 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot dY.v\\
\end{array}
Initial program 76.0%
Applied rewrites76.0%
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-*.f3276.0%
Applied rewrites76.0%
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-*.f3276.0%
Applied rewrites76.0%
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-*.f3275.9%
Applied rewrites75.9%
Applied rewrites75.9%
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
Applied rewrites75.9%
lift-fma.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
swap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
+-commutativeN/A
Applied rewrites75.9%
herbie shell --seed 2025183
(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))))