
(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}
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}
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_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}
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}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor h) dX.v))
(t_2
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h))))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5 (* dX.u (floor w)))
(t_6
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* t_5 dX.u) (floor w)))))
(if (>= (+ (* t_4 t_4) (* t_1 t_1)) (+ (* t_3 t_3) (* t_0 t_0)))
(/ 1.0 (/ (sqrt (fmax t_2 t_6)) t_5))
(/ (- (- dY.u)) (/ (sqrt (fmax t_6 t_2)) (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) * dY_46_v;
float t_1 = floorf(h) * dX_46_v;
float t_2 = fmaf(((dY_46_u * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)));
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(w) * dX_46_u;
float t_5 = dX_46_u * floorf(w);
float t_6 = fmaf(((dX_46_v * floorf(h)) * dX_46_v), floorf(h), ((t_5 * dX_46_u) * floorf(w)));
float tmp;
if (((t_4 * t_4) + (t_1 * t_1)) >= ((t_3 * t_3) + (t_0 * t_0))) {
tmp = 1.0f / (sqrtf(fmaxf(t_2, t_6)) / t_5);
} else {
tmp = -(-dY_46_u) / (sqrtf(fmaxf(t_6, t_2)) / 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) * dY_46_v) t_1 = Float32(floor(h) * dX_46_v) t_2 = 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_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(dX_46_u * floor(w)) t_6 = fma(Float32(Float32(dX_46_v * floor(h)) * dX_46_v), floor(h), Float32(Float32(t_5 * dX_46_u) * floor(w))) tmp = Float32(0.0) if (Float32(Float32(t_4 * t_4) + Float32(t_1 * t_1)) >= Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) tmp = Float32(Float32(1.0) / Float32(sqrt(fmax(t_2, t_6)) / t_5)); else tmp = Float32(Float32(-Float32(-dY_46_u)) / Float32(sqrt(fmax(t_6, t_2)) / floor(w))); end return tmp end
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \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_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_6 := \mathsf{fma}\left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v, \left\lfloor h\right\rfloor , \left(t\_5 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;t\_4 \cdot t\_4 + t\_1 \cdot t\_1 \geq t\_3 \cdot t\_3 + t\_0 \cdot t\_0:\\
\;\;\;\;\frac{1}{\frac{\sqrt{\mathsf{max}\left(t\_2, t\_6\right)}}{t\_5}}\\
\mathbf{else}:\\
\;\;\;\;\frac{-\left(-dY.u\right)}{\frac{\sqrt{\mathsf{max}\left(t\_6, t\_2\right)}}{\left\lfloor w\right\rfloor }}\\
\end{array}
Initial program 77.2%
Applied rewrites77.2%
Applied rewrites77.2%
Applied rewrites77.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4
(sqrt
(fmax
(fma
(* (* dX.v (floor h)) dX.v)
(floor h)
(* (* (* dX.u (floor w)) dX.u) (floor w)))
(fma
(* (* dY.u (floor w)) dY.u)
(floor w)
(* (* (* dY.v (floor h)) dY.v) (floor h)))))))
(if (>= (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))
(* (floor w) (/ dX.u t_4))
(/ (- (- dY.u)) (/ t_4 (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) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = sqrtf(fmaxf(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 * floorf(w)) * dY_46_u), floorf(w), (((dY_46_v * floorf(h)) * dY_46_v) * floorf(h)))));
float tmp;
if (((t_3 * t_3) + (t_0 * t_0)) >= ((t_1 * t_1) + (t_2 * t_2))) {
tmp = floorf(w) * (dX_46_u / t_4);
} else {
tmp = -(-dY_46_u) / (t_4 / 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) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = sqrt(fmax(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 * floor(w)) * dY_46_u), floor(w), Float32(Float32(Float32(dY_46_v * floor(h)) * dY_46_v) * floor(h))))) tmp = Float32(0.0) if (Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) >= Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) tmp = Float32(floor(w) * Float32(dX_46_u / t_4)); else tmp = Float32(Float32(-Float32(-dY_46_u)) / Float32(t_4 / floor(w))); end return 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 h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \sqrt{\mathsf{max}\left(\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), \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)\right)}\\
\mathbf{if}\;t\_3 \cdot t\_3 + t\_0 \cdot t\_0 \geq t\_1 \cdot t\_1 + t\_2 \cdot t\_2:\\
\;\;\;\;\left\lfloor w\right\rfloor \cdot \frac{dX.u}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{-\left(-dY.u\right)}{\frac{t\_4}{\left\lfloor w\right\rfloor }}\\
\end{array}
Initial program 77.2%
Applied rewrites77.2%
Applied rewrites77.2%
Applied rewrites77.2%
Applied rewrites77.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0
(fma
(* (* (floor h) (floor h)) dY.v)
dY.v
(* (* (* dY.u (floor w)) dY.u) (floor w))))
(t_1 (* dX.u (floor w)))
(t_2 (fma t_1 t_1 (* (* (* dX.v (floor h)) dX.v) (floor h))))
(t_3 (/ 1.0 (sqrt (fmax t_2 t_0)))))
(if (>= t_2 t_0) (* t_3 (* (floor w) dX.u)) (* t_3 (* (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 = fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, (((dY_46_u * floorf(w)) * dY_46_u) * floorf(w)));
float t_1 = dX_46_u * floorf(w);
float t_2 = fmaf(t_1, t_1, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)));
float t_3 = 1.0f / sqrtf(fmaxf(t_2, t_0));
float tmp;
if (t_2 >= t_0) {
tmp = t_3 * (floorf(w) * dX_46_u);
} else {
tmp = t_3 * (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 = fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(Float32(Float32(dY_46_u * floor(w)) * dY_46_u) * floor(w))) t_1 = Float32(dX_46_u * floor(w)) t_2 = fma(t_1, t_1, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) t_3 = Float32(Float32(1.0) / sqrt(fmax(t_2, t_0))) tmp = Float32(0.0) if (t_2 >= t_0) tmp = Float32(t_3 * Float32(floor(w) * dX_46_u)); else tmp = Float32(t_3 * Float32(floor(w) * dY_46_u)); end return tmp end
\begin{array}{l}
t_0 := \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \left(\left(dY.u \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_1 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1, t\_1, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_3 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2, t\_0\right)}}\\
\mathbf{if}\;t\_2 \geq t\_0:\\
\;\;\;\;t\_3 \cdot \left(\left\lfloor w\right\rfloor \cdot dX.u\right)\\
\mathbf{else}:\\
\;\;\;\;t\_3 \cdot \left(\left\lfloor w\right\rfloor \cdot dY.u\right)\\
\end{array}
Initial program 77.2%
lift-+.f32N/A
+-commutativeN/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
distribute-lft-neg-outN/A
distribute-lft-neg-inN/A
remove-double-negN/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites77.2%
lift-+.f32N/A
+-commutativeN/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
distribute-lft-neg-outN/A
distribute-lft-neg-inN/A
remove-double-negN/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites77.2%
lift-+.f32N/A
+-commutativeN/A
add-flipN/A
sub-flipN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
distribute-lft-neg-outN/A
distribute-lft-neg-inN/A
remove-double-negN/A
lift-*.f32N/A
lower-fma.f32N/A
Applied rewrites77.1%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
Applied rewrites77.1%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
Applied rewrites77.1%
lift-+.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-fma.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
Applied rewrites77.1%
(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
(* (* (floor w) (floor w)) dX.u)
dX.u
(* (* (* dX.v (floor h)) dX.v) (floor h))))
(t_2 (/ (floor w) (sqrt (fmax t_0 t_1)))))
(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(((floorf(w) * floorf(w)) * dX_46_u), dX_46_u, (((dX_46_v * floorf(h)) * dX_46_v) * floorf(h)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_0, t_1));
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(floor(w) * floor(w)) * dX_46_u), dX_46_u, Float32(Float32(Float32(dX_46_v * floor(h)) * dX_46_v) * floor(h))) t_2 = Float32(floor(w) / sqrt(fmax(t_0, t_1))) 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}
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(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dX.u, dX.u, \left(\left(dX.v \cdot \left\lfloor h\right\rfloor \right) \cdot dX.v\right) \cdot \left\lfloor h\right\rfloor \right)\\
t_2 := \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_0, t\_1\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
Initial program 77.2%
Applied rewrites77.2%
Applied rewrites77.1%
(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_0 t_1)))))
(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_0, t_1));
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_0, t_1))) 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}
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\_0, t\_1\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
Initial program 77.2%
Applied rewrites77.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.1
Applied rewrites77.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.0
Applied rewrites77.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
(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 dX.u) (floor w)) (floor w))))
(t_2 (/ (floor w) (sqrt (fmax t_0 t_1)))))
(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 * dX_46_u) * floorf(w)) * floorf(w)));
float t_2 = floorf(w) / sqrtf(fmaxf(t_0, t_1));
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 * dX_46_u) * floor(w)) * floor(w))) t_2 = Float32(floor(w) / sqrt(fmax(t_0, t_1))) 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}
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 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\_0, t\_1\right)}}\\
\mathbf{if}\;t\_1 \geq t\_0:\\
\;\;\;\;t\_2 \cdot dX.u\\
\mathbf{else}:\\
\;\;\;\;t\_2 \cdot dY.u\\
\end{array}
Initial program 77.2%
Applied rewrites77.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.1
Applied rewrites77.1%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.0
Applied rewrites77.0%
lift-*.f32N/A
lift-*.f32N/A
associate-*l*N/A
*-commutativeN/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
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-*.f3277.0
Applied rewrites77.0%
herbie shell --seed 2025173
(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))))