
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
Herbie found 7 alternatives:
| Alternative | Accuracy | Speedup |
|---|
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* (floor w) dY.u))
(t_2 (* (floor w) dX.u))
(t_3 (+ (* t_2 t_2) (* t_0 t_0)))
(t_4 (* (floor h) dY.v))
(t_5 (+ (* t_1 t_1) (* t_4 t_4)))
(t_6 (/ 1.0 (sqrt (fmax t_3 t_5)))))
(if (>= t_3 t_5) (* t_6 t_0) (* t_6 t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(w) * dX_46_u;
float t_3 = (t_2 * t_2) + (t_0 * t_0);
float t_4 = floorf(h) * dY_46_v;
float t_5 = (t_1 * t_1) + (t_4 * t_4);
float t_6 = 1.0f / sqrtf(fmaxf(t_3, t_5));
float tmp;
if (t_3 >= t_5) {
tmp = t_6 * t_0;
} else {
tmp = t_6 * t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(h) * dX_46_v) t_1 = Float32(floor(w) * dY_46_u) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(Float32(t_2 * t_2) + Float32(t_0 * t_0)) t_4 = Float32(floor(h) * dY_46_v) t_5 = Float32(Float32(t_1 * t_1) + Float32(t_4 * t_4)) t_6 = Float32(Float32(1.0) / sqrt(fmax(t_3, t_5))) tmp = Float32(0.0) if (t_3 >= t_5) tmp = Float32(t_6 * t_0); else tmp = Float32(t_6 * t_4); end return tmp end
function tmp_2 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dX_46_v; t_1 = floor(w) * dY_46_u; t_2 = floor(w) * dX_46_u; t_3 = (t_2 * t_2) + (t_0 * t_0); t_4 = floor(h) * dY_46_v; t_5 = (t_1 * t_1) + (t_4 * t_4); t_6 = single(1.0) / sqrt(max(t_3, t_5)); tmp = single(0.0); if (t_3 >= t_5) tmp = t_6 * t_0; else tmp = t_6 * t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := t\_2 \cdot t\_2 + t\_0 \cdot t\_0\\
t_4 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_5 := t\_1 \cdot t\_1 + t\_4 \cdot t\_4\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3, t\_5\right)}}\\
\mathbf{if}\;t\_3 \geq t\_5:\\
\;\;\;\;t\_6 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_4\\
\end{array}
\end{array}
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dX.v (floor h)))
(t_1 (* (floor w) (floor w)))
(t_2 (fma (* t_1 dY.u) dY.u (* (* dY.v dY.v) (* (floor h) (floor h)))))
(t_3 (fma (* t_1 dX.u) dX.u (* t_0 t_0)))
(t_4 (sqrt (fmax t_3 t_2))))
(if (>= t_3 t_2) (/ (* (floor h) dX.v) t_4) (/ (* (floor h) dY.v) t_4))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = dX_46_v * floorf(h);
float t_1 = floorf(w) * floorf(w);
float t_2 = fmaf((t_1 * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * (floorf(h) * floorf(h))));
float t_3 = fmaf((t_1 * dX_46_u), dX_46_u, (t_0 * t_0));
float t_4 = sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = (floorf(h) * dX_46_v) / t_4;
} else {
tmp = (floorf(h) * dY_46_v) / t_4;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dX_46_v * floor(h)) t_1 = Float32(floor(w) * floor(w)) t_2 = fma(Float32(t_1 * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * Float32(floor(h) * floor(h)))) t_3 = fma(Float32(t_1 * dX_46_u), dX_46_u, Float32(t_0 * t_0)) t_4 = sqrt(fmax(t_3, t_2)) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / t_4); else tmp = Float32(Float32(floor(h) * dY_46_v) / t_4); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.u, dX.u, t\_0 \cdot t\_0\right)\\
t_4 := \sqrt{\mathsf{max}\left(t\_3, t\_2\right)}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_4}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_4}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) (floor h)))
(t_1 (fma (* (* (floor w) (floor w)) dY.u) dY.u (* (* dY.v dY.v) t_0)))
(t_2 (fma (* (* (floor w) dX.u) (floor w)) dX.u (* t_0 (* dX.v dX.v))))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1) (/ (* (floor h) dX.v) t_3) (/ (* (floor h) dY.v) 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 = floorf(h) * floorf(h);
float t_1 = fmaf(((floorf(w) * floorf(w)) * dY_46_u), dY_46_u, ((dY_46_v * dY_46_v) * t_0));
float t_2 = fmaf(((floorf(w) * dX_46_u) * floorf(w)), dX_46_u, (t_0 * (dX_46_v * dX_46_v)));
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = (floorf(h) * dX_46_v) / t_3;
} else {
tmp = (floorf(h) * dY_46_v) / t_3;
}
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 = fma(Float32(Float32(floor(w) * floor(w)) * dY_46_u), dY_46_u, Float32(Float32(dY_46_v * dY_46_v) * t_0)) t_2 = fma(Float32(Float32(floor(w) * dX_46_u) * floor(w)), dX_46_u, Float32(t_0 * Float32(dX_46_v * dX_46_v))) t_3 = sqrt(fmax(t_2, t_1)) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(Float32(floor(h) * dX_46_v) / t_3); else tmp = Float32(Float32(floor(h) * dY_46_v) / t_3); 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 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \cdot dY.u, dY.u, \left(dY.v \cdot dY.v\right) \cdot t\_0\right)\\
t_2 := \mathsf{fma}\left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor , dX.u, t\_0 \cdot \left(dX.v \cdot dX.v\right)\right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dY.v}{t\_3}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
associate-*l*N/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lift-floor.f3276.3
Applied rewrites76.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* (floor h) (floor h)))
(t_2 (fma (* t_1 dY.v) dY.v (* (* dY.u dY.u) t_0)))
(t_3 (fma (* t_1 dX.v) dX.v (* (* dX.u dX.u) t_0)))
(t_4 (/ (floor h) (sqrt (fmax t_3 t_2)))))
(if (>= t_3 t_2) (* t_4 dX.v) (* t_4 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) * floorf(w);
float t_1 = floorf(h) * floorf(h);
float t_2 = fmaf((t_1 * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_0));
float t_3 = fmaf((t_1 * dX_46_v), dX_46_v, ((dX_46_u * dX_46_u) * t_0));
float t_4 = floorf(h) / sqrtf(fmaxf(t_3, t_2));
float tmp;
if (t_3 >= t_2) {
tmp = t_4 * dX_46_v;
} else {
tmp = t_4 * 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) * floor(w)) t_1 = Float32(floor(h) * floor(h)) t_2 = fma(Float32(t_1 * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0)) t_3 = fma(Float32(t_1 * dX_46_v), dX_46_v, Float32(Float32(dX_46_u * dX_46_u) * t_0)) t_4 = Float32(floor(h) / sqrt(fmax(t_3, t_2))) tmp = Float32(0.0) if (t_3 >= t_2) tmp = Float32(t_4 * dX_46_v); else tmp = Float32(t_4 * dY_46_v); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := \left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_1 \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\\
t_3 := \mathsf{fma}\left(t\_1 \cdot dX.v, dX.v, \left(dX.u \cdot dX.u\right) \cdot t\_0\right)\\
t_4 := \frac{\left\lfloor h\right\rfloor }{\sqrt{\mathsf{max}\left(t\_3, t\_2\right)}}\\
\mathbf{if}\;t\_3 \geq t\_2:\\
\;\;\;\;t\_4 \cdot dX.v\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot dY.v\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.3
Applied rewrites76.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
unswap-sqrN/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
lift-*.f32N/A
*-commutativeN/A
lower-*.f3276.4
Applied rewrites76.4%
Applied rewrites76.3%
Applied rewrites76.2%
(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 h) dY.v))
(t_3 (* (floor w) dX.u))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (/ 1.0 (sqrt (fmax t_4 t_1))))
(t_6 (if (>= t_4 t_1) (* t_5 t_0) (* t_5 t_2)))
(t_7 (* (floor w) dY.u))
(t_8
(/
1.0
(sqrt
(fmax
(* (* dX.u dX.u) (* (floor w) (floor w)))
(+ (* t_7 t_7) (* t_2 t_2)))))))
(if (<= dX.v -5000000.0)
t_6
(if (<= dX.v 4.0)
(if (>=
(* (* t_3 dX.u) (floor w))
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* t_7 dY.u) (floor w))))
(* t_8 t_0)
(* t_8 t_2))
t_6))))
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(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = 1.0f / sqrtf(fmaxf(t_4, t_1));
float tmp;
if (t_4 >= t_1) {
tmp = t_5 * t_0;
} else {
tmp = t_5 * t_2;
}
float t_6 = tmp;
float t_7 = floorf(w) * dY_46_u;
float t_8 = 1.0f / sqrtf(fmaxf(((dX_46_u * dX_46_u) * (floorf(w) * floorf(w))), ((t_7 * t_7) + (t_2 * t_2))));
float tmp_1;
if (dX_46_v <= -5000000.0f) {
tmp_1 = t_6;
} else if (dX_46_v <= 4.0f) {
float tmp_2;
if (((t_3 * dX_46_u) * floorf(w)) >= fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), ((t_7 * dY_46_u) * floorf(w)))) {
tmp_2 = t_8 * t_0;
} else {
tmp_2 = t_8 * t_2;
}
tmp_1 = tmp_2;
} else {
tmp_1 = t_6;
}
return tmp_1;
}
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(Float32(-dY_46_v) * dY_46_v) * Float32(floor(h) * floor(h))) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(Float32(1.0) / sqrt(fmax(t_4, t_1))) tmp = Float32(0.0) if (t_4 >= t_1) tmp = Float32(t_5 * t_0); else tmp = Float32(t_5 * t_2); end t_6 = tmp t_7 = Float32(floor(w) * dY_46_u) t_8 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(dX_46_u * dX_46_u) * Float32(floor(w) * floor(w))), Float32(Float32(t_7 * t_7) + Float32(t_2 * t_2))))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(-5000000.0)) tmp_1 = t_6; elseif (dX_46_v <= Float32(4.0)) tmp_2 = Float32(0.0) if (Float32(Float32(t_3 * dX_46_u) * floor(w)) >= fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(t_7 * dY_46_u) * floor(w)))) tmp_2 = Float32(t_8 * t_0); else tmp_2 = Float32(t_8 * t_2); end tmp_1 = tmp_2; else tmp_1 = t_6; end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := \left(\left(-dY.v\right) \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \frac{1}{\sqrt{\mathsf{max}\left(t\_4, t\_1\right)}}\\
t_6 := \begin{array}{l}
\mathbf{if}\;t\_4 \geq t\_1:\\
\;\;\;\;t\_5 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot t\_2\\
\end{array}\\
t_7 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right), t\_7 \cdot t\_7 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;dX.v \leq -5000000:\\
\;\;\;\;t\_6\\
\mathbf{elif}\;dX.v \leq 4:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(t\_3 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \geq \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(t\_7 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right):\\
\;\;\;\;t\_8 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_2\\
\end{array}\\
\mathbf{else}:\\
\;\;\;\;t\_6\\
\end{array}
\end{array}
if dX.v < -5e6 or 4 < dX.v Initial program 76.3%
Applied rewrites67.7%
Applied rewrites70.2%
Applied rewrites69.8%
Taylor expanded in dY.u around 0
Applied rewrites45.2%
Taylor expanded in dY.u around 0
Applied rewrites47.6%
Taylor expanded in dY.u around 0
Applied rewrites47.6%
if -5e6 < dX.v < 4Initial program 76.3%
Applied rewrites67.9%
Applied rewrites67.3%
Applied rewrites67.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites56.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites58.8%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites62.4%
Applied rewrites62.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (* (- dY.v) dY.v) (* (floor h) (floor h))))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor h) dY.v))
(t_3 (* (floor w) dY.u))
(t_4 (* (floor w) dX.u))
(t_5
(sqrt
(/
1.0
(fmax
(fma t_1 t_1 (* (* t_4 dX.u) (floor w)))
(fma
(* (* dY.v (floor h)) dY.v)
(floor h)
(* (* t_3 dY.u) (floor w)))))))
(t_6 (+ (* t_4 t_4) (* t_1 t_1)))
(t_7 (/ 1.0 (sqrt (fmax t_6 t_0)))))
(if (<= dX.v 50000000.0)
(if (>=
(* (* dX.u dX.u) (* (floor w) (floor w)))
(+ (* t_3 t_3) (* t_2 t_2)))
(* t_5 t_1)
(* t_5 t_2))
(if (>= t_6 t_0) (* t_7 t_1) (* t_7 t_2)))))
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 * dY_46_v) * (floorf(h) * floorf(h));
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * dY_46_v;
float t_3 = floorf(w) * dY_46_u;
float t_4 = floorf(w) * dX_46_u;
float t_5 = sqrtf((1.0f / fmaxf(fmaf(t_1, t_1, ((t_4 * dX_46_u) * floorf(w))), fmaf(((dY_46_v * floorf(h)) * dY_46_v), floorf(h), ((t_3 * dY_46_u) * floorf(w))))));
float t_6 = (t_4 * t_4) + (t_1 * t_1);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, t_0));
float tmp_1;
if (dX_46_v <= 50000000.0f) {
float tmp_2;
if (((dX_46_u * dX_46_u) * (floorf(w) * floorf(w))) >= ((t_3 * t_3) + (t_2 * t_2))) {
tmp_2 = t_5 * t_1;
} else {
tmp_2 = t_5 * t_2;
}
tmp_1 = tmp_2;
} else if (t_6 >= t_0) {
tmp_1 = t_7 * t_1;
} else {
tmp_1 = t_7 * t_2;
}
return tmp_1;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(Float32(Float32(-dY_46_v) * dY_46_v) * Float32(floor(h) * floor(h))) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(floor(w) * dX_46_u) t_5 = sqrt(Float32(Float32(1.0) / fmax(fma(t_1, t_1, Float32(Float32(t_4 * dX_46_u) * floor(w))), fma(Float32(Float32(dY_46_v * floor(h)) * dY_46_v), floor(h), Float32(Float32(t_3 * dY_46_u) * floor(w)))))) t_6 = Float32(Float32(t_4 * t_4) + Float32(t_1 * t_1)) t_7 = Float32(Float32(1.0) / sqrt(fmax(t_6, t_0))) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(50000000.0)) tmp_2 = Float32(0.0) if (Float32(Float32(dX_46_u * dX_46_u) * Float32(floor(w) * floor(w))) >= Float32(Float32(t_3 * t_3) + Float32(t_2 * t_2))) tmp_2 = Float32(t_5 * t_1); else tmp_2 = Float32(t_5 * t_2); end tmp_1 = tmp_2; elseif (t_6 >= t_0) tmp_1 = Float32(t_7 * t_1); else tmp_1 = Float32(t_7 * t_2); end return tmp_1 end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left(\left(-dY.v\right) \cdot dY.v\right) \cdot \left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right)\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := \sqrt{\frac{1}{\mathsf{max}\left(\mathsf{fma}\left(t\_1, t\_1, \left(t\_4 \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right), \mathsf{fma}\left(\left(dY.v \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, \left\lfloor h\right\rfloor , \left(t\_3 \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\right)}}\\
t_6 := t\_4 \cdot t\_4 + t\_1 \cdot t\_1\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_0\right)}}\\
\mathbf{if}\;dX.v \leq 50000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(dX.u \cdot dX.u\right) \cdot \left(\left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \right) \geq t\_3 \cdot t\_3 + t\_2 \cdot t\_2:\\
\;\;\;\;t\_5 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_5 \cdot t\_2\\
\end{array}\\
\mathbf{elif}\;t\_6 \geq t\_0:\\
\;\;\;\;t\_7 \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_2\\
\end{array}
\end{array}
if dX.v < 5e7Initial program 76.3%
Applied rewrites67.9%
Applied rewrites67.3%
Applied rewrites67.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites56.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites58.8%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites62.4%
Taylor expanded in w around 0
Applied rewrites68.2%
Taylor expanded in w around 0
Applied rewrites64.5%
if 5e7 < dX.v Initial program 76.3%
Applied rewrites67.7%
Applied rewrites70.2%
Applied rewrites69.8%
Taylor expanded in dY.u around 0
Applied rewrites45.2%
Taylor expanded in dY.u around 0
Applied rewrites47.6%
Taylor expanded in dY.u around 0
Applied rewrites47.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) (floor w)))
(t_1 (* dY.v (floor h)))
(t_2 (* (* (* (floor w) dX.u) dX.u) (floor w)))
(t_3
(fma
(* t_1 dY.v)
(floor h)
(* (* (* (floor w) dY.u) dY.u) (floor w)))))
(if (>= t_2 t_3)
(*
(/
1.0
(sqrt
(fmax
(* (* dX.u dX.u) t_0)
(fma (* (* (floor h) (floor h)) dY.v) dY.v (* (* dY.u dY.u) t_0)))))
(* (floor h) dX.v))
(/ t_1 (sqrt (fmax t_2 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 = floorf(w) * floorf(w);
float t_1 = dY_46_v * floorf(h);
float t_2 = ((floorf(w) * dX_46_u) * dX_46_u) * floorf(w);
float t_3 = fmaf((t_1 * dY_46_v), floorf(h), (((floorf(w) * dY_46_u) * dY_46_u) * floorf(w)));
float tmp;
if (t_2 >= t_3) {
tmp = (1.0f / sqrtf(fmaxf(((dX_46_u * dX_46_u) * t_0), fmaf(((floorf(h) * floorf(h)) * dY_46_v), dY_46_v, ((dY_46_u * dY_46_u) * t_0))))) * (floorf(h) * dX_46_v);
} else {
tmp = t_1 / sqrtf(fmaxf(t_2, t_3));
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * floor(w)) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32(Float32(Float32(floor(w) * dX_46_u) * dX_46_u) * floor(w)) t_3 = fma(Float32(t_1 * dY_46_v), floor(h), Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * floor(w))) tmp = Float32(0.0) if (t_2 >= t_3) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(dX_46_u * dX_46_u) * t_0), fma(Float32(Float32(floor(h) * floor(h)) * dY_46_v), dY_46_v, Float32(Float32(dY_46_u * dY_46_u) * t_0))))) * Float32(floor(h) * dX_46_v)); else tmp = Float32(t_1 / sqrt(fmax(t_2, t_3))); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot \left\lfloor w\right\rfloor \\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := \left(\left(\left\lfloor w\right\rfloor \cdot dX.u\right) \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \\
t_3 := \mathsf{fma}\left(t\_1 \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
\mathbf{if}\;t\_2 \geq t\_3:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot t\_0, \mathsf{fma}\left(\left(\left\lfloor h\right\rfloor \cdot \left\lfloor h\right\rfloor \right) \cdot dY.v, dY.v, \left(dY.u \cdot dY.u\right) \cdot t\_0\right)\right)}} \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right)\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1}{\sqrt{\mathsf{max}\left(t\_2, t\_3\right)}}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites67.9%
Applied rewrites67.3%
Applied rewrites67.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites56.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites58.8%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites62.4%
Applied rewrites62.5%
Applied rewrites62.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (* (* (* dX.u dX.u) (floor w)) (floor w)))
(t_2
(fma
(* t_0 dY.v)
(floor h)
(* (* (* (floor w) dY.u) dY.u) (floor w))))
(t_3 (sqrt (fmax t_1 t_2))))
(if (>= t_1 t_2) (/ (* (floor h) dX.v) t_3) (/ t_0 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);
float t_1 = ((dX_46_u * dX_46_u) * floorf(w)) * floorf(w);
float t_2 = fmaf((t_0 * dY_46_v), floorf(h), (((floorf(w) * dY_46_u) * dY_46_u) * floorf(w)));
float t_3 = sqrtf(fmaxf(t_1, t_2));
float tmp;
if (t_1 >= t_2) {
tmp = (floorf(h) * dX_46_v) / t_3;
} else {
tmp = t_0 / t_3;
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(dY_46_v * floor(h)) t_1 = Float32(Float32(Float32(dX_46_u * dX_46_u) * floor(w)) * floor(w)) t_2 = fma(Float32(t_0 * dY_46_v), floor(h), Float32(Float32(Float32(floor(w) * dY_46_u) * dY_46_u) * floor(w))) t_3 = sqrt(fmax(t_1, t_2)) tmp = Float32(0.0) if (t_1 >= t_2) tmp = Float32(Float32(floor(h) * dX_46_v) / t_3); else tmp = Float32(t_0 / t_3); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := \left(\left(dX.u \cdot dX.u\right) \cdot \left\lfloor w\right\rfloor \right) \cdot \left\lfloor w\right\rfloor \\
t_2 := \mathsf{fma}\left(t\_0 \cdot dY.v, \left\lfloor h\right\rfloor , \left(\left(\left\lfloor w\right\rfloor \cdot dY.u\right) \cdot dY.u\right) \cdot \left\lfloor w\right\rfloor \right)\\
t_3 := \sqrt{\mathsf{max}\left(t\_1, t\_2\right)}\\
\mathbf{if}\;t\_1 \geq t\_2:\\
\;\;\;\;\frac{\left\lfloor h\right\rfloor \cdot dX.v}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_0}{t\_3}\\
\end{array}
\end{array}
Initial program 76.3%
Applied rewrites67.9%
Applied rewrites67.3%
Applied rewrites67.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites56.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites58.8%
Taylor expanded in dX.u around inf
*-commutativeN/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
lift-fabs.f32N/A
associate-*l*N/A
*-commutativeN/A
lift-floor.f32N/A
lift-*.f32N/A
*-commutativeN/A
lower-special-*N/A
lift-fabs.f32N/A
rem-sqrt-square-revN/A
pow2N/A
sqrt-pow1N/A
lower-special-/N/A
metadata-evalN/A
Applied rewrites62.4%
Applied rewrites62.5%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f3262.5
Applied rewrites62.5%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f3262.5
Applied rewrites62.5%
lift-*.f32N/A
lift-*.f32N/A
lift-floor.f32N/A
associate-*l*N/A
pow2N/A
*-commutativeN/A
lower-*.f32N/A
pow2N/A
lift-*.f32N/A
lift-floor.f3262.5
Applied rewrites62.5%
herbie shell --seed 2025132
(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))))