
(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}
Sampling outcomes in binary32 precision:
Herbie found 5 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 (pow (* dX.v (floor h)) 2.0))
(t_1 (* (floor h) dX.v))
(t_2 (+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_3 (* (floor w) dY.u))
(t_4 (* dX.u (floor w)))
(t_5 (pow t_4 2.0))
(t_6 (* (floor h) dY.v)))
(if (>= (+ t_5 (* t_1 t_1)) (+ (* t_3 t_3) (* t_6 t_6)))
(/ t_4 (sqrt (fmax (+ t_0 t_5) t_2)))
(* (/ 1.0 (sqrt (fmax (+ t_5 t_0) 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 = powf((dX_46_v * floorf(h)), 2.0f);
float t_1 = floorf(h) * dX_46_v;
float t_2 = powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_3 = floorf(w) * dY_46_u;
float t_4 = dX_46_u * floorf(w);
float t_5 = powf(t_4, 2.0f);
float t_6 = floorf(h) * dY_46_v;
float tmp;
if ((t_5 + (t_1 * t_1)) >= ((t_3 * t_3) + (t_6 * t_6))) {
tmp = t_4 / sqrtf(fmaxf((t_0 + t_5), t_2));
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + t_0), 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(dX_46_v * floor(h)) ^ Float32(2.0) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(dX_46_u * floor(w)) t_5 = t_4 ^ Float32(2.0) t_6 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (Float32(t_5 + Float32(t_1 * t_1)) >= Float32(Float32(t_3 * t_3) + Float32(t_6 * t_6))) tmp = Float32(t_4 / sqrt(fmax(Float32(t_0 + t_5), t_2))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_5 + t_0), t_2))) * t_3); 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 = (dX_46_v * floor(h)) ^ single(2.0); t_1 = floor(h) * dX_46_v; t_2 = ((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_3 = floor(w) * dY_46_u; t_4 = dX_46_u * floor(w); t_5 = t_4 ^ single(2.0); t_6 = floor(h) * dY_46_v; tmp = single(0.0); if ((t_5 + (t_1 * t_1)) >= ((t_3 * t_3) + (t_6 * t_6))) tmp = t_4 / sqrt(max((t_0 + t_5), t_2)); else tmp = (single(1.0) / sqrt(max((t_5 + t_0), t_2))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := dX.u \cdot \left\lfloor w\right\rfloor \\
t_5 := {t\_4}^{2}\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;t\_5 + t\_1 \cdot t\_1 \geq t\_3 \cdot t\_3 + t\_6 \cdot t\_6:\\
\;\;\;\;\frac{t\_4}{\sqrt{\mathsf{max}\left(t\_0 + t\_5, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_0, t\_2\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 76.6%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites76.8%
Applied rewrites76.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3276.8
Applied rewrites76.8%
Final simplification76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (pow (* dY.u (floor w)) 2.0))
(t_2 (pow (* dY.v (floor h)) 2.0))
(t_3 (* (floor w) dY.u))
(t_4 (pow (* dX.u (floor w)) 2.0))
(t_5 (+ t_4 (pow (* dX.v (floor h)) 2.0)))
(t_6 (* (floor h) dY.v)))
(if (>= (+ t_4 (* t_0 t_0)) (+ (* t_3 t_3) (* t_6 t_6)))
(* dX.u (/ (floor w) (sqrt (fmax t_5 (- t_1 t_2)))))
(* (/ 1.0 (sqrt (fmax t_5 (+ t_2 t_1)))) 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) * dX_46_v;
float t_1 = powf((dY_46_u * floorf(w)), 2.0f);
float t_2 = powf((dY_46_v * floorf(h)), 2.0f);
float t_3 = floorf(w) * dY_46_u;
float t_4 = powf((dX_46_u * floorf(w)), 2.0f);
float t_5 = t_4 + powf((dX_46_v * floorf(h)), 2.0f);
float t_6 = floorf(h) * dY_46_v;
float tmp;
if ((t_4 + (t_0 * t_0)) >= ((t_3 * t_3) + (t_6 * t_6))) {
tmp = dX_46_u * (floorf(w) / sqrtf(fmaxf(t_5, (t_1 - t_2))));
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, (t_2 + t_1)))) * 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) * dX_46_v) t_1 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_2 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_5 = Float32(t_4 + (Float32(dX_46_v * floor(h)) ^ Float32(2.0))) t_6 = Float32(floor(h) * dY_46_v) tmp = Float32(0.0) if (Float32(t_4 + Float32(t_0 * t_0)) >= Float32(Float32(t_3 * t_3) + Float32(t_6 * t_6))) tmp = Float32(dX_46_u * Float32(floor(w) / sqrt(fmax(t_5, Float32(t_1 - t_2))))); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, Float32(t_2 + t_1)))) * t_3); 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 = (dY_46_u * floor(w)) ^ single(2.0); t_2 = (dY_46_v * floor(h)) ^ single(2.0); t_3 = floor(w) * dY_46_u; t_4 = (dX_46_u * floor(w)) ^ single(2.0); t_5 = t_4 + ((dX_46_v * floor(h)) ^ single(2.0)); t_6 = floor(h) * dY_46_v; tmp = single(0.0); if ((t_4 + (t_0 * t_0)) >= ((t_3 * t_3) + (t_6 * t_6))) tmp = dX_46_u * (floor(w) / sqrt(max(t_5, (t_1 - t_2)))); else tmp = (single(1.0) / sqrt(max(t_5, (t_2 + t_1)))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_1 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := t\_4 + {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_6 := \left\lfloor h\right\rfloor \cdot dY.v\\
\mathbf{if}\;t\_4 + t\_0 \cdot t\_0 \geq t\_3 \cdot t\_3 + t\_6 \cdot t\_6:\\
\;\;\;\;dX.u \cdot \frac{\left\lfloor w\right\rfloor }{\sqrt{\mathsf{max}\left(t\_5, t\_1 - t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_2 + t\_1\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 76.6%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites76.8%
Applied rewrites76.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3276.8
Applied rewrites76.8%
Applied rewrites76.5%
Final simplification76.5%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (* (floor w) dX.u))
(t_2 (* t_1 t_1))
(t_3 (* (floor w) dY.u))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (* (floor h) dX.v))
(t_6 (/ 1.0 (sqrt (fmax (+ t_2 (* t_5 t_5)) t_4))))
(t_7 (* t_6 t_1)))
(if (<= dX.v 1.5)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_4) t_7 (* t_6 t_3))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
t_7
(*
(/
1.0
(sqrt (fmax (+ t_2 (exp (* (log (* (- dX.v) (floor h))) 2.0))) t_4)))
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) * dY_46_v;
float t_1 = floorf(w) * dX_46_u;
float t_2 = t_1 * t_1;
float t_3 = floorf(w) * dY_46_u;
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = floorf(h) * dX_46_v;
float t_6 = 1.0f / sqrtf(fmaxf((t_2 + (t_5 * t_5)), t_4));
float t_7 = t_6 * t_1;
float tmp_1;
if (dX_46_v <= 1.5f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_4) {
tmp_2 = t_7;
} else {
tmp_2 = t_6 * t_3;
}
tmp_1 = tmp_2;
} else if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp_1 = t_7;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf((t_2 + expf((logf((-dX_46_v * floorf(h))) * 2.0f))), t_4))) * t_3;
}
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) * dY_46_v) t_1 = Float32(floor(w) * dX_46_u) t_2 = Float32(t_1 * t_1) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(Float32(1.0) / sqrt(fmax(Float32(t_2 + Float32(t_5 * t_5)), t_4))) t_7 = Float32(t_6 * t_1) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(1.5)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_2 = t_7; else tmp_2 = Float32(t_6 * t_3); end tmp_1 = tmp_2; elseif (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp_1 = t_7; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_2 + exp(Float32(log(Float32(Float32(-dX_46_v) * floor(h))) * Float32(2.0)))), t_4))) * t_3); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dY_46_v; t_1 = floor(w) * dX_46_u; t_2 = t_1 * t_1; t_3 = floor(w) * dY_46_u; t_4 = (t_3 * t_3) + (t_0 * t_0); t_5 = floor(h) * dX_46_v; t_6 = single(1.0) / sqrt(max((t_2 + (t_5 * t_5)), t_4)); t_7 = t_6 * t_1; tmp_2 = single(0.0); if (dX_46_v <= single(1.5)) tmp_3 = single(0.0); if ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_3 = t_7; else tmp_3 = t_6 * t_3; end tmp_2 = tmp_3; elseif ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp_2 = t_7; else tmp_2 = (single(1.0) / sqrt(max((t_2 + exp((log((-dX_46_v * floor(h))) * single(2.0)))), t_4))) * t_3; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_2 := t\_1 \cdot t\_1\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_2 + t\_5 \cdot t\_5, t\_4\right)}}\\
t_7 := t\_6 \cdot t\_1\\
\mathbf{if}\;dX.v \leq 1.5:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_4:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;t\_6 \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_7\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_2 + e^{\log \left(\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot 2}, t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
if dX.v < 1.5Initial program 78.5%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3270.3
Applied rewrites70.3%
if 1.5 < dX.v Initial program 70.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3266.2
Applied rewrites66.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.2
Applied rewrites66.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.2
Applied rewrites66.2%
lift-*.f32N/A
sqr-neg-revN/A
pow2N/A
pow-to-expN/A
lower-exp.f32N/A
lower-*.f32N/A
lower-log.f32N/A
lift-*.f32N/A
*-commutativeN/A
distribute-lft-neg-inN/A
lift-neg.f32N/A
lower-*.f3267.6
Applied rewrites67.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dY.v))
(t_1 (pow (floor h) 2.0))
(t_2 (* (floor w) dX.u))
(t_3 (* (floor w) dY.u))
(t_4 (* t_3 t_3))
(t_5 (+ t_4 (* t_0 t_0)))
(t_6 (* (floor h) dX.v))
(t_7 (+ (* t_2 t_2) (* t_6 t_6)))
(t_8 (/ 1.0 (sqrt (fmax t_7 t_5))))
(t_9 (* t_8 t_2)))
(if (<= dX.v 100.0)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_5) t_9 (* t_8 t_3))
(if (>=
(* (* t_1 dX.v) dX.v)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
t_9
(* (/ 1.0 (sqrt (fmax t_7 (+ t_4 (* t_1 (* dY.v 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) * dY_46_v;
float t_1 = powf(floorf(h), 2.0f);
float t_2 = floorf(w) * dX_46_u;
float t_3 = floorf(w) * dY_46_u;
float t_4 = t_3 * t_3;
float t_5 = t_4 + (t_0 * t_0);
float t_6 = floorf(h) * dX_46_v;
float t_7 = (t_2 * t_2) + (t_6 * t_6);
float t_8 = 1.0f / sqrtf(fmaxf(t_7, t_5));
float t_9 = t_8 * t_2;
float tmp_1;
if (dX_46_v <= 100.0f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_5) {
tmp_2 = t_9;
} else {
tmp_2 = t_8 * t_3;
}
tmp_1 = tmp_2;
} else if (((t_1 * dX_46_v) * dX_46_v) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp_1 = t_9;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf(t_7, (t_4 + (t_1 * (dY_46_v * dY_46_v)))))) * t_3;
}
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) * dY_46_v) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(floor(w) * dX_46_u) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(t_3 * t_3) t_5 = Float32(t_4 + Float32(t_0 * t_0)) t_6 = Float32(floor(h) * dX_46_v) t_7 = Float32(Float32(t_2 * t_2) + Float32(t_6 * t_6)) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_5))) t_9 = Float32(t_8 * t_2) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(100.0)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_5) tmp_2 = t_9; else tmp_2 = Float32(t_8 * t_3); end tmp_1 = tmp_2; elseif (Float32(Float32(t_1 * dX_46_v) * dX_46_v) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp_1 = t_9; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(fmax(t_7, Float32(t_4 + Float32(t_1 * Float32(dY_46_v * dY_46_v)))))) * t_3); end return tmp_1 end
function tmp_4 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(h) * dY_46_v; t_1 = floor(h) ^ single(2.0); t_2 = floor(w) * dX_46_u; t_3 = floor(w) * dY_46_u; t_4 = t_3 * t_3; t_5 = t_4 + (t_0 * t_0); t_6 = floor(h) * dX_46_v; t_7 = (t_2 * t_2) + (t_6 * t_6); t_8 = single(1.0) / sqrt(max(t_7, t_5)); t_9 = t_8 * t_2; tmp_2 = single(0.0); if (dX_46_v <= single(100.0)) tmp_3 = single(0.0); if ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_5) tmp_3 = t_9; else tmp_3 = t_8 * t_3; end tmp_2 = tmp_3; elseif (((t_1 * dX_46_v) * dX_46_v) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp_2 = t_9; else tmp_2 = (single(1.0) / sqrt(max(t_7, (t_4 + (t_1 * (dY_46_v * dY_46_v)))))) * t_3; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_4 + t\_0 \cdot t\_0\\
t_6 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_7 := t\_2 \cdot t\_2 + t\_6 \cdot t\_6\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_5\right)}}\\
t_9 := t\_8 \cdot t\_2\\
\mathbf{if}\;dX.v \leq 100:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor w\right\rfloor \right)}^{2} \cdot dX.u\right) \cdot dX.u \geq t\_5:\\
\;\;\;\;t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;\left(t\_1 \cdot dX.v\right) \cdot dX.v \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_9\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_4 + t\_1 \cdot \left(dY.v \cdot dY.v\right)\right)}} \cdot t\_3\\
\end{array}
\end{array}
if dX.v < 100Initial program 79.1%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3270.7
Applied rewrites70.7%
if 100 < dX.v Initial program 67.5%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3265.8
Applied rewrites65.8%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.8
Applied rewrites65.8%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.8
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.8
Applied rewrites65.8%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lift-pow.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3265.9
Applied rewrites65.9%
(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
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))))))
(if (>=
(pow (* dX.v (floor h)) 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* t_4 t_3)
(* t_4 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(h) * dY_46_v;
float t_3 = floorf(w) * dX_46_u;
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2))));
float tmp;
if (powf((dX_46_v * floorf(h)), 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = t_4 * t_3;
} else {
tmp = t_4 * 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(h) * dY_46_v) t_3 = Float32(floor(w) * dX_46_u) t_4 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))))) tmp = Float32(0.0) if ((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(t_4 * t_3); else tmp = Float32(t_4 * 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(h) * dY_46_v; t_3 = floor(w) * dX_46_u; t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_1 * t_1) + (t_2 * t_2)))); tmp = single(0.0); if (((dX_46_v * floor(h)) ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = t_4 * t_3; else tmp = t_4 * 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 h\right\rfloor \cdot dY.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_1 \cdot t\_1 + t\_2 \cdot t\_2\right)}}\\
\mathbf{if}\;{\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} \geq {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_1\\
\end{array}
\end{array}
Initial program 76.6%
Taylor expanded in dX.u around 0
*-commutativeN/A
unpow2N/A
associate-*r*N/A
lower-*.f32N/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3262.0
Applied rewrites62.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3262.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.0
Applied rewrites62.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3262.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.0
Applied rewrites62.0%
Applied rewrites62.0%
herbie shell --seed 2024352
(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))))