
(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}
Sampling outcomes in binary32 precision:
Herbie found 4 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 (pow (* dY.u (floor w)) 2.0))
(t_1 (* dY.v (floor h)))
(t_2 (+ (pow (* dX.v (floor h)) 2.0) (pow (* dX.u (floor w)) 2.0)))
(t_3 (pow t_1 2.0))
(t_4 (sqrt (fmax t_2 (+ t_3 t_0)))))
(if (>= t_2 (+ t_0 t_3))
(* (/ 1.0 t_4) (* (floor h) dX.v))
(/ (* t_1 (- -1.0)) 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 = powf((dY_46_u * floorf(w)), 2.0f);
float t_1 = dY_46_v * floorf(h);
float t_2 = powf((dX_46_v * floorf(h)), 2.0f) + powf((dX_46_u * floorf(w)), 2.0f);
float t_3 = powf(t_1, 2.0f);
float t_4 = sqrtf(fmaxf(t_2, (t_3 + t_0)));
float tmp;
if (t_2 >= (t_0 + t_3)) {
tmp = (1.0f / t_4) * (floorf(h) * dX_46_v);
} else {
tmp = (t_1 * -(-1.0f)) / 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)) ^ Float32(2.0) t_1 = Float32(dY_46_v * floor(h)) t_2 = Float32((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))) t_3 = t_1 ^ Float32(2.0) t_4 = sqrt(fmax(t_2, Float32(t_3 + t_0))) tmp = Float32(0.0) if (t_2 >= Float32(t_0 + t_3)) tmp = Float32(Float32(Float32(1.0) / t_4) * Float32(floor(h) * dX_46_v)); else tmp = Float32(Float32(t_1 * Float32(-Float32(-1.0))) / 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 = (dY_46_u * floor(w)) ^ single(2.0); t_1 = dY_46_v * floor(h); t_2 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((dX_46_u * floor(w)) ^ single(2.0)); t_3 = t_1 ^ single(2.0); t_4 = sqrt(max(t_2, (t_3 + t_0))); tmp = single(0.0); if (t_2 >= (t_0 + t_3)) tmp = (single(1.0) / t_4) * (floor(h) * dX_46_v); else tmp = (t_1 * -single(-1.0)) / t_4; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_2 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_3 := {t\_1}^{2}\\
t_4 := \sqrt{\mathsf{max}\left(t\_2, t\_3 + t\_0\right)}\\
\mathbf{if}\;t\_2 \geq t\_0 + t\_3:\\
\;\;\;\;\frac{1}{t\_4} \cdot \left(\left\lfloor h\right\rfloor \cdot dX.v\right)\\
\mathbf{else}:\\
\;\;\;\;\frac{t\_1 \cdot \left(--1\right)}{t\_4}\\
\end{array}
\end{array}
Initial program 77.2%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites77.3%
lift-+.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
+-commutativeN/A
lift-+.f3277.3
Applied rewrites77.3%
lift-+.f32N/A
+-commutativeN/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f32N/A
pow2N/A
lower-pow.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-+.f3277.3
Applied rewrites77.3%
Final simplification77.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor h) dY.v))
(t_3 (+ (* t_0 t_0) (* t_2 t_2)))
(t_4 (+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (* t_5 t_5))
(t_7 (+ t_6 (* t_1 t_1)))
(t_8 (/ 1.0 (sqrt (fmax t_7 t_3))))
(t_9 (* t_8 t_2))
(t_10 (* t_8 t_1))
(t_11 (if (>= t_7 t_3) t_10 t_9)))
(if (or (<= t_11 -0.9999989867210388) (not (<= t_11 4.999999987376214e-7)))
(if (>= (pow t_1 2.0) t_4) t_10 t_9)
(if (>= (pow (* dX.u (floor w)) 2.0) t_4)
(*
(/ 1.0 (sqrt (fmax (+ t_6 (* (pow (floor h) 2.0) (* dX.v dX.v))) t_3)))
t_1)
t_9))))
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) * dY_46_u;
float t_1 = floorf(h) * dX_46_v;
float t_2 = floorf(h) * dY_46_v;
float t_3 = (t_0 * t_0) + (t_2 * t_2);
float t_4 = powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = t_5 * t_5;
float t_7 = t_6 + (t_1 * t_1);
float t_8 = 1.0f / sqrtf(fmaxf(t_7, t_3));
float t_9 = t_8 * t_2;
float t_10 = t_8 * t_1;
float tmp;
if (t_7 >= t_3) {
tmp = t_10;
} else {
tmp = t_9;
}
float t_11 = tmp;
float tmp_2;
if ((t_11 <= -0.9999989867210388f) || !(t_11 <= 4.999999987376214e-7f)) {
float tmp_3;
if (powf(t_1, 2.0f) >= t_4) {
tmp_3 = t_10;
} else {
tmp_3 = t_9;
}
tmp_2 = tmp_3;
} else if (powf((dX_46_u * floorf(w)), 2.0f) >= t_4) {
tmp_2 = (1.0f / sqrtf(fmaxf((t_6 + (powf(floorf(h), 2.0f) * (dX_46_v * dX_46_v))), t_3))) * t_1;
} else {
tmp_2 = t_9;
}
return tmp_2;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(h) * dY_46_v) t_3 = Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)) t_4 = Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(t_5 * t_5) t_7 = Float32(t_6 + Float32(t_1 * t_1)) t_8 = Float32(Float32(1.0) / sqrt(fmax(t_7, t_3))) t_9 = Float32(t_8 * t_2) t_10 = Float32(t_8 * t_1) tmp = Float32(0.0) if (t_7 >= t_3) tmp = t_10; else tmp = t_9; end t_11 = tmp tmp_2 = Float32(0.0) if ((t_11 <= Float32(-0.9999989867210388)) || !(t_11 <= Float32(4.999999987376214e-7))) tmp_3 = Float32(0.0) if ((t_1 ^ Float32(2.0)) >= t_4) tmp_3 = t_10; else tmp_3 = t_9; end tmp_2 = tmp_3; elseif ((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) >= t_4) tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_6 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))), t_3))) * t_1); else tmp_2 = t_9; end return tmp_2 end
function tmp_5 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = floor(h) * dX_46_v; t_2 = floor(h) * dY_46_v; t_3 = (t_0 * t_0) + (t_2 * t_2); t_4 = ((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = t_5 * t_5; t_7 = t_6 + (t_1 * t_1); t_8 = single(1.0) / sqrt(max(t_7, t_3)); t_9 = t_8 * t_2; t_10 = t_8 * t_1; tmp = single(0.0); if (t_7 >= t_3) tmp = t_10; else tmp = t_9; end t_11 = tmp; tmp_3 = single(0.0); if ((t_11 <= single(-0.9999989867210388)) || ~((t_11 <= single(4.999999987376214e-7)))) tmp_4 = single(0.0); if ((t_1 ^ single(2.0)) >= t_4) tmp_4 = t_10; else tmp_4 = t_9; end tmp_3 = tmp_4; elseif (((dX_46_u * floor(w)) ^ single(2.0)) >= t_4) tmp_3 = (single(1.0) / sqrt(max((t_6 + ((floor(h) ^ single(2.0)) * (dX_46_v * dX_46_v))), t_3))) * t_1; else tmp_3 = t_9; end tmp_5 = tmp_3; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_3 := t\_0 \cdot t\_0 + t\_2 \cdot t\_2\\
t_4 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2} + {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_5 \cdot t\_5\\
t_7 := t\_6 + t\_1 \cdot t\_1\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_3\right)}}\\
t_9 := t\_8 \cdot t\_2\\
t_10 := t\_8 \cdot t\_1\\
t_11 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_3:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{if}\;t\_11 \leq -0.9999989867210388 \lor \neg \left(t\_11 \leq 4.999999987376214 \cdot 10^{-7}\right):\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_1}^{2} \geq t\_4:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{elif}\;{\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2} \geq t\_4:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6 + {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot \left(dX.v \cdot dX.v\right), t\_3\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}
\end{array}
if (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < -0.999998987 or 4.99999999e-7 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) Initial program 99.6%
Taylor expanded in dX.u around inf
Applied rewrites67.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3267.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.7
Applied rewrites67.7%
lift-+.f32N/A
+-commutativeN/A
lift-pow.f32N/A
lift-*.f32N/A
Applied rewrites67.7%
Taylor expanded in dX.u around 0
Applied rewrites99.6%
if -0.999998987 < (if (>=.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dX.v)) (*.f32 (/.f32 #s(literal 1 binary32) (sqrt.f32 (fmax.f32 (+.f32 (*.f32 (*.f32 (floor.f32 w) dX.u) (*.f32 (floor.f32 w) dX.u)) (*.f32 (*.f32 (floor.f32 h) dX.v) (*.f32 (floor.f32 h) dX.v))) (+.f32 (*.f32 (*.f32 (floor.f32 w) dY.u) (*.f32 (floor.f32 w) dY.u)) (*.f32 (*.f32 (floor.f32 h) dY.v) (*.f32 (floor.f32 h) dY.v)))))) (*.f32 (floor.f32 h) dY.v))) < 4.99999999e-7Initial program 65.0%
Taylor expanded in dX.u around inf
Applied rewrites65.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3265.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3265.0
Applied rewrites65.0%
lift-+.f32N/A
+-commutativeN/A
lift-pow.f32N/A
lift-*.f32N/A
Applied rewrites65.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lower-*.f32N/A
lower-pow.f32N/A
pow2N/A
lower-*.f3265.0
Applied rewrites65.0%
Final simplification77.2%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* (floor h) dY.v))
(t_2 (+ (* t_0 t_0) (* t_1 t_1)))
(t_3 (* (floor h) dX.v))
(t_4 (* (floor w) dX.u))
(t_5 (* t_4 t_4)))
(if (>=
(pow (* dX.u (floor w)) 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(*
(/ 1.0 (sqrt (fmax (+ t_5 (* (pow (floor h) 2.0) (* dX.v dX.v))) t_2)))
t_3)
(* (/ 1.0 (sqrt (fmax (+ t_5 (* t_3 t_3)) t_2))) 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(w) * dY_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = (t_0 * t_0) + (t_1 * t_1);
float t_3 = floorf(h) * dX_46_v;
float t_4 = floorf(w) * dX_46_u;
float t_5 = t_4 * t_4;
float tmp;
if (powf((dX_46_u * floorf(w)), 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (powf(floorf(h), 2.0f) * (dX_46_v * dX_46_v))), t_2))) * t_3;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_5 + (t_3 * t_3)), t_2))) * 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(w) * dY_46_u) t_1 = Float32(floor(h) * dY_46_v) t_2 = Float32(Float32(t_0 * t_0) + Float32(t_1 * t_1)) t_3 = Float32(floor(h) * dX_46_v) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(t_4 * t_4) tmp = Float32(0.0) if ((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_5 + Float32((floor(h) ^ Float32(2.0)) * Float32(dX_46_v * dX_46_v))), t_2))) * t_3); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(t_5 + Float32(t_3 * t_3)), t_2))) * 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(w) * dY_46_u; t_1 = floor(h) * dY_46_v; t_2 = (t_0 * t_0) + (t_1 * t_1); t_3 = floor(h) * dX_46_v; t_4 = floor(w) * dX_46_u; t_5 = t_4 * t_4; tmp = single(0.0); if (((dX_46_u * floor(w)) ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((t_5 + ((floor(h) ^ single(2.0)) * (dX_46_v * dX_46_v))), t_2))) * t_3; else tmp = (single(1.0) / sqrt(max((t_5 + (t_3 * t_3)), t_2))) * t_1; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_0 \cdot t\_0 + t\_1 \cdot t\_1\\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4\\
\mathbf{if}\;{\left(dX.u \cdot \left\lfloor w\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}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + {\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot \left(dX.v \cdot dX.v\right), t\_2\right)}} \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5 + t\_3 \cdot t\_3, t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in dX.u around inf
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.0
Applied rewrites66.0%
lift-+.f32N/A
+-commutativeN/A
lift-pow.f32N/A
lift-*.f32N/A
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
lower-*.f32N/A
lower-pow.f32N/A
pow2N/A
lower-*.f3266.0
Applied rewrites66.0%
(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 (* t_2 t_2))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_4 t_4) (* t_0 t_0))))
(if (>=
(pow (* dX.u (floor w)) 2.0)
(+ (pow (* dY.u (floor w)) 2.0) (pow (* dY.v (floor h)) 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_1 t_1) t_3)))) t_0)
(*
(/
1.0
(sqrt (fmax t_5 (+ (* (floor w) (* (floor w) (* dY.u dY.u))) t_3))))
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 = floorf(h) * dX_46_v;
float t_1 = floorf(w) * dY_46_u;
float t_2 = floorf(h) * dY_46_v;
float t_3 = t_2 * t_2;
float t_4 = floorf(w) * dX_46_u;
float t_5 = (t_4 * t_4) + (t_0 * t_0);
float tmp;
if (powf((dX_46_u * floorf(w)), 2.0f) >= (powf((dY_46_u * floorf(w)), 2.0f) + powf((dY_46_v * floorf(h)), 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((t_1 * t_1) + t_3)))) * t_0;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((floorf(w) * (floorf(w) * (dY_46_u * dY_46_u))) + t_3)))) * t_2;
}
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(t_2 * t_2) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(t_4 * t_4) + Float32(t_0 * t_0)) tmp = Float32(0.0) if ((Float32(dX_46_u * floor(w)) ^ Float32(2.0)) >= Float32((Float32(dY_46_u * floor(w)) ^ Float32(2.0)) + (Float32(dY_46_v * floor(h)) ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, Float32(Float32(t_1 * t_1) + t_3)))) * t_0); else tmp = Float32(Float32(Float32(1.0) / sqrt(fmax(t_5, Float32(Float32(floor(w) * Float32(floor(w) * Float32(dY_46_u * dY_46_u))) + t_3)))) * t_2); 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 = t_2 * t_2; t_4 = floor(w) * dX_46_u; t_5 = (t_4 * t_4) + (t_0 * t_0); tmp = single(0.0); if (((dX_46_u * floor(w)) ^ single(2.0)) >= (((dY_46_u * floor(w)) ^ single(2.0)) + ((dY_46_v * floor(h)) ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(t_5, ((t_1 * t_1) + t_3)))) * t_0; else tmp = (single(1.0) / sqrt(max(t_5, ((floor(w) * (floor(w) * (dY_46_u * dY_46_u))) + t_3)))) * t_2; 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 := t\_2 \cdot t\_2\\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_4 \cdot t\_4 + t\_0 \cdot t\_0\\
\mathbf{if}\;{\left(dX.u \cdot \left\lfloor w\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}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_1 \cdot t\_1 + t\_3\right)}} \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dY.u \cdot dY.u\right)\right) + t\_3\right)}} \cdot t\_2\\
\end{array}
\end{array}
Initial program 77.2%
Taylor expanded in dX.u around inf
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3266.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3266.0
Applied rewrites66.0%
lift-+.f32N/A
+-commutativeN/A
lift-pow.f32N/A
lift-*.f32N/A
Applied rewrites66.0%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
unpow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3266.0
Applied rewrites66.0%
herbie shell --seed 2025019
(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))))