
(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(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(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 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_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(((t_3 != t_3) ? t_5 : ((t_5 != t_5) ? t_3 : max(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 (* dY.v (floor h)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow (* (floor w) dX.u) 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (sqrt (fmax t_1 (+ (pow t_2 2.0) (pow t_0 2.0))))))
(if (>= t_1 (+ (* t_0 t_0) (* t_2 t_2)))
(/ (/ (* (- dX.v) (floor h)) -1.0) t_3)
(* (/ 1.0 t_3) t_0))))
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 = powf((dX_46_v * floorf(h)), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = sqrtf(fmaxf(t_1, (powf(t_2, 2.0f) + powf(t_0, 2.0f))));
float tmp;
if (t_1 >= ((t_0 * t_0) + (t_2 * t_2))) {
tmp = ((-dX_46_v * floorf(h)) / -1.0f) / t_3;
} else {
tmp = (1.0f / t_3) * t_0;
}
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(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = sqrt(((t_1 != t_1) ? Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : ((Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? t_1 : max(t_1, Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))))))) tmp = Float32(0.0) if (t_1 >= Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2))) tmp = Float32(Float32(Float32(Float32(-dX_46_v) * floor(h)) / Float32(-1.0)) / t_3); else tmp = Float32(Float32(Float32(1.0) / t_3) * t_0); 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_v * floor(h); t_1 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = sqrt(max(t_1, ((t_2 ^ single(2.0)) + (t_0 ^ single(2.0))))); tmp = single(0.0); if (t_1 >= ((t_0 * t_0) + (t_2 * t_2))) tmp = ((-dX_46_v * floor(h)) / single(-1.0)) / t_3; else tmp = (single(1.0) / t_3) * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \sqrt{\mathsf{max}\left(t\_1, {t\_2}^{2} + {t\_0}^{2}\right)}\\
\mathbf{if}\;t\_1 \geq t\_0 \cdot t\_0 + t\_2 \cdot t\_2:\\
\;\;\;\;\frac{\frac{\left(-dX.v\right) \cdot \left\lfloor h\right\rfloor }{-1}}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_3} \cdot t\_0\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites77.7%
Applied rewrites77.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3277.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3277.7
Applied rewrites77.7%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.7
Applied rewrites77.7%
Final simplification77.7%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (pow (floor w) 2.0))
(t_2 (* dY.v (floor h)))
(t_3 (* t_2 t_2))
(t_4 (* dY.u (floor w)))
(t_5 (+ t_3 (* t_4 t_4)))
(t_6 (+ (pow t_4 2.0) (pow t_2 2.0)))
(t_7 (* dX.v (floor h)))
(t_8 (+ (* t_7 t_7) (* t_0 t_0)))
(t_9 (/ 1.0 (sqrt (fmax t_8 t_5))))
(t_10 (* t_9 t_2))
(t_11 (* t_9 t_7))
(t_12 (if (>= t_8 t_5) t_11 t_10)))
(if (<= t_12 -0.9999989867210388)
(if (>= (pow t_7 2.0) t_6) t_11 t_10)
(if (<= t_12 9.999999747378752e-5)
(if (>= (* (* t_1 dX.u) dX.u) t_6) t_11 t_10)
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) t_6)
t_11
(* (/ 1.0 (sqrt (fmax t_8 (+ (* (* t_1 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(w) * dX_46_u;
float t_1 = powf(floorf(w), 2.0f);
float t_2 = dY_46_v * floorf(h);
float t_3 = t_2 * t_2;
float t_4 = dY_46_u * floorf(w);
float t_5 = t_3 + (t_4 * t_4);
float t_6 = powf(t_4, 2.0f) + powf(t_2, 2.0f);
float t_7 = dX_46_v * floorf(h);
float t_8 = (t_7 * t_7) + (t_0 * t_0);
float t_9 = 1.0f / sqrtf(fmaxf(t_8, t_5));
float t_10 = t_9 * t_2;
float t_11 = t_9 * t_7;
float tmp;
if (t_8 >= t_5) {
tmp = t_11;
} else {
tmp = t_10;
}
float t_12 = tmp;
float tmp_2;
if (t_12 <= -0.9999989867210388f) {
float tmp_3;
if (powf(t_7, 2.0f) >= t_6) {
tmp_3 = t_11;
} else {
tmp_3 = t_10;
}
tmp_2 = tmp_3;
} else if (t_12 <= 9.999999747378752e-5f) {
float tmp_4;
if (((t_1 * dX_46_u) * dX_46_u) >= t_6) {
tmp_4 = t_11;
} else {
tmp_4 = t_10;
}
tmp_2 = tmp_4;
} else if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= t_6) {
tmp_2 = t_11;
} else {
tmp_2 = (1.0f / sqrtf(fmaxf(t_8, (((t_1 * dY_46_u) * dY_46_u) + t_3)))) * t_2;
}
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) * dX_46_u) t_1 = floor(w) ^ Float32(2.0) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(t_2 * t_2) t_4 = Float32(dY_46_u * floor(w)) t_5 = Float32(t_3 + Float32(t_4 * t_4)) t_6 = Float32((t_4 ^ Float32(2.0)) + (t_2 ^ Float32(2.0))) t_7 = Float32(dX_46_v * floor(h)) t_8 = Float32(Float32(t_7 * t_7) + Float32(t_0 * t_0)) t_9 = Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? t_5 : ((t_5 != t_5) ? t_8 : max(t_8, t_5))))) t_10 = Float32(t_9 * t_2) t_11 = Float32(t_9 * t_7) tmp = Float32(0.0) if (t_8 >= t_5) tmp = t_11; else tmp = t_10; end t_12 = tmp tmp_2 = Float32(0.0) if (t_12 <= Float32(-0.9999989867210388)) tmp_3 = Float32(0.0) if ((t_7 ^ Float32(2.0)) >= t_6) tmp_3 = t_11; else tmp_3 = t_10; end tmp_2 = tmp_3; elseif (t_12 <= Float32(9.999999747378752e-5)) tmp_4 = Float32(0.0) if (Float32(Float32(t_1 * dX_46_u) * dX_46_u) >= t_6) tmp_4 = t_11; else tmp_4 = t_10; end tmp_2 = tmp_4; elseif (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= t_6) tmp_2 = t_11; else tmp_2 = Float32(Float32(Float32(1.0) / sqrt(((t_8 != t_8) ? Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) + t_3) : ((Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) + t_3) != Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) + t_3)) ? t_8 : max(t_8, Float32(Float32(Float32(t_1 * dY_46_u) * dY_46_u) + t_3)))))) * t_2); end return tmp_2 end
function tmp_6 = code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dX_46_u; t_1 = floor(w) ^ single(2.0); t_2 = dY_46_v * floor(h); t_3 = t_2 * t_2; t_4 = dY_46_u * floor(w); t_5 = t_3 + (t_4 * t_4); t_6 = (t_4 ^ single(2.0)) + (t_2 ^ single(2.0)); t_7 = dX_46_v * floor(h); t_8 = (t_7 * t_7) + (t_0 * t_0); t_9 = single(1.0) / sqrt(max(t_8, t_5)); t_10 = t_9 * t_2; t_11 = t_9 * t_7; tmp = single(0.0); if (t_8 >= t_5) tmp = t_11; else tmp = t_10; end t_12 = tmp; tmp_3 = single(0.0); if (t_12 <= single(-0.9999989867210388)) tmp_4 = single(0.0); if ((t_7 ^ single(2.0)) >= t_6) tmp_4 = t_11; else tmp_4 = t_10; end tmp_3 = tmp_4; elseif (t_12 <= single(9.999999747378752e-5)) tmp_5 = single(0.0); if (((t_1 * dX_46_u) * dX_46_u) >= t_6) tmp_5 = t_11; else tmp_5 = t_10; end tmp_3 = tmp_5; elseif ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= t_6) tmp_3 = t_11; else tmp_3 = (single(1.0) / sqrt(max(t_8, (((t_1 * dY_46_u) * dY_46_u) + t_3)))) * t_2; end tmp_6 = tmp_3; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := t\_2 \cdot t\_2\\
t_4 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_5 := t\_3 + t\_4 \cdot t\_4\\
t_6 := {t\_4}^{2} + {t\_2}^{2}\\
t_7 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_8 := t\_7 \cdot t\_7 + t\_0 \cdot t\_0\\
t_9 := \frac{1}{\sqrt{\mathsf{max}\left(t\_8, t\_5\right)}}\\
t_10 := t\_9 \cdot t\_2\\
t_11 := t\_9 \cdot t\_7\\
t_12 := \begin{array}{l}
\mathbf{if}\;t\_8 \geq t\_5:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}\\
\mathbf{if}\;t\_12 \leq -0.9999989867210388:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{t\_7}^{2} \geq t\_6:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}\\
\mathbf{elif}\;t\_12 \leq 9.999999747378752 \cdot 10^{-5}:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left(t\_1 \cdot dX.u\right) \cdot dX.u \geq t\_6:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;t\_10\\
\end{array}\\
\mathbf{elif}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_6:\\
\;\;\;\;t\_11\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_8, \left(t\_1 \cdot dY.u\right) \cdot dY.u + t\_3\right)}} \cdot t\_2\\
\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))) < -0.999998987Initial program 99.5%
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.f3299.5
Applied rewrites99.5%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3299.5
Applied rewrites99.5%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3299.5
Applied rewrites99.5%
Applied rewrites99.5%
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))) < 9.99999975e-5Initial program 65.7%
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.f3249.9
Applied rewrites49.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3249.9
Applied rewrites49.9%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3249.9
Applied rewrites49.9%
Taylor expanded in dX.u around inf
*-commutativeN/A
unpow2N/A
associate-*r*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3265.7
Applied rewrites65.7%
if 9.99999975e-5 < (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 99.3%
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.f3299.3
Applied rewrites99.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3299.3
Applied rewrites99.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3299.3
Applied rewrites99.3%
lift-*.f32N/A
lift-*.f32N/A
associate-*r*N/A
lift-*.f32N/A
*-commutativeN/A
associate-*r*N/A
unpow2N/A
lift-pow.f32N/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f3299.4
Applied rewrites99.4%
Final simplification77.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.v (floor h)))
(t_1 (+ (pow (* dX.v (floor h)) 2.0) (pow (* (floor w) dX.u) 2.0)))
(t_2 (* dY.u (floor w)))
(t_3 (sqrt (fmax t_1 (+ (pow t_2 2.0) (pow t_0 2.0))))))
(if (>= t_1 (+ (* t_0 t_0) (* t_2 t_2)))
(* (/ dX.v t_3) (floor h))
(* (/ 1.0 t_3) t_0))))
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 = powf((dX_46_v * floorf(h)), 2.0f) + powf((floorf(w) * dX_46_u), 2.0f);
float t_2 = dY_46_u * floorf(w);
float t_3 = sqrtf(fmaxf(t_1, (powf(t_2, 2.0f) + powf(t_0, 2.0f))));
float tmp;
if (t_1 >= ((t_0 * t_0) + (t_2 * t_2))) {
tmp = (dX_46_v / t_3) * floorf(h);
} else {
tmp = (1.0f / t_3) * t_0;
}
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(dX_46_v * floor(h)) ^ Float32(2.0)) + (Float32(floor(w) * dX_46_u) ^ Float32(2.0))) t_2 = Float32(dY_46_u * floor(w)) t_3 = sqrt(((t_1 != t_1) ? Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) : ((Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) != Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0)))) ? t_1 : max(t_1, Float32((t_2 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))))))) tmp = Float32(0.0) if (t_1 >= Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2))) tmp = Float32(Float32(dX_46_v / t_3) * floor(h)); else tmp = Float32(Float32(Float32(1.0) / t_3) * t_0); 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_v * floor(h); t_1 = ((dX_46_v * floor(h)) ^ single(2.0)) + ((floor(w) * dX_46_u) ^ single(2.0)); t_2 = dY_46_u * floor(w); t_3 = sqrt(max(t_1, ((t_2 ^ single(2.0)) + (t_0 ^ single(2.0))))); tmp = single(0.0); if (t_1 >= ((t_0 * t_0) + (t_2 * t_2))) tmp = (dX_46_v / t_3) * floor(h); else tmp = (single(1.0) / t_3) * t_0; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_1 := {\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(\left\lfloor w\right\rfloor \cdot dX.u\right)}^{2}\\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \sqrt{\mathsf{max}\left(t\_1, {t\_2}^{2} + {t\_0}^{2}\right)}\\
\mathbf{if}\;t\_1 \geq t\_0 \cdot t\_0 + t\_2 \cdot t\_2:\\
\;\;\;\;\frac{dX.v}{t\_3} \cdot \left\lfloor h\right\rfloor \\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_3} \cdot t\_0\\
\end{array}
\end{array}
Initial program 77.6%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
*-rgt-identityN/A
Applied rewrites77.6%
Applied rewrites77.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3277.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3277.6
Applied rewrites77.6%
lift-*.f32N/A
pow2N/A
lift-pow.f3277.6
Applied rewrites77.6%
Final simplification77.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* dX.v (floor h)))
(t_2 (* t_1 t_1))
(t_3 (* dY.v (floor h)))
(t_4 (+ (* t_3 t_3) (* t_0 t_0)))
(t_5 (* (floor w) dX.u)))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow t_0 2.0) (pow t_3 2.0)))
(*
(/ 1.0 (sqrt (fmax (+ (* (* dX.u dX.u) (pow (floor w) 2.0)) t_2) t_4)))
t_1)
(* (/ 1.0 (sqrt (fmax (+ t_2 (* t_5 t_5)) 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 = dY_46_u * floorf(w);
float t_1 = dX_46_v * floorf(h);
float t_2 = t_1 * t_1;
float t_3 = dY_46_v * floorf(h);
float t_4 = (t_3 * t_3) + (t_0 * t_0);
float t_5 = floorf(w) * dX_46_u;
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf(t_0, 2.0f) + powf(t_3, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf((((dX_46_u * dX_46_u) * powf(floorf(w), 2.0f)) + t_2), t_4))) * t_1;
} else {
tmp = (1.0f / sqrtf(fmaxf((t_2 + (t_5 * t_5)), t_4))) * 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_u * floor(w)) t_1 = Float32(dX_46_v * floor(h)) t_2 = Float32(t_1 * t_1) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) t_5 = Float32(floor(w) * dX_46_u) tmp = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32((t_0 ^ Float32(2.0)) + (t_3 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(Float32(Float32(dX_46_u * dX_46_u) * (floor(w) ^ Float32(2.0))) + t_2) != Float32(Float32(Float32(dX_46_u * dX_46_u) * (floor(w) ^ Float32(2.0))) + t_2)) ? t_4 : ((t_4 != t_4) ? Float32(Float32(Float32(dX_46_u * dX_46_u) * (floor(w) ^ Float32(2.0))) + t_2) : max(Float32(Float32(Float32(dX_46_u * dX_46_u) * (floor(w) ^ Float32(2.0))) + t_2), t_4))))) * t_1); else tmp = Float32(Float32(Float32(1.0) / sqrt(((Float32(t_2 + Float32(t_5 * t_5)) != Float32(t_2 + Float32(t_5 * t_5))) ? t_4 : ((t_4 != t_4) ? Float32(t_2 + Float32(t_5 * t_5)) : max(Float32(t_2 + Float32(t_5 * t_5)), t_4))))) * 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 = dY_46_u * floor(w); t_1 = dX_46_v * floor(h); t_2 = t_1 * t_1; t_3 = dY_46_v * floor(h); t_4 = (t_3 * t_3) + (t_0 * t_0); t_5 = floor(w) * dX_46_u; tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= ((t_0 ^ single(2.0)) + (t_3 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max((((dX_46_u * dX_46_u) * (floor(w) ^ single(2.0))) + t_2), t_4))) * t_1; else tmp = (single(1.0) / sqrt(max((t_2 + (t_5 * t_5)), t_4))) * t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := t\_3 \cdot t\_3 + t\_0 \cdot t\_0\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq {t\_0}^{2} + {t\_3}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left(dX.u \cdot dX.u\right) \cdot {\left(\left\lfloor w\right\rfloor \right)}^{2} + t\_2, t\_4\right)}} \cdot t\_1\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_2 + t\_5 \cdot t\_5, t\_4\right)}} \cdot t\_3\\
\end{array}
\end{array}
Initial program 77.6%
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.f3267.3
Applied rewrites67.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3267.3
Applied rewrites67.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3267.3
Applied rewrites67.3%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
lift-*.f32N/A
lower-*.f3267.3
Applied rewrites67.3%
Final simplification67.3%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.u (floor w)))
(t_2 (* dY.v (floor h)))
(t_3 (* dX.v (floor h)))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_2 t_2) (* t_1 t_1)))))))
(if (>= (pow t_3 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* t_4 t_3)
(* t_4 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(w) * dX_46_u;
float t_1 = dY_46_u * floorf(w);
float t_2 = dY_46_v * floorf(h);
float t_3 = dX_46_v * floorf(h);
float t_4 = 1.0f / sqrtf(fmaxf(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1))));
float tmp;
if (powf(t_3, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * t_3;
} else {
tmp = t_4 * 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(w) * dX_46_u) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(dY_46_v * floor(h)) t_3 = Float32(dX_46_v * floor(h)) t_4 = Float32(Float32(1.0) / sqrt(((Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) != Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0))) ? Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) : ((Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1)) != Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))) ? Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)) : max(Float32(Float32(t_3 * t_3) + Float32(t_0 * t_0)), Float32(Float32(t_2 * t_2) + Float32(t_1 * t_1))))))) tmp = Float32(0.0) if ((t_3 ^ Float32(2.0)) >= Float32((t_1 ^ Float32(2.0)) + (t_2 ^ Float32(2.0)))) tmp = Float32(t_4 * t_3); else tmp = Float32(t_4 * 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(w) * dX_46_u; t_1 = dY_46_u * floor(w); t_2 = dY_46_v * floor(h); t_3 = dX_46_v * floor(h); t_4 = single(1.0) / sqrt(max(((t_3 * t_3) + (t_0 * t_0)), ((t_2 * t_2) + (t_1 * t_1)))); tmp = single(0.0); if ((t_3 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * t_3; else tmp = t_4 * t_2; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \frac{1}{\sqrt{\mathsf{max}\left(t\_3 \cdot t\_3 + t\_0 \cdot t\_0, t\_2 \cdot t\_2 + t\_1 \cdot t\_1\right)}}\\
\mathbf{if}\;{t\_3}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_3\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\end{array}
\end{array}
Initial program 77.6%
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.f3267.3
Applied rewrites67.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3267.3
Applied rewrites67.3%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3267.3
Applied rewrites67.3%
Applied rewrites67.3%
Final simplification67.3%
herbie shell --seed 2024276
(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))))