
(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(((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_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(((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_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 (* (floor w) dX.u))
(t_1 (+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0)))
(t_2 (+ (pow (* (floor h) dX.v) 2.0) (pow t_0 2.0)))
(t_3 (sqrt (fmax t_2 t_1))))
(if (>= t_2 t_1)
(* t_0 (/ 1.0 t_3))
(/ (/ (* (- dY.u) (floor w)) -1.0) 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 = floorf(w) * dX_46_u;
float t_1 = powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f);
float t_2 = powf((floorf(h) * dX_46_v), 2.0f) + powf(t_0, 2.0f);
float t_3 = sqrtf(fmaxf(t_2, t_1));
float tmp;
if (t_2 >= t_1) {
tmp = t_0 * (1.0f / t_3);
} else {
tmp = ((-dY_46_u * floorf(w)) / -1.0f) / 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) * dX_46_u) t_1 = Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))) t_2 = Float32((Float32(floor(h) * dX_46_v) ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_3 = sqrt(((t_2 != t_2) ? t_1 : ((t_1 != t_1) ? t_2 : max(t_2, t_1)))) tmp = Float32(0.0) if (t_2 >= t_1) tmp = Float32(t_0 * Float32(Float32(1.0) / t_3)); else tmp = Float32(Float32(Float32(Float32(-dY_46_u) * floor(w)) / Float32(-1.0)) / 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(w) * dX_46_u; t_1 = ((dY_46_v * floor(h)) ^ single(2.0)) + ((dY_46_u * floor(w)) ^ single(2.0)); t_2 = ((floor(h) * dX_46_v) ^ single(2.0)) + (t_0 ^ single(2.0)); t_3 = sqrt(max(t_2, t_1)); tmp = single(0.0); if (t_2 >= t_1) tmp = t_0 * (single(1.0) / t_3); else tmp = ((-dY_46_u * floor(w)) / single(-1.0)) / t_3; end tmp_2 = tmp; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := {\left(\left\lfloor h\right\rfloor \cdot dX.v\right)}^{2} + {t\_0}^{2}\\
t_3 := \sqrt{\mathsf{max}\left(t\_2, t\_1\right)}\\
\mathbf{if}\;t\_2 \geq t\_1:\\
\;\;\;\;t\_0 \cdot \frac{1}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{\frac{\left(-dY.u\right) \cdot \left\lfloor w\right\rfloor }{-1}}{t\_3}\\
\end{array}
\end{array}
Initial program 71.2%
lift-*.f32N/A
lift-/.f32N/A
associate-*l/N/A
frac-2negN/A
Applied rewrites71.4%
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-+.f3271.4
Applied rewrites71.4%
lift-*.f32N/A
pow2N/A
lower-pow.f3271.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3271.4
Applied rewrites71.4%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lift-pow.f3271.4
lift-*.f32N/A
pow2N/A
lift-pow.f3271.4
lift-*.f32N/A
pow2N/A
lower-pow.f3271.4
lift-*.f32N/A
*-commutativeN/A
lift-*.f3271.4
Applied rewrites71.4%
Final simplification71.4%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* dY.u (floor w)))
(t_1 (* t_0 t_0))
(t_2 (* (floor h) dX.v))
(t_3 (* dY.v (floor h)))
(t_4 (+ (pow t_3 2.0) (pow t_0 2.0)))
(t_5 (* (floor w) dX.u))
(t_6 (+ (* t_2 t_2) (* t_5 t_5)))
(t_7 (/ 1.0 (sqrt (fmax t_6 (+ (* t_3 t_3) t_1)))))
(t_8 (* t_7 t_5)))
(if (<= dX.v 0.14000000059604645)
(if (>= (* (* (pow (floor w) 2.0) dX.u) dX.u) t_4) t_8 (* t_7 t_0))
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) t_4)
t_8
(*
(/
1.0
(sqrt (fmax t_6 (+ (* (* (* dY.v dY.v) (floor h)) (floor h)) t_1))))
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_u * floorf(w);
float t_1 = t_0 * t_0;
float t_2 = floorf(h) * dX_46_v;
float t_3 = dY_46_v * floorf(h);
float t_4 = powf(t_3, 2.0f) + powf(t_0, 2.0f);
float t_5 = floorf(w) * dX_46_u;
float t_6 = (t_2 * t_2) + (t_5 * t_5);
float t_7 = 1.0f / sqrtf(fmaxf(t_6, ((t_3 * t_3) + t_1)));
float t_8 = t_7 * t_5;
float tmp_1;
if (dX_46_v <= 0.14000000059604645f) {
float tmp_2;
if (((powf(floorf(w), 2.0f) * dX_46_u) * dX_46_u) >= t_4) {
tmp_2 = t_8;
} else {
tmp_2 = t_7 * t_0;
}
tmp_1 = tmp_2;
} else if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= t_4) {
tmp_1 = t_8;
} else {
tmp_1 = (1.0f / sqrtf(fmaxf(t_6, ((((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)) + t_1)))) * t_0;
}
return tmp_1;
}
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(t_0 * t_0) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32((t_3 ^ Float32(2.0)) + (t_0 ^ Float32(2.0))) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(t_2 * t_2) + Float32(t_5 * t_5)) t_7 = Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(t_3 * t_3) + t_1) : ((Float32(Float32(t_3 * t_3) + t_1) != Float32(Float32(t_3 * t_3) + t_1)) ? t_6 : max(t_6, Float32(Float32(t_3 * t_3) + t_1)))))) t_8 = Float32(t_7 * t_5) tmp_1 = Float32(0.0) if (dX_46_v <= Float32(0.14000000059604645)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(w) ^ Float32(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_2 = t_8; else tmp_2 = Float32(t_7 * t_0); end tmp_1 = tmp_2; elseif (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= t_4) tmp_1 = t_8; else tmp_1 = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_1) : ((Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_1) != Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_1)) ? t_6 : max(t_6, Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_1)))))) * t_0); 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 = dY_46_u * floor(w); t_1 = t_0 * t_0; t_2 = floor(h) * dX_46_v; t_3 = dY_46_v * floor(h); t_4 = (t_3 ^ single(2.0)) + (t_0 ^ single(2.0)); t_5 = floor(w) * dX_46_u; t_6 = (t_2 * t_2) + (t_5 * t_5); t_7 = single(1.0) / sqrt(max(t_6, ((t_3 * t_3) + t_1))); t_8 = t_7 * t_5; tmp_2 = single(0.0); if (dX_46_v <= single(0.14000000059604645)) tmp_3 = single(0.0); if ((((floor(w) ^ single(2.0)) * dX_46_u) * dX_46_u) >= t_4) tmp_3 = t_8; else tmp_3 = t_7 * t_0; end tmp_2 = tmp_3; elseif ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= t_4) tmp_2 = t_8; else tmp_2 = (single(1.0) / sqrt(max(t_6, ((((dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_1)))) * t_0; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {t\_3}^{2} + {t\_0}^{2}\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_2 \cdot t\_2 + t\_5 \cdot t\_5\\
t_7 := \frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_3 \cdot t\_3 + t\_1\right)}}\\
t_8 := t\_7 \cdot t\_5\\
\mathbf{if}\;dX.v \leq 0.14000000059604645:\\
\;\;\;\;\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\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7 \cdot t\_0\\
\end{array}\\
\mathbf{elif}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor + t\_1\right)}} \cdot t\_0\\
\end{array}
\end{array}
if dX.v < 0.140000001Initial program 75.7%
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.f3263.7
Applied rewrites63.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3263.7
Applied rewrites63.7%
lift-*.f32N/A
pow2N/A
lower-pow.f3263.7
lift-*.f32N/A
*-commutativeN/A
lift-*.f3263.7
Applied rewrites63.7%
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.f3268.4
Applied rewrites68.4%
if 0.140000001 < dX.v Initial program 61.7%
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.f3257.2
Applied rewrites57.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3257.2
Applied rewrites57.2%
lift-*.f32N/A
pow2N/A
lower-pow.f3257.2
lift-*.f32N/A
*-commutativeN/A
lift-*.f3257.2
Applied rewrites57.2%
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-*.f3257.2
Applied rewrites57.2%
Final simplification64.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.u (floor w)))
(t_2 (* t_1 t_1))
(t_3 (pow (floor h) 2.0))
(t_4 (* dY.v (floor h)))
(t_5 (* (floor w) dX.u))
(t_6 (+ (* t_0 t_0) (* t_5 t_5))))
(if (>= (* (* t_3 dX.v) dX.v) (+ (pow t_4 2.0) (pow t_1 2.0)))
(* (/ 1.0 (sqrt (fmax t_6 (+ (* t_4 t_4) t_2)))) t_5)
(* (/ 1.0 (sqrt (fmax t_6 (+ (* (* t_3 dY.v) dY.v) 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(h) * dX_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = t_1 * t_1;
float t_3 = powf(floorf(h), 2.0f);
float t_4 = dY_46_v * floorf(h);
float t_5 = floorf(w) * dX_46_u;
float t_6 = (t_0 * t_0) + (t_5 * t_5);
float tmp;
if (((t_3 * dX_46_v) * dX_46_v) >= (powf(t_4, 2.0f) + powf(t_1, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(t_6, ((t_4 * t_4) + t_2)))) * t_5;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_6, (((t_3 * dY_46_v) * dY_46_v) + 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(h) * dX_46_v) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(t_1 * t_1) t_3 = floor(h) ^ Float32(2.0) t_4 = Float32(dY_46_v * floor(h)) t_5 = Float32(floor(w) * dX_46_u) t_6 = Float32(Float32(t_0 * t_0) + Float32(t_5 * t_5)) tmp = Float32(0.0) if (Float32(Float32(t_3 * dX_46_v) * dX_46_v) >= Float32((t_4 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(t_4 * t_4) + t_2) : ((Float32(Float32(t_4 * t_4) + t_2) != Float32(Float32(t_4 * t_4) + t_2)) ? t_6 : max(t_6, Float32(Float32(t_4 * t_4) + t_2)))))) * t_5); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_6 != t_6) ? Float32(Float32(Float32(t_3 * dY_46_v) * dY_46_v) + t_2) : ((Float32(Float32(Float32(t_3 * dY_46_v) * dY_46_v) + t_2) != Float32(Float32(Float32(t_3 * dY_46_v) * dY_46_v) + t_2)) ? t_6 : max(t_6, Float32(Float32(Float32(t_3 * dY_46_v) * dY_46_v) + 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(h) * dX_46_v; t_1 = dY_46_u * floor(w); t_2 = t_1 * t_1; t_3 = floor(h) ^ single(2.0); t_4 = dY_46_v * floor(h); t_5 = floor(w) * dX_46_u; t_6 = (t_0 * t_0) + (t_5 * t_5); tmp = single(0.0); if (((t_3 * dX_46_v) * dX_46_v) >= ((t_4 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(t_6, ((t_4 * t_4) + t_2)))) * t_5; else tmp = (single(1.0) / sqrt(max(t_6, (((t_3 * dY_46_v) * dY_46_v) + t_2)))) * 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 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_4 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := t\_0 \cdot t\_0 + t\_5 \cdot t\_5\\
\mathbf{if}\;\left(t\_3 \cdot dX.v\right) \cdot dX.v \geq {t\_4}^{2} + {t\_1}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, t\_4 \cdot t\_4 + t\_2\right)}} \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_6, \left(t\_3 \cdot dY.v\right) \cdot dY.v + t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 71.2%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
swap-sqrN/A
unpow2N/A
lift-pow.f32N/A
associate-*r*N/A
*-commutativeN/A
lift-*.f32N/A
lift-*.f3261.6
lift-*.f32N/A
*-commutativeN/A
lower-*.f3261.6
Applied rewrites61.6%
Final simplification61.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor h) dX.v))
(t_1 (* dY.u (floor w)))
(t_2 (* t_1 t_1))
(t_3 (* dY.v (floor h)))
(t_4 (* (floor w) dX.u))
(t_5 (+ (* t_0 t_0) (* t_4 t_4))))
(if (>=
(* (* (pow (floor h) 2.0) dX.v) dX.v)
(+ (pow t_3 2.0) (pow t_1 2.0)))
(* (/ 1.0 (sqrt (fmax t_5 (+ (* t_3 t_3) t_2)))) t_4)
(*
(/
1.0
(sqrt (fmax t_5 (+ (* (* (* dY.v dY.v) (floor h)) (floor h)) 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(h) * dX_46_v;
float t_1 = dY_46_u * floorf(w);
float t_2 = t_1 * t_1;
float t_3 = dY_46_v * floorf(h);
float t_4 = floorf(w) * dX_46_u;
float t_5 = (t_0 * t_0) + (t_4 * t_4);
float tmp;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (powf(t_3, 2.0f) + powf(t_1, 2.0f))) {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((t_3 * t_3) + t_2)))) * t_4;
} else {
tmp = (1.0f / sqrtf(fmaxf(t_5, ((((dY_46_v * dY_46_v) * floorf(h)) * floorf(h)) + 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(h) * dX_46_v) t_1 = Float32(dY_46_u * floor(w)) t_2 = Float32(t_1 * t_1) t_3 = Float32(dY_46_v * floor(h)) t_4 = Float32(floor(w) * dX_46_u) t_5 = Float32(Float32(t_0 * t_0) + Float32(t_4 * t_4)) tmp = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32((t_3 ^ Float32(2.0)) + (t_1 ^ Float32(2.0)))) tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(t_3 * t_3) + t_2) : ((Float32(Float32(t_3 * t_3) + t_2) != Float32(Float32(t_3 * t_3) + t_2)) ? t_5 : max(t_5, Float32(Float32(t_3 * t_3) + t_2)))))) * t_4); else tmp = Float32(Float32(Float32(1.0) / sqrt(((t_5 != t_5) ? Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2) : ((Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2) != Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2)) ? t_5 : max(t_5, Float32(Float32(Float32(Float32(dY_46_v * dY_46_v) * floor(h)) * floor(h)) + 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(h) * dX_46_v; t_1 = dY_46_u * floor(w); t_2 = t_1 * t_1; t_3 = dY_46_v * floor(h); t_4 = floor(w) * dX_46_u; t_5 = (t_0 * t_0) + (t_4 * t_4); tmp = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= ((t_3 ^ single(2.0)) + (t_1 ^ single(2.0)))) tmp = (single(1.0) / sqrt(max(t_5, ((t_3 * t_3) + t_2)))) * t_4; else tmp = (single(1.0) / sqrt(max(t_5, ((((dY_46_v * dY_46_v) * floor(h)) * floor(h)) + t_2)))) * 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 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_2 := t\_1 \cdot t\_1\\
t_3 := dY.v \cdot \left\lfloor h\right\rfloor \\
t_4 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_5 := t\_0 \cdot t\_0 + t\_4 \cdot t\_4\\
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq {t\_3}^{2} + {t\_1}^{2}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, t\_3 \cdot t\_3 + t\_2\right)}} \cdot t\_4\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_5, \left(\left(dY.v \cdot dY.v\right) \cdot \left\lfloor h\right\rfloor \right) \cdot \left\lfloor h\right\rfloor + t\_2\right)}} \cdot t\_1\\
\end{array}
\end{array}
Initial program 71.2%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
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-*.f3261.6
Applied rewrites61.6%
Final simplification61.6%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* dY.v (floor h)))
(t_2 (* dY.u (floor w)))
(t_3 (* (floor h) dX.v))
(t_4
(/
1.0
(sqrt
(fmax (+ (* t_3 t_3) (* t_0 t_0)) (+ (* t_1 t_1) (* t_2 t_2)))))))
(if (>= (pow t_3 2.0) (+ (pow t_1 2.0) (pow t_2 2.0)))
(* t_4 t_0)
(* 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_v * floorf(h);
float t_2 = dY_46_u * floorf(w);
float t_3 = floorf(h) * dX_46_v;
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(t_3, 2.0f) >= (powf(t_1, 2.0f) + powf(t_2, 2.0f))) {
tmp = t_4 * t_0;
} 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_v * floor(h)) t_2 = Float32(dY_46_u * floor(w)) t_3 = Float32(floor(h) * dX_46_v) 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_1 * t_1) + Float32(t_2 * t_2)) : ((Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2)) != Float32(Float32(t_1 * t_1) + Float32(t_2 * t_2))) ? 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_1 * t_1) + Float32(t_2 * t_2))))))) 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_0); 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_v * floor(h); t_2 = dY_46_u * floor(w); t_3 = floor(h) * dX_46_v; 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 ((t_3 ^ single(2.0)) >= ((t_1 ^ single(2.0)) + (t_2 ^ single(2.0)))) tmp = t_4 * t_0; 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.v \cdot \left\lfloor h\right\rfloor \\
t_2 := dY.u \cdot \left\lfloor w\right\rfloor \\
t_3 := \left\lfloor h\right\rfloor \cdot dX.v\\
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}\;{t\_3}^{2} \geq {t\_1}^{2} + {t\_2}^{2}:\\
\;\;\;\;t\_4 \cdot t\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\end{array}
\end{array}
Initial program 71.2%
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.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
Applied rewrites61.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3261.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3261.6
Applied rewrites61.6%
Taylor expanded in dX.u around 0
unpow2N/A
unpow2N/A
unswap-sqrN/A
unpow2N/A
lower-pow.f32N/A
lower-*.f32N/A
lower-floor.f3261.6
Applied rewrites61.6%
Final simplification61.6%
herbie shell --seed 2024271
(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))))