
(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 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(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 (floor w) 2.0))
(t_1 (pow (floor h) 2.0))
(t_2 (* dX.v (floor h)))
(t_3
(sqrt
(fmax
(+ (pow t_2 2.0) (pow (* dX.u (floor w)) 2.0))
(+ (pow (* dY.v (floor h)) 2.0) (pow (* dY.u (floor w)) 2.0))))))
(if (>=
(fma (* t_1 dX.v) dX.v (* (* t_0 dX.u) dX.u))
(fma (* t_1 dY.v) dY.v (* (* t_0 dY.u) dY.u)))
(/ (* t_2 (- -1.0)) t_3)
(* (/ 1.0 t_3) (* (floor h) dY.v)))))
float code(float w, float h, float dX_46_u, float dX_46_v, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = powf(floorf(w), 2.0f);
float t_1 = powf(floorf(h), 2.0f);
float t_2 = dX_46_v * floorf(h);
float t_3 = sqrtf(fmaxf((powf(t_2, 2.0f) + powf((dX_46_u * floorf(w)), 2.0f)), (powf((dY_46_v * floorf(h)), 2.0f) + powf((dY_46_u * floorf(w)), 2.0f))));
float tmp;
if (fmaf((t_1 * dX_46_v), dX_46_v, ((t_0 * dX_46_u) * dX_46_u)) >= fmaf((t_1 * dY_46_v), dY_46_v, ((t_0 * dY_46_u) * dY_46_u))) {
tmp = (t_2 * -(-1.0f)) / t_3;
} else {
tmp = (1.0f / t_3) * (floorf(h) * dY_46_v);
}
return tmp;
}
function code(w, h, dX_46_u, dX_46_v, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) ^ Float32(2.0) t_1 = floor(h) ^ Float32(2.0) t_2 = Float32(dX_46_v * floor(h)) t_3 = sqrt(fmax(Float32((t_2 ^ Float32(2.0)) + (Float32(dX_46_u * floor(w)) ^ Float32(2.0))), Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) + (Float32(dY_46_u * floor(w)) ^ Float32(2.0))))) tmp = Float32(0.0) if (fma(Float32(t_1 * dX_46_v), dX_46_v, Float32(Float32(t_0 * dX_46_u) * dX_46_u)) >= fma(Float32(t_1 * dY_46_v), dY_46_v, Float32(Float32(t_0 * dY_46_u) * dY_46_u))) tmp = Float32(Float32(t_2 * Float32(-Float32(-1.0))) / t_3); else tmp = Float32(Float32(Float32(1.0) / t_3) * Float32(floor(h) * dY_46_v)); end return tmp end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := {\left(\left\lfloor w\right\rfloor \right)}^{2}\\
t_1 := {\left(\left\lfloor h\right\rfloor \right)}^{2}\\
t_2 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_3 := \sqrt{\mathsf{max}\left({t\_2}^{2} + {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}, {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} + {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\right)}\\
\mathbf{if}\;\mathsf{fma}\left(t\_1 \cdot dX.v, dX.v, \left(t\_0 \cdot dX.u\right) \cdot dX.u\right) \geq \mathsf{fma}\left(t\_1 \cdot dY.v, dY.v, \left(t\_0 \cdot dY.u\right) \cdot dY.u\right):\\
\;\;\;\;\frac{t\_2 \cdot \left(--1\right)}{t\_3}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_3} \cdot \left(\left\lfloor h\right\rfloor \cdot dY.v\right)\\
\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%
lift-sqrt.f32N/A
pow1/2N/A
Applied rewrites76.8%
Taylor expanded in w around 0
lower->=.f32N/A
Applied rewrites76.8%
Final simplification76.8%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dY.v))
(t_2 (* t_1 t_1))
(t_3 (* dX.v (floor h)))
(t_4 (pow (* dY.u (floor w)) 2.0))
(t_5 (* (floor h) dX.v))
(t_6 (* t_5 t_5))
(t_7 (pow (* dX.u (floor w)) 2.0))
(t_8 (* (floor w) dY.u))
(t_9 (+ (* t_8 t_8) t_2))
(t_10 (pow (* dY.v (floor h)) 2.0))
(t_11 (sqrt (fmax (+ (pow t_3 2.0) t_7) (+ t_10 t_4)))))
(if (<= dX.u 600000000.0)
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) (+ t_4 t_2))
(* (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_6) t_9))) t_5)
(*
(/
1.0
(sqrt (fmax (+ (* (floor w) (* (floor w) (* dX.u dX.u))) t_6) t_9)))
t_1))
(if (>= (+ (pow t_5 2.0) t_7) (- t_10 t_4))
(/ (* t_3 (- -1.0)) t_11)
(* (/ 1.0 t_11) 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) * dX_46_u;
float t_1 = floorf(h) * dY_46_v;
float t_2 = t_1 * t_1;
float t_3 = dX_46_v * floorf(h);
float t_4 = powf((dY_46_u * floorf(w)), 2.0f);
float t_5 = floorf(h) * dX_46_v;
float t_6 = t_5 * t_5;
float t_7 = powf((dX_46_u * floorf(w)), 2.0f);
float t_8 = floorf(w) * dY_46_u;
float t_9 = (t_8 * t_8) + t_2;
float t_10 = powf((dY_46_v * floorf(h)), 2.0f);
float t_11 = sqrtf(fmaxf((powf(t_3, 2.0f) + t_7), (t_10 + t_4)));
float tmp_1;
if (dX_46_u <= 600000000.0f) {
float tmp_2;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (t_4 + t_2)) {
tmp_2 = (1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_6), t_9))) * t_5;
} else {
tmp_2 = (1.0f / sqrtf(fmaxf(((floorf(w) * (floorf(w) * (dX_46_u * dX_46_u))) + t_6), t_9))) * t_1;
}
tmp_1 = tmp_2;
} else if ((powf(t_5, 2.0f) + t_7) >= (t_10 - t_4)) {
tmp_1 = (t_3 * -(-1.0f)) / t_11;
} else {
tmp_1 = (1.0f / t_11) * t_1;
}
return tmp_1;
}
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(floor(h) * dY_46_v) t_2 = Float32(t_1 * t_1) t_3 = Float32(dX_46_v * floor(h)) t_4 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_5 = Float32(floor(h) * dX_46_v) t_6 = Float32(t_5 * t_5) t_7 = Float32(dX_46_u * floor(w)) ^ Float32(2.0) t_8 = Float32(floor(w) * dY_46_u) t_9 = Float32(Float32(t_8 * t_8) + t_2) t_10 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_11 = sqrt(fmax(Float32((t_3 ^ Float32(2.0)) + t_7), Float32(t_10 + t_4))) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(600000000.0)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32(t_4 + t_2)) tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_0 * t_0) + t_6), t_9))) * t_5); else tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_6), t_9))) * t_1); end tmp_1 = tmp_2; elseif (Float32((t_5 ^ Float32(2.0)) + t_7) >= Float32(t_10 - t_4)) tmp_1 = Float32(Float32(t_3 * Float32(-Float32(-1.0))) / t_11); else tmp_1 = Float32(Float32(Float32(1.0) / t_11) * t_1); 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(w) * dX_46_u; t_1 = floor(h) * dY_46_v; t_2 = t_1 * t_1; t_3 = dX_46_v * floor(h); t_4 = (dY_46_u * floor(w)) ^ single(2.0); t_5 = floor(h) * dX_46_v; t_6 = t_5 * t_5; t_7 = (dX_46_u * floor(w)) ^ single(2.0); t_8 = floor(w) * dY_46_u; t_9 = (t_8 * t_8) + t_2; t_10 = (dY_46_v * floor(h)) ^ single(2.0); t_11 = sqrt(max(((t_3 ^ single(2.0)) + t_7), (t_10 + t_4))); tmp_2 = single(0.0); if (dX_46_u <= single(600000000.0)) tmp_3 = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (t_4 + t_2)) tmp_3 = (single(1.0) / sqrt(max(((t_0 * t_0) + t_6), t_9))) * t_5; else tmp_3 = (single(1.0) / sqrt(max(((floor(w) * (floor(w) * (dX_46_u * dX_46_u))) + t_6), t_9))) * t_1; end tmp_2 = tmp_3; elseif (((t_5 ^ single(2.0)) + t_7) >= (t_10 - t_4)) tmp_2 = (t_3 * -single(-1.0)) / t_11; else tmp_2 = (single(1.0) / t_11) * t_1; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_2 := t\_1 \cdot t\_1\\
t_3 := dX.v \cdot \left\lfloor h\right\rfloor \\
t_4 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_5 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_6 := t\_5 \cdot t\_5\\
t_7 := {\left(dX.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_8 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_9 := t\_8 \cdot t\_8 + t\_2\\
t_10 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_11 := \sqrt{\mathsf{max}\left({t\_3}^{2} + t\_7, t\_10 + t\_4\right)}\\
\mathbf{if}\;dX.u \leq 600000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_4 + t\_2:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_6, t\_9\right)}} \cdot t\_5\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot dX.u\right)\right) + t\_6, t\_9\right)}} \cdot t\_1\\
\end{array}\\
\mathbf{elif}\;{t\_5}^{2} + t\_7 \geq t\_10 - t\_4:\\
\;\;\;\;\frac{t\_3 \cdot \left(--1\right)}{t\_11}\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{t\_11} \cdot t\_1\\
\end{array}
\end{array}
if dX.u < 6e8Initial program 78.9%
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.f3267.6
Applied rewrites67.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3267.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.6
Applied rewrites67.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3267.6
Applied rewrites67.6%
if 6e8 < dX.u Initial program 63.5%
lift-*.f32N/A
*-commutativeN/A
lift-/.f32N/A
frac-2negN/A
metadata-evalN/A
Applied rewrites63.8%
lift-sqrt.f32N/A
pow1/2N/A
Applied rewrites63.8%
Taylor expanded in w around 0
lower->=.f32N/A
Applied rewrites63.8%
Applied rewrites63.8%
Final simplification67.1%
(FPCore (w h dX.u dX.v dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dX.u))
(t_1 (* (floor h) dX.v))
(t_2 (* (floor w) dY.u))
(t_3 (* (floor h) dY.v))
(t_4 (* t_3 t_3))
(t_5 (+ (* t_2 t_2) t_4))
(t_6 (pow (* dY.u (floor w)) 2.0))
(t_7 (* t_1 t_1))
(t_8 (/ 1.0 (sqrt (fmax (+ (* t_0 t_0) t_7) t_5))))
(t_9 (* t_8 t_1)))
(if (<= dX.u 600000000.0)
(if (>= (* (* (pow (floor h) 2.0) dX.v) dX.v) (+ t_6 t_4))
t_9
(*
(/
1.0
(sqrt (fmax (+ (* (floor w) (* (floor w) (* dX.u dX.u))) t_7) t_5)))
t_3))
(if (>= (pow t_1 2.0) (- (pow (* dY.v (floor h)) 2.0) t_6))
t_9
(* t_8 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 = floorf(h) * dX_46_v;
float t_2 = floorf(w) * dY_46_u;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_3 * t_3;
float t_5 = (t_2 * t_2) + t_4;
float t_6 = powf((dY_46_u * floorf(w)), 2.0f);
float t_7 = t_1 * t_1;
float t_8 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + t_7), t_5));
float t_9 = t_8 * t_1;
float tmp_1;
if (dX_46_u <= 600000000.0f) {
float tmp_2;
if (((powf(floorf(h), 2.0f) * dX_46_v) * dX_46_v) >= (t_6 + t_4)) {
tmp_2 = t_9;
} else {
tmp_2 = (1.0f / sqrtf(fmaxf(((floorf(w) * (floorf(w) * (dX_46_u * dX_46_u))) + t_7), t_5))) * t_3;
}
tmp_1 = tmp_2;
} else if (powf(t_1, 2.0f) >= (powf((dY_46_v * floorf(h)), 2.0f) - t_6)) {
tmp_1 = t_9;
} else {
tmp_1 = t_8 * 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(w) * dX_46_u) t_1 = Float32(floor(h) * dX_46_v) t_2 = Float32(floor(w) * dY_46_u) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_3 * t_3) t_5 = Float32(Float32(t_2 * t_2) + t_4) t_6 = Float32(dY_46_u * floor(w)) ^ Float32(2.0) t_7 = Float32(t_1 * t_1) t_8 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_0 * t_0) + t_7), t_5))) t_9 = Float32(t_8 * t_1) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(600000000.0)) tmp_2 = Float32(0.0) if (Float32(Float32((floor(h) ^ Float32(2.0)) * dX_46_v) * dX_46_v) >= Float32(t_6 + t_4)) tmp_2 = t_9; else tmp_2 = Float32(Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(floor(w) * Float32(floor(w) * Float32(dX_46_u * dX_46_u))) + t_7), t_5))) * t_3); end tmp_1 = tmp_2; elseif ((t_1 ^ Float32(2.0)) >= Float32((Float32(dY_46_v * floor(h)) ^ Float32(2.0)) - t_6)) tmp_1 = t_9; else tmp_1 = Float32(t_8 * 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(w) * dX_46_u; t_1 = floor(h) * dX_46_v; t_2 = floor(w) * dY_46_u; t_3 = floor(h) * dY_46_v; t_4 = t_3 * t_3; t_5 = (t_2 * t_2) + t_4; t_6 = (dY_46_u * floor(w)) ^ single(2.0); t_7 = t_1 * t_1; t_8 = single(1.0) / sqrt(max(((t_0 * t_0) + t_7), t_5)); t_9 = t_8 * t_1; tmp_2 = single(0.0); if (dX_46_u <= single(600000000.0)) tmp_3 = single(0.0); if ((((floor(h) ^ single(2.0)) * dX_46_v) * dX_46_v) >= (t_6 + t_4)) tmp_3 = t_9; else tmp_3 = (single(1.0) / sqrt(max(((floor(w) * (floor(w) * (dX_46_u * dX_46_u))) + t_7), t_5))) * t_3; end tmp_2 = tmp_3; elseif ((t_1 ^ single(2.0)) >= (((dY_46_v * floor(h)) ^ single(2.0)) - t_6)) tmp_2 = t_9; else tmp_2 = t_8 * t_3; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_2 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_3 \cdot t\_3\\
t_5 := t\_2 \cdot t\_2 + t\_4\\
t_6 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_7 := t\_1 \cdot t\_1\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_7, t\_5\right)}}\\
t_9 := t\_8 \cdot t\_1\\
\mathbf{if}\;dX.u \leq 600000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;\left({\left(\left\lfloor h\right\rfloor \right)}^{2} \cdot dX.v\right) \cdot dX.v \geq t\_6 + t\_4:\\
\;\;\;\;t\_9\\
\mathbf{else}:\\
\;\;\;\;\frac{1}{\sqrt{\mathsf{max}\left(\left\lfloor w\right\rfloor \cdot \left(\left\lfloor w\right\rfloor \cdot \left(dX.u \cdot dX.u\right)\right) + t\_7, t\_5\right)}} \cdot t\_3\\
\end{array}\\
\mathbf{elif}\;{t\_1}^{2} \geq {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2} - t\_6:\\
\;\;\;\;t\_9\\
\mathbf{else}:\\
\;\;\;\;t\_8 \cdot t\_3\\
\end{array}
\end{array}
if dX.u < 6e8Initial program 78.9%
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.f3267.6
Applied rewrites67.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3267.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.6
Applied rewrites67.6%
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
unpow-prod-downN/A
pow2N/A
associate-*l*N/A
lower-*.f32N/A
lower-*.f32N/A
pow2N/A
lower-*.f3267.6
Applied rewrites67.6%
if 6e8 < dX.u Initial program 63.5%
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.f3232.0
Applied rewrites32.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3232.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3232.0
Applied rewrites32.0%
lift->=.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift->=.f3232.0
Applied rewrites32.0%
Applied rewrites51.0%
(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.u (floor w)) 2.0))
(t_2 (* (floor h) dX.v))
(t_3 (* (floor w) dY.u))
(t_4 (pow (* dY.v (floor h)) 2.0))
(t_5 (* (floor h) dY.v))
(t_6
(/
1.0
(sqrt
(fmax (+ (* t_0 t_0) (* t_2 t_2)) (+ (* t_3 t_3) (* t_5 t_5))))))
(t_7 (* t_6 t_5))
(t_8 (* t_6 t_2)))
(if (<= dX.u 600000000.0)
(if (>= (pow (* dX.v (floor h)) 2.0) (+ t_1 t_4)) t_8 t_7)
(if (>= (pow t_2 2.0) (- t_4 t_1)) t_8 t_7))))
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_u * floorf(w)), 2.0f);
float t_2 = floorf(h) * dX_46_v;
float t_3 = floorf(w) * dY_46_u;
float t_4 = powf((dY_46_v * floorf(h)), 2.0f);
float t_5 = floorf(h) * dY_46_v;
float t_6 = 1.0f / sqrtf(fmaxf(((t_0 * t_0) + (t_2 * t_2)), ((t_3 * t_3) + (t_5 * t_5))));
float t_7 = t_6 * t_5;
float t_8 = t_6 * t_2;
float tmp_1;
if (dX_46_u <= 600000000.0f) {
float tmp_2;
if (powf((dX_46_v * floorf(h)), 2.0f) >= (t_1 + t_4)) {
tmp_2 = t_8;
} else {
tmp_2 = t_7;
}
tmp_1 = tmp_2;
} else if (powf(t_2, 2.0f) >= (t_4 - t_1)) {
tmp_1 = t_8;
} else {
tmp_1 = t_7;
}
return tmp_1;
}
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)) ^ Float32(2.0) t_2 = Float32(floor(h) * dX_46_v) t_3 = Float32(floor(w) * dY_46_u) t_4 = Float32(dY_46_v * floor(h)) ^ Float32(2.0) t_5 = Float32(floor(h) * dY_46_v) t_6 = Float32(Float32(1.0) / sqrt(fmax(Float32(Float32(t_0 * t_0) + Float32(t_2 * t_2)), Float32(Float32(t_3 * t_3) + Float32(t_5 * t_5))))) t_7 = Float32(t_6 * t_5) t_8 = Float32(t_6 * t_2) tmp_1 = Float32(0.0) if (dX_46_u <= Float32(600000000.0)) tmp_2 = Float32(0.0) if ((Float32(dX_46_v * floor(h)) ^ Float32(2.0)) >= Float32(t_1 + t_4)) tmp_2 = t_8; else tmp_2 = t_7; end tmp_1 = tmp_2; elseif ((t_2 ^ Float32(2.0)) >= Float32(t_4 - t_1)) tmp_1 = t_8; else tmp_1 = t_7; 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(w) * dX_46_u; t_1 = (dY_46_u * floor(w)) ^ single(2.0); t_2 = floor(h) * dX_46_v; t_3 = floor(w) * dY_46_u; t_4 = (dY_46_v * floor(h)) ^ single(2.0); t_5 = floor(h) * dY_46_v; t_6 = single(1.0) / sqrt(max(((t_0 * t_0) + (t_2 * t_2)), ((t_3 * t_3) + (t_5 * t_5)))); t_7 = t_6 * t_5; t_8 = t_6 * t_2; tmp_2 = single(0.0); if (dX_46_u <= single(600000000.0)) tmp_3 = single(0.0); if (((dX_46_v * floor(h)) ^ single(2.0)) >= (t_1 + t_4)) tmp_3 = t_8; else tmp_3 = t_7; end tmp_2 = tmp_3; elseif ((t_2 ^ single(2.0)) >= (t_4 - t_1)) tmp_2 = t_8; else tmp_2 = t_7; end tmp_4 = tmp_2; end
\begin{array}{l}
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_1 := {\left(dY.u \cdot \left\lfloor w\right\rfloor \right)}^{2}\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\\
t_3 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_4 := {\left(dY.v \cdot \left\lfloor h\right\rfloor \right)}^{2}\\
t_5 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_6 := \frac{1}{\sqrt{\mathsf{max}\left(t\_0 \cdot t\_0 + t\_2 \cdot t\_2, t\_3 \cdot t\_3 + t\_5 \cdot t\_5\right)}}\\
t_7 := t\_6 \cdot t\_5\\
t_8 := t\_6 \cdot t\_2\\
\mathbf{if}\;dX.u \leq 600000000:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;{\left(dX.v \cdot \left\lfloor h\right\rfloor \right)}^{2} \geq t\_1 + t\_4:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}\\
\mathbf{elif}\;{t\_2}^{2} \geq t\_4 - t\_1:\\
\;\;\;\;t\_8\\
\mathbf{else}:\\
\;\;\;\;t\_7\\
\end{array}
\end{array}
if dX.u < 6e8Initial program 78.9%
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.f3267.6
Applied rewrites67.6%
lift-*.f32N/A
pow2N/A
lower-pow.f3267.6
lift-*.f32N/A
*-commutativeN/A
lift-*.f3267.6
Applied rewrites67.6%
lift->=.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift->=.f3267.6
Applied rewrites67.6%
if 6e8 < dX.u Initial program 63.5%
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.f3232.0
Applied rewrites32.0%
lift-*.f32N/A
pow2N/A
lower-pow.f3232.0
lift-*.f32N/A
*-commutativeN/A
lift-*.f3232.0
Applied rewrites32.0%
lift->=.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift->=.f3232.0
Applied rewrites32.0%
Applied rewrites51.0%
(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 (* (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_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(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_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(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_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(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_0; else tmp = t_4 * 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 := \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\_0\\
\mathbf{else}:\\
\;\;\;\;t\_4 \cdot t\_2\\
\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.f3262.5
Applied rewrites62.5%
lift-*.f32N/A
pow2N/A
lower-pow.f3262.5
lift-*.f32N/A
*-commutativeN/A
lift-*.f3262.5
Applied rewrites62.5%
lift->=.f32N/A
lift-*.f32N/A
*-commutativeN/A
lift-*.f32N/A
lower-pow.f32N/A
pow2N/A
lift-*.f32N/A
lift->=.f3262.5
Applied rewrites62.5%
herbie shell --seed 2024352
(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))))