
(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 9 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}
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h dX.u dX.v_m 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 w) dX.u))
(t_4 (* t_3 t_3))
(t_5 (pow t_0 2.0))
(t_6 (pow t_1 2.0))
(t_7 (* (floor h) dX.v_m))
(t_8 (- t_5 t_6)))
(if (>= (+ (pow t_3 2.0) (pow t_7 2.0)) t_2)
(* t_3 (/ 1.0 (sqrt (fmax (+ t_4 (* t_7 t_7)) t_2))))
(*
t_0
(/
1.0
(sqrt
(fmax
(+ t_4 (pow (exp 2.0) (log t_7)))
(* (+ t_5 t_6) (* t_8 (/ 1.0 t_8))))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float dX_46_u, float dX_46_v_m, 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(w) * dX_46_u;
float t_4 = t_3 * t_3;
float t_5 = powf(t_0, 2.0f);
float t_6 = powf(t_1, 2.0f);
float t_7 = floorf(h) * dX_46_v_m;
float t_8 = t_5 - t_6;
float tmp;
if ((powf(t_3, 2.0f) + powf(t_7, 2.0f)) >= t_2) {
tmp = t_3 * (1.0f / sqrtf(fmaxf((t_4 + (t_7 * t_7)), t_2)));
} else {
tmp = t_0 * (1.0f / sqrtf(fmaxf((t_4 + powf(expf(2.0f), logf(t_7))), ((t_5 + t_6) * (t_8 * (1.0f / t_8))))));
}
return tmp;
}
dX.v_m = abs(dX_46_v) function code(w, h, dX_46_u, dX_46_v_m, 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(w) * dX_46_u) t_4 = Float32(t_3 * t_3) t_5 = t_0 ^ Float32(2.0) t_6 = t_1 ^ Float32(2.0) t_7 = Float32(floor(h) * dX_46_v_m) t_8 = Float32(t_5 - t_6) tmp = Float32(0.0) if (Float32((t_3 ^ Float32(2.0)) + (t_7 ^ Float32(2.0))) >= t_2) tmp = Float32(t_3 * Float32(Float32(1.0) / sqrt(((Float32(t_4 + Float32(t_7 * t_7)) != Float32(t_4 + Float32(t_7 * t_7))) ? t_2 : ((t_2 != t_2) ? Float32(t_4 + Float32(t_7 * t_7)) : max(Float32(t_4 + Float32(t_7 * t_7)), t_2)))))); else tmp = Float32(t_0 * Float32(Float32(1.0) / sqrt(((Float32(t_4 + (exp(Float32(2.0)) ^ log(t_7))) != Float32(t_4 + (exp(Float32(2.0)) ^ log(t_7)))) ? Float32(Float32(t_5 + t_6) * Float32(t_8 * Float32(Float32(1.0) / t_8))) : ((Float32(Float32(t_5 + t_6) * Float32(t_8 * Float32(Float32(1.0) / t_8))) != Float32(Float32(t_5 + t_6) * Float32(t_8 * Float32(Float32(1.0) / t_8)))) ? Float32(t_4 + (exp(Float32(2.0)) ^ log(t_7))) : max(Float32(t_4 + (exp(Float32(2.0)) ^ log(t_7))), Float32(Float32(t_5 + t_6) * Float32(t_8 * Float32(Float32(1.0) / t_8))))))))); end return tmp end
dX.v_m = abs(dX_46_v); function tmp_2 = code(w, h, dX_46_u, dX_46_v_m, 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(w) * dX_46_u; t_4 = t_3 * t_3; t_5 = t_0 ^ single(2.0); t_6 = t_1 ^ single(2.0); t_7 = floor(h) * dX_46_v_m; t_8 = t_5 - t_6; tmp = single(0.0); if (((t_3 ^ single(2.0)) + (t_7 ^ single(2.0))) >= t_2) tmp = t_3 * (single(1.0) / sqrt(max((t_4 + (t_7 * t_7)), t_2))); else tmp = t_0 * (single(1.0) / sqrt(max((t_4 + (exp(single(2.0)) ^ log(t_7))), ((t_5 + t_6) * (t_8 * (single(1.0) / t_8)))))); end tmp_2 = tmp; end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\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 w\right\rfloor \cdot dX.u\\
t_4 := t\_3 \cdot t\_3\\
t_5 := {t\_0}^{2}\\
t_6 := {t\_1}^{2}\\
t_7 := \left\lfloor h\right\rfloor \cdot dX.v\_m\\
t_8 := t\_5 - t\_6\\
\mathbf{if}\;{t\_3}^{2} + {t\_7}^{2} \geq t\_2:\\
\;\;\;\;t\_3 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + t\_7 \cdot t\_7, t\_2\right)}}\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_4 + {\left(e^{2}\right)}^{\log t\_7}, \left(t\_5 + t\_6\right) \cdot \left(t\_8 \cdot \frac{1}{t\_8}\right)\right)}}\\
\end{array}
\end{array}
Initial program 73.8%
lift-floor.f32N/A
lift-*.f32N/A
*-lft-identityN/A
lift-floor.f32N/A
lift-*.f32N/A
*-lft-identityN/A
*-lft-identityN/A
*-lft-identityN/A
lift-*.f32N/A
lift-floor.f32N/A
lift-floor.f32N/A
lift-*.f32N/A
lift-*.f32N/A
lift-*.f32N/A
Applied egg-rr74.5%
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
pow-to-expN/A
*-commutativeN/A
exp-prodN/A
unpow1N/A
pow-to-expN/A
rem-log-expN/A
lower-pow.f32N/A
lower-exp.f32N/A
rem-log-expN/A
pow-to-expN/A
unpow1N/A
lower-log.f3274.5
Applied egg-rr74.5%
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lift-pow.f3274.5
Applied egg-rr74.5%
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lift-pow.f3274.5
Applied egg-rr74.5%
Final simplification74.5%
dX.v_m = (fabs.f32 dX.v)
(FPCore (w h dX.u dX.v_m dY.u dY.v maxAniso)
:precision binary32
(let* ((t_0 (* (floor w) dY.u))
(t_1 (* t_0 t_0))
(t_2 (* (floor h) dX.v_m))
(t_3 (* (floor h) dY.v))
(t_4 (+ t_1 (* t_3 t_3)))
(t_5 (* (floor w) dX.u))
(t_6 (>= (pow t_5 2.0) (+ (pow t_0 2.0) (pow t_3 2.0))))
(t_7 (+ (* t_5 t_5) (* t_2 t_2)))
(t_8 (/ 1.0 (sqrt (fmax t_7 t_4))))
(t_9 (* t_0 t_8))
(t_10 (* t_5 t_8))
(t_11 (if (>= t_7 t_4) t_10 t_9)))
(if (<= t_11 -0.9999998807907104)
(if t_6 t_10 t_9)
(if (<= t_11 0.00215999991632998)
(if (>= t_7 (* (* dY.v dY.v) (pow (floor h) 2.0))) t_10 t_9)
(if t_6
t_10
(*
t_0
(/
1.0
(sqrt
(fmax
t_7
(+ t_1 (* (floor h) (* (floor h) (* dY.v dY.v)))))))))))))dX.v_m = fabs(dX_46_v);
float code(float w, float h, float dX_46_u, float dX_46_v_m, float dY_46_u, float dY_46_v, float maxAniso) {
float t_0 = floorf(w) * dY_46_u;
float t_1 = t_0 * t_0;
float t_2 = floorf(h) * dX_46_v_m;
float t_3 = floorf(h) * dY_46_v;
float t_4 = t_1 + (t_3 * t_3);
float t_5 = floorf(w) * dX_46_u;
int t_6 = powf(t_5, 2.0f) >= (powf(t_0, 2.0f) + powf(t_3, 2.0f));
float t_7 = (t_5 * t_5) + (t_2 * t_2);
float t_8 = 1.0f / sqrtf(fmaxf(t_7, t_4));
float t_9 = t_0 * t_8;
float t_10 = t_5 * t_8;
float tmp;
if (t_7 >= t_4) {
tmp = t_10;
} else {
tmp = t_9;
}
float t_11 = tmp;
float tmp_2;
if (t_11 <= -0.9999998807907104f) {
float tmp_3;
if (t_6) {
tmp_3 = t_10;
} else {
tmp_3 = t_9;
}
tmp_2 = tmp_3;
} else if (t_11 <= 0.00215999991632998f) {
float tmp_4;
if (t_7 >= ((dY_46_v * dY_46_v) * powf(floorf(h), 2.0f))) {
tmp_4 = t_10;
} else {
tmp_4 = t_9;
}
tmp_2 = tmp_4;
} else if (t_6) {
tmp_2 = t_10;
} else {
tmp_2 = t_0 * (1.0f / sqrtf(fmaxf(t_7, (t_1 + (floorf(h) * (floorf(h) * (dY_46_v * dY_46_v)))))));
}
return tmp_2;
}
dX.v_m = abs(dX_46_v) function code(w, h, dX_46_u, dX_46_v_m, dY_46_u, dY_46_v, maxAniso) t_0 = Float32(floor(w) * dY_46_u) t_1 = Float32(t_0 * t_0) t_2 = Float32(floor(h) * dX_46_v_m) t_3 = Float32(floor(h) * dY_46_v) t_4 = Float32(t_1 + Float32(t_3 * t_3)) t_5 = Float32(floor(w) * dX_46_u) t_6 = (t_5 ^ Float32(2.0)) >= Float32((t_0 ^ Float32(2.0)) + (t_3 ^ Float32(2.0))) t_7 = Float32(Float32(t_5 * t_5) + Float32(t_2 * t_2)) t_8 = Float32(Float32(1.0) / sqrt(((t_7 != t_7) ? t_4 : ((t_4 != t_4) ? t_7 : max(t_7, t_4))))) t_9 = Float32(t_0 * t_8) t_10 = Float32(t_5 * t_8) tmp = Float32(0.0) if (t_7 >= t_4) tmp = t_10; else tmp = t_9; end t_11 = tmp tmp_2 = Float32(0.0) if (t_11 <= Float32(-0.9999998807907104)) tmp_3 = Float32(0.0) if (t_6) tmp_3 = t_10; else tmp_3 = t_9; end tmp_2 = tmp_3; elseif (t_11 <= Float32(0.00215999991632998)) tmp_4 = Float32(0.0) if (t_7 >= Float32(Float32(dY_46_v * dY_46_v) * (floor(h) ^ Float32(2.0)))) tmp_4 = t_10; else tmp_4 = t_9; end tmp_2 = tmp_4; elseif (t_6) tmp_2 = t_10; else tmp_2 = Float32(t_0 * Float32(Float32(1.0) / sqrt(((t_7 != t_7) ? Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) : ((Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))) != Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v))))) ? t_7 : max(t_7, Float32(t_1 + Float32(floor(h) * Float32(floor(h) * Float32(dY_46_v * dY_46_v)))))))))); end return tmp_2 end
dX.v_m = abs(dX_46_v); function tmp_6 = code(w, h, dX_46_u, dX_46_v_m, dY_46_u, dY_46_v, maxAniso) t_0 = floor(w) * dY_46_u; t_1 = t_0 * t_0; t_2 = floor(h) * dX_46_v_m; t_3 = floor(h) * dY_46_v; t_4 = t_1 + (t_3 * t_3); t_5 = floor(w) * dX_46_u; t_6 = (t_5 ^ single(2.0)) >= ((t_0 ^ single(2.0)) + (t_3 ^ single(2.0))); t_7 = (t_5 * t_5) + (t_2 * t_2); t_8 = single(1.0) / sqrt(max(t_7, t_4)); t_9 = t_0 * t_8; t_10 = t_5 * t_8; tmp = single(0.0); if (t_7 >= t_4) tmp = t_10; else tmp = t_9; end t_11 = tmp; tmp_3 = single(0.0); if (t_11 <= single(-0.9999998807907104)) tmp_4 = single(0.0); if (t_6) tmp_4 = t_10; else tmp_4 = t_9; end tmp_3 = tmp_4; elseif (t_11 <= single(0.00215999991632998)) tmp_5 = single(0.0); if (t_7 >= ((dY_46_v * dY_46_v) * (floor(h) ^ single(2.0)))) tmp_5 = t_10; else tmp_5 = t_9; end tmp_3 = tmp_5; elseif (t_6) tmp_3 = t_10; else tmp_3 = t_0 * (single(1.0) / sqrt(max(t_7, (t_1 + (floor(h) * (floor(h) * (dY_46_v * dY_46_v))))))); end tmp_6 = tmp_3; end
\begin{array}{l}
dX.v_m = \left|dX.v\right|
\\
\begin{array}{l}
t_0 := \left\lfloor w\right\rfloor \cdot dY.u\\
t_1 := t\_0 \cdot t\_0\\
t_2 := \left\lfloor h\right\rfloor \cdot dX.v\_m\\
t_3 := \left\lfloor h\right\rfloor \cdot dY.v\\
t_4 := t\_1 + t\_3 \cdot t\_3\\
t_5 := \left\lfloor w\right\rfloor \cdot dX.u\\
t_6 := {t\_5}^{2} \geq {t\_0}^{2} + {t\_3}^{2}\\
t_7 := t\_5 \cdot t\_5 + t\_2 \cdot t\_2\\
t_8 := \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_4\right)}}\\
t_9 := t\_0 \cdot t\_8\\
t_10 := t\_5 \cdot t\_8\\
t_11 := \begin{array}{l}
\mathbf{if}\;t\_7 \geq t\_4:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{if}\;t\_11 \leq -0.9999998807907104:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_6:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{elif}\;t\_11 \leq 0.00215999991632998:\\
\;\;\;\;\begin{array}{l}
\mathbf{if}\;t\_7 \geq \left(dY.v \cdot dY.v\right) \cdot {\left(\left\lfloor h\right\rfloor \right)}^{2}:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_9\\
\end{array}\\
\mathbf{elif}\;t\_6:\\
\;\;\;\;t\_10\\
\mathbf{else}:\\
\;\;\;\;t\_0 \cdot \frac{1}{\sqrt{\mathsf{max}\left(t\_7, t\_1 + \left\lfloor h\right\rfloor \cdot \left(\left\lfloor h\right\rfloor \cdot \left(dY.v \cdot dY.v\right)\right)\right)}}\\
\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 w) dX.u)) (*.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 w) dY.u))) < -0.999999881Initial program 99.6%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3299.6
Simplified99.6%
lift-floor.f32N/A
lift-pow.f32N/A
associate-*r*N/A
pow2N/A
lift-pow.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
Applied egg-rr99.6%
if -0.999999881 < (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 w) dX.u)) (*.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 w) dY.u))) < 0.00215999992Initial program 62.6%
Taylor expanded in dY.u around 0
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f32N/A
unpow2N/A
lower-*.f3262.3
Simplified62.3%
if 0.00215999992 < (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 w) dX.u)) (*.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 w) dY.u))) Initial program 99.3%
Taylor expanded in dX.u around inf
unpow2N/A
associate-*l*N/A
*-commutativeN/A
lower-*.f32N/A
*-commutativeN/A
lower-*.f32N/A
lower-pow.f32N/A
lower-floor.f3298.9
Simplified98.9%
lift-floor.f32N/A
lift-pow.f32N/A
associate-*r*N/A
pow2N/A
lift-pow.f32N/A
unpow-prod-downN/A
*-commutativeN/A
lift-*.f32N/A
pow2N/A
lift-*.f32N/A
Applied egg-rr98.9%
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
lift-floor.f32N/A
lift-*.f32N/A
unpow1N/A
metadata-evalN/A
sqr-powN/A
lift-*.f32N/A
*-commutativeN/A
unpow-prod-downN/A
pow2N/A
lift-*.f32N/A
pow2N/A
associate-*l*N/A
lift-*.f32N/A
lift-*.f3298.9
lift-*.f32N/A
*-commutativeN/A
lower-*.f3298.9
Applied egg-rr98.9%
Final simplification76.0%
herbie shell --seed 2024218
(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))))