2 This file is part of PulseAudio.
4 Copyright 2007 Lennart Poettering
6 PulseAudio is free software; you can redistribute it and/or modify
7 it under the terms of the GNU Lesser General Public License as
8 published by the Free Software Foundation; either version 2.1 of the
9 License, or (at your option) any later version.
11 PulseAudio is distributed in the hope that it will be useful, but
12 WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with PulseAudio; if not, see <http://www.gnu.org/licenses/>.
27 #include <pulse/sample.h>
28 #include <pulse/xmalloc.h>
30 #include <pulsecore/macro.h>
32 #include "time-smoother.h"
34 #define HISTORY_MAX 64
37 * Implementation of a time smoothing algorithm to synchronize remote
38 * clocks to a local one. Evens out noise, adjusts to clock skew and
39 * allows cheap estimations of the remote time while clock updates may
40 * be seldom and received in non-equidistant intervals.
42 * Basically, we estimate the gradient of received clock samples in a
43 * certain history window (of size 'history_time') with linear
44 * regression. With that info we estimate the remote time in
45 * 'adjust_time' ahead and smoothen our current estimation function
46 * towards that point with a 3rd order polynomial interpolation with
47 * fitting derivatives. (more or less a b-spline)
49 * The larger 'history_time' is chosen the better we will suppress
50 * noise -- but we'll adjust to clock skew slower..
52 * The larger 'adjust_time' is chosen the smoother our estimation
53 * function will be -- but we'll adjust to clock skew slower, too.
55 * If 'monotonic' is true the resulting estimation function is
56 * guaranteed to be monotonic.
60 pa_usec_t adjust_time, history_time;
62 pa_usec_t time_offset;
64 pa_usec_t px, py; /* Point p, where we want to reach stability */
65 double dp; /* Gradient we want at point p */
67 pa_usec_t ex, ey; /* Point e, which we estimated before and need to smooth to */
68 double de; /* Gradient we estimated for point e */
69 pa_usec_t ry; /* The original y value for ex */
71 /* History of last measurements */
72 pa_usec_t history_x[HISTORY_MAX], history_y[HISTORY_MAX];
73 unsigned history_idx, n_history;
75 /* To even out for monotonicity */
76 pa_usec_t last_y, last_x;
78 /* Cached parameters for our interpolation polynomial y=ax^3+b^2+cx */
84 bool smoothing:1; /* If false we skip the polynomial interpolation step */
91 pa_smoother* pa_smoother_new(
92 pa_usec_t adjust_time,
93 pa_usec_t history_time,
97 pa_usec_t time_offset,
102 pa_assert(adjust_time > 0);
103 pa_assert(history_time > 0);
104 pa_assert(min_history >= 2);
105 pa_assert(min_history <= HISTORY_MAX);
107 s = pa_xnew(pa_smoother, 1);
108 s->adjust_time = adjust_time;
109 s->history_time = history_time;
110 s->min_history = min_history;
111 s->monotonic = monotonic;
112 s->smoothing = smoothing;
114 pa_smoother_reset(s, time_offset, paused);
119 void pa_smoother_free(pa_smoother* s) {
127 x = (x) % HISTORY_MAX; \
130 #define REDUCE_INC(x) \
132 x = ((x)+1) % HISTORY_MAX; \
135 static void drop_old(pa_smoother *s, pa_usec_t x) {
137 /* Drop items from history which are too old, but make sure to
138 * always keep min_history in the history */
140 while (s->n_history > s->min_history) {
142 if (s->history_x[s->history_idx] + s->history_time >= x)
143 /* This item is still valid, and thus all following ones
144 * are too, so let's quit this loop */
147 /* Item is too old, let's drop it */
148 REDUCE_INC(s->history_idx);
154 static void add_to_history(pa_smoother *s, pa_usec_t x, pa_usec_t y) {
158 /* First try to update an existing history entry */
160 for (j = s->n_history; j > 0; j--) {
162 if (s->history_x[i] == x) {
170 /* Drop old entries */
173 /* Calculate position for new entry */
174 j = s->history_idx + s->n_history;
184 /* And make sure we don't store more entries than fit in */
185 if (s->n_history > HISTORY_MAX) {
186 s->history_idx += s->n_history - HISTORY_MAX;
187 REDUCE(s->history_idx);
188 s->n_history = HISTORY_MAX;
192 static double avg_gradient(pa_smoother *s, pa_usec_t x) {
193 unsigned i, j, c = 0;
194 int64_t ax = 0, ay = 0, k, t;
197 /* FIXME: Optimization: Jason Newton suggested that instead of
198 * going through the history on each iteration we could calculated
199 * avg_gradient() as we go.
201 * Second idea: it might make sense to weight history entries:
202 * more recent entries should matter more than old ones. */
204 /* Too few measurements, assume gradient of 1 */
205 if (s->n_history < s->min_history)
208 /* First, calculate average of all measurements */
210 for (j = s->n_history; j > 0; j--) {
212 ax += (int64_t) s->history_x[i];
213 ay += (int64_t) s->history_y[i];
219 pa_assert(c >= s->min_history);
223 /* Now, do linear regression */
227 for (j = s->n_history; j > 0; j--) {
230 dx = (int64_t) s->history_x[i] - ax;
231 dy = (int64_t) s->history_y[i] - ay;
239 r = (double) k / (double) t;
241 return (s->monotonic && r < 0) ? 0 : r;
244 static void calc_abc(pa_smoother *s) {
245 pa_usec_t ex, ey, px, py;
254 /* We have two points: (ex|ey) and (px|py) with two gradients at
255 * these points de and dp. We do a polynomial
256 * interpolation of degree 3 with these 6 values */
258 ex = s->ex; ey = s->ey;
259 px = s->px; py = s->py;
260 de = s->de; dp = s->dp;
264 /* To increase the dynamic range and simplify calculation, we
265 * move these values to the origin */
266 kx = (int64_t) px - (int64_t) ex;
267 ky = (int64_t) py - (int64_t) ey;
269 /* Calculate a, b, c for y=ax^3+bx^2+cx */
271 s->b = (((double) (3*ky)/ (double) kx - dp - (double) (2*de))) / (double) kx;
272 s->a = (dp/(double) kx - 2*s->b - de/(double) kx) / (double) (3*kx);
277 static void estimate(pa_smoother *s, pa_usec_t x, pa_usec_t *y, double *deriv) {
282 /* Linear interpolation right from px */
285 /* The requested point is right of the point where we wanted
286 * to be on track again, thus just linearly estimate */
288 t = (int64_t) s->py + (int64_t) llrint(s->dp * (double) (x - s->px));
298 } else if (x <= s->ex) {
299 /* Linear interpolation left from ex */
302 t = (int64_t) s->ey - (int64_t) llrint(s->de * (double) (s->ex - x));
313 /* Spline interpolation between ex and px */
316 /* Ok, we're not yet on track, thus let's interpolate, and
317 * make sure that the first derivative is smooth */
322 tx = (double) (x - s->ex);
325 ty = (tx * (s->c + tx * (s->b + tx * s->a)));
327 /* Move back from origin */
328 ty += (double) s->ey;
330 *y = ty >= 0 ? (pa_usec_t) llrint(ty) : 0;
334 *deriv = s->c + (tx * (s->b*2 + tx * s->a*3));
337 /* Guarantee monotonicity */
340 if (deriv && *deriv < 0)
345 void pa_smoother_put(pa_smoother *s, pa_usec_t x, pa_usec_t y) {
356 x = PA_LIKELY(x >= s->time_offset) ? x - s->time_offset : 0;
361 /* First, we calculate the position we'd estimate for x, so that
362 * we can adjust our position smoothly from this one */
363 estimate(s, x, &ney, &nde);
364 s->ex = x; s->ey = ney; s->de = nde;
368 /* Then, we add the new measurement to our history */
369 add_to_history(s, x, y);
371 /* And determine the average gradient of the history */
372 s->dp = avg_gradient(s, x);
374 /* And calculate when we want to be on track again */
376 s->px = s->ex + s->adjust_time;
377 s->py = s->ry + (pa_usec_t) llrint(s->dp * (double) s->adjust_time);
383 s->abc_valid = false;
386 pa_log_debug("%p, put(%llu | %llu) = %llu", s, (unsigned long long) (x + s->time_offset), (unsigned long long) x, (unsigned long long) y);
390 pa_usec_t pa_smoother_get(pa_smoother *s, pa_usec_t x) {
399 x = PA_LIKELY(x >= s->time_offset) ? x - s->time_offset : 0;
405 estimate(s, x, &y, NULL);
409 /* Make sure the querier doesn't jump forth and back. */
419 pa_log_debug("%p, get(%llu | %llu) = %llu", s, (unsigned long long) (x + s->time_offset), (unsigned long long) x, (unsigned long long) y);
425 void pa_smoother_set_time_offset(pa_smoother *s, pa_usec_t offset) {
428 s->time_offset = offset;
431 pa_log_debug("offset(%llu)", (unsigned long long) offset);
435 void pa_smoother_pause(pa_smoother *s, pa_usec_t x) {
442 pa_log_debug("pause(%llu)", (unsigned long long) x);
449 void pa_smoother_resume(pa_smoother *s, pa_usec_t x, bool fix_now) {
455 if (x < s->pause_time)
459 pa_log_debug("resume(%llu)", (unsigned long long) x);
463 s->time_offset += x - s->pause_time;
466 pa_smoother_fix_now(s);
469 void pa_smoother_fix_now(pa_smoother *s) {
476 pa_usec_t pa_smoother_translate(pa_smoother *s, pa_usec_t x, pa_usec_t y_delay) {
486 x = PA_LIKELY(x >= s->time_offset) ? x - s->time_offset : 0;
488 estimate(s, x, &ney, &nde);
490 /* Play safe and take the larger gradient, so that we wakeup
491 * earlier when this is used for sleeping */
496 pa_log_debug("translate(%llu) = %llu (%0.2f)", (unsigned long long) y_delay, (unsigned long long) ((double) y_delay / nde), nde);
499 return (pa_usec_t) llrint((double) y_delay / nde);
502 void pa_smoother_reset(pa_smoother *s, pa_usec_t time_offset, bool paused) {
508 s->ex = s->ey = s->ry = 0;
514 s->last_y = s->last_x = 0;
516 s->abc_valid = false;
519 s->time_offset = s->pause_time = time_offset;
522 pa_log_debug("reset()");