]> git.tue.mpg.de Git - paraslash.git/commitdiff
speed up the compress filter
authorAndre <maan@p133.(none)>
Mon, 13 Mar 2006 00:35:25 +0000 (01:35 +0100)
committerAndre <maan@p133.(none)>
Mon, 13 Mar 2006 00:35:25 +0000 (01:35 +0100)
Complete rewrite of the core loop of the compress filter.  The gprof
output of para_audiod (see below) indicates that CPU usage goes down
to about 40%.

new:
~~~~

Each sample counts as 0.01 seconds.
%   cumulative   self              self     total
time   seconds   seconds    calls   s/call   s/call  name

35.94      1.47     1.47     7414     0.00     0.00  compress
22.74      2.40     0.93    57319     0.00     0.00  mp3dec
16.87      3.09     0.69        1     0.69     4.04  audiod_mainloop

34.59      1.47     1.47     7414     0.00     0.00  compress
22.59      2.43     0.96        1     0.96     4.20  audiod_mainloop
17.53      3.17     0.74    61285     0.00     0.00  mp3dec

35.41      1.48     1.48     7414     0.00     0.00  compress
20.10      2.32     0.84        1     0.84     4.15  audiod_mainloop
17.82      3.06     0.74    58675     0.00     0.00  mp3dec

Average of self-seconds (3rd column):
compress 1.47
audiod_mainloop 0.83
mp3dec 0.80

0.2.11:
~~~~~~~

Each sample counts as 0.01 seconds.
%   cumulative   self              self     total
time   seconds   seconds    calls   s/call   s/call  name

56.38      3.31     3.31     7414     0.00     0.00  compress
12.93      4.08     0.76    57566     0.00     0.00  mp3dec
11.39      4.75     0.67        1     0.67     5.86  audiod_mainloop

57.59      3.38     3.38     7414     0.00     0.00  compress
14.08      4.20     0.82    63095     0.00     0.00  mp3dec
11.26      4.86     0.66        1     0.66     5.82  audiod_mainloop

57.79      4.12     4.12     7414     0.00     0.00  compress
13.62      5.08     0.97        1     0.97     7.07  audiod_mainloop
 8.08      5.66     0.57    57196     0.00     0.00  mp3dec

Average of self-seconds (3rd column):
compress  3.60
audiod_mainloop 0.76
mp3dec 0.71

compress.c
compress_filter.ggo

index bad4ada4c6a78f712a01bd8a635ef2328b0a3115..b7ddc28534106eede2ba8b7e6ad1b1d60a3b5510 100644 (file)
@@ -19,7 +19,7 @@
 /** \file compress.c paraslash's dynamic audio range compressor */
 
 /*
- * Based on AudioCompress, (C) 2002-2004  M. Hari Nezumi <magenta@trikuare.cx>
+ * Used ideas of AudioCompress, (C) 2002-2004  M. Hari Nezumi <magenta@trikuare.cx>
  */
 
 #include "para.h"
 #include "filter.h"
 #include "string.h"
 
-/** how fine-grained the gain is */
-#define GAINSHIFT 10
 /** the size of the output data buffer */
 #define COMPRESS_CHUNK_SIZE 40960
 
 /** data specific to the compress filter */
 struct private_compress_data {
-       /** an array holding the previous peak values */
-       int *peaks;
-       /** current bucket number to be modified */
-       unsigned pn;
-       /** number of times clipping occured */
-       unsigned clip;
        /** the current multiplier */
-       int current_gain;
-       /** the target multiplier */
-       int target_gain;
-       /** pointer to the configuration data for this instance of the compress filter */
+       unsigned current_gain;
+       /** points to the configuration data for this instance of the compress filter */
        struct gengetopt_args_info *conf;
+       /** minimal admissible gain */
+       unsigned min_gain;
+       /** maximal admissible gain */
+       unsigned max_gain;
+       /** number of samples already seen */
+       unsigned num_samples;
+       /** absolute value of the maximal sample in the current block */
+       unsigned peak;
 };
 
 static ssize_t compress(char *inbuf, size_t inbuf_len, struct filter_node *fn)
 {
-       int16_t *audio = (int16_t *) inbuf, *ip = audio, *op;
-       int peak = 1, pos = 0, i, gr, gf, gn;
-       size_t length = MIN((inbuf_len / 2) * 2, (fn->bufsize - fn->loaded) / 2 * 2);
+       size_t i, length = MIN((inbuf_len / 2) * 2, (fn->bufsize - fn->loaded) / 2 * 2);
        struct private_compress_data *pcd = fn->private_data;
+       int16_t *ip = (int16_t *)inbuf, *op = (int16_t *)(fn->buf + fn->loaded);
+       unsigned gain_shift = pcd->conf->inertia_arg + pcd->conf->damp_arg,
+               mask = (1 << pcd->conf->blocksize_arg) - 1;
 
        if (!length)
                return 0;
-       /* determine peak's value and position */
-       for (i = 0; i < length / 2; i++, ip++) {
-               int val = ABS(*ip);
-               if (val > peak) {
-                       peak = val;
-                       pos = i;
-               }
-       }
-       pcd->peaks[pcd->pn] = peak;
-       for (i = 0; i < pcd->conf->buckets_arg; i++) {
-               if (pcd->peaks[i] > peak) {
-                       peak = pcd->peaks[i];
-                       pos = 0;
-               }
-       }
-       /* determine target gain */
-       gn = (1 << GAINSHIFT) * pcd->conf->target_level_arg / peak;
-       if (gn < (1 << GAINSHIFT))
-               gn = 1 << GAINSHIFT;
-       pcd->target_gain = (pcd->target_gain * ((1 << pcd->conf->gain_smooth_arg) - 1) + gn)
-               >> pcd->conf->gain_smooth_arg;
-       /* give it an extra insignificant nudge to counteract possible
-        * rounding error
-        */
-       if (gn < pcd->target_gain)
-               pcd->target_gain--;
-       else if (gn > pcd->target_gain)
-               pcd->target_gain++;
-       if (pcd->target_gain > pcd->conf->gain_max_arg << GAINSHIFT)
-               pcd->target_gain = pcd->conf->gain_max_arg << GAINSHIFT;
-       /* see if a peak is going to clip */
-       gn = (1 << GAINSHIFT) * 32768 / peak;
-       if (gn < pcd->target_gain) {
-               pcd->target_gain = gn;
-               if (pcd->conf->anticlip_given)
-                       pos = 0;
-       } else
-               /* we're ramping up, so draw it out over the whole frame */
-               pos = length;
-       /* determine gain rate necessary to make target */
-       if (!pos)
-               pos = 1;
-       gr = ((pcd->target_gain - pcd->current_gain) << 16) / pos;
-       gf = pcd->current_gain << 16;
-       ip = audio;
-       op = (int16_t *)(fn->buf + fn->loaded);
        for (i = 0; i < length / 2; i++) {
-               int sample;
-               /* interpolate the gain */
-               pcd->current_gain = gf >> 16;
-               if (i < pos)
-                       gf += gr;
-               else if (i == pos)
-                       gf = pcd->target_gain << 16;
-               /* amplify */
-               sample = (*ip++) * pcd->current_gain * pcd->conf->volume_arg / 10 >> GAINSHIFT;
-               if (sample < -32768) {
-                       pcd->clip++;
-                       sample = -32768;
-               } else if (sample > 32767) {
-                       pcd->clip++;
-                       sample = 32767;
+               /* be careful in that heat, my dear */
+               int sample = *ip++, adjusted_sample;
+
+               if (sample > 0) {
+                       adjusted_sample = (sample * pcd->current_gain)
+                               >> gain_shift;
+                       if (unlikely(adjusted_sample > 32767)) {
+                               adjusted_sample = 32767;
+                               pcd->current_gain = (pcd->current_gain +
+                                       (1 << pcd->conf->inertia_arg)) / 2;
+                               pcd->peak = 0;
+                       } else
+                               if (adjusted_sample > pcd->peak)
+                                       pcd->peak = sample;
+               } else {
+                       adjusted_sample = -((-sample * pcd->current_gain)
+                               >> gain_shift);
+                       if (unlikely(adjusted_sample < -32768)) {
+                               adjusted_sample = -32768;
+                               pcd->current_gain = (pcd->current_gain +
+                                       (1 << pcd->conf->inertia_arg)) / 2;
+                               pcd->peak = 0;
+                       } else
+                               if (-adjusted_sample > pcd->peak)
+                                       pcd->peak = -adjusted_sample;
+               }
+               *op++ = adjusted_sample;
+               if (likely(++pcd->num_samples & mask))
+                       continue;
+               if (pcd->peak < pcd->conf->target_level_arg) {
+                       if (pcd->current_gain < pcd->max_gain)
+                               pcd->current_gain++;
+               } else {
+                       if (pcd->current_gain > pcd->min_gain + 1)
+                               pcd->current_gain -= 2;
                }
-               *op++ = sample;
+//             PARA_DEBUG_LOG("gain: %lu, peak: %d\n", pcd->current_gain,
+//                     pcd->peak);
+               pcd->peak = 0;
+//             PARA_INFO_LOG("sample: %lu\n", ABS(sample));
        }
-       pcd->pn = (pcd->pn + 1) % pcd->conf->buckets_arg;
-       PARA_DEBUG_LOG("bucket: %03i, input len: %zd, length: %zd, peak: %05i, "
-               "current gain: %03i, clipped: %d\n", pcd->pn, inbuf_len,
-               length, peak, pcd->current_gain, pcd->clip);
-       fn->loaded = length;
+       fn->loaded += length;
        return length;
 }
 
 static void close_compress(struct filter_node *fn)
 {
-       struct private_compress_data *pcd = fn->private_data;
-       free(pcd->peaks);
        free(fn->private_data);
        free(fn->buf);
 }
@@ -152,13 +122,13 @@ static void open_compress(struct filter_node *fn)
 {
        struct private_compress_data *pcd = para_calloc(
                sizeof(struct private_compress_data));
-//     compress_cmdline_parser(fn->argc, fn->argv, &pcd->conf);
        pcd->conf = fn->conf;
-       pcd->peaks = para_calloc(pcd->conf->buckets_arg * sizeof(int));
        fn->private_data = pcd;
        fn->bufsize = COMPRESS_CHUNK_SIZE;
        fn->buf = para_malloc(fn->bufsize);
-       fn->loaded = 0;
+       pcd->current_gain = 1 << pcd->conf->inertia_arg;
+       pcd->min_gain = 1 << (pcd->conf->inertia_arg - pcd->conf->aggressiveness_arg);
+       pcd->max_gain = 1 << (pcd->conf->inertia_arg + pcd->conf->aggressiveness_arg);
 }
 
 /** the init function of the compress filter */
index 6a6d76306b179cdf12b107299a32b0fb304c1945..f4ba7a83044ca9447c58f1e05b1cb6e93f925b21 100644 (file)
@@ -1,7 +1,7 @@
 section "The dynamic audio range compressor"
-option "anticlip" c "enable clipping protection" flag on
-option "buckets" b "history length" int typestr="number" default="400" no
-option "target_level" t "target signal level (1-32767)" int typestr="number" default="25000" no
-option "gain_max" g "maximum amount to amplify by" int typestr="number" default="4" no
-option "gain_smooth" i "how much inertia ramping has" int typestr="number" default="5" no
-option "volume" v "set soft volume (0-10)" int typestr="number" default="10" no
+
+option "blocksize" b "larger blocksize means fewer volume adjustments per time unit" int typestr="number" default="16" no
+option "aggressiveness" a "controls the maximum amount to amplify by" int typestr="number" default="2" no
+option "inertia" i "how much inertia ramping has" int typestr="number" default="6" no
+option "target_level" t "target signal level (0-32768)" int typestr="number" default="25000" no
+option "damp" d "if non-zero, scale down after normalizing" int typestr="number" default="0" no