dsp.c: optimize goerztzel sample loops, in dtmf_detect, mf_detect and tone_detect

use a temporary short int when repeatedly used to call goertzel_sample.

alecdavis (license 585)
Reported by: alecdavis
Tested by: alecdavis

Review: https://reviewboard.asterisk.org/r/2093/
........

Merged revisions 372212 from http://svn.asterisk.org/svn/asterisk/branches/1.8

git-svn-id: https://origsvn.digium.com/svn/asterisk/branches/10@372213 65c4cc65-6c06-0410-ace0-fbb531ad65f3

diff --git a/main/dsp.c b/main/dsp.c
index 96a101f0be220a82d66f14aa05935de46f1d23c0..9ba4775adc1fd3d85c797bdecbd524b05b516462 100644 (file)
--- a/main/dsp.c
+++ b/main/dsp.c
@@ -530,6 +530,7 @@ static int tone_detect(struct ast_dsp *dsp, tone_detect_state_t *s, int16_t *amp
        int limit;
        int res = 0;
        int16_t *ptr;
+       short samp;
        int start, end;
        fragment_t mute = {0, 0};
 
@@ -547,10 +548,11 @@ static int tone_detect(struct ast_dsp *dsp, tone_detect_state_t *s, int16_t *amp
                end = start + limit;
 
                for (i = limit, ptr = amp ; i > 0; i--, ptr++) {
+                       samp = *ptr;
                        /* signed 32 bit int should be enough to suqare any possible signed 16 bit value */
-                       s->energy += (int32_t) *ptr * (int32_t) *ptr;
+                       s->energy += (int32_t) samp * (int32_t) samp;
 
-                       goertzel_sample(&s->tone, *ptr);
+                       goertzel_sample(&s->tone, samp);
                }
 
                s->samples_pending -= limit;
@@ -643,10 +645,10 @@ static int dtmf_detect(struct ast_dsp *dsp, digit_detect_state_t *s, int16_t amp
 {
        float row_energy[4];
        float col_energy[4];
-       float famp;
        int i;
        int j;
        int sample;
+       short samp;
        int best_row;
        int best_col;
        int hit;
@@ -669,18 +671,18 @@ static int dtmf_detect(struct ast_dsp *dsp, digit_detect_state_t *s, int16_t amp
                /* The following unrolled loop takes only 35% (rough estimate) of the 
                   time of a rolled loop on the machine on which it was developed */
                for (j = sample; j < limit; j++) {
-                       famp = amp[j];
-                       s->td.dtmf.energy += famp*famp;
+                       samp = amp[j];
+                       s->td.dtmf.energy += (int32_t) samp * (int32_t) samp;
                        /* With GCC 2.95, the following unrolled code seems to take about 35%
                           (rough estimate) as long as a neat little 0-3 loop */
-                       goertzel_sample(s->td.dtmf.row_out, amp[j]);
-                       goertzel_sample(s->td.dtmf.col_out, amp[j]);
-                       goertzel_sample(s->td.dtmf.row_out + 1, amp[j]);
-                       goertzel_sample(s->td.dtmf.col_out + 1, amp[j]);
-                       goertzel_sample(s->td.dtmf.row_out + 2, amp[j]);
-                       goertzel_sample(s->td.dtmf.col_out + 2, amp[j]);
-                       goertzel_sample(s->td.dtmf.row_out + 3, amp[j]);
-                       goertzel_sample(s->td.dtmf.col_out + 3, amp[j]);
+                       goertzel_sample(s->td.dtmf.row_out, samp);
+                       goertzel_sample(s->td.dtmf.col_out, samp);
+                       goertzel_sample(s->td.dtmf.row_out + 1, samp);
+                       goertzel_sample(s->td.dtmf.col_out + 1, samp);
+                       goertzel_sample(s->td.dtmf.row_out + 2, samp);
+                       goertzel_sample(s->td.dtmf.col_out + 2, samp);
+                       goertzel_sample(s->td.dtmf.row_out + 3, samp);
+                       goertzel_sample(s->td.dtmf.col_out + 3, samp);
                }
                s->td.dtmf.current_sample += (limit - sample);
                if (s->td.dtmf.current_sample < DTMF_GSIZE) {
@@ -798,6 +800,7 @@ static int mf_detect(struct ast_dsp *dsp, digit_detect_state_t *s, int16_t amp[]
        int i;
        int j;
        int sample;
+       short samp;
        int hit;
        int limit;
        fragment_t mute = {0, 0};
@@ -821,12 +824,13 @@ static int mf_detect(struct ast_dsp *dsp, digit_detect_state_t *s, int16_t amp[]
                for (j = sample;  j < limit;  j++) {
                        /* With GCC 2.95, the following unrolled code seems to take about 35%
                           (rough estimate) as long as a neat little 0-3 loop */
-                       goertzel_sample(s->td.mf.tone_out, amp[j]);
-                       goertzel_sample(s->td.mf.tone_out + 1, amp[j]);
-                       goertzel_sample(s->td.mf.tone_out + 2, amp[j]);
-                       goertzel_sample(s->td.mf.tone_out + 3, amp[j]);
-                       goertzel_sample(s->td.mf.tone_out + 4, amp[j]);
-                       goertzel_sample(s->td.mf.tone_out + 5, amp[j]);
+                       samp = amp[j];
+                       goertzel_sample(s->td.mf.tone_out, samp);
+                       goertzel_sample(s->td.mf.tone_out + 1, samp);
+                       goertzel_sample(s->td.mf.tone_out + 2, samp);
+                       goertzel_sample(s->td.mf.tone_out + 3, samp);
+                       goertzel_sample(s->td.mf.tone_out + 4, samp);
+                       goertzel_sample(s->td.mf.tone_out + 5, samp);
                }
                s->td.mf.current_sample += (limit - sample);
                if (s->td.mf.current_sample < MF_GSIZE) {