Use specialized memcpy wrapper to avoid missing a sizeof(short). Also get rid off floats around the innerloop. Use precalc table.

1 files changed, 18 insertions, 10 deletions

diff --git a/timestretch.c b/timestretch.c
index 63ba059..7f9bcd6 100644
--- a/timestretch.c
+++ b/timestretch.c
@@ -10,6 +10,7 @@
 
 // global values
 static short * g_overlap_buffer;
+static short * g_overlap_heuristic;
 static size_t  g_overlap;
 
 static size_t  g_input_length;  // Minimal length 
@@ -25,8 +26,11 @@ static size_t  g_firstframe;    // If this is set, the caller should provide ini
 #define CORR_WINDOW        15   // overlapping correlation window length (music default = 28 ms)
 #define OUTPUT_LEN         30   // one processing sequence length in milliseconds (music default = 82 ms)
 
+#define sampcpy(A,B,C) memcpy((A),(B),(C)*sizeof(short))
+
 // Returns the length of one output frame
 static size_t calc_convert_values( int sample_rate, float tempo ) {
+    unsigned int i;
     g_overlap = ( sample_rate * OVERLAP ) / 1000;
 
     free( g_overlap_buffer );
@@ -38,6 +42,11 @@ static size_t calc_convert_values( int sample_rate, float tempo ) {
         g_output_length = g_overlap;
     g_corr_length = ( sample_rate * CORR_WINDOW ) / 1000;
 
+    free( g_overlap_heuristic );
+    g_overlap_heuristic = malloc( sizeof(short) * g_corr_length );
+    for( i = 0; i < g_corr_length; ++i )
+        g_overlap_heuristic[i] = (short)16384.0*(-(double)i*(double)i/((double)g_corr_length*(double)g_corr_length)+(double)i/(double)g_corr_length+0.75f);
+
     g_skip = (size_t)( tempo * (float)g_output_length * 65536.0 );
     g_input_length = g_corr_length + g_output_length + g_overlap;
     if( g_skip / 65536 > g_input_length )
@@ -62,22 +71,21 @@ static unsigned int find_corr_max(const short *input, const short *mixbuf) {
 
     // Scans for the best correlation value by testing each possible position
     for (i = 0; i < g_corr_length; ++i) {
-        double acc, i_f = (double)i, sl_f = (double)g_corr_length;
-        double heur = 0.75 - ( i_f * i_f ) / ( sl_f * sl_f ) + i_f / sl_f;
-        int64_t acc_i;
+        int64_t acc = 0;
 
-        for(j = 0, acc_i = 0; j < g_overlap; ++j )
-            acc_i += input[i+j] * mixbuf[j];
+        for(j = 0; j < g_overlap; ++j )
+            acc += input[i+j] * mixbuf[j];
 
         // boost middle of sequence by top of a flat parabola
         // slope stolen from soundtouch
-        acc = (double)acc_i * heur;
+        acc *= g_overlap_heuristic[i];
         if ( corr_max < acc ) {
             offs = i;
             corr_max = acc;
         }
     }
-// printf( "%03d %014.0lf\n", best_offs, best_corr );
+
+printf( "%03d %016llX\n", offs, (int64_t)corr_max );
     return offs;
 }
 
@@ -89,7 +97,7 @@ size_t process_frame( short *input, short *output, short *overlap ) {
     // The first frame needs to be copied verbatim,
     // we do not have anything to mix, yet.
     if( g_firstframe ) {
-        memcpy( output, input, g_output_length * sizeof(short) );
+        sampcpy( output, input, g_output_length );
         g_firstframe = 0;
     } else {
         offset = find_corr_max( input, overlap );
@@ -99,11 +107,11 @@ size_t process_frame( short *input, short *output, short *overlap ) {
             output[i] = ( i_ * overlap[i] + i * input[i+offset] ) / (int)g_overlap;
 
         // Copy rest of the input verbatim
-        memcpy( output + g_overlap, input + offset + g_overlap, ( g_output_length - g_overlap ) * sizeof(short) );
+        sampcpy( output + g_overlap, input + offset + g_overlap, g_output_length - g_overlap );
     }
 
     // Remember end of this frame for next frame
-    memcpy( overlap, input + offset + g_output_length, g_overlap * sizeof(short) );
+    sampcpy( overlap, input + offset + g_output_length, g_overlap );
 
     // Remove the processed samples from the input buffer.
     g_offset &= 0xffff;