Makes analysis work when encoding more than 20 ms at a time
[opus.git] / src / analysis.c
index 547e5a4..53247df 100644 (file)
@@ -327,6 +327,10 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
        {
           float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
                      + out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
+#ifdef FIXED_POINT
+          /* FIXME: It's probably best to change the BFCC filter initial state instead */
+          binE *= 5.55e-17f;
+#endif
           E += binE;
           tE += binE*tonality[i];
           nE += binE*2.f*(.5f-noisiness[i]);
@@ -343,7 +347,7 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
           tonal->highE[b]+=.5f;
           tonal->lowE[b]-=.5f;
        }
-       relativeE += (logE[b]-tonal->lowE[b])/(EPSILON+tonal->highE[b]-tonal->lowE[b]);
+       relativeE += (logE[b]-tonal->lowE[b])/(1e-15+tonal->highE[b]-tonal->lowE[b]);
 
        L1=L2=0;
        for (i=0;i<NB_FRAMES;i++)
@@ -352,12 +356,12 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
           L2 += tonal->E[i][b];
        }
 
-       stationarity = MIN16(0.99f,L1/sqrt(EPSILON+NB_FRAMES*L2));
+       stationarity = MIN16(0.99f,L1/sqrt(1e-15+NB_FRAMES*L2));
        stationarity *= stationarity;
        stationarity *= stationarity;
        frame_stationarity += stationarity;
        /*band_tonality[b] = tE/(1e-15+E)*/;
-       band_tonality[b] = MAX16(tE/(EPSILON+E), stationarity*tonal->prev_band_tonality[b]);
+       band_tonality[b] = MAX16(tE/(1e-15+E), stationarity*tonal->prev_band_tonality[b]);
 #if 0
        if (b>=NB_TONAL_SKIP_BANDS)
        {
@@ -479,7 +483,7 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
     features[23] = info->tonality_slope;
     features[24] = tonal->lowECount;
 
-#ifndef FIXED_POINT
+#ifndef DISABLE_FLOAT_API
     mlp_process(&net, features, frame_probs);
     frame_probs[0] = .5f*(frame_probs[0]+1);
     /* Curve fitting between the MLP probability and the actual probability */
@@ -614,44 +618,30 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, con
     RESTORE_STACK;
 }
 
-int run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, const opus_val16 *pcm,
-                        const void *analysis_pcm, int frame_size, int variable_duration, int c1, int c2, int C, opus_int32 Fs, int bitrate_bps,
-                        int delay_compensation, int lsb_depth, downmix_func downmix, AnalysisInfo *analysis_info)
+void run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, const void *analysis_pcm,
+                 int analysis_frame_size, int frame_size, int c1, int c2, int C, opus_int32 Fs,
+                 int lsb_depth, downmix_func downmix, AnalysisInfo *analysis_info)
 {
    int offset;
    int pcm_len;
 
-   /* Avoid overflow/wrap-around of the analysis buffer */
-   frame_size = IMIN((DETECT_SIZE-5)*Fs/100, frame_size);
-
-   pcm_len = frame_size - analysis->analysis_offset;
-   offset = analysis->analysis_offset;
-   do {
-      tonality_analysis(analysis, NULL, celt_mode, analysis_pcm, IMIN(480, pcm_len), offset, c1, c2, C, lsb_depth, downmix);
-      offset += 480;
-      pcm_len -= 480;
-   } while (pcm_len>0);
-   analysis->analysis_offset = frame_size;
-
-   if (variable_duration == OPUS_FRAMESIZE_VARIABLE && frame_size >= Fs/200)
+   if (analysis_pcm != NULL)
    {
-      int LM = 3;
-      LM = optimize_framesize(pcm, frame_size, C, Fs, bitrate_bps,
-            analysis->prev_tonality, analysis->subframe_mem, delay_compensation, downmix);
-      while ((Fs/400<<LM)>frame_size)
-         LM--;
-      frame_size = (Fs/400<<LM);
-   } else {
-      frame_size = frame_size_select(frame_size, variable_duration, Fs);
+      /* Avoid overflow/wrap-around of the analysis buffer */
+      analysis_frame_size = IMIN((DETECT_SIZE-5)*Fs/100, analysis_frame_size);
+
+      pcm_len = analysis_frame_size - analysis->analysis_offset;
+      offset = analysis->analysis_offset;
+      do {
+         tonality_analysis(analysis, NULL, celt_mode, analysis_pcm, IMIN(480, pcm_len), offset, c1, c2, C, lsb_depth, downmix);
+         offset += 480;
+         pcm_len -= 480;
+      } while (pcm_len>0);
+      analysis->analysis_offset = analysis_frame_size;
+
+      analysis->analysis_offset -= frame_size;
    }
-   if (frame_size<0)
-      return -1;
-   analysis->analysis_offset -= frame_size;
 
-   /* Only perform analysis up to 20-ms frames. Longer ones will be split if
-      they're in CELT-only mode. */
    analysis_info->valid = 0;
    tonality_get_info(analysis, analysis_info, frame_size);
-
-   return frame_size;
 }