[FFmpeg-devel] [PATCH 1/3] avcodec/vorbisenc: Separate copying audio samples from windowing

Tyler Jones tdjones879 at gmail.com
Wed Jun 14 23:58:22 EEST 2017


Audio samples are shifted around when copying from the frame queue so that
analysis can be done without negatively impacting calculation of the MDCT.

Window coefficients are applied to the current two overlapped windows
simultaneously instead of applying overlap for the next frame ahead of time.
This improves readability when applying windows of varying lengths.

Signed-off-by: Tyler Jones <tdjones879 at gmail.com>
---
 libavcodec/vorbisenc.c | 76 +++++++++++++++++++++-----------------------------
 1 file changed, 32 insertions(+), 44 deletions(-)

diff --git a/libavcodec/vorbisenc.c b/libavcodec/vorbisenc.c
index afded40..9b66d56 100644
--- a/libavcodec/vorbisenc.c
+++ b/libavcodec/vorbisenc.c
@@ -453,7 +453,7 @@ static int create_vorbis_context(vorbis_enc_context *venc,
     venc->samples    = av_malloc_array(sizeof(float) * venc->channels, (1 << venc->log2_blocksize[1]));
     venc->floor      = av_malloc_array(sizeof(float) * venc->channels, (1 << venc->log2_blocksize[1]) / 2);
     venc->coeffs     = av_malloc_array(sizeof(float) * venc->channels, (1 << venc->log2_blocksize[1]) / 2);
-    venc->scratch    = av_malloc_array(sizeof(float) * venc->channels, (1 << venc->log2_blocksize[1]) / 2);
+    venc->scratch    = av_malloc_array(sizeof(float) * venc->channels, (1 << venc->log2_blocksize[1]));
 
     if (!venc->saved || !venc->samples || !venc->floor || !venc->coeffs || !venc->scratch)
         return AVERROR(ENOMEM);
@@ -994,8 +994,7 @@ static int residue_encode(vorbis_enc_context *venc, vorbis_enc_residue *rc,
     return 0;
 }
 
-static int apply_window_and_mdct(vorbis_enc_context *venc,
-                                 float *audio, int samples)
+static int apply_window_and_mdct(vorbis_enc_context *venc, int samples)
 {
     int channel;
     const float * win = venc->win[0];
@@ -1003,46 +1002,19 @@ static int apply_window_and_mdct(vorbis_enc_context *venc,
     float n = (float)(1 << venc->log2_blocksize[0]) / 4.0;
     AVFloatDSPContext *fdsp = venc->fdsp;
 
-    if (!venc->have_saved && !samples)
-        return 0;
+    for (channel = 0; channel < venc->channels; channel++) {
+        float *offset = venc->samples + channel * window_len * 2;
 
-    if (venc->have_saved) {
-        for (channel = 0; channel < venc->channels; channel++)
-            memcpy(venc->samples + channel * window_len * 2,
-                   venc->saved + channel * window_len, sizeof(float) * window_len);
-    } else {
-        for (channel = 0; channel < venc->channels; channel++)
-            memset(venc->samples + channel * window_len * 2, 0,
-                   sizeof(float) * window_len);
-    }
+        fdsp->vector_fmul(offset, offset, win, samples);
+        fdsp->vector_fmul_scalar(offset, offset, 1/n, samples);
 
-    if (samples) {
-        for (channel = 0; channel < venc->channels; channel++) {
-            float *offset = venc->samples + channel * window_len * 2 + window_len;
+        offset += window_len;
 
-            fdsp->vector_fmul_reverse(offset, audio + channel * window_len, win, samples);
-            fdsp->vector_fmul_scalar(offset, offset, 1/n, samples);
-        }
-    } else {
-        for (channel = 0; channel < venc->channels; channel++)
-            memset(venc->samples + channel * window_len * 2 + window_len,
-                   0, sizeof(float) * window_len);
-    }
+        fdsp->vector_fmul_reverse(offset, offset, win, samples);
+        fdsp->vector_fmul_scalar(offset, offset, 1/n, samples);
 
-    for (channel = 0; channel < venc->channels; channel++)
         venc->mdct[0].mdct_calc(&venc->mdct[0], venc->coeffs + channel * window_len,
                      venc->samples + channel * window_len * 2);
-
-    if (samples) {
-        for (channel = 0; channel < venc->channels; channel++) {
-            float *offset = venc->saved + channel * window_len;
-
-            fdsp->vector_fmul(offset, audio + channel * window_len, win, samples);
-            fdsp->vector_fmul_scalar(offset, offset, 1/n, samples);
-        }
-        venc->have_saved = 1;
-    } else {
-        venc->have_saved = 0;
     }
     return 1;
 }
@@ -1071,24 +1043,40 @@ static AVFrame *spawn_empty_frame(AVCodecContext *avctx, int channels)
     return f;
 }
 
-/* Concatenate audio frames into an appropriately sized array of samples */
-static void move_audio(vorbis_enc_context *venc, float *audio, int *samples, int sf_size)
+/* Set up audio samples for psy analysis and window/mdct */
+static void move_audio(vorbis_enc_context *venc, int *samples, int sf_size)
 {
     AVFrame *cur = NULL;
     int frame_size = 1 << (venc->log2_blocksize[1] - 1);
     int subframes = frame_size / sf_size;
+    int sf, ch;
 
-    for (int sf = 0; sf < subframes; sf++) {
+    /* Copy samples from last frame into current frame */
+    if (venc->have_saved)
+        for (ch = 0; ch < venc->channels; ch++)
+            memcpy(venc->samples + 2 * ch * frame_size,
+                   venc->saved + ch * frame_size, sizeof(float) * frame_size);
+    else
+        for (ch = 0; ch < venc->channels; ch++)
+            memset(venc->samples + 2 * ch * frame_size, 0, sizeof(float) * frame_size);
+
+    for (sf = 0; sf < subframes; sf++) {
         cur = ff_bufqueue_get(&venc->bufqueue);
         *samples += cur->nb_samples;
 
-        for (int ch = 0; ch < venc->channels; ch++) {
+        for (ch = 0; ch < venc->channels; ch++) {
+            float *offset = venc->samples + 2 * ch * frame_size + frame_size;
+            float *save = venc->saved + ch * frame_size;
             const float *input = (float *) cur->extended_data[ch];
             const size_t len  = cur->nb_samples * sizeof(float);
-            memcpy(audio + ch*frame_size + sf*sf_size, input, len);
+
+            memcpy(offset + sf*sf_size, input, len);
+            memcpy(save + sf*sf_size, input, len);   // Move samples for next frame
         }
         av_frame_free(&cur);
     }
+    venc->have_saved = 1;
+    memcpy(venc->scratch, venc->samples, 2 * venc->channels * frame_size);
 }
 
 static int vorbis_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
@@ -1129,9 +1117,9 @@ static int vorbis_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
         }
     }
 
-    move_audio(venc, venc->scratch, &samples, avctx->frame_size);
+    move_audio(venc, &samples, avctx->frame_size);
 
-    if (!apply_window_and_mdct(venc, venc->scratch, samples))
+    if (!apply_window_and_mdct(venc, samples))
         return 0;
 
     if ((ret = ff_alloc_packet2(avctx, avpkt, 8192, 0)) < 0)
-- 
2.7.4

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 473 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20170614/796f33ce/attachment.sig>


More information about the ffmpeg-devel mailing list