On Fri, 21 Jan 2022, Andreas Rheinhardt wrote:

Martin Storsjö:
Also trim off delay samples at the start instead of adjusting pts
to compensate for them; this avoids unwanted offsets if working
with raw samples without considering their pts.
---
 libavcodec/libfdk-aacdec.c | 80 +++++++++++++++++++++++++++++++-------
 1 file changed, 65 insertions(+), 15 deletions(-)

diff --git a/libavcodec/libfdk-aacdec.c b/libavcodec/libfdk-aacdec.c
index 93b52023b0..d560e313ca 100644
--- a/libavcodec/libfdk-aacdec.c
+++ b/libavcodec/libfdk-aacdec.c
@@ -58,7 +58,11 @@ typedef struct FDKAACDecContext {
     int drc_cut;
     int album_mode;
     int level_limit;
-    int output_delay;
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+    int output_delay_set;
+    int flush_samples;
+    int delay_samples;
+#endif
 } FDKAACDecContext;


@@ -123,7 +127,12 @@ static int get_stream_info(AVCodecContext *avctx)
     avctx->sample_rate = info->sampleRate;
     avctx->frame_size  = info->frameSize;
 #if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
-    s->output_delay    = info->outputDelay;
+    if (!s->output_delay_set && info->outputDelay) {
+        // Set this only once.
+        s->flush_samples    = info->outputDelay;
+        s->delay_samples    = info->outputDelay;
+        s->output_delay_set = 1;
+    }
 #endif

     for (i = 0; i < info->numChannels; i++) {
@@ -367,14 +376,31 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, 
void *data,
     int ret;
     AAC_DECODER_ERROR err;
     UINT valid = avpkt->size;
+    UINT flags = 0;
+    int input_offset = 0;

-    err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
-    if (err != AAC_DEC_OK) {
-        av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
-        return AVERROR_INVALIDDATA;
+    if (avpkt->size) {
+        err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
+        if (err != AAC_DEC_OK) {
+            av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
+            return AVERROR_INVALIDDATA;
+        }
+    } else {
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+        /* Handle decoder draining */
+        if (s->flush_samples > 0) {
+            flags |= AACDEC_FLUSH;
+        } else {
+            return AVERROR_EOF;
+        }
+#else
+        return AVERROR_EOF;
+#endif
     }

-    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer, 
s->decoder_buffer_size / sizeof(INT_PCM), 0);
+    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer,
+                                 s->decoder_buffer_size / sizeof(INT_PCM),
+                                 flags);
     if (err == AAC_DEC_NOT_ENOUGH_BITS) {
         ret = avpkt->size - valid;
         goto end;
@@ -390,16 +416,36 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, 
void *data,
         goto end;
     frame->nb_samples = avctx->frame_size;

+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+    if (flags & AACDEC_FLUSH) {
+        // Only return the right amount of samples at the end; if calling the
+        // decoder with AACDEC_FLUSH, it will keep returning frames 
indefinitely.
+        frame->nb_samples = FFMIN(s->flush_samples, frame->nb_samples);
+        av_log(s, AV_LOG_DEBUG, "Returning %d/%d delayed samples.\n",
+                                frame->nb_samples, s->flush_samples);
+        s->flush_samples -= frame->nb_samples;
+    } else {
+        // Trim off samples from the start to compensate for extra decoder
+        // delay. We could also just adjust the pts, but this avoids
+        // including the extra samples in the output altogether.
+        if (s->delay_samples) {
+            int drop_samples = FFMIN(s->delay_samples, frame->nb_samples);
+            av_log(s, AV_LOG_DEBUG, "Dropping %d/%d delayed samples.\n",
+                                    drop_samples, s->delay_samples);
+            s->delay_samples  -= drop_samples;
+            frame->nb_samples -= drop_samples;
+            input_offset = drop_samples * avctx->channels;
+            if (frame->nb_samples <= 0)
+                return 0;
+        }
+    }
+#endif
+
     if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
         goto end;

-    if (frame->pts != AV_NOPTS_VALUE)
-        frame->pts -= av_rescale_q(s->output_delay,
-                                   (AVRational){1, avctx->sample_rate},
-                                   avctx->time_base);
-
-    memcpy(frame->extended_data[0], s->decoder_buffer,
-           avctx->channels * avctx->frame_size *
+    memcpy(frame->extended_data[0], s->decoder_buffer + input_offset,
+           avctx->channels * frame->nb_samples *
            av_get_bytes_per_sample(avctx->sample_fmt));

     *got_frame_ptr = 1;
@@ -432,7 +478,11 @@ const AVCodec ff_libfdk_aac_decoder = {
     .decode         = fdk_aac_decode_frame,
     .close          = fdk_aac_decode_close,
     .flush          = fdk_aac_decode_flush,
-    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
+    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+                      | AV_CODEC_CAP_DELAY
+#endif
+    ,
     .priv_class     = &fdk_aac_dec_class,
     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
                       FF_CODEC_CAP_INIT_CLEANUP,


When I use the libfdk-aac decoder I get the exact number of samples like
with the native aac decoder (namely number of frames * 1024, as
expected). What makes you believe this is necessary?

The fdk-aac decoder can have, depending on combination of options, some amount of extra internal delay, that the libavcodec internal aac decoder doesn't have. (It's also possible to set the options in a state where the fdk-aac decoder doesn't induce any extra delay.)

Currently, we compensate for that extra delay by just offsetting pts backwards, so for a stream with N packets, we return samples with timestamps [-delay,N*framesize-delay].

In order not to lose data at the end, we must make the decoder flushable and flush up to (delay) samples at the end. And since one doesn't normally expect extra delay samples at the start of an AAC decoder output, we also trim out the same amount of samples at the start (to simplify for users that don't observe the pts, who otherwise are surprised by the stream starting from pts -delay instead of at pts 0).

// Martin
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".

Reply via email to