smart padding for better gapless playback

Taken smart padding code using LPC extrapolation from vorbis/opus.
Padding is done on both beginning and ending, but enc_delay and padding
remains the same (we discard extra padding frame introduced on our side
after encoding).

MSVC/fdkaac.vcxproj

@@ -98,6 +98,8 @@ copy ..\fdk-aac\libSYS\include\machine_type.h include\fdk-aac\ </Command>
     <ClCompile Include="..\src\aacenc.c" />
     <ClCompile Include="..\src\caf_reader.c" />
     <ClCompile Include="..\src\compat_win32.c" />
+    <ClCompile Include="..\src\extrapolater.c" />
+    <ClCompile Include="..\src\lpc.c" />
     <ClCompile Include="..\src\lpcm.c" />
     <ClCompile Include="..\src\m4af.c" />
     <ClCompile Include="..\src\main.c" />
@@ -114,6 +116,7 @@ copy ..\fdk-aac\libSYS\include\machine_type.h include\fdk-aac\ </Command>
     <ClInclude Include="..\src\caf_reader.h" />
     <ClInclude Include="..\src\catypes.h" />
     <ClInclude Include="..\src\compat.h" />
+    <ClInclude Include="..\src\lpc.h" />
     <ClInclude Include="..\src\lpcm.h" />
     <ClInclude Include="..\src\m4af.h" />
     <ClInclude Include="..\src\m4af_endian.h" />

MSVC/fdkaac.vcxproj.filters

@@ -24,6 +24,12 @@
     <ClCompile Include="..\src\compat_win32.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\src\extrapolater.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
+    <ClCompile Include="..\src\lpc.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
     <ClCompile Include="..\src\lpcm.c">
       <Filter>Source Files</Filter>
     </ClCompile>
@@ -62,6 +68,9 @@
     <ClInclude Include="..\src\compat.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\src\lpc.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
     <ClInclude Include="..\src\lpcm.h">
       <Filter>Header Files</Filter>
     </ClInclude>

Makefile.am

@@ -6,6 +6,8 @@ bin_PROGRAMS = fdkaac
 fdkaac_SOURCES = \
     src/aacenc.c               \
     src/caf_reader.c           \
+    src/extrapolater.c         \
+    src/lpc.c                  \
     src/lpcm.c                 \
     src/m4af.c                 \
     src/main.c                 \

src/aacenc.c

@@ -231,5 +231,5 @@ int aac_encode_frame(HANDLE_AACENCODER encoder,
         return -1;
     }
     *olen = oargs.numOutBytes;
-    return oargs.numInSamples;
+    return oargs.numInSamples / format->channels_per_frame;
 }

src/extrapolater.c

@@ -0,0 +1,204 @@
+#if HAVE_CONFIG_H
+#  include "config.h"
+#endif
+#if HAVE_STDINT_H
+#  include <stdint.h>
+#endif
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+#include "pcm_reader.h"
+#include "lpc.h"
+
+typedef int16_t sample_t;
+
+typedef struct buffer_t {
+    sample_t *data;
+    unsigned count;    /* count in frames */
+    unsigned capacity; /* size in bytes   */
+} buffer_t;
+
+typedef struct extrapolater_t {
+    pcm_reader_vtbl_t *vtbl;
+    pcm_reader_t *src;
+    pcm_sample_description_t format;
+    buffer_t buffer[2];
+    unsigned nbuffer;
+    int (*process)(struct extrapolater_t *, void *, unsigned);
+} extrapolater_t;
+
+#define LPC_ORDER 32
+
+static inline pcm_reader_t *get_source(pcm_reader_t *reader)
+{
+    return ((extrapolater_t *)reader)->src;
+}
+
+static const
+pcm_sample_description_t *get_format(pcm_reader_t *reader)
+{
+    return pcm_get_format(get_source(reader));
+}
+
+static int64_t get_length(pcm_reader_t *reader)
+{
+    return pcm_get_length(get_source(reader));
+}
+
+static int64_t get_position(pcm_reader_t *reader)
+{
+    return pcm_get_position(get_source(reader));
+}
+
+static int realloc_buffer(buffer_t *bp, size_t size)
+{
+    if (bp->capacity < size) {
+        void *p = realloc(bp->data, size);
+        if (!p) return -1;
+        bp->data = p;
+        bp->capacity = size;
+    }
+    return 0;
+}
+
+static void reverse_buffer(sample_t *data, unsigned nframes, unsigned nchannels)
+{
+    unsigned i = 0, j = nchannels * (nframes - 1), n;
+
+    for (; i < j; i += nchannels, j -= nchannels) {
+        for (n = 0; n < nchannels; ++n) {
+            sample_t tmp = data[i + n];
+            data[i + n] = data[j + n];
+            data[j + n] = tmp;
+        }
+    }
+}
+
+static int fetch(extrapolater_t *self, unsigned nframes)
+{
+    const pcm_sample_description_t *sfmt = pcm_get_format(self->src);
+    buffer_t *bp = &self->buffer[self->nbuffer];
+    int rc = 0;
+
+    if (realloc_buffer(bp, nframes * sfmt->bytes_per_frame) == 0) {
+        rc = pcm_read_frames(self->src, bp->data, nframes);
+        bp->count = rc > 0 ? rc : 0;
+    }
+    if (rc > 0)
+        self->nbuffer ^= 1;
+    return bp->count;
+}
+
+static int extrapolate(extrapolater_t *self, const buffer_t *bp,
+                        void *dst, unsigned nframes)
+{
+    const pcm_sample_description_t *sfmt = pcm_get_format(self->src);
+    unsigned i, n = sfmt->channels_per_frame;
+    float lpc[LPC_ORDER];
+
+    for (i = 0; i < n; ++i) {
+        vorbis_lpc_from_data(bp->data + i, lpc, bp->count, LPC_ORDER, n);
+        vorbis_lpc_predict(lpc, &bp->data[i + n * (bp->count - LPC_ORDER)],
+                           LPC_ORDER, (sample_t*)dst + i, nframes, n);
+    }
+    return nframes;
+}
+
+static int process1(extrapolater_t *self, void *buffer, unsigned nframes);
+static int process2(extrapolater_t *self, void *buffer, unsigned nframes);
+static int process3(extrapolater_t *self, void *buffer, unsigned nframes);
+
+static int process0(extrapolater_t *self, void *buffer, unsigned nframes)
+{
+    const pcm_sample_description_t *sfmt = pcm_get_format(self->src);
+    unsigned nchannels = sfmt->channels_per_frame;
+    buffer_t *bp = &self->buffer[self->nbuffer];
+
+    if (fetch(self, nframes) < 2 * LPC_ORDER)
+        memset(buffer, 0, nframes * sfmt->bytes_per_frame);
+    else {
+        reverse_buffer(bp->data, bp->count, nchannels);
+        extrapolate(self, bp, buffer, nframes);
+        reverse_buffer(buffer, nframes, nchannels);
+        reverse_buffer(bp->data, bp->count, nchannels);
+    }
+    self->process = bp->count ? process1 : process2;
+    return nframes;
+}
+
+static int process1(extrapolater_t *self, void *buffer, unsigned nframes)
+{
+    const pcm_sample_description_t *sfmt = pcm_get_format(self->src);
+    buffer_t *bp = &self->buffer[self->nbuffer ^ 1];
+
+    assert(bp->count <= nframes);
+    memcpy(buffer, bp->data, bp->count * sfmt->bytes_per_frame); 
+    if (!fetch(self, nframes))
+        self->process = process2;
+    return bp->count;
+}
+
+static int process2(extrapolater_t *self, void *buffer, unsigned nframes)
+{
+    const pcm_sample_description_t *sfmt = pcm_get_format(self->src);
+    buffer_t *bp = &self->buffer[self->nbuffer];
+    buffer_t *bbp = &self->buffer[self->nbuffer ^ 1];
+
+    if (bp->count < 2 * LPC_ORDER) {
+        size_t total = bp->count + bbp->count;
+        if (bbp->count &&
+            realloc_buffer(bbp, total * sfmt->bytes_per_frame) == 0)
+        {
+            memcpy(bbp->data + bbp->count * sfmt->channels_per_frame,
+                   bp->data, bp->count * sfmt->bytes_per_frame);
+            bbp->count = total;
+            bp->count = 0;
+            bp = bbp;
+            self->nbuffer ^= 1;
+        }
+    }
+    self->process = process3;
+
+    if (bp->count >= 2 * LPC_ORDER)
+        extrapolate(self, bp, buffer, nframes);
+    else
+        memset(buffer, 0, nframes * sfmt->bytes_per_frame);
+    return nframes;
+}
+
+static int process3(extrapolater_t *self, void *buffer, unsigned nframes)
+{
+    return 0;
+}
+
+static int read_frames(pcm_reader_t *reader, void *buffer, unsigned nframes)
+{
+    extrapolater_t *self = (extrapolater_t *)reader;
+    return self->process(self, buffer, nframes);
+}
+
+static void teardown(pcm_reader_t **reader)
+{
+    extrapolater_t *self = (extrapolater_t *)*reader;
+    pcm_teardown(&self->src);
+    free(self->buffer[0].data);
+    free(self->buffer[1].data);
+    free(self);
+    *reader = 0;
+}
+
+static pcm_reader_vtbl_t my_vtable = {
+    get_format, get_length, get_position, read_frames, teardown
+};
+
+pcm_reader_t *extrapolater_open(pcm_reader_t *reader)
+{
+    extrapolater_t *self = 0;
+
+    if ((self = calloc(1, sizeof(extrapolater_t))) == 0)
+        return 0;
+    self->src = reader;
+    self->vtbl = &my_vtable;
+    self->process = process0;
+    return (pcm_reader_t *)self;
+}

src/lpc.c

@@ -0,0 +1,169 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2009             *
+ * by the Xiph.Org Foundation http://www.xiph.org/                  *
+ *                                                                  *
+ ********************************************************************
+
+  function: LPC low level routines
+  last mod: $Id: lpc.c 16227 2009-07-08 06:58:46Z xiphmont $
+
+ ********************************************************************/
+
+/* Some of these routines (autocorrelator, LPC coefficient estimator)
+   are derived from code written by Jutta Degener and Carsten Bormann;
+   thus we include their copyright below.  The entirety of this file
+   is freely redistributable on the condition that both of these
+   copyright notices are preserved without modification.  */
+
+/* Preserved Copyright: *********************************************/
+
+/* Copyright 1992, 1993, 1994 by Jutta Degener and Carsten Bormann,
+Technische Universita"t Berlin
+
+Any use of this software is permitted provided that this notice is not
+removed and that neither the authors nor the Technische Universita"t
+Berlin are deemed to have made any representations as to the
+suitability of this software for any purpose nor are held responsible
+for any defects of this software. THERE IS ABSOLUTELY NO WARRANTY FOR
+THIS SOFTWARE.
+
+As a matter of courtesy, the authors request to be informed about uses
+this software has found, about bugs in this software, and about any
+improvements that may be of general interest.
+
+Berlin, 28.11.1994
+Jutta Degener
+Carsten Bormann
+
+*********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#if HAVE_STDINT_H
+# include <stdint.h>
+#endif
+
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "lpc.h"
+#include "lpcm.h"
+
+/* Autocorrelation LPC coeff generation algorithm invented by
+   N. Levinson in 1947, modified by J. Durbin in 1959. */
+
+/* Input : n elements of time doamin data
+   Output: m lpc coefficients, excitation energy */
+
+float vorbis_lpc_from_data(short *data,float *lpci,int n,int m,int stride){
+  double *aut=malloc(sizeof(*aut)*(m+1));
+  double *lpc=malloc(sizeof(*lpc)*(m));
+  double error;
+  double epsilon;
+  int i,j;
+
+  /* autocorrelation, p+1 lag coefficients */
+  j=m+1;
+  while(j--){
+    double d=0; /* double needed for accumulator depth */
+    for(i=j;i<n;i++)d+=(double)data[i*stride]*data[(i-j)*stride]/1073741824.0;
+    aut[j]=d;
+  }
+
+  /* Generate lpc coefficients from autocorr values */
+
+  /* set our noise floor to about -100dB */
+  error=aut[0] * (1. + 1e-10);
+  epsilon=1e-9*aut[0]+1e-10;
+
+  for(i=0;i<m;i++){
+    double r= -aut[i+1];
+
+    if(error<epsilon){
+      memset(lpc+i,0,(m-i)*sizeof(*lpc));
+      goto done;
+    }
+
+    /* Sum up this iteration's reflection coefficient; note that in
+       Vorbis we don't save it.  If anyone wants to recycle this code
+       and needs reflection coefficients, save the results of 'r' from
+       each iteration. */
+
+    for(j=0;j<i;j++)r-=lpc[j]*aut[i-j];
+    r/=error;
+
+    /* Update LPC coefficients and total error */
+
+    lpc[i]=r;
+    for(j=0;j<i/2;j++){
+      double tmp=lpc[j];
+
+      lpc[j]+=r*lpc[i-1-j];
+      lpc[i-1-j]+=r*tmp;
+    }
+    if(i&1)lpc[j]+=lpc[j]*r;
+
+    error*=1.-r*r;
+
+  }
+
+ done:
+
+  /* slightly damp the filter */
+  {
+    double g = .99;
+    double damp = g;
+    for(j=0;j<m;j++){
+      lpc[j]*=damp;
+      damp*=g;
+    }
+  }
+
+  for(j=0;j<m;j++)lpci[j]=(float)lpc[j];
+
+  /* we need the error value to know how big an impulse to hit the
+     filter with later */
+
+  free(aut);
+  free(lpc);
+  return error;
+}
+
+void vorbis_lpc_predict(float *coeff,short *prime,int m,
+                        short *data,long n,int stride){
+
+  /* in: coeff[0...m-1] LPC coefficients
+         prime[0...m-1] initial values (allocated size of n+m-1)
+    out: data[0...n-1] data samples */
+
+  long i,j,o,p;
+  float y;
+  float *work=malloc(sizeof(*work)*(m+n));
+
+  if(!prime)
+    for(i=0;i<m;i++)
+      work[i]=0.f;
+  else
+    for(i=0;i<m;i++)
+      work[i]=prime[i*stride]/32768.0f;
+
+  for(i=0;i<n;i++){
+    y=0;
+    o=i;
+    p=m;
+    for(j=0;j<m;j++)
+      y-=work[o++]*coeff[--p];
+
+    work[o]=y;
+    data[i*stride]=lrint(pcm_clip(y*32768.0,-32768.0,32767.0));
+  }
+  free(work);
+}

src/lpc.h

@@ -0,0 +1,27 @@
+/********************************************************************
+ *                                                                  *
+ * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE.   *
+ * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
+ * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
+ * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
+ *                                                                  *
+ * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2007             *
+ * by the Xiph.Org Foundation http://www.xiph.org/                  *
+ *                                                                  *
+ ********************************************************************
+
+  function: LPC low level routines
+  last mod: $Id: lpc.h 16037 2009-05-26 21:10:58Z xiphmont $
+
+ ********************************************************************/
+
+#ifndef _V_LPC_H_
+#define _V_LPC_H_
+
+/* simple linear scale LPC code */
+extern float vorbis_lpc_from_data(short *data,float *lpc,int n,int m,int stride);
+
+extern void vorbis_lpc_predict(float *coeff,short *prime,int m,
+                               short *data,long n,int stride);
+
+#endif

src/lpcm.c

@@ -13,35 +13,6 @@
 #include "lpcm.h"
 #include "m4af_endian.h"
 
-#if defined(_MSC_VER) && _MSC_VER < 1800
-#  ifdef _M_IX86
-inline int lrint(double x)
-{
-    int n;
-    _asm {
-        fld x
-        fistp n
-    }
-    return n;
-}
-#  else
-#    include <emmintrin.h>
-inline int lrint(double x)
-{
-    return _mm_cvtsd_si32(_mm_load_sd(&x));
-}
-#  endif
-#endif
-
-static
-inline double pcm_clip(double n, double min_value, double max_value)
-{
-    if (n < min_value)
-        return min_value;
-    else if (n > max_value)
-        return max_value;
-    return n;
-}
 static
 inline float pcm_i2f(int32_t n)
 {

src/lpcm.h

@@ -31,6 +31,36 @@ typedef struct pcm_sample_description_t {
 #define PCM_BYTES_PER_CHANNEL(desc) \
     ((desc)->bytes_per_frame / (desc)->channels_per_frame)
 
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#  ifdef _M_IX86
+static inline int lrint(double x)
+{
+    int n;
+    _asm {
+        fld x
+        fistp n
+    }
+    return n;
+}
+#  else
+#    include <emmintrin.h>
+static inline int lrint(double x)
+{
+    return _mm_cvtsd_si32(_mm_load_sd(&x));
+}
+#  endif
+#endif
+
+static
+inline double pcm_clip(double n, double min_value, double max_value)
+{
+    if (n < min_value)
+        return min_value;
+    else if (n > max_value)
+        return max_value;
+    return n;
+}
+
 int pcm_convert_to_native_sint16(const pcm_sample_description_t *format,
                                  const void *input, uint32_t nframes,
                                  int16_t *result);

src/main.c

@@ -489,51 +489,82 @@ int write_sample(FILE *ofp, m4af_ctx_t *m4af,
 }
 
 static
-int encode(pcm_reader_t *reader, HANDLE_AACENCODER encoder,
-           uint32_t frame_length, FILE *ofp, m4af_ctx_t *m4af,
-           int show_progress)
+int encode(aacenc_param_ex_t *params, pcm_reader_t *reader,
+           HANDLE_AACENCODER encoder, uint32_t frame_length, 
+           m4af_ctx_t *m4af)
 {
-    int16_t *ibuf = 0;
-    uint8_t *obuf = 0;
-    uint32_t olen;
-    uint32_t osize = 0;
+    struct buffer_t {
+        uint8_t *data;
+        uint32_t len, size;
+    };
+    int16_t *ibuf = 0, *ip;
+    struct buffer_t obuf[2] = {{ 0 }}, *obp;
+    unsigned flip = 0;
     int nread = 1;
-    int consumed;
     int rc = -1;
-    int frames_written = 0;
+    int remaining, consumed;
+    int frames_written = 0, encoded = 0;
     aacenc_progress_t progress = { 0 };
     const pcm_sample_description_t *fmt = pcm_get_format(reader);
 
     ibuf = malloc(frame_length * fmt->bytes_per_frame);
     aacenc_progress_init(&progress, pcm_get_length(reader), fmt->sample_rate);
-    do {
+
+    for (;;) {
+        /*
+         * Since we delay the write, we cannot just exit loop when interrupted.
+         * Instead, we regard it as EOF.
+         */
         if (g_interrupted)
             nread = 0;
-        else if (nread) {
+        if (nread > 0) {
             if ((nread = pcm_read_frames(reader, ibuf, frame_length)) < 0) {
                 fprintf(stderr, "ERROR: read failed\n");
                 goto END;
             }
-            if (show_progress)
+            if (!params->silent)
                 aacenc_progress_update(&progress, pcm_get_position(reader),
                                        fmt->sample_rate * 2);
         }
-        if ((consumed = aac_encode_frame(encoder, fmt, ibuf, nread,
-                                         &obuf, &olen, &osize)) < 0)
-            goto END;
-        if (olen > 0) {
-            if (write_sample(ofp, m4af, obuf, olen, frame_length) < 0)
+        ip = ibuf;
+        remaining = nread;
+        do {
+            obp = &obuf[flip];
+            consumed = aac_encode_frame(encoder, fmt, ip, remaining,
+                                        &obp->data, &obp->len, &obp->size);
+            if (consumed < 0) goto END;
+            if (consumed == 0 && obp->len == 0) goto DONE;
+            if (obp->len == 0) break;
+
+            remaining -= consumed;
+            ip += consumed * fmt->channels_per_frame;
+            flip ^= 1;
+            /*
+             * As we pad 1 frame at beginning and ending by our extrapolator,
+             * we want to drop them.
+             * We delay output by 1 frame by double buffering, and discard
+             * second frame and final frame from the encoder.
+             * Since sbr_header is included in the first frame (in case of
+             * SBR), we cannot discard first frame. So we pick second instead.
+             */
+            ++encoded;
+            if (encoded == 1 || encoded == 3)
+                continue;
+            obp = &obuf[flip];
+            if (write_sample(params->output_fp, m4af, obp->data, obp->len,
+                             frame_length) < 0)
                 goto END;
             ++frames_written;
+        } while (remaining > 0);
     }
-    } while (nread > 0 || olen > 0);
-
-    if (show_progress)
+DONE:
+    if (!params->silent)
         aacenc_progress_finish(&progress, pcm_get_position(reader));
     rc = frames_written;
 END:
     if (ibuf) free(ibuf);
-    if (obuf) free(obuf);
+    if (obuf[0].data) free(obuf[0].data);
+    if (obuf[1].data) free(obuf[1].data);
     return rc;
 }
 
@@ -709,10 +740,13 @@ pcm_reader_t *open_input(aacenc_param_ex_t *params)
             }
             break;
         default:
+            fprintf(stderr, "ERROR: unsupported input file\n");
             goto END;
         }
     }
-    return pcm_open_sint16_converter(reader);
+    if ((reader = pcm_open_sint16_converter(reader)) != 0)
+        reader = extrapolater_open(reader);
+    return reader;
 END:
     return 0;
 }
@@ -794,8 +828,7 @@ int main(int argc, char **argv)
         m4af_set_priming_mode(m4af, params.gapless_mode + 1);
         m4af_begin_write(m4af);
     }
-    frame_count = encode(reader, encoder, aacinfo.frameLength,
-                         params.output_fp, m4af, !params.silent);
+    frame_count = encode(&params, reader, encoder, aacinfo.frameLength, m4af);
     if (frame_count < 0)
         goto END;
     if (m4af) {

src/pcm_reader.h

@@ -109,4 +109,6 @@ int apple_chan_chunk(pcm_io_context_t *io, uint32_t chunk_size,
 
 pcm_reader_t *pcm_open_sint16_converter(pcm_reader_t *reader);
 
+pcm_reader_t *extrapolater_open(pcm_reader_t *reader);
+
 #endif