Commit last ok'ed parts of QCELP decoder and enable it.
authorKenan Gillet <kenan.gillet@gmail.com>
Tue, 2 Dec 2008 16:48:05 +0000 (16:48 +0000)
committerVitor Sessak <vitor1001@gmail.com>
Tue, 2 Dec 2008 16:48:05 +0000 (16:48 +0000)
patch by Kenan Gillet, kenan.gillet gmail com

Originally committed as revision 15978 to svn://svn.ffmpeg.org/ffmpeg/trunk

Changelog
doc/general.texi
libavcodec/Makefile
libavcodec/allcodecs.c
libavcodec/avcodec.h
libavcodec/qcelp_lsp.c [new file with mode: 0644]
libavcodec/qcelpdec.c
libavformat/mov.c

index dd1b847..50d1896 100644 (file)
--- a/Changelog
+++ b/Changelog
@@ -141,6 +141,7 @@ version <next>
 - Speex decoding via libspeex
 - Electronic Arts TGQ decoder
 - RV40 decoder
+- QCELP / PureVoice decoder
 
 version 0.4.9-pre1:
 
index 140338d..9ceb097 100644 (file)
@@ -397,6 +397,7 @@ following image formats are supported:
 @item Musepack               @tab     @tab  X
     @tab SV7 and SV8 are supported.
 @item Nellymoser ASAO        @tab  X  @tab  X
+@item QCELP / PureVoice      @tab     @tab  X
 @item Qdesign QDM2           @tab     @tab  X
     @tab There are still some distortions.
 @item QT IMA ADPCM           @tab  X  @tab  X
index 425066c..b7cce5f 100644 (file)
@@ -158,6 +158,7 @@ OBJS-$(CONFIG_PNG_DECODER)             += png.o pngdec.o
 OBJS-$(CONFIG_PNG_ENCODER)             += png.o pngenc.o
 OBJS-$(CONFIG_PPM_ENCODER)             += pnmenc.o pnm.o
 OBJS-$(CONFIG_PTX_DECODER)             += ptx.o
+OBJS-$(CONFIG_QCELP_DECODER)           += qcelpdec.o qcelp_lsp.o celp_math.o celp_filters.o
 OBJS-$(CONFIG_QDM2_DECODER)            += qdm2.o mpegaudiodec.o mpegaudiodecheader.o mpegaudio.o mpegaudiodata.o
 OBJS-$(CONFIG_QDRAW_DECODER)           += qdrw.o
 OBJS-$(CONFIG_QPEG_DECODER)            += qpeg.o
index 6880d00..d5fe8b5 100644 (file)
@@ -204,6 +204,7 @@ void avcodec_register_all(void)
     REGISTER_DECODER (MPC7, mpc7);
     REGISTER_DECODER (MPC8, mpc8);
     REGISTER_ENCDEC  (NELLYMOSER, nellymoser);
+    REGISTER_DECODER (QCELP, qcelp);
     REGISTER_DECODER (QDM2, qdm2);
     REGISTER_DECODER (RA_144, ra_144);
     REGISTER_DECODER (RA_288, ra_288);
index 53ded30..2afcce9 100644 (file)
@@ -30,7 +30,7 @@
 #include "libavutil/avutil.h"
 
 #define LIBAVCODEC_VERSION_MAJOR 52
-#define LIBAVCODEC_VERSION_MINOR  5
+#define LIBAVCODEC_VERSION_MINOR  6
 #define LIBAVCODEC_VERSION_MICRO  0
 
 #define LIBAVCODEC_VERSION_INT  AV_VERSION_INT(LIBAVCODEC_VERSION_MAJOR, \
diff --git a/libavcodec/qcelp_lsp.c b/libavcodec/qcelp_lsp.c
new file mode 100644 (file)
index 0000000..a4be004
--- /dev/null
@@ -0,0 +1,100 @@
+/*
+ * QCELP decoder
+ * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file qcelp_lsp.c
+ * QCELP decoder
+ * @author Reynaldo H. Verdejo Pinochet
+ */
+
+#include "libavutil/mathematics.h"
+
+/**
+ * initial coefficient to perform bandwidth expansion on LPC
+ *
+ * @note: 0.9883 looks like an approximation of 253/256.
+ *
+ * TIA/EIA/IS-733 2.4.3.3.6 6
+ */
+#define QCELP_BANDWITH_EXPANSION_COEFF 0.9883
+
+/**
+ * Computes the Pa / (1 + z(-1)) or Qa / (1 - z(-1)) coefficients
+ * needed for LSP to LPC conversion.
+ * We only need to calculate the 6 first elements of the polynomial.
+ *
+ * @param lspf line spectral pair frequencies
+ * @param f [out] polynomial input/output as a vector
+ *
+ * TIA/EIA/IS-733 2.4.3.3.5-1/2
+ */
+static void lsp2polyf(const float *lspf, double *f, int lp_half_order)
+{
+    int i, j;
+
+    f[0] = 1.0;
+    f[1] = -2 * cos(M_PI * lspf[0]);
+    lspf -= 2;
+    for(i=2; i<=lp_half_order; i++)
+    {
+        double val = -2 * cos(M_PI * lspf[2*i]);
+        f[i] = val * f[i-1] + 2*f[i-2];
+        for(j=i-1; j>1; j--)
+            f[j] += f[j-1] * val + f[j-2];
+        f[1] += val;
+    }
+}
+
+/**
+ * Reconstructs LPC coefficients from the line spectral pair frequencies
+ * and performs bandwidth expansion.
+ *
+ * @param lspf line spectral pair frequencies
+ * @param lpc linear predictive coding coefficients
+ *
+ * @note: bandwith_expansion_coeff could be precalculated into a table
+ *        but it seems to be slower on x86
+ *
+ * TIA/EIA/IS-733 2.4.3.3.5
+ */
+void ff_qcelp_lspf2lpc(const float *lspf, float *lpc)
+{
+    double pa[6], qa[6];
+    int   i;
+    double bandwith_expansion_coeff = -QCELP_BANDWITH_EXPANSION_COEFF * 0.5;
+
+    lsp2polyf(lspf,     pa, 5);
+    lsp2polyf(lspf + 1, qa, 5);
+
+    for (i=4; i>=0; i--)
+    {
+        double paf = pa[i+1] + pa[i];
+        double qaf = qa[i+1] - qa[i];
+
+        lpc[i  ] = paf + qaf;
+        lpc[9-i] = paf - qaf;
+    }
+    for (i=0; i<10; i++)
+    {
+        lpc[i] *= bandwith_expansion_coeff;
+        bandwith_expansion_coeff *= QCELP_BANDWITH_EXPANSION_COEFF;
+    }
+}
index a759b1b..4ff2c47 100644 (file)
@@ -65,8 +65,8 @@ typedef struct {
     float             last_codebook_gain;
     int               prev_g1[2];
     int               prev_bitrate;
-    float             prev_pitch_gain[4];
-    uint8_t           prev_pitch_lag[4];
+    float             pitch_gain[4];
+    uint8_t           pitch_lag[4];
     uint16_t          first16bits;
 } QCELPContext;
 
@@ -75,7 +75,7 @@ typedef struct {
  *
  * TIA/EIA/IS-733 2.4.3.3.5
  */
-void qcelp_lspf2lpc(const float *lspf, float *lpc);
+void ff_qcelp_lspf2lpc(const float *lspf, float *lpc);
 
 static void weighted_vector_sumf(float *out, const float *in_a,
                                  const float *in_b, float weight_coeff_a,
@@ -494,6 +494,58 @@ static const float *do_pitchfilter(float memory[303], const float v_in[160],
     return memory + 143;
 }
 
+/**
+ * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
+ * TIA/EIA/IS-733 2.4.5.2
+ *
+ * @param q the context
+ * @param cdn_vector the scaled codebook vector
+ */
+static void apply_pitch_filters(QCELPContext *q,
+                                float *cdn_vector) {
+    int         i;
+    const float *v_synthesis_filtered, *v_pre_filtered;
+
+    if (q->bitrate >= RATE_HALF ||
+       (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
+
+        if (q->bitrate >= RATE_HALF) {
+
+            // Compute gain & lag for the whole frame.
+            for (i = 0; i < 4; i++) {
+                q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
+
+                q->pitch_lag[i] = q->frame.plag[i] + 16;
+            }
+        } else {
+            float max_pitch_gain = q->erasure_count < 3 ? 0.9 - 0.3 * (q->erasure_count - 1)
+                                                        : 0.0;
+            for (i = 0; i < 4; i++)
+                q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
+
+            memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
+        }
+
+        // pitch synthesis filter
+        v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem, cdn_vector,
+                                              q->pitch_gain, q->pitch_lag, q->frame.pfrac);
+
+        // pitch prefilter update
+        for (i = 0; i < 4; i++)
+            q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
+
+        v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem, v_synthesis_filtered,
+                                        q->pitch_gain, q->pitch_lag, q->frame.pfrac);
+
+        apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
+    } else {
+        memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
+        memcpy(q->pitch_pre_filter_mem,       cdn_vector + 17, 143 * sizeof(float));
+        memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
+        memset(q->pitch_lag,  0, sizeof(q->pitch_lag));
+    }
+}
+
 /**
  * Interpolates LSP frequencies and computes LPC coefficients
  * for a given bitrate & pitch subframe.
@@ -522,9 +574,9 @@ void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
     {
         weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
                              weight, 1.0 - weight, 10);
-        qcelp_lspf2lpc(interpolated_lspf, lpc);
+        ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
     }else if(q->bitrate >= RATE_QUARTER || (q->bitrate == I_F_Q && !subframe_num))
-        qcelp_lspf2lpc(curr_lspf, lpc);
+        ff_qcelp_lspf2lpc(curr_lspf, lpc);
 }
 
 static int buf_size2bitrate(const int buf_size)
index 44fb3f2..f015fda 100644 (file)
@@ -988,7 +988,9 @@ static int mov_read_stsd(MOVContext *c, ByteIOContext *pb, MOV_atom_t atom)
 #endif
     /* no ifdef since parameters are always those */
     case CODEC_ID_QCELP:
+        st->codec->frame_size= 160;
         st->codec->channels= 1; /* really needed */
+        break;
     case CODEC_ID_AMR_NB:
     case CODEC_ID_AMR_WB:
         st->codec->frame_size= sc->samples_per_frame;