Make the targa encoder return sane error values rather than -1.
[ffmpeg.git] / libavcodec / qcelpdec.c
index 87baee4..0441e1f 100644 (file)
@@ -20,7 +20,7 @@
  */
 
 /**
- * @file qcelpdec.c
+ * @file
  * QCELP decoder
  * @author Reynaldo H. Verdejo Pinochet
  * @remark FFmpeg merging spearheaded by Kenan Gillet
 #include <stddef.h>
 
 #include "avcodec.h"
-#include "bitstream.h"
+#include "internal.h"
+#include "get_bits.h"
 
 #include "qcelpdata.h"
 
 #include "celp_math.h"
 #include "celp_filters.h"
+#include "acelp_filters.h"
+#include "acelp_vectors.h"
+#include "lsp.h"
 
 #undef NDEBUG
 #include <assert.h>
@@ -70,25 +74,13 @@ typedef struct
     float    pitch_gain[4];
     uint8_t  pitch_lag[4];
     uint16_t first16bits;
-} QCELPContext;
+    uint8_t  warned_buf_mismatch_bitrate;
 
-/**
- * Reconstructs LPC coefficients from the line spectral pair frequencies.
- *
- * TIA/EIA/IS-733 2.4.3.3.5
- */
-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,
-                                 float weight_coeff_b, int length)
-{
-    int i;
-
-    for(i=0; i<length; i++)
-        out[i] = weight_coeff_a * in_a[i]
-               + weight_coeff_b * in_b[i];
-}
+    /* postfilter */
+    float    postfilter_synth_mem[10];
+    float    postfilter_agc_mem;
+    float    postfilter_tilt_mem;
+} QCELPContext;
 
 /**
  * Initialize the speech codec according to the specification.
@@ -109,8 +101,8 @@ static av_cold int qcelp_decode_init(AVCodecContext *avctx)
 }
 
 /**
- * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
- * transmission codes of any bitrate and checks for badly received packets.
+ * Decode the 10 quantized LSP frequencies from the LSPV/LSP
+ * transmission codes of any bitrate and check for badly received packets.
  *
  * @param q the context
  * @param lspf line spectral pair frequencies
@@ -172,7 +164,7 @@ static int decode_lspf(QCELPContext *q, float *lspf)
             lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
 
         // Low-pass filter the LSP frequencies.
-        weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
+        ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
     }else
     {
         q->octave_count = 0;
@@ -205,7 +197,7 @@ static int decode_lspf(QCELPContext *q, float *lspf)
 }
 
 /**
- * Converts codebook transmission codes to GAIN and INDEX.
+ * Convert codebook transmission codes to GAIN and INDEX.
  *
  * @param q the context
  * @param gain array holding the decoded gain
@@ -257,7 +249,7 @@ static void decode_gain_and_index(QCELPContext  *q,
             gain[2] =     gain[1];
             gain[1] = 0.6*gain[0] + 0.4*gain[1];
         }
-    }else
+    }else if (q->bitrate != SILENCE)
     {
         if(q->bitrate == RATE_OCTAVE)
         {
@@ -317,7 +309,7 @@ static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
 }
 
 /**
- * Computes the scaled codebook vector Cdn From INDEX and GAIN
+ * Compute the scaled codebook vector Cdn From INDEX and GAIN
  * for all rates.
  *
  * The specification lacks some information here.
@@ -413,6 +405,9 @@ static void compute_svector(QCELPContext *q, const float *gain,
                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
             }
         break;
+        case SILENCE:
+            memset(cdn_vector, 0, 160 * sizeof(float));
+        break;
     }
 }
 
@@ -423,29 +418,18 @@ static void compute_svector(QCELPContext *q, const float *gain,
  * @param v_in gain-controlled vector
  * @param v_ref vector to control gain of
  *
- * FIXME: If v_ref is a zero vector, it energy is zero
- *        and the behavior of the gain control is
- *        undefined in the specs.
- *
- * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
+ * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
  */
 static void apply_gain_ctrl(float *v_out, const float *v_ref,
                             const float *v_in)
 {
-    int   i, j, len;
-    float scalefactor;
+    int i;
 
-    for(i=0, j=0; i<4; i++)
-    {
-        scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
-        if(scalefactor)
-            scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
-                        / scalefactor);
-        else
-            av_log_missing_feature(NULL, "Zero energy for gain control", 1);
-        for(len=j+40; j<len; j++)
-            v_out[j] = scalefactor * v_in[j];
-    }
+    for (i = 0; i < 160; i += 40)
+        ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i,
+                                                ff_dot_productf(v_ref + i,
+                                                                v_ref + i, 40),
+                                                40);
 }
 
 /**
@@ -508,7 +492,7 @@ static const float *do_pitchfilter(float memory[303], const float v_in[160],
 
 /**
  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
- * TIA/EIA/IS-733 2.4.5.2
+ * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
  *
  * @param q the context
  * @param cdn_vector the scaled codebook vector
@@ -519,6 +503,7 @@ static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
     const float *v_synthesis_filtered, *v_pre_filtered;
 
     if(q->bitrate >= RATE_HALF ||
+       q->bitrate == SILENCE ||
        (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
     {
 
@@ -534,7 +519,19 @@ static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
             }
         }else
         {
-            float max_pitch_gain = q->erasure_count < 3 ? 0.9 - 0.3 * (q->erasure_count - 1) : 0.0;
+            float max_pitch_gain;
+
+            if (q->bitrate == I_F_Q)
+            {
+                  if (q->erasure_count < 3)
+                      max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
+                  else
+                      max_pitch_gain = 0.0;
+            }else
+            {
+                assert(q->bitrate == SILENCE);
+                max_pitch_gain = 1.0;
+            }
             for(i=0; i<4; i++)
                 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
 
@@ -567,18 +564,48 @@ static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
 }
 
 /**
- * Interpolates LSP frequencies and computes LPC coefficients
+ * Reconstruct LPC coefficients from the line spectral pair frequencies
+ * and perform bandwidth expansion.
+ *
+ * @param lspf line spectral pair frequencies
+ * @param lpc linear predictive coding coefficients
+ *
+ * @note: bandwidth_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
+ */
+static void lspf2lpc(const float *lspf, float *lpc)
+{
+    double lsp[10];
+    double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
+    int   i;
+
+    for (i=0; i<10; i++)
+        lsp[i] = cos(M_PI * lspf[i]);
+
+    ff_acelp_lspd2lpc(lsp, lpc, 5);
+
+    for (i=0; i<10; i++)
+    {
+        lpc[i] *= bandwidth_expansion_coeff;
+        bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
+    }
+}
+
+/**
+ * Interpolate LSP frequencies and compute LPC coefficients
  * for a given bitrate & pitch subframe.
  *
- * TIA/EIA/IS-733 2.4.3.3.4
+ * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
  *
  * @param q the context
  * @param curr_lspf LSP frequencies vector of the current frame
  * @param lpc float vector for the resulting LPC
  * @param subframe_num frame number in decoded stream
  */
-void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
-                     const int subframe_num)
+static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
+                            float *lpc, const int subframe_num)
 {
     float interpolated_lspf[10];
     float weight;
@@ -592,15 +619,17 @@ void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
 
     if(weight != 1.0)
     {
-        weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
-                             weight, 1.0 - weight, 10);
-        ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
+        ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
+                                weight, 1.0 - weight, 10);
+        lspf2lpc(interpolated_lspf, lpc);
     }else if(q->bitrate >= RATE_QUARTER ||
              (q->bitrate == I_F_Q && !subframe_num))
-        ff_qcelp_lspf2lpc(curr_lspf, lpc);
+        lspf2lpc(curr_lspf, lpc);
+    else if(q->bitrate == SILENCE && !subframe_num)
+        lspf2lpc(q->prev_lspf, lpc);
 }
 
-static int buf_size2bitrate(const int buf_size)
+static qcelp_packet_rate buf_size2bitrate(const int buf_size)
 {
     switch(buf_size)
     {
@@ -611,7 +640,7 @@ static int buf_size2bitrate(const int buf_size)
         case  1: return SILENCE;
     }
 
-    return -1;
+    return I_F_Q;
 }
 
 /**
@@ -626,7 +655,7 @@ static int buf_size2bitrate(const int buf_size)
  *
  * TIA/EIA/IS-733 2.4.8.7.1
  */
-static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
+static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
                              const uint8_t **buf)
 {
     qcelp_packet_rate bitrate;
@@ -635,8 +664,13 @@ static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
     {
         if(bitrate > **buf)
         {
+            QCELPContext *q = avctx->priv_data;
+            if (!q->warned_buf_mismatch_bitrate)
+            {
             av_log(avctx, AV_LOG_WARNING,
                    "Claimed bitrate and buffer size mismatch.\n");
+                q->warned_buf_mismatch_bitrate = 1;
+            }
             bitrate = **buf;
         }else if(bitrate < **buf)
         {
@@ -654,9 +688,8 @@ static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
 
     if(bitrate == SILENCE)
     {
-        // FIXME: the decoder should not handle SILENCE frames as I_F_Q frames
-        av_log_missing_feature(avctx, "Blank frame", 1);
-        bitrate = I_F_Q;
+        //FIXME: Remove experimental warning when tested with samples.
+        av_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
     }
     return bitrate;
 }
@@ -668,9 +701,41 @@ static void warn_insufficient_frame_quality(AVCodecContext *avctx,
            message);
 }
 
+static void postfilter(QCELPContext *q, float *samples, float *lpc)
+{
+    static const float pow_0_775[10] = {
+        0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
+        0.216676, 0.167924, 0.130141, 0.100859, 0.078166
+    }, pow_0_625[10] = {
+        0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
+        0.059605, 0.037253, 0.023283, 0.014552, 0.009095
+    };
+    float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
+    int n;
+
+    for (n = 0; n < 10; n++) {
+        lpc_s[n] = lpc[n] * pow_0_625[n];
+        lpc_p[n] = lpc[n] * pow_0_775[n];
+    }
+
+    ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
+                                      q->formant_mem + 10, 160, 10);
+    memcpy(pole_out, q->postfilter_synth_mem,       sizeof(float) * 10);
+    ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
+    memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
+
+    ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
+
+    ff_adaptive_gain_control(samples, pole_out + 10,
+        ff_dot_productf(q->formant_mem + 10, q->formant_mem + 10, 160),
+        160, 0.9375, &q->postfilter_agc_mem);
+}
+
 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
-                              const uint8_t *buf, int buf_size)
+                              AVPacket *avpkt)
 {
+    const uint8_t *buf = avpkt->data;
+    int buf_size = avpkt->size;
     QCELPContext *q = avctx->priv_data;
     float *outbuffer = data;
     int   i;
@@ -763,15 +828,11 @@ erasure:
                                      10);
         formant_mem += 40;
     }
-    memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
 
-    // FIXME: postfilter and final gain control should be here.
-    // TIA/EIA/IS-733 2.4.8.6
+    // postfilter, as per TIA/EIA/IS-733 2.4.8.6
+    postfilter(q, outbuffer, lpc);
 
-    formant_mem = q->formant_mem + 10;
-    for(i=0; i<160; i++)
-        *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
-                                QCELP_CLIP_UPPER_BOUND);
+    memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
 
     memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
     q->prev_bitrate = q->bitrate;
@@ -784,7 +845,7 @@ erasure:
 AVCodec qcelp_decoder =
 {
     .name   = "qcelp",
-    .type   = CODEC_TYPE_AUDIO,
+    .type   = AVMEDIA_TYPE_AUDIO,
     .id     = CODEC_ID_QCELP,
     .init   = qcelp_decode_init,
     .decode = qcelp_decode_frame,