avcodec/dcaenc: Initial implementation of ADPCM encoding for DCA encoder

libavcodec/Makefile

@@ -244,7 +244,8 @@ OBJS-$(CONFIG_CYUV_DECODER)            += cyuv.o
 OBJS-$(CONFIG_DCA_DECODER)             += dcadec.o dca.o dcadata.o dcahuff.o \
                                           dca_core.o dca_exss.o dca_xll.o dca_lbr.o \
                                           dcadsp.o dcadct.o synth_filter.o
-OBJS-$(CONFIG_DCA_ENCODER)             += dcaenc.o dca.o dcadata.o dcahuff.o
+OBJS-$(CONFIG_DCA_ENCODER)             += dcaenc.o dca.o dcadata.o dcahuff.o \
+                                          dcaadpcm.o
 OBJS-$(CONFIG_DDS_DECODER)             += dds.o
 OBJS-$(CONFIG_DIRAC_DECODER)           += diracdec.o dirac.o diracdsp.o diractab.o \
                                           dirac_arith.o dirac_dwt.o dirac_vlc.o

libavcodec/dca_core.c

@@ -18,6 +18,7 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
+#include "dcaadpcm.h"
 #include "dcadec.h"
 #include "dcadata.h"
 #include "dcahuff.h"
@@ -670,46 +671,21 @@ static inline int extract_audio(DCACoreDecoder *s, int32_t *audio, int abits, in
     return 0;
 }
 
-static inline void dequantize(int32_t *output, const int32_t *input,
-                              int32_t step_size, int32_t scale, int residual)
-{
-    // Account for quantizer step size
-    int64_t step_scale = (int64_t)step_size * scale;
-    int n, shift = 0;
-
-    // Limit scale factor resolution to 22 bits
-    if (step_scale > (1 << 23)) {
-        shift = av_log2(step_scale >> 23) + 1;
-        step_scale >>= shift;
-    }
-
-    // Scale the samples
-    if (residual) {
-        for (n = 0; n < DCA_SUBBAND_SAMPLES; n++)
-            output[n] += clip23(norm__(input[n] * step_scale, 22 - shift));
-    } else {
-        for (n = 0; n < DCA_SUBBAND_SAMPLES; n++)
-            output[n]  = clip23(norm__(input[n] * step_scale, 22 - shift));
-    }
-}
-
 static inline void inverse_adpcm(int32_t **subband_samples,
                                  const int16_t *vq_index,
                                  const int8_t *prediction_mode,
                                  int sb_start, int sb_end,
                                  int ofs, int len)
 {
-    int i, j, k;
+    int i, j;
 
     for (i = sb_start; i < sb_end; i++) {
         if (prediction_mode[i]) {
-            const int16_t *coeff = ff_dca_adpcm_vb[vq_index[i]];
+            const int pred_id = vq_index[i];
             int32_t *ptr = subband_samples[i] + ofs;
             for (j = 0; j < len; j++) {
-                int64_t err = 0;
+                int32_t x = ff_dcaadpcm_predict(pred_id, ptr + j - DCA_ADPCM_COEFFS);
-                for (k = 0; k < DCA_ADPCM_COEFFS; k++)
+                ptr[j] = clip23(ptr[j] + x);
-                    err += (int64_t)ptr[j - k - 1] * coeff[k];
-                ptr[j] = clip23(ptr[j] + clip23(norm13(err)));
             }
         }
     }
@@ -817,8 +793,8 @@ static int parse_subframe_audio(DCACoreDecoder *s, int sf, enum HeaderType heade
                     scale = clip23(adj * scale >> 22);
                 }
 
-                dequantize(s->subband_samples[ch][band] + ofs,
+                ff_dca_core_dequantize(s->subband_samples[ch][band] + ofs,
-                           audio, step_size, scale, 0);
+                           audio, step_size, scale, 0, DCA_SUBBAND_SAMPLES);
             }
         }
 
@@ -1146,8 +1122,8 @@ static int parse_xbr_subframe(DCACoreDecoder *s, int xbr_base_ch, int xbr_nchann
                 else
                     scale = xbr_scale_factors[ch][band][1];
 
-                dequantize(s->subband_samples[ch][band] + ofs,
+                ff_dca_core_dequantize(s->subband_samples[ch][band] + ofs,
-                           audio, step_size, scale, 1);
+                           audio, step_size, scale, 1, DCA_SUBBAND_SAMPLES);
             }
         }
 
@@ -1326,8 +1302,8 @@ static int parse_x96_subframe_audio(DCACoreDecoder *s, int sf, int xch_base, int
                 // Get the scale factor
                 scale = s->scale_factors[ch][band >> 1][band & 1];
 
-                dequantize(s->x96_subband_samples[ch][band] + ofs,
+                ff_dca_core_dequantize(s->x96_subband_samples[ch][band] + ofs,
-                           audio, step_size, scale, 0);
+                           audio, step_size, scale, 0, DCA_SUBBAND_SAMPLES);
             }
         }
 

libavcodec/dca_core.h

@@ -33,6 +33,7 @@
 #include "dca_exss.h"
 #include "dcadsp.h"
 #include "dcadct.h"
+#include "dcamath.h"
 #include "dcahuff.h"
 #include "fft.h"
 #include "synth_filter.h"
@@ -43,7 +44,6 @@
 #define DCA_SUBFRAMES           16
 #define DCA_SUBBAND_SAMPLES     8
 #define DCA_PCMBLOCK_SAMPLES    32
-#define DCA_ADPCM_COEFFS        4
 #define DCA_LFE_HISTORY         8
 #define DCA_ABITS_MAX           26
 
@@ -195,6 +195,29 @@ static inline int ff_dca_core_map_spkr(DCACoreDecoder *core, int spkr)
     return -1;
 }
 
+static inline void ff_dca_core_dequantize(int32_t *output, const int32_t *input,
+                                          int32_t step_size, int32_t scale, int residual, int len)
+{
+    // Account for quantizer step size
+    int64_t step_scale = (int64_t)step_size * scale;
+    int n, shift = 0;
+
+    // Limit scale factor resolution to 22 bits
+    if (step_scale > (1 << 23)) {
+        shift = av_log2(step_scale >> 23) + 1;
+        step_scale >>= shift;
+    }
+
+    // Scale the samples
+    if (residual) {
+        for (n = 0; n < len; n++)
+            output[n] += clip23(norm__(input[n] * step_scale, 22 - shift));
+    } else {
+        for (n = 0; n < len; n++)
+            output[n]  = clip23(norm__(input[n] * step_scale, 22 - shift));
+    }
+}
+
 int ff_dca_core_parse(DCACoreDecoder *s, uint8_t *data, int size);
 int ff_dca_core_parse_exss(DCACoreDecoder *s, uint8_t *data, DCAExssAsset *asset);
 int ff_dca_core_filter_fixed(DCACoreDecoder *s, int x96_synth);

libavcodec/dcaadpcm.c

@@ -0,0 +1,228 @@
+/*
+ * DCA ADPCM engine
+ * Copyright (C) 2017 Daniil Cherednik
+ *
+ * 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
+ */
+
+
+#include "dcaadpcm.h"
+#include "dcaenc.h"
+#include "dca_core.h"
+#include "mathops.h"
+
+typedef int32_t premultiplied_coeffs[10];
+
+//assume we have DCA_ADPCM_COEFFS values before x
+static inline int64_t calc_corr(const int32_t *x, int len, int j, int k)
+{
+    int n;
+    int64_t s = 0;
+    for (n = 0; n < len; n++)
+        s += MUL64(x[n-j], x[n-k]);
+    return s;
+}
+
+static inline int64_t apply_filter(const int16_t a[DCA_ADPCM_COEFFS], const int64_t corr[15], const int32_t aa[10])
+{
+    int64_t err = 0;
+    int64_t tmp = 0;
+
+    err = corr[0];
+
+    tmp += MUL64(a[0], corr[1]);
+    tmp += MUL64(a[1], corr[2]);
+    tmp += MUL64(a[2], corr[3]);
+    tmp += MUL64(a[3], corr[4]);
+
+    tmp = norm__(tmp, 13);
+    tmp += tmp;
+
+    err -= tmp;
+    tmp = 0;
+
+    tmp += MUL64(corr[5], aa[0]);
+    tmp += MUL64(corr[6], aa[1]);
+    tmp += MUL64(corr[7], aa[2]);
+    tmp += MUL64(corr[8], aa[3]);
+
+    tmp += MUL64(corr[9], aa[4]);
+    tmp += MUL64(corr[10], aa[5]);
+    tmp += MUL64(corr[11], aa[6]);
+
+    tmp += MUL64(corr[12], aa[7]);
+    tmp += MUL64(corr[13], aa[8]);
+
+    tmp += MUL64(corr[14], aa[9]);
+
+    tmp = norm__(tmp, 26);
+
+    err += tmp;
+
+    return llabs(err);
+}
+
+static int64_t find_best_filter(const DCAADPCMEncContext *s, const int32_t *in, int len)
+{
+    const premultiplied_coeffs *precalc_data = s->private_data;
+    int i, j, k = 0;
+    int vq;
+    int64_t err;
+    int64_t min_err = 1ll << 62;
+    int64_t corr[15];
+
+    for (i = 0; i <= DCA_ADPCM_COEFFS; i++)
+        for (j = i; j <= DCA_ADPCM_COEFFS; j++)
+            corr[k++] = calc_corr(in+4, len, i, j);
+
+    for (i = 0; i < DCA_ADPCM_VQCODEBOOK_SZ; i++) {
+        err = apply_filter(ff_dca_adpcm_vb[i], corr, *precalc_data);
+        if (err < min_err) {
+            min_err = err;
+            vq = i;
+        }
+        precalc_data++;
+    }
+
+    return vq;
+}
+
+static inline int64_t calc_prediction_gain(int pred_vq, const int32_t *in, int32_t *out, int len)
+{
+    int i;
+    int32_t error;
+
+    int64_t signal_energy = 0;
+    int64_t error_energy = 0;
+
+    for (i = 0; i < len; i++) {
+        error = in[DCA_ADPCM_COEFFS + i] - ff_dcaadpcm_predict(pred_vq, in + i);
+        out[i] = error;
+        signal_energy += MUL64(in[DCA_ADPCM_COEFFS + i], in[DCA_ADPCM_COEFFS + i]);
+        error_energy += MUL64(error, error);
+    }
+
+    if (!error_energy)
+        return -1;
+
+    return signal_energy / error_energy;
+}
+
+int ff_dcaadpcm_subband_analysis(const DCAADPCMEncContext *s, const int32_t *in, int len, int *diff)
+{
+    int pred_vq, i;
+    int32_t input_buffer[16 + DCA_ADPCM_COEFFS];
+    int32_t input_buffer2[16 + DCA_ADPCM_COEFFS];
+
+    int32_t max = 0;
+    int shift_bits;
+    uint64_t pg = 0;
+
+    for (i = 0; i < len + DCA_ADPCM_COEFFS; i++)
+        max |= FFABS(in[i]);
+
+    // normalize input to simplify apply_filter
+    shift_bits = av_log2(max) - 11;
+
+    for (i = 0; i < len + DCA_ADPCM_COEFFS; i++) {
+        input_buffer[i] = norm__(in[i], 7);
+        input_buffer2[i] = norm__(in[i], shift_bits);
+    }
+
+    pred_vq = find_best_filter(s, input_buffer2, len);
+
+    if (pred_vq < 0)
+        return -1;
+
+    pg = calc_prediction_gain(pred_vq, input_buffer, diff, len);
+
+    // Greater than 10db (10*log(10)) prediction gain to use ADPCM.
+    // TODO: Tune it.
+    if (pg < 10)
+        return -1;
+
+    for (i = 0; i < len; i++)
+        diff[i] <<= 7;
+
+    return pred_vq;
+}
+
+static void precalc(premultiplied_coeffs *data)
+{
+    int i, j, k;
+
+    for (i = 0; i < DCA_ADPCM_VQCODEBOOK_SZ; i++) {
+        int id = 0;
+        int32_t t = 0;
+        for (j = 0; j < DCA_ADPCM_COEFFS; j++) {
+            for (k = j; k < DCA_ADPCM_COEFFS; k++) {
+                t = (int32_t)ff_dca_adpcm_vb[i][j] * (int32_t)ff_dca_adpcm_vb[i][k];
+                if (j != k)
+                    t *= 2;
+                (*data)[id++] = t;
+             }
+        }
+        data++;
+    }
+}
+
+int ff_dcaadpcm_do_real(int pred_vq_index,
+                        softfloat quant, int32_t scale_factor, int32_t step_size,
+                        const int32_t *prev_hist, const int32_t *in, int32_t *next_hist, int32_t *out,
+                        int len, int32_t peak)
+{
+    int i;
+    int64_t delta;
+    int32_t dequant_delta;
+    int32_t work_bufer[16 + DCA_ADPCM_COEFFS];
+
+    memcpy(work_bufer, prev_hist, sizeof(int32_t) * DCA_ADPCM_COEFFS);
+
+    for (i = 0; i < len; i++) {
+        work_bufer[DCA_ADPCM_COEFFS + i] = ff_dcaadpcm_predict(pred_vq_index, &work_bufer[i]);
+
+        delta = (int64_t)in[i] - ((int64_t)work_bufer[DCA_ADPCM_COEFFS + i] << 7);
+
+        out[i] = quantize_value(av_clip64(delta, -peak, peak), quant);
+
+        ff_dca_core_dequantize(&dequant_delta, &out[i], step_size, scale_factor, 0, 1);
+
+        work_bufer[DCA_ADPCM_COEFFS+i] += dequant_delta;
+    }
+
+    memcpy(next_hist, &work_bufer[len], sizeof(int32_t) * DCA_ADPCM_COEFFS);
+
+    return 0;
+}
+
+av_cold int ff_dcaadpcm_init(DCAADPCMEncContext *s)
+{
+    if (!s)
+        return -1;
+
+    s->private_data = av_malloc(sizeof(premultiplied_coeffs) * DCA_ADPCM_VQCODEBOOK_SZ);
+    precalc(s->private_data);
+    return 0;
+}
+
+av_cold void ff_dcaadpcm_free(DCAADPCMEncContext *s)
+{
+    if (!s)
+        return;
+
+    av_freep(&s->private_data);
+}

libavcodec/dcaadpcm.h

@@ -0,0 +1,54 @@
+/*
+ * DCA ADPCM engine
+ * Copyright (C) 2017 Daniil Cherednik
+ *
+ * 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
+ */
+
+#ifndef AVCODEC_DCAADPCM_H
+#define AVCODEC_DCAADPCM_H
+
+#include "dcamath.h"
+#include "dcadata.h"
+#include "dcaenc.h"
+
+typedef struct DCAADPCMEncContext {
+    void *private_data;
+} DCAADPCMEncContext;
+
+static inline int64_t ff_dcaadpcm_predict(int pred_vq_index, const int32_t *input)
+{
+    int i;
+    const int16_t *coeff = ff_dca_adpcm_vb[pred_vq_index];
+    int64_t pred = 0;
+    for (i = 0; i < DCA_ADPCM_COEFFS; i++)
+        pred += (int64_t)input[DCA_ADPCM_COEFFS - 1 - i] * coeff[i];
+
+    return clip23(norm13(pred));
+}
+
+int ff_dcaadpcm_subband_analysis(const DCAADPCMEncContext *s, const int32_t *input, int len, int *diff);
+
+int ff_dcaadpcm_do_real(int pred_vq_index,
+                        softfloat quant, int32_t scale_factor, int32_t step_size,
+                        const int32_t *prev_hist, const int32_t *in, int32_t *next_hist, int32_t *out,
+                        int len, int32_t peak);
+
+av_cold int ff_dcaadpcm_init(DCAADPCMEncContext *s);
+av_cold void ff_dcaadpcm_free(DCAADPCMEncContext *s);
+
+#endif /* AVCODEC_DCAADPCM_H */

libavcodec/dcadata.c

@@ -61,7 +61,7 @@ const uint8_t ff_dca_quant_index_group_size[DCA_CODE_BOOKS] = {
 /* ADPCM data */
 
 /* 16 bits signed fractional Q13 binary codes */
-const int16_t ff_dca_adpcm_vb[4096][4] = {
+const int16_t ff_dca_adpcm_vb[DCA_ADPCM_VQCODEBOOK_SZ][DCA_ADPCM_COEFFS] = {
     {   9928,  -2618,  -1093, -1263 },
     {  11077,  -2876,  -1747,  -308 },
     {  10503,  -1082,  -1426, -1167 },

libavcodec/dcadata.h

@@ -25,6 +25,9 @@
 
 #include "dcahuff.h"
 
+#define DCA_ADPCM_COEFFS        4
+#define DCA_ADPCM_VQCODEBOOK_SZ 4096
+
 extern const uint32_t ff_dca_bit_rates[32];
 
 extern const uint8_t ff_dca_channels[16];
@@ -36,7 +39,7 @@ extern const uint8_t ff_dca_dmix_primary_nch[8];
 extern const uint8_t ff_dca_quant_index_sel_nbits[DCA_CODE_BOOKS];
 extern const uint8_t ff_dca_quant_index_group_size[DCA_CODE_BOOKS];
 
-extern const int16_t ff_dca_adpcm_vb[4096][4];
+extern const int16_t ff_dca_adpcm_vb[DCA_ADPCM_VQCODEBOOK_SZ][DCA_ADPCM_COEFFS];
 
 extern const uint32_t ff_dca_scale_factor_quant6[64];
 extern const uint32_t ff_dca_scale_factor_quant7[128];

libavcodec/dcaenc.c

@@ -25,8 +25,12 @@
 #include "libavutil/channel_layout.h"
 #include "libavutil/common.h"
 #include "libavutil/ffmath.h"
+#include "libavutil/opt.h"
 #include "avcodec.h"
 #include "dca.h"
+#include "dcaadpcm.h"
+#include "dcamath.h"
+#include "dca_core.h"
 #include "dcadata.h"
 #include "dcaenc.h"
 #include "internal.h"
@@ -44,8 +48,15 @@
 #define SUBBAND_SAMPLES (SUBFRAMES * SUBSUBFRAMES * 8)
 #define AUBANDS 25
 
+typedef struct CompressionOptions {
+    int adpcm_mode;
+} CompressionOptions;
+
 typedef struct DCAEncContext {
+    AVClass *class;
     PutBitContext pb;
+    DCAADPCMEncContext adpcm_ctx;
+    CompressionOptions options;
     int frame_size;
     int frame_bits;
     int fullband_channels;
@@ -61,10 +72,13 @@ typedef struct DCAEncContext {
     int32_t lfe_peak_cb;
     const int8_t *channel_order_tab;  ///< channel reordering table, lfe and non lfe
 
+    int32_t prediction_mode[MAX_CHANNELS][DCAENC_SUBBANDS];
+    int32_t adpcm_history[MAX_CHANNELS][DCAENC_SUBBANDS][DCA_ADPCM_COEFFS * 2];
     int32_t history[MAX_CHANNELS][512]; /* This is a circular buffer */
-    int32_t subband[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
+    int32_t *subband[MAX_CHANNELS][DCAENC_SUBBANDS];
     int32_t quantized[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
     int32_t peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS];
+    int32_t diff_peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS]; ///< expected peak of residual signal
     int32_t downsampled_lfe[DCA_LFE_SAMPLES];
     int32_t masking_curve_cb[SUBSUBFRAMES][256];
     int32_t bit_allocation_sel[MAX_CHANNELS];
@@ -77,6 +91,7 @@ typedef struct DCAEncContext {
     int32_t worst_quantization_noise;
     int32_t worst_noise_ever;
     int consumed_bits;
+    int consumed_adpcm_bits; ///< Number of bits to transmit ADPCM related info
 } DCAEncContext;
 
 static int32_t cos_table[2048];
@@ -107,18 +122,52 @@ static double gammafilter(int i, double f)
     return 20 * log10(h);
 }
 
+static int subband_bufer_alloc(DCAEncContext *c)
+{
+    int ch, band;
+    int32_t *bufer = av_calloc(MAX_CHANNELS * DCAENC_SUBBANDS *
+                               (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS),
+                               sizeof(int32_t));
+    if (!bufer)
+        return -1;
+
+    /* we need a place for DCA_ADPCM_COEFF samples from previous frame
+     * to calc prediction coefficients for each subband */
+    for (ch = 0; ch < MAX_CHANNELS; ch++) {
+        for (band = 0; band < DCAENC_SUBBANDS; band++) {
+            c->subband[ch][band] = bufer +
+                                   ch * DCAENC_SUBBANDS * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) +
+                                   band * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) + DCA_ADPCM_COEFFS;
+        }
+    }
+    return 0;
+}
+
+static void subband_bufer_free(DCAEncContext *c)
+{
+    int32_t *bufer = c->subband[0][0] - DCA_ADPCM_COEFFS;
+    av_freep(&bufer);
+}
+
 static int encode_init(AVCodecContext *avctx)
 {
     DCAEncContext *c = avctx->priv_data;
     uint64_t layout = avctx->channel_layout;
     int i, j, min_frame_bits;
 
+    if (subband_bufer_alloc(c))
+        return AVERROR(ENOMEM);
+
     c->fullband_channels = c->channels = avctx->channels;
     c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
     c->band_interpolation = band_interpolation[1];
     c->band_spectrum = band_spectrum[1];
     c->worst_quantization_noise = -2047;
     c->worst_noise_ever = -2047;
+    c->consumed_adpcm_bits = 0;
+
+    if (ff_dcaadpcm_init(&c->adpcm_ctx))
+        return AVERROR(ENOMEM);
 
     if (!layout) {
         av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
@@ -150,6 +199,12 @@ static int encode_init(AVCodecContext *avctx)
         }
         /* 6 - no Huffman */
         c->bit_allocation_sel[i] = 6;
+
+        for (j = 0; j < DCAENC_SUBBANDS; j++) {
+            /* -1 - no ADPCM */
+            c->prediction_mode[i][j] = -1;
+            memset(c->adpcm_history[i][j], 0, sizeof(int32_t)*DCA_ADPCM_COEFFS);
+        }
     }
 
     for (i = 0; i < 9; i++) {
@@ -238,6 +293,16 @@ static int encode_init(AVCodecContext *avctx)
     return 0;
 }
 
+static av_cold int encode_close(AVCodecContext *avctx)
+{
+    if (avctx->priv_data) {
+        DCAEncContext *c = avctx->priv_data;
+        subband_bufer_free(c);
+        ff_dcaadpcm_free(&c->adpcm_ctx);
+    }
+    return 0;
+}
+
 static inline int32_t cos_t(int x)
 {
     return cos_table[x & 2047];
@@ -253,12 +318,6 @@ static inline int32_t half32(int32_t a)
     return (a + 1) >> 1;
 }
 
-static inline int32_t mul32(int32_t a, int32_t b)
-{
-    int64_t r = (int64_t)a * b + 0x80000000ULL;
-    return r >> 32;
-}
-
 static void subband_transform(DCAEncContext *c, const int32_t *input)
 {
     int ch, subs, i, k, j;
@@ -545,31 +604,53 @@ static void calc_masking(DCAEncContext *c, const int32_t *input)
     }
 }
 
+static inline int32_t find_peak(const int32_t *in, int len) {
+    int sample;
+    int32_t m = 0;
+    for (sample = 0; sample < len; sample++) {
+        int32_t s = abs(in[sample]);
+        if (m < s) {
+            m = s;
+        }
+    }
+    return get_cb(m);
+}
+
 static void find_peaks(DCAEncContext *c)
 {
     int band, ch;
 
-    for (ch = 0; ch < c->fullband_channels; ch++)
+    for (ch = 0; ch < c->fullband_channels; ch++) {
         for (band = 0; band < 32; band++) {
-            int sample;
+            c->peak_cb[ch][band] = find_peak(c->subband[ch][band], SUBBAND_SAMPLES);
-            int32_t m = 0;
-
-            for (sample = 0; sample < SUBBAND_SAMPLES; sample++) {
-                int32_t s = abs(c->subband[ch][band][sample]);
-                if (m < s)
-                    m = s;
         }
-            c->peak_cb[ch][band] = get_cb(m);
     }
 
     if (c->lfe_channel) {
-        int sample;
+        c->lfe_peak_cb = find_peak(c->downsampled_lfe, DCA_LFE_SAMPLES);
-        int32_t m = 0;
+    }
+}
 
-        for (sample = 0; sample < DCA_LFE_SAMPLES; sample++)
+static void adpcm_analysis(DCAEncContext *c)
-            if (m < abs(c->downsampled_lfe[sample]))
+{
-                m = abs(c->downsampled_lfe[sample]);
+    int ch, band;
-        c->lfe_peak_cb = get_cb(m);
+    int pred_vq_id;
+    int32_t *samples;
+    int32_t estimated_diff[SUBBAND_SAMPLES];
+
+    c->consumed_adpcm_bits = 0;
+    for (ch = 0; ch < c->fullband_channels; ch++) {
+        for (band = 0; band < 32; band++) {
+            samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
+            pred_vq_id = ff_dcaadpcm_subband_analysis(&c->adpcm_ctx, samples, SUBBAND_SAMPLES, estimated_diff);
+            if (pred_vq_id >= 0) {
+                c->prediction_mode[ch][band] = pred_vq_id;
+                c->consumed_adpcm_bits += 12; //12 bits to transmit prediction vq index
+                c->diff_peak_cb[ch][band] = find_peak(estimated_diff, 16);
+            } else {
+                c->prediction_mode[ch][band] = -1;
+            }
+        }
     }
 }
 
@@ -578,13 +659,16 @@ static const int snr_fudge = 128;
 #define USED_NABITS 2
 #define USED_26ABITS 4
 
-static int32_t quantize_value(int32_t value, softfloat quant)
+static inline int32_t get_step_size(const DCAEncContext *c, int ch, int band)
 {
-    int32_t offset = 1 << (quant.e - 1);
+    int32_t step_size;
 
-    value = mul32(value, quant.m) + offset;
+    if (c->bitrate_index == 3)
-    value = value >> quant.e;
+        step_size = ff_dca_lossless_quant[c->abits[ch][band]];
-    return value;
+    else
+        step_size = ff_dca_lossy_quant[c->abits[ch][band]];
+
+    return step_size;
 }
 
 static int calc_one_scale(int32_t peak_cb, int abits, softfloat *quant)
@@ -619,12 +703,38 @@ static int calc_one_scale(int32_t peak_cb, int abits, softfloat *quant)
     return our_nscale;
 }
 
-static void quantize_all(DCAEncContext *c)
+static inline void quantize_adpcm_subband(DCAEncContext *c, int ch, int band)
+{
+    int32_t step_size;
+    int32_t diff_peak_cb = c->diff_peak_cb[ch][band];
+    c->scale_factor[ch][band] = calc_one_scale(diff_peak_cb,
+                                               c->abits[ch][band],
+                                               &c->quant[ch][band]);
+
+    step_size = get_step_size(c, ch, band);
+    ff_dcaadpcm_do_real(c->prediction_mode[ch][band],
+                        c->quant[ch][band], ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], step_size,
+                        c->adpcm_history[ch][band], c->subband[ch][band], c->adpcm_history[ch][band]+4, c->quantized[ch][band],
+                        SUBBAND_SAMPLES, cb_to_level[-diff_peak_cb]);
+}
+
+static void quantize_adpcm(DCAEncContext *c)
+{
+    int band, ch;
+
+    for (ch = 0; ch < c->fullband_channels; ch++)
+        for (band = 0; band < 32; band++)
+            if (c->prediction_mode[ch][band] >= 0)
+                quantize_adpcm_subband(c, ch, band);
+}
+
+static void quantize_pcm(DCAEncContext *c)
 {
     int sample, band, ch;
 
     for (ch = 0; ch < c->fullband_channels; ch++)
         for (band = 0; band < 32; band++)
+            if (c->prediction_mode[ch][band] == -1)
                 for (sample = 0; sample < SUBBAND_SAMPLES; sample++)
                     c->quantized[ch][band][sample] = quantize_value(c->subband[ch][band][sample], c->quant[ch][band]);
 }
@@ -710,6 +820,7 @@ static int init_quantization_noise(DCAEncContext *c, int noise)
     uint32_t bits_counter = 0;
 
     c->consumed_bits = 132 + 333 * c->fullband_channels;
+    c->consumed_bits += c->consumed_adpcm_bits;
     if (c->lfe_channel)
         c->consumed_bits += 72;
 
@@ -740,12 +851,15 @@ static int init_quantization_noise(DCAEncContext *c, int noise)
     /* TODO: May be cache scaled values */
     for (ch = 0; ch < c->fullband_channels; ch++) {
         for (band = 0; band < 32; band++) {
+            if (c->prediction_mode[ch][band] == -1) {
                 c->scale_factor[ch][band] = calc_one_scale(c->peak_cb[ch][band],
                                                            c->abits[ch][band],
                                                            &c->quant[ch][band]);
             }
         }
-    quantize_all(c);
+    }
+    quantize_adpcm(c);
+    quantize_pcm(c);
 
     memset(huff_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * 7 * sizeof(uint32_t));
     memset(clc_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * sizeof(uint32_t));
@@ -819,6 +933,41 @@ static void shift_history(DCAEncContext *c, const int32_t *input)
         }
 }
 
+static void fill_in_adpcm_bufer(DCAEncContext *c)
+{
+     int ch, band;
+     int32_t step_size;
+     /* We fill in ADPCM work buffer for subbands which hasn't been ADPCM coded
+      * in current frame - we need this data if subband of next frame is
+      * ADPCM
+      */
+     for (ch = 0; ch < c->channels; ch++) {
+        for (band = 0; band < 32; band++) {
+            int32_t *samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
+            if (c->prediction_mode[ch][band] == -1) {
+                step_size = get_step_size(c, ch, band);
+
+                ff_dca_core_dequantize(c->adpcm_history[ch][band],
+                                       c->quantized[ch][band]+12, step_size, ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], 0, 4);
+            } else {
+                AV_COPY128U(c->adpcm_history[ch][band], c->adpcm_history[ch][band]+4);
+            }
+            /* Copy dequantized values for LPC analysis.
+             * It reduces artifacts in case of extreme quantization,
+             * example: in current frame abits is 1 and has no prediction flag,
+             * but end of this frame is sine like signal. In this case, if LPC analysis uses
+             * original values, likely LPC analysis returns good prediction gain, and sets prediction flag.
+             * But there are no proper value in decoder history, so likely result will be no good.
+             * Bitstream has "Predictor history flag switch", but this flag disables history for all subbands
+             */
+            samples[0] = c->adpcm_history[ch][band][0] << 7;
+            samples[1] = c->adpcm_history[ch][band][1] << 7;
+            samples[2] = c->adpcm_history[ch][band][2] << 7;
+            samples[3] = c->adpcm_history[ch][band][3] << 7;
+        }
+     }
+}
+
 static void calc_lfe_scales(DCAEncContext *c)
 {
     if (c->lfe_channel)
@@ -1001,9 +1150,14 @@ static void put_subframe(DCAEncContext *c, int subframe)
     /* Prediction mode: no ADPCM, in each channel and subband */
     for (ch = 0; ch < c->fullband_channels; ch++)
         for (band = 0; band < DCAENC_SUBBANDS; band++)
-            put_bits(&c->pb, 1, 0);
+            put_bits(&c->pb, 1, !(c->prediction_mode[ch][band] == -1));
+
+    /* Prediction VQ address */
+    for (ch = 0; ch < c->fullband_channels; ch++)
+        for (band = 0; band < DCAENC_SUBBANDS; band++)
+            if (c->prediction_mode[ch][band] >= 0)
+                put_bits(&c->pb, 12, c->prediction_mode[ch][band]);
 
-    /* Prediction VQ address: not transmitted */
     /* Bit allocation index */
     for (ch = 0; ch < c->fullband_channels; ch++) {
         if (c->bit_allocation_sel[ch] == 6) {
@@ -1068,12 +1222,15 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
         lfe_downsample(c, samples);
 
     calc_masking(c, samples);
+    if (c->options.adpcm_mode)
+        adpcm_analysis(c);
     find_peaks(c);
     assign_bits(c);
     calc_lfe_scales(c);
     shift_history(c, samples);
 
     init_put_bits(&c->pb, avpkt->data, avpkt->size);
+    fill_in_adpcm_bufer(c);
     put_frame_header(c);
     put_primary_audio_header(c);
     for (i = 0; i < SUBFRAMES; i++)
@@ -1092,6 +1249,20 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
     return 0;
 }
 
+#define DCAENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
+
+static const AVOption options[] = {
+    { "dca_adpcm", "Use ADPCM encoding", offsetof(DCAEncContext, options.adpcm_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, DCAENC_FLAGS },
+    { NULL },
+};
+
+static const AVClass dcaenc_class = {
+    .class_name = "DCA (DTS Coherent Acoustics)",
+    .item_name = av_default_item_name,
+    .option = options,
+    .version = LIBAVUTIL_VERSION_INT,
+};
+
 static const AVCodecDefault defaults[] = {
     { "b",          "1411200" },
     { NULL },
@@ -1104,6 +1275,7 @@ AVCodec ff_dca_encoder = {
     .id                    = AV_CODEC_ID_DTS,
     .priv_data_size        = sizeof(DCAEncContext),
     .init                  = encode_init,
+    .close                 = encode_close,
     .encode2               = encode_frame,
     .capabilities          = AV_CODEC_CAP_EXPERIMENTAL,
     .sample_fmts           = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32,
@@ -1116,4 +1288,5 @@ AVCodec ff_dca_encoder = {
                                                   AV_CH_LAYOUT_5POINT1,
                                                   0 },
     .defaults              = defaults,
+    .priv_class            = &dcaenc_class,
 };

libavcodec/dcaenc.h

@@ -24,6 +24,8 @@
 
 #include <stdint.h>
 
+#include "dcamath.h"
+
 typedef struct {
     int32_t m;
     int32_t e;
@@ -144,4 +146,13 @@ static const int8_t channel_reorder_nolfe[16][9] = {
     { 3,  2,  4,  0,  1,  5,  7,  6, -1 },
 };
 
+static inline int32_t quantize_value(int32_t value, softfloat quant)
+{
+    int32_t offset = 1 << (quant.e - 1);
+
+    value = mul32(value, quant.m) + offset;
+    value = value >> quant.e;
+    return value;
+}
+
 #endif /* AVCODEC_DCAENC_H */

libavcodec/dcamath.h

@@ -49,6 +49,7 @@ static inline int32_t mul17(int32_t a, int32_t b) { return mul__(a, b, 17); }
 static inline int32_t mul22(int32_t a, int32_t b) { return mul__(a, b, 22); }
 static inline int32_t mul23(int32_t a, int32_t b) { return mul__(a, b, 23); }
 static inline int32_t mul31(int32_t a, int32_t b) { return mul__(a, b, 31); }
+static inline int32_t mul32(int32_t a, int32_t b) { return mul__(a, b, 32); }
 
 static inline int32_t clip23(int32_t a) { return av_clip_intp2(a, 23); }