From e96c467499a26614bc7b5f79056ffb599ff616d7 Mon Sep 17 00:00:00 2001
From: Daniil Cherednik <dan.cherednik@gmail.com>
Date: Sun, 8 Jan 2017 00:16:49 +0300
Subject: [PATCH 3/3] avcodec/dcaenc: Implementation of Huffman codes for DCA
encoder
To: ffmpeg-devel@ffmpeg.org
---
libavcodec/Makefile | 2 +-
libavcodec/dca_core.c | 16 ++---
libavcodec/dcadata.c | 8 +++
libavcodec/dcadata.h | 5 ++
libavcodec/dcaenc.c | 159 ++++++++++++++++++++++++++++++++++++++------------
libavcodec/dcaenc.h | 1 -
libavcodec/dcahuff.c | 22 +++++++
libavcodec/dcahuff.h | 3 +
tests/fate/acodec.mak | 2 +-
9 files changed, 166 insertions(+), 52 deletions(-)
@@ -234,7 +234,7 @@ 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-$(CONFIG_DCA_ENCODER) += dcaenc.o dca.o dcadata.o
+OBJS-$(CONFIG_DCA_ENCODER) += dcaenc.o dca.o dcadata.o dcahuff.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
@@ -92,14 +92,6 @@ static const uint8_t block_code_nbits[7] = {
7, 10, 12, 13, 15, 17, 19
};
-static const uint8_t quant_index_sel_nbits[DCA_CODE_BOOKS] = {
- 1, 2, 2, 2, 2, 3, 3, 3, 3, 3
-};
-
-static const uint8_t quant_index_group_size[DCA_CODE_BOOKS] = {
- 1, 3, 3, 3, 3, 7, 7, 7, 7, 7
-};
-
static int dca_get_vlc(GetBitContext *s, DCAVLC *v, int i)
{
return get_vlc2(s, v->vlc[i].table, v->vlc[i].bits, v->max_depth) + v->offset;
@@ -400,12 +392,12 @@ static int parse_coding_header(DCACoreDecoder *s, enum HeaderType header, int xc
// Quantization index codebook select
for (n = 0; n < DCA_CODE_BOOKS; n++)
for (ch = xch_base; ch < s->nchannels; ch++)
- s->quant_index_sel[ch][n] = get_bits(&s->gb, quant_index_sel_nbits[n]);
+ s->quant_index_sel[ch][n] = get_bits(&s->gb, ff_dca_quant_index_sel_nbits[n]);
// Scale factor adjustment index
for (n = 0; n < DCA_CODE_BOOKS; n++)
for (ch = xch_base; ch < s->nchannels; ch++)
- if (s->quant_index_sel[ch][n] < quant_index_group_size[n])
+ if (s->quant_index_sel[ch][n] < ff_dca_quant_index_group_size[n])
s->scale_factor_adj[ch][n] = ff_dca_scale_factor_adj[get_bits(&s->gb, 2)];
if (header == HEADER_XXCH) {
@@ -663,7 +655,7 @@ static inline int extract_audio(DCACoreDecoder *s, int32_t *audio, int abits, in
if (abits <= DCA_CODE_BOOKS) {
int sel = s->quant_index_sel[ch][abits - 1];
- if (sel < quant_index_group_size[abits - 1]) {
+ if (sel < ff_dca_quant_index_group_size[abits - 1]) {
// Huffman codes
return parse_huffman_codes(s, audio, abits, sel);
}
@@ -1562,7 +1554,7 @@ static int parse_x96_coding_header(DCACoreDecoder *s, int exss, int xch_base)
// Quantization index codebook select
for (n = 0; n < 6 + 4 * s->x96_high_res; n++)
for (ch = xch_base; ch < s->x96_nchannels; ch++)
- s->quant_index_sel[ch][n] = get_bits(&s->gb, quant_index_sel_nbits[n]);
+ s->quant_index_sel[ch][n] = get_bits(&s->gb, ff_dca_quant_index_sel_nbits[n]);
if (exss) {
// Reserved
@@ -50,6 +50,14 @@ const uint8_t ff_dca_dmix_primary_nch[8] = {
1, 2, 2, 3, 3, 4, 4, 0
};
+const uint8_t ff_dca_quant_index_sel_nbits[DCA_CODE_BOOKS] = {
+ 1, 2, 2, 2, 2, 3, 3, 3, 3, 3
+};
+
+const uint8_t ff_dca_quant_index_group_size[DCA_CODE_BOOKS] = {
+ 1, 3, 3, 3, 3, 7, 7, 7, 7, 7
+};
+
/* ADPCM data */
/* 16 bits signed fractional Q13 binary codes */
@@ -23,6 +23,8 @@
#include <stdint.h>
+#include "dcahuff.h"
+
extern const uint32_t ff_dca_bit_rates[32];
extern const uint8_t ff_dca_channels[16];
@@ -31,6 +33,9 @@ extern const uint8_t ff_dca_bits_per_sample[8];
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 uint32_t ff_dca_scale_factor_quant6[64];
@@ -70,6 +70,7 @@ typedef struct DCAEncContext {
int abits[MAX_CHANNELS][DCAENC_SUBBANDS];
int scale_factor[MAX_CHANNELS][DCAENC_SUBBANDS];
softfloat quant[MAX_CHANNELS][DCAENC_SUBBANDS];
+ int32_t quant_index_sel[MAX_CHANNELS][DCA_CODE_BOOKS];
int32_t eff_masking_curve_cb[256];
int32_t band_masking_cb[32];
int32_t worst_quantization_noise;
@@ -109,7 +110,7 @@ static int encode_init(AVCodecContext *avctx)
{
DCAEncContext *c = avctx->priv_data;
uint64_t layout = avctx->channel_layout;
- int i, min_frame_bits;
+ int i, j, min_frame_bits;
c->fullband_channels = c->channels = avctx->channels;
c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
@@ -142,6 +143,12 @@ static int encode_init(AVCodecContext *avctx)
c->channel_order_tab = channel_reorder_nolfe[c->channel_config];
}
+ for (i = 0; i < MAX_CHANNELS; i++) {
+ for (j = 0; j < DCA_CODE_BOOKS; j++) {
+ c->quant_index_sel[i][j] = ff_dca_quant_index_group_size[j];
+ }
+ }
+
for (i = 0; i < 9; i++) {
if (sample_rates[i] == avctx->sample_rate)
break;
@@ -619,9 +626,58 @@ static void quantize_all(DCAEncContext *c)
c->quantized[ch][band][sample] = quantize_value(c->subband[ch][band][sample], c->quant[ch][band]);
}
+static void accumulate_huff_bit_consumption(int abits, int32_t *quantized, uint32_t *result)
+{
+ uint8_t sel, id = abits - 1;
+ for (sel = 0; sel < ff_dca_quant_index_group_size[id]; sel++)
+ result[sel] += ff_dca_vlc_calc_quant_bits(quantized, SUBBAND_SAMPLES, sel, id);
+}
+
+static uint32_t set_best_code(uint32_t vlc_bits[DCA_CODE_BOOKS][7], uint32_t clc_bits[DCA_CODE_BOOKS], int32_t res[DCA_CODE_BOOKS])
+{
+ uint8_t i, sel;
+ uint32_t best_sel_bits[DCA_CODE_BOOKS];
+ int32_t best_sel_id[DCA_CODE_BOOKS];
+ uint32_t t, bits = 0;
+
+ for (i = 0; i < DCA_CODE_BOOKS; i++) {
+
+ av_assert0(!((!!vlc_bits[i][0]) ^ (!!clc_bits[i])));
+ if (vlc_bits[i][0] == 0) {
+ /* do not transmit adjustment index for empty codebooks */
+ res[i] = ff_dca_quant_index_group_size[i];
+ /* and skip it */
+ continue;
+ }
+
+ best_sel_bits[i] = vlc_bits[i][0];
+ best_sel_id[i] = 0;
+ for (sel = 0; sel < ff_dca_quant_index_group_size[i]; sel++) {
+ if (best_sel_bits[i] > vlc_bits[i][sel] && vlc_bits[i][sel]) {
+ best_sel_bits[i] = vlc_bits[i][sel];
+ best_sel_id[i] = sel;
+ }
+ }
+
+ /* 2 bits to transmit scale factor adjustment index */
+ t = best_sel_bits[i] + 2;
+ if (t < clc_bits[i]) {
+ res[i] = best_sel_id[i];
+ bits += t;
+ } else {
+ res[i] = ff_dca_quant_index_group_size[i];
+ bits += clc_bits[i];
+ }
+ }
+ return bits;
+}
+
static int init_quantization_noise(DCAEncContext *c, int noise)
{
int ch, band, ret = 0;
+ uint32_t huff_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS][7];
+ uint32_t clc_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS];
+ uint32_t bits_counter = 0;
c->consumed_bits = 132 + 493 * c->fullband_channels;
if (c->lfe_channel)
@@ -648,10 +704,36 @@ static int init_quantization_noise(DCAEncContext *c, int noise)
}
}
- for (ch = 0; ch < c->fullband_channels; ch++)
+ /* Recalc scale_factor each time to get bits consumption in case of Huffman coding.
+ It is suboptimal solution */
+ /* TODO: May be cache scaled values */
+ for (ch = 0; ch < c->fullband_channels; ch++) {
+ for (band = 0; band < 32; band++) {
+ c->scale_factor[ch][band] = calc_one_scale(c->peak_cb[ch][band],
+ c->abits[ch][band],
+ &c->quant[ch][band]);
+ }
+ }
+ quantize_all(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));
+ for (ch = 0; ch < c->fullband_channels; ch++) {
for (band = 0; band < 32; band++) {
- c->consumed_bits += bit_consumption[c->abits[ch][band]];
+ if (c->abits[ch][band] && c->abits[ch][band] <= DCA_CODE_BOOKS) {
+ accumulate_huff_bit_consumption(c->abits[ch][band], c->quantized[ch][band], huff_bit_count_accum[ch][c->abits[ch][band] - 1]);
+ clc_bit_count_accum[ch][c->abits[ch][band] - 1] += bit_consumption[c->abits[ch][band]];
+ } else {
+ bits_counter += bit_consumption[c->abits[ch][band]];
+ }
}
+ }
+
+ for (ch = 0; ch < c->fullband_channels; ch++) {
+ bits_counter += set_best_code(huff_bit_count_accum[ch], clc_bit_count_accum[ch], c->quant_index_sel[ch]);
+ }
+
+ c->consumed_bits += bits_counter;
return ret;
}
@@ -706,16 +788,8 @@ static void shift_history(DCAEncContext *c, const int32_t *input)
}
}
-static void calc_scales(DCAEncContext *c)
+static void calc_lfe_scales(DCAEncContext *c)
{
- int band, ch;
-
- for (ch = 0; ch < c->fullband_channels; ch++)
- for (band = 0; band < 32; band++)
- c->scale_factor[ch][band] = calc_one_scale(c->peak_cb[ch][band],
- c->abits[ch][band],
- &c->quant[ch][band]);
-
if (c->lfe_channel)
c->lfe_scale_factor = calc_one_scale(c->lfe_peak_cb, 11, &c->lfe_quant);
}
@@ -805,9 +879,6 @@ static void put_frame_header(DCAEncContext *c)
static void put_primary_audio_header(DCAEncContext *c)
{
- static const int bitlen[11] = { 0, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3 };
- static const int thr[11] = { 0, 1, 3, 3, 3, 3, 7, 7, 7, 7, 7 };
-
int ch, i;
/* Number of subframes */
put_bits(&c->pb, 4, SUBFRAMES - 1);
@@ -839,36 +910,51 @@ static void put_primary_audio_header(DCAEncContext *c)
for (ch = 0; ch < c->fullband_channels; ch++)
put_bits(&c->pb, 3, 6);
- /* Quantization index codebook select: dummy data
- to avoid transmission of scale factor adjustment */
- for (i = 1; i < 11; i++)
+ /* Quantization index codebook select */
+ for (i = 0; i < DCA_CODE_BOOKS; i++)
for (ch = 0; ch < c->fullband_channels; ch++)
- put_bits(&c->pb, bitlen[i], thr[i]);
+ put_bits(&c->pb, ff_dca_quant_index_sel_nbits[i], c->quant_index_sel[ch][i]);
+
+ /* Scale factor adjustment index: transmitted in case of Huffman coding */
+ for (i = 0; i < DCA_CODE_BOOKS; i++)
+ for (ch = 0; ch < c->fullband_channels; ch++)
+ if (c->quant_index_sel[ch][i] < ff_dca_quant_index_group_size[i])
+ put_bits(&c->pb, 2, 0);
- /* Scale factor adjustment index: not transmitted */
/* Audio header CRC check word: not transmitted */
}
static void put_subframe_samples(DCAEncContext *c, int ss, int band, int ch)
{
- if (c->abits[ch][band] <= 7) {
- int sum, i, j;
- for (i = 0; i < 8; i += 4) {
- sum = 0;
- for (j = 3; j >= 0; j--) {
- sum *= ff_dca_quant_levels[c->abits[ch][band]];
- sum += c->quantized[ch][band][ss * 8 + i + j];
- sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
- }
- put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
+ int i, j, sum, bits, sel;
+ if (c->abits[ch][band] <= DCA_CODE_BOOKS) {
+ av_assert0(c->abits[ch][band] > 0);
+ sel = c->quant_index_sel[ch][c->abits[ch][band] - 1];
+ // Huffman codes
+ if (sel < ff_dca_quant_index_group_size[c->abits[ch][band] - 1]) {
+ ff_dca_vlc_enc_quant(&c->pb, &c->quantized[ch][band][ss * 8], 8, sel, c->abits[ch][band] - 1);
+ return;
}
- } else {
- int i;
- for (i = 0; i < 8; i++) {
- int bits = bit_consumption[c->abits[ch][band]] / 16;
- put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
+
+ // Block codes
+ if (c->abits[ch][band] <= 7) {
+ for (i = 0; i < 8; i += 4) {
+ sum = 0;
+ for (j = 3; j >= 0; j--) {
+ sum *= ff_dca_quant_levels[c->abits[ch][band]];
+ sum += c->quantized[ch][band][ss * 8 + i + j];
+ sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
+ }
+ put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
+ }
+ return;
}
}
+
+ for (i = 0; i < 8; i++) {
+ bits = bit_consumption[c->abits[ch][band]] / 16;
+ put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
+ }
}
static void put_subframe(DCAEncContext *c, int subframe)
@@ -947,8 +1033,7 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
calc_masking(c, samples);
find_peaks(c);
assign_bits(c);
- calc_scales(c);
- quantize_all(c);
+ calc_lfe_scales(c);
shift_history(c, samples);
init_put_bits(&c->pb, avpkt->data, avpkt->size);
@@ -95,7 +95,6 @@ static const softfloat scalefactor_inv[128] = {
/* manually derived from
* Table B.5: Selection of quantization levels and codebooks
- * FIXME: will become invalid when Huffman codes are introduced.
*/
static const int bit_consumption[27] = {
-8, 28, 40, 48, 52, 60, 68, 76, 80, 96,
@@ -1335,3 +1335,25 @@ av_cold void ff_dca_init_vlcs(void)
vlcs_initialized = 1;
}
+
+uint32_t ff_dca_vlc_calc_quant_bits(int *values, uint8_t n, uint8_t sel, uint8_t table)
+{
+ uint8_t i, id;
+ uint32_t sum = 0;
+ for (i = 0; i < n; i++) {
+ id = values[i] - bitalloc_offsets[table];
+ av_assert0(id < bitalloc_sizes[table]);
+ sum += bitalloc_bits[table][sel][id];
+ }
+ return sum;
+}
+
+void ff_dca_vlc_enc_quant(PutBitContext *pb, int *values, uint8_t n, uint8_t sel, uint8_t table)
+{
+ uint8_t i, id;
+ for (i = 0; i < n; i++) {
+ id = values[i] - bitalloc_offsets[table];
+ av_assert0(id < bitalloc_sizes[table]);
+ put_bits(pb, bitalloc_bits[table][sel][id], bitalloc_codes[table][sel][id]);
+ }
+}
@@ -27,6 +27,7 @@
#include "avcodec.h"
#include "get_bits.h"
+#include "put_bits.h"
#define DCA_CODE_BOOKS 10
@@ -55,5 +56,7 @@ extern VLC ff_dca_vlc_grid_3;
extern VLC ff_dca_vlc_rsd;
av_cold void ff_dca_init_vlcs(void);
+uint32_t ff_dca_vlc_calc_quant_bits(int *values, uint8_t n, uint8_t sel, uint8_t abits);
+void ff_dca_vlc_enc_quant(PutBitContext *pb, int *values, uint8_t n, uint8_t sel, uint8_t abits);
#endif /* AVCODEC_DCAHUFF_H */
@@ -104,7 +104,7 @@ fate-acodec-dca: tests/data/asynth-44100-2.wav
fate-acodec-dca: SRC = tests/data/asynth-44100-2.wav
fate-acodec-dca: CMD = md5 -i $(TARGET_PATH)/$(SRC) -c:a dca -strict -2 -f dts -flags +bitexact
fate-acodec-dca: CMP = oneline
-fate-acodec-dca: REF = 7ffdefdf47069289990755c79387cc90
+fate-acodec-dca: REF = 7cd79a3717943a06b217f1130223a86f
FATE_ACODEC-$(call ENCDEC, DCA, WAV) += fate-acodec-dca2
fate-acodec-dca2: CMD = enc_dec_pcm dts wav s16le $(SRC) -c:a dca -strict -2 -flags +bitexact
--
2.1.4