diff options
author | Rostislav Pehlivanov <atomnuker@gmail.com> | 2017-12-04 08:55:45 +0000 |
---|---|---|
committer | Rostislav Pehlivanov <atomnuker@gmail.com> | 2017-12-04 10:51:19 +0000 |
commit | ce87e630fa009d0bf78cd7a7599e6efc221e6543 (patch) | |
tree | a0dab6973b6b5c9b0ff994b28fc51afcaac0152c | |
parent | 86fda8be3f3892c48474a319e0ef7509dc137e3e (diff) | |
download | ffmpeg-ce87e630fa009d0bf78cd7a7599e6efc221e6543.tar.gz |
opus_celt: deduplicate band quantization/dequantization function
No point in having the same code twice to do exactly the same thing.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
-rw-r--r-- | libavcodec/opus.c | 111 | ||||
-rw-r--r-- | libavcodec/opus.h | 3 | ||||
-rw-r--r-- | libavcodec/opus_celt.c | 114 | ||||
-rw-r--r-- | libavcodec/opus_pvq.c | 58 | ||||
-rw-r--r-- | libavcodec/opus_pvq.h | 6 | ||||
-rw-r--r-- | libavcodec/opusenc.c | 110 | ||||
-rw-r--r-- | libavcodec/opusenc.h | 2 | ||||
-rw-r--r-- | libavcodec/opusenc_psy.c | 2 |
8 files changed, 162 insertions, 244 deletions
diff --git a/libavcodec/opus.c b/libavcodec/opus.c index 5847e88e11..b791d749b4 100644 --- a/libavcodec/opus.c +++ b/libavcodec/opus.c @@ -29,7 +29,8 @@ #include "libavutil/error.h" #include "libavutil/ffmath.h" -#include "opus.h" +#include "opus_celt.h" +#include "opustab.h" #include "vorbis.h" static const uint16_t opus_frame_duration[32] = { @@ -438,3 +439,111 @@ av_cold int ff_opus_parse_extradata(AVCodecContext *avctx, return 0; } + +void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc) +{ + float lowband_scratch[8 * 22]; + float norm1[2 * 8 * 100]; + float *norm2 = norm1 + 8 * 100; + + int totalbits = (f->framebits << 3) - f->anticollapse_needed; + + int update_lowband = 1; + int lowband_offset = 0; + + int i, j; + + for (i = f->start_band; i < f->end_band; i++) { + uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 }; + int band_offset = ff_celt_freq_bands[i] << f->size; + int band_size = ff_celt_freq_range[i] << f->size; + float *X = f->block[0].coeffs + band_offset; + float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL; + float *norm_loc1, *norm_loc2; + + int consumed = opus_rc_tell_frac(rc); + int effective_lowband = -1; + int b = 0; + + /* Compute how many bits we want to allocate to this band */ + if (i != f->start_band) + f->remaining -= consumed; + f->remaining2 = totalbits - consumed - 1; + if (i <= f->coded_bands - 1) { + int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i); + b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14); + } + + if ((ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] || + i == f->start_band + 1) && (update_lowband || lowband_offset == 0)) + lowband_offset = i; + + if (i == f->start_band + 1) { + /* Special Hybrid Folding (RFC 8251 section 9). Copy the first band into + the second to ensure the second band never has to use the LCG. */ + int offset = 8 * ff_celt_freq_bands[i]; + int count = 8 * (ff_celt_freq_range[i] - ff_celt_freq_range[i-1]); + + memcpy(&norm1[offset], &norm1[offset - count], count * sizeof(float)); + + if (f->channels == 2) + memcpy(&norm2[offset], &norm2[offset - count], count * sizeof(float)); + } + + /* Get a conservative estimate of the collapse_mask's for the bands we're + going to be folding from. */ + if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE || + f->blocks > 1 || f->tf_change[i] < 0)) { + int foldstart, foldend; + + /* This ensures we never repeat spectral content within one band */ + effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band], + ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]); + foldstart = lowband_offset; + while (ff_celt_freq_bands[--foldstart] > effective_lowband); + foldend = lowband_offset - 1; + while (++foldend < i && ff_celt_freq_bands[foldend] < effective_lowband + ff_celt_freq_range[i]); + + cm[0] = cm[1] = 0; + for (j = foldstart; j < foldend; j++) { + cm[0] |= f->block[0].collapse_masks[j]; + cm[1] |= f->block[f->channels - 1].collapse_masks[j]; + } + } + + if (f->dual_stereo && i == f->intensity_stereo) { + /* Switch off dual stereo to do intensity */ + f->dual_stereo = 0; + for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++) + norm1[j] = (norm1[j] + norm2[j]) / 2; + } + + norm_loc1 = effective_lowband != -1 ? norm1 + (effective_lowband << f->size) : NULL; + norm_loc2 = effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL; + + if (f->dual_stereo) { + cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, NULL, band_size, b >> 1, + f->blocks, norm_loc1, f->size, + norm1 + band_offset, 0, 1.0f, + lowband_scratch, cm[0]); + + cm[1] = f->pvq->quant_band(f->pvq, f, rc, i, Y, NULL, band_size, b >> 1, + f->blocks, norm_loc2, f->size, + norm2 + band_offset, 0, 1.0f, + lowband_scratch, cm[1]); + } else { + cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, Y, band_size, b >> 0, + f->blocks, norm_loc1, f->size, + norm1 + band_offset, 0, 1.0f, + lowband_scratch, cm[0] | cm[1]); + cm[1] = cm[0]; + } + + f->block[0].collapse_masks[i] = (uint8_t)cm[0]; + f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1]; + f->remaining += f->pulses[i] + consumed; + + /* Update the folding position only as long as we have 1 bit/sample depth */ + update_lowband = (b > band_size << 3); + } +} diff --git a/libavcodec/opus.h b/libavcodec/opus.h index 90b87ba5c4..a10db7f0e8 100644 --- a/libavcodec/opus.h +++ b/libavcodec/opus.h @@ -191,4 +191,7 @@ int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc, enum OpusBandwidth bandwidth, int coded_channels, int duration_ms); +/* Encode or decode CELT bands */ +void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc); + #endif /* AVCODEC_OPUS_H */ diff --git a/libavcodec/opus_celt.c b/libavcodec/opus_celt.c index 72b299a19c..ff74e2f067 100644 --- a/libavcodec/opus_celt.c +++ b/libavcodec/opus_celt.c @@ -676,110 +676,6 @@ static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X) } } -static void celt_decode_bands(CeltFrame *f, OpusRangeCoder *rc) -{ - float lowband_scratch[8 * 22]; - float norm[2 * 8 * 100]; - - int totalbits = (f->framebits << 3) - f->anticollapse_needed; - - int update_lowband = 1; - int lowband_offset = 0; - - int i, j; - - memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs)); - memset(f->block[1].coeffs, 0, sizeof(f->block[0].coeffs)); - - for (i = f->start_band; i < f->end_band; i++) { - uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 }; - int band_offset = ff_celt_freq_bands[i] << f->size; - int band_size = ff_celt_freq_range[i] << f->size; - float *X = f->block[0].coeffs + band_offset; - float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL; - - int consumed = opus_rc_tell_frac(rc); - float *norm2 = norm + 8 * 100; - int effective_lowband = -1; - int b = 0; - - /* Compute how many bits we want to allocate to this band */ - if (i != f->start_band) - f->remaining -= consumed; - f->remaining2 = totalbits - consumed - 1; - if (i <= f->coded_bands - 1) { - int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i); - b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14); - } - - if ((ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] || - i == f->start_band + 1) && (update_lowband || lowband_offset == 0)) - lowband_offset = i; - - if (i == f->start_band + 1) { - /* Special Hybrid Folding (RFC 8251 section 9). Copy the first band into - the second to ensure the second band never has to use the LCG. */ - int offset = 8 * ff_celt_freq_bands[i]; - int count = 8 * (ff_celt_freq_range[i] - ff_celt_freq_range[i-1]); - - memcpy(&norm[offset], &norm[offset - count], count * sizeof(float)); - - if (f->channels == 2) - memcpy(&norm2[offset], &norm2[offset - count], count * sizeof(float)); - } - - /* Get a conservative estimate of the collapse_mask's for the bands we're - going to be folding from. */ - if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE || - f->blocks > 1 || f->tf_change[i] < 0)) { - int foldstart, foldend; - - /* This ensures we never repeat spectral content within one band */ - effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band], - ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]); - foldstart = lowband_offset; - while (ff_celt_freq_bands[--foldstart] > effective_lowband); - foldend = lowband_offset - 1; - while (++foldend < i && ff_celt_freq_bands[foldend] < effective_lowband + ff_celt_freq_range[i]); - - cm[0] = cm[1] = 0; - for (j = foldstart; j < foldend; j++) { - cm[0] |= f->block[0].collapse_masks[j]; - cm[1] |= f->block[f->channels - 1].collapse_masks[j]; - } - } - - if (f->dual_stereo && i == f->intensity_stereo) { - /* Switch off dual stereo to do intensity */ - f->dual_stereo = 0; - for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++) - norm[j] = (norm[j] + norm2[j]) / 2; - } - - if (f->dual_stereo) { - cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, NULL, band_size, b / 2, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]); - - cm[1] = f->pvq->decode_band(f->pvq, f, rc, i, Y, NULL, band_size, b/2, f->blocks, - effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size, - norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]); - } else { - cm[0] = f->pvq->decode_band(f->pvq, f, rc, i, X, Y, band_size, b, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]|cm[1]); - cm[1] = cm[0]; - } - - f->block[0].collapse_masks[i] = (uint8_t)cm[0]; - f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1]; - f->remaining += f->pulses[i] + consumed; - - /* Update the folding position only as long as we have 1 bit/sample depth */ - update_lowband = (b > band_size << 3); - } -} - int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, float **output, int channels, int frame_size, int start_band, int end_band) @@ -819,8 +715,10 @@ int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, if (!f->output_channels) f->output_channels = channels; - memset(f->block[0].collapse_masks, 0, sizeof(f->block[0].collapse_masks)); - memset(f->block[1].collapse_masks, 0, sizeof(f->block[1].collapse_masks)); + for (i = 0; i < f->channels; i++) { + memset(f->block[i].coeffs, 0, sizeof(f->block[i].coeffs)); + memset(f->block[i].collapse_masks, 0, sizeof(f->block[i].collapse_masks)); + } consumed = opus_rc_tell(rc); @@ -857,7 +755,7 @@ int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc, celt_decode_tf_changes (f, rc); celt_decode_allocation (f, rc); celt_decode_fine_energy (f, rc); - celt_decode_bands (f, rc); + ff_celt_quant_bands (f, rc); if (f->anticollapse_needed) f->anticollapse = ff_opus_rc_get_raw(rc, 1); @@ -1021,7 +919,7 @@ int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels, if ((ret = ff_mdct15_init(&frm->imdct[i], 1, i + 3, -1.0f/32768)) < 0) goto fail; - if ((ret = ff_celt_pvq_init(&frm->pvq)) < 0) + if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0) goto fail; frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT); diff --git a/libavcodec/opus_pvq.c b/libavcodec/opus_pvq.c index 449215f814..f8668de379 100644 --- a/libavcodec/opus_pvq.c +++ b/libavcodec/opus_pvq.c @@ -486,8 +486,7 @@ static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, int duration, float *lowband_out, int level, float gain, float *lowband_scratch, - int fill, int quant, - QUANT_FN(*rec)) + int fill, int quant) { int i; const uint8_t *cache; @@ -700,8 +699,8 @@ static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, sign = 1 - 2 * sign; /* We use orig_fill here because we want to fold the side, but if itheta==16384, we'll have cleared the low bits of fill. */ - cm = rec(pvq, f, rc, band, x2, NULL, N, mbits, blocks, lowband, duration, - lowband_out, level, gain, lowband_scratch, orig_fill); + cm = pvq->quant_band(pvq, f, rc, band, x2, NULL, N, mbits, blocks, lowband, duration, + lowband_out, level, gain, lowband_scratch, orig_fill); /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse), and there's no need to worry about mixing with the other channel. */ y2[0] = -sign * x2[1]; @@ -753,24 +752,25 @@ static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, if (mbits >= sbits) { /* In stereo mode, we do not apply a scaling to the mid * because we need the normalized mid for folding later */ - cm = rec(pvq, f, rc, band, X, NULL, N, mbits, blocks, lowband, - duration, next_lowband_out1, next_level, - stereo ? 1.0f : (gain * mid), lowband_scratch, fill); + cm = pvq->quant_band(pvq, f, rc, band, X, NULL, N, mbits, blocks, + lowband, duration, next_lowband_out1, next_level, + stereo ? 1.0f : (gain * mid), lowband_scratch, fill); rebalance = mbits - (rebalance - f->remaining2); if (rebalance > 3 << 3 && itheta != 0) sbits += rebalance - (3 << 3); /* For a stereo split, the high bits of fill are always zero, * so no folding will be done to the side. */ - cmt = rec(pvq, f, rc, band, Y, NULL, N, sbits, blocks, next_lowband2, - duration, NULL, next_level, gain * side, NULL, - fill >> blocks); + cmt = pvq->quant_band(pvq, f, rc, band, Y, NULL, N, sbits, blocks, + next_lowband2, duration, NULL, next_level, + gain * side, NULL, fill >> blocks); cm |= cmt << ((B0 >> 1) & (stereo - 1)); } else { /* For a stereo split, the high bits of fill are always zero, * so no folding will be done to the side. */ - cm = rec(pvq, f, rc, band, Y, NULL, N, sbits, blocks, next_lowband2, - duration, NULL, next_level, gain * side, NULL, fill >> blocks); + cm = pvq->quant_band(pvq, f, rc, band, Y, NULL, N, sbits, blocks, + next_lowband2, duration, NULL, next_level, + gain * side, NULL, fill >> blocks); cm <<= ((B0 >> 1) & (stereo - 1)); rebalance = sbits - (rebalance - f->remaining2); if (rebalance > 3 << 3 && itheta != 16384) @@ -778,9 +778,9 @@ static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, /* In stereo mode, we do not apply a scaling to the mid because * we need the normalized mid for folding later */ - cm |= rec(pvq, f, rc, band, X, NULL, N, mbits, blocks, lowband, duration, - next_lowband_out1, next_level, stereo ? 1.0f : (gain * mid), - lowband_scratch, fill); + cm |= pvq->quant_band(pvq, f, rc, band, X, NULL, N, mbits, blocks, + lowband, duration, next_lowband_out1, next_level, + stereo ? 1.0f : (gain * mid), lowband_scratch, fill); } } } else { @@ -874,19 +874,16 @@ static av_always_inline uint32_t quant_band_template(CeltPVQ *pvq, CeltFrame *f, return cm; } - static QUANT_FN(pvq_decode_band) { return quant_band_template(pvq, f, rc, band, X, Y, N, b, blocks, lowband, duration, - lowband_out, level, gain, lowband_scratch, fill, 0, - pvq->decode_band); + lowband_out, level, gain, lowband_scratch, fill, 0); } static QUANT_FN(pvq_encode_band) { return quant_band_template(pvq, f, rc, band, X, Y, N, b, blocks, lowband, duration, - lowband_out, level, gain, lowband_scratch, fill, 1, - pvq->encode_band); + lowband_out, level, gain, lowband_scratch, fill, 1); } static float pvq_band_cost(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int band, @@ -914,14 +911,14 @@ static float pvq_band_cost(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int b } if (f->dual_stereo) { - pvq->encode_band(pvq, f, rc, band, X, NULL, band_size, b / 2, f->blocks, NULL, - f->size, norm1, 0, 1.0f, lowband_scratch, cm[0]); + pvq->quant_band(pvq, f, rc, band, X, NULL, band_size, b / 2, f->blocks, NULL, + f->size, norm1, 0, 1.0f, lowband_scratch, cm[0]); - pvq->encode_band(pvq, f, rc, band, Y, NULL, band_size, b / 2, f->blocks, NULL, - f->size, norm2, 0, 1.0f, lowband_scratch, cm[1]); + pvq->quant_band(pvq, f, rc, band, Y, NULL, band_size, b / 2, f->blocks, NULL, + f->size, norm2, 0, 1.0f, lowband_scratch, cm[1]); } else { - pvq->encode_band(pvq, f, rc, band, X, Y, band_size, b, f->blocks, NULL, f->size, - norm1, 0, 1.0f, lowband_scratch, cm[0] | cm[1]); + pvq->quant_band(pvq, f, rc, band, X, Y, band_size, b, f->blocks, NULL, f->size, + norm1, 0, 1.0f, lowband_scratch, cm[0] | cm[1]); } for (i = 0; i < band_size; i++) { @@ -939,16 +936,15 @@ static float pvq_band_cost(CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int b return lambda*dist*cost; } -int av_cold ff_celt_pvq_init(CeltPVQ **pvq) +int av_cold ff_celt_pvq_init(CeltPVQ **pvq, int encode) { CeltPVQ *s = av_malloc(sizeof(CeltPVQ)); if (!s) return AVERROR(ENOMEM); - s->pvq_search = ppp_pvq_search_c; - s->decode_band = pvq_decode_band; - s->encode_band = pvq_encode_band; - s->band_cost = pvq_band_cost; + s->pvq_search = ppp_pvq_search_c; + s->quant_band = encode ? pvq_encode_band : pvq_decode_band; + s->band_cost = pvq_band_cost; if (ARCH_X86) ff_opus_dsp_init_x86(s); diff --git a/libavcodec/opus_pvq.h b/libavcodec/opus_pvq.h index 9246337360..63cc8c921e 100644 --- a/libavcodec/opus_pvq.h +++ b/libavcodec/opus_pvq.h @@ -38,14 +38,14 @@ struct CeltPVQ { float (*pvq_search)(float *X, int *y, int K, int N); - QUANT_FN(*decode_band); - QUANT_FN(*encode_band); + QUANT_FN(*quant_band); float (*band_cost)(struct CeltPVQ *pvq, CeltFrame *f, OpusRangeCoder *rc, int band, float *bits, float lambda); }; -int ff_celt_pvq_init (struct CeltPVQ **pvq); void ff_opus_dsp_init_x86(struct CeltPVQ *s); + +int ff_celt_pvq_init(struct CeltPVQ **pvq, int encode); void ff_celt_pvq_uninit(struct CeltPVQ **pvq); #endif /* AVCODEC_OPUS_PVQ_H */ diff --git a/libavcodec/opusenc.c b/libavcodec/opusenc.c index 79d20dc6e6..93b2116e7b 100644 --- a/libavcodec/opusenc.c +++ b/libavcodec/opusenc.c @@ -255,7 +255,7 @@ static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f) } } -static void celt_enc_tf(OpusRangeCoder *rc, CeltFrame *f) +static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc) { int i, tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed; int bits = f->transient ? 2 : 4; @@ -282,7 +282,7 @@ static void celt_enc_tf(OpusRangeCoder *rc, CeltFrame *f) f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]]; } -void ff_celt_enc_bitalloc(OpusRangeCoder *rc, CeltFrame *f) +void ff_celt_enc_bitalloc(CeltFrame *f, OpusRangeCoder *rc) { int i, j, low, high, total, done, bandbits, remaining, tbits_8ths; int skip_startband = f->start_band; @@ -690,7 +690,7 @@ static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f, } } -static void celt_quant_coarse(OpusRangeCoder *rc, CeltFrame *f, +static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc, float last_energy[][CELT_MAX_BANDS]) { uint32_t inter, intra; @@ -710,7 +710,7 @@ static void celt_quant_coarse(OpusRangeCoder *rc, CeltFrame *f, } } -static void celt_quant_fine(OpusRangeCoder *rc, CeltFrame *f) +static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc) { int i, ch; for (i = f->start_band; i < f->end_band; i++) { @@ -747,95 +747,6 @@ static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f } } -static void celt_quant_bands(OpusRangeCoder *rc, CeltFrame *f) -{ - float lowband_scratch[8 * 22]; - float norm[2 * 8 * 100]; - - int totalbits = (f->framebits << 3) - f->anticollapse_needed; - - int update_lowband = 1; - int lowband_offset = 0; - - int i, j; - - for (i = f->start_band; i < f->end_band; i++) { - uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 }; - int band_offset = ff_celt_freq_bands[i] << f->size; - int band_size = ff_celt_freq_range[i] << f->size; - float *X = f->block[0].coeffs + band_offset; - float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL; - - int consumed = opus_rc_tell_frac(rc); - float *norm2 = norm + 8 * 100; - int effective_lowband = -1; - int b = 0; - - /* Compute how many bits we want to allocate to this band */ - if (i != f->start_band) - f->remaining -= consumed; - f->remaining2 = totalbits - consumed - 1; - if (i <= f->coded_bands - 1) { - int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i); - b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14); - } - - if (ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] && - (update_lowband || lowband_offset == 0)) - lowband_offset = i; - - /* Get a conservative estimate of the collapse_mask's for the bands we're - going to be folding from. */ - if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE || - f->blocks > 1 || f->tf_change[i] < 0)) { - int foldstart, foldend; - - /* This ensures we never repeat spectral content within one band */ - effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band], - ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]); - foldstart = lowband_offset; - while (ff_celt_freq_bands[--foldstart] > effective_lowband); - foldend = lowband_offset - 1; - while (ff_celt_freq_bands[++foldend] < effective_lowband + ff_celt_freq_range[i]); - - cm[0] = cm[1] = 0; - for (j = foldstart; j < foldend; j++) { - cm[0] |= f->block[0].collapse_masks[j]; - cm[1] |= f->block[f->channels - 1].collapse_masks[j]; - } - } - - if (f->dual_stereo && i == f->intensity_stereo) { - /* Switch off dual stereo to do intensity */ - f->dual_stereo = 0; - for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++) - norm[j] = (norm[j] + norm2[j]) / 2; - } - - if (f->dual_stereo) { - cm[0] = f->pvq->encode_band(f->pvq, f, rc, i, X, NULL, band_size, b / 2, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]); - - cm[1] = f->pvq->encode_band(f->pvq, f, rc, i, Y, NULL, band_size, b / 2, f->blocks, - effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size, - norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]); - } else { - cm[0] = f->pvq->encode_band(f->pvq, f, rc, i, X, Y, band_size, b, f->blocks, - effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size, - norm + band_offset, 0, 1.0f, lowband_scratch, cm[0] | cm[1]); - cm[1] = cm[0]; - } - - f->block[0].collapse_masks[i] = (uint8_t)cm[0]; - f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1]; - f->remaining += f->pulses[i] + consumed; - - /* Update the folding position only as long as we have 1 bit/sample depth */ - update_lowband = (b > band_size << 3); - } -} - static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f, int index) { @@ -883,11 +794,11 @@ static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc, ff_opus_rc_enc_log(rc, f->transient, 3); /* Main encoding */ - celt_quant_coarse(rc, f, s->last_quantized_energy); - celt_enc_tf (rc, f); - ff_celt_enc_bitalloc(rc, f); - celt_quant_fine (rc, f); - celt_quant_bands (rc, f); + celt_quant_coarse (f, rc, s->last_quantized_energy); + celt_enc_tf (f, rc); + ff_celt_enc_bitalloc(f, rc); + celt_quant_fine (f, rc); + ff_celt_quant_bands (f, rc); /* Anticollapse bit */ if (f->anticollapse_needed) @@ -1080,7 +991,7 @@ static av_cold int opus_encode_init(AVCodecContext *avctx) ff_af_queue_init(avctx, &s->afq); - if ((ret = ff_celt_pvq_init(&s->pvq)) < 0) + if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0) return ret; if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT))) @@ -1117,6 +1028,7 @@ static av_cold int opus_encode_init(AVCodecContext *avctx) s->frame[i].avctx = s->avctx; s->frame[i].seed = 0; s->frame[i].pvq = s->pvq; + s->frame[i].apply_phase_inv = 1; s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f; } diff --git a/libavcodec/opusenc.h b/libavcodec/opusenc.h index 3273d0a9a2..56e4af67e3 100644 --- a/libavcodec/opusenc.h +++ b/libavcodec/opusenc.h @@ -51,6 +51,6 @@ typedef struct OpusPacketInfo { int frames; } OpusPacketInfo; -void ff_celt_enc_bitalloc(OpusRangeCoder *rc, CeltFrame *f); +void ff_celt_enc_bitalloc(CeltFrame *f, OpusRangeCoder *rc); #endif /* AVCODEC_OPUSENC_H */ diff --git a/libavcodec/opusenc_psy.c b/libavcodec/opusenc_psy.c index 1b108ecb87..dc549dc9a9 100644 --- a/libavcodec/opusenc_psy.c +++ b/libavcodec/opusenc_psy.c @@ -316,7 +316,7 @@ static int bands_dist(OpusPsyContext *s, CeltFrame *f, float *total_dist) OpusRangeCoder dump; ff_opus_rc_enc_init(&dump); - ff_celt_enc_bitalloc(&dump, f); + ff_celt_enc_bitalloc(f, &dump); for (i = 0; i < CELT_MAX_BANDS; i++) { float bits = 0.0f; |