yading@10: /* yading@10: * Windows Media Audio Voice decoder. yading@10: * Copyright (c) 2009 Ronald S. Bultje yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * @brief Windows Media Audio Voice compatible decoder yading@10: * @author Ronald S. Bultje yading@10: */ yading@10: yading@10: #include yading@10: yading@10: #include "libavutil/channel_layout.h" yading@10: #include "libavutil/float_dsp.h" yading@10: #include "libavutil/mem.h" yading@10: #include "avcodec.h" yading@10: #include "internal.h" yading@10: #include "get_bits.h" yading@10: #include "put_bits.h" yading@10: #include "wmavoice_data.h" yading@10: #include "celp_filters.h" yading@10: #include "acelp_vectors.h" yading@10: #include "acelp_filters.h" yading@10: #include "lsp.h" yading@10: #include "dct.h" yading@10: #include "rdft.h" yading@10: #include "sinewin.h" yading@10: yading@10: #define MAX_BLOCKS 8 ///< maximum number of blocks per frame yading@10: #define MAX_LSPS 16 ///< maximum filter order yading@10: #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple yading@10: ///< of 16 for ASM input buffer alignment yading@10: #define MAX_FRAMES 3 ///< maximum number of frames per superframe yading@10: #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame yading@10: #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history yading@10: #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES) yading@10: ///< maximum number of samples per superframe yading@10: #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that yading@10: ///< was split over two packets yading@10: #define VLC_NBITS 6 ///< number of bits to read per VLC iteration yading@10: yading@10: /** yading@10: * Frame type VLC coding. yading@10: */ yading@10: static VLC frame_type_vlc; yading@10: yading@10: /** yading@10: * Adaptive codebook types. yading@10: */ yading@10: enum { yading@10: ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed) yading@10: ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which yading@10: ///< we interpolate to get a per-sample pitch. yading@10: ///< Signal is generated using an asymmetric sinc yading@10: ///< window function yading@10: ///< @note see #wmavoice_ipol1_coeffs yading@10: ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using yading@10: ///< a Hamming sinc window function yading@10: ///< @note see #wmavoice_ipol2_coeffs yading@10: }; yading@10: yading@10: /** yading@10: * Fixed codebook types. yading@10: */ yading@10: enum { yading@10: FCB_TYPE_SILENCE = 0, ///< comfort noise during silence yading@10: ///< generated from a hardcoded (fixed) codebook yading@10: ///< with per-frame (low) gain values yading@10: FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block yading@10: ///< gain values yading@10: FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals, yading@10: ///< used in particular for low-bitrate streams yading@10: FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in yading@10: ///< combinations of either single pulses or yading@10: ///< pulse pairs yading@10: }; yading@10: yading@10: /** yading@10: * Description of frame types. yading@10: */ yading@10: static const struct frame_type_desc { yading@10: uint8_t n_blocks; ///< amount of blocks per frame (each block yading@10: ///< (contains 160/#n_blocks samples) yading@10: uint8_t log_n_blocks; ///< log2(#n_blocks) yading@10: uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*) yading@10: uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*) yading@10: uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs yading@10: ///< (rather than just one single pulse) yading@10: ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES yading@10: uint16_t frame_size; ///< the amount of bits that make up the block yading@10: ///< data (per frame) yading@10: } frame_descs[17] = { yading@10: { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 }, yading@10: { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 }, yading@10: { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 }, yading@10: { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 }, yading@10: { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 }, yading@10: { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 }, yading@10: { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 }, yading@10: { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 }, yading@10: { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 }, yading@10: { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 }, yading@10: { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 }, yading@10: { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 }, yading@10: { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 }, yading@10: { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 }, yading@10: { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 }, yading@10: { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 }, yading@10: { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 } yading@10: }; yading@10: yading@10: /** yading@10: * WMA Voice decoding context. yading@10: */ yading@10: typedef struct { yading@10: /** yading@10: * @name Global values specified in the stream header / extradata or used all over. yading@10: * @{ yading@10: */ yading@10: GetBitContext gb; ///< packet bitreader. During decoder init, yading@10: ///< it contains the extradata from the yading@10: ///< demuxer. During decoding, it contains yading@10: ///< packet data. yading@10: int8_t vbm_tree[25]; ///< converts VLC codes to frame type yading@10: yading@10: int spillover_bitsize; ///< number of bits used to specify yading@10: ///< #spillover_nbits in the packet header yading@10: ///< = ceil(log2(ctx->block_align << 3)) yading@10: int history_nsamples; ///< number of samples in history for signal yading@10: ///< prediction (through ACB) yading@10: yading@10: /* postfilter specific values */ yading@10: int do_apf; ///< whether to apply the averaged yading@10: ///< projection filter (APF) yading@10: int denoise_strength; ///< strength of denoising in Wiener filter yading@10: ///< [0-11] yading@10: int denoise_tilt_corr; ///< Whether to apply tilt correction to the yading@10: ///< Wiener filter coefficients (postfilter) yading@10: int dc_level; ///< Predicted amount of DC noise, based yading@10: ///< on which a DC removal filter is used yading@10: yading@10: int lsps; ///< number of LSPs per frame [10 or 16] yading@10: int lsp_q_mode; ///< defines quantizer defaults [0, 1] yading@10: int lsp_def_mode; ///< defines different sets of LSP defaults yading@10: ///< [0, 1] yading@10: int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded yading@10: ///< per-frame (independent coding) yading@10: int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded yading@10: ///< per superframe (residual coding) yading@10: yading@10: int min_pitch_val; ///< base value for pitch parsing code yading@10: int max_pitch_val; ///< max value + 1 for pitch parsing yading@10: int pitch_nbits; ///< number of bits used to specify the yading@10: ///< pitch value in the frame header yading@10: int block_pitch_nbits; ///< number of bits used to specify the yading@10: ///< first block's pitch value yading@10: int block_pitch_range; ///< range of the block pitch yading@10: int block_delta_pitch_nbits; ///< number of bits used to specify the yading@10: ///< delta pitch between this and the last yading@10: ///< block's pitch value, used in all but yading@10: ///< first block yading@10: int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is yading@10: ///< from -this to +this-1) yading@10: uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale yading@10: ///< conversion yading@10: yading@10: /** yading@10: * @} yading@10: * yading@10: * @name Packet values specified in the packet header or related to a packet. yading@10: * yading@10: * A packet is considered to be a single unit of data provided to this yading@10: * decoder by the demuxer. yading@10: * @{ yading@10: */ yading@10: int spillover_nbits; ///< number of bits of the previous packet's yading@10: ///< last superframe preceding this yading@10: ///< packet's first full superframe (useful yading@10: ///< for re-synchronization also) yading@10: int has_residual_lsps; ///< if set, superframes contain one set of yading@10: ///< LSPs that cover all frames, encoded as yading@10: ///< independent and residual LSPs; if not yading@10: ///< set, each frame contains its own, fully yading@10: ///< independent, LSPs yading@10: int skip_bits_next; ///< number of bits to skip at the next call yading@10: ///< to #wmavoice_decode_packet() (since yading@10: ///< they're part of the previous superframe) yading@10: yading@10: uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE]; yading@10: ///< cache for superframe data split over yading@10: ///< multiple packets yading@10: int sframe_cache_size; ///< set to >0 if we have data from an yading@10: ///< (incomplete) superframe from a previous yading@10: ///< packet that spilled over in the current yading@10: ///< packet; specifies the amount of bits in yading@10: ///< #sframe_cache yading@10: PutBitContext pb; ///< bitstream writer for #sframe_cache yading@10: yading@10: /** yading@10: * @} yading@10: * yading@10: * @name Frame and superframe values yading@10: * Superframe and frame data - these can change from frame to frame, yading@10: * although some of them do in that case serve as a cache / history for yading@10: * the next frame or superframe. yading@10: * @{ yading@10: */ yading@10: double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous yading@10: ///< superframe yading@10: int last_pitch_val; ///< pitch value of the previous frame yading@10: int last_acb_type; ///< frame type [0-2] of the previous frame yading@10: int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val) yading@10: ///< << 16) / #MAX_FRAMESIZE yading@10: float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE yading@10: yading@10: int aw_idx_is_ext; ///< whether the AW index was encoded in yading@10: ///< 8 bits (instead of 6) yading@10: int aw_pulse_range; ///< the range over which #aw_pulse_set1() yading@10: ///< can apply the pulse, relative to the yading@10: ///< value in aw_first_pulse_off. The exact yading@10: ///< position of the first AW-pulse is within yading@10: ///< [pulse_off, pulse_off + this], and yading@10: ///< depends on bitstream values; [16 or 24] yading@10: int aw_n_pulses[2]; ///< number of AW-pulses in each block; note yading@10: ///< that this number can be negative (in yading@10: ///< which case it basically means "zero") yading@10: int aw_first_pulse_off[2]; ///< index of first sample to which to yading@10: ///< apply AW-pulses, or -0xff if unset yading@10: int aw_next_pulse_off_cache; ///< the position (relative to start of the yading@10: ///< second block) at which pulses should yading@10: ///< start to be positioned, serves as a yading@10: ///< cache for pitch-adaptive window pulses yading@10: ///< between blocks yading@10: yading@10: int frame_cntr; ///< current frame index [0 - 0xFFFE]; is yading@10: ///< only used for comfort noise in #pRNG() yading@10: float gain_pred_err[6]; ///< cache for gain prediction yading@10: float excitation_history[MAX_SIGNAL_HISTORY]; yading@10: ///< cache of the signal of previous yading@10: ///< superframes, used as a history for yading@10: ///< signal generation yading@10: float synth_history[MAX_LSPS]; ///< see #excitation_history yading@10: /** yading@10: * @} yading@10: * yading@10: * @name Postfilter values yading@10: * yading@10: * Variables used for postfilter implementation, mostly history for yading@10: * smoothing and so on, and context variables for FFT/iFFT. yading@10: * @{ yading@10: */ yading@10: RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the yading@10: ///< postfilter (for denoise filter) yading@10: DCTContext dct, dst; ///< contexts for phase shift (in Hilbert yading@10: ///< transform, part of postfilter) yading@10: float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi] yading@10: ///< range yading@10: float postfilter_agc; ///< gain control memory, used in yading@10: ///< #adaptive_gain_control() yading@10: float dcf_mem[2]; ///< DC filter history yading@10: float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE]; yading@10: ///< zero filter output (i.e. excitation) yading@10: ///< by postfilter yading@10: float denoise_filter_cache[MAX_FRAMESIZE]; yading@10: int denoise_filter_cache_size; ///< samples in #denoise_filter_cache yading@10: DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80]; yading@10: ///< aligned buffer for LPC tilting yading@10: DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80]; yading@10: ///< aligned buffer for denoise coefficients yading@10: DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16]; yading@10: ///< aligned buffer for postfilter speech yading@10: ///< synthesis yading@10: /** yading@10: * @} yading@10: */ yading@10: } WMAVoiceContext; yading@10: yading@10: /** yading@10: * Set up the variable bit mode (VBM) tree from container extradata. yading@10: * @param gb bit I/O context. yading@10: * The bit context (s->gb) should be loaded with byte 23-46 of the yading@10: * container extradata (i.e. the ones containing the VBM tree). yading@10: * @param vbm_tree pointer to array to which the decoded VBM tree will be yading@10: * written. yading@10: * @return 0 on success, <0 on error. yading@10: */ yading@10: static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25]) yading@10: { yading@10: static const uint8_t bits[] = { yading@10: 2, 2, 2, 4, 4, 4, yading@10: 6, 6, 6, 8, 8, 8, yading@10: 10, 10, 10, 12, 12, 12, yading@10: 14, 14, 14, 14 yading@10: }; yading@10: static const uint16_t codes[] = { yading@10: 0x0000, 0x0001, 0x0002, // 00/01/10 yading@10: 0x000c, 0x000d, 0x000e, // 11+00/01/10 yading@10: 0x003c, 0x003d, 0x003e, // 1111+00/01/10 yading@10: 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10 yading@10: 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10 yading@10: 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10 yading@10: 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx yading@10: }; yading@10: int cntr[8] = { 0 }, n, res; yading@10: yading@10: memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25); yading@10: for (n = 0; n < 17; n++) { yading@10: res = get_bits(gb, 3); yading@10: if (cntr[res] > 3) // should be >= 3 + (res == 7)) yading@10: return -1; yading@10: vbm_tree[res * 3 + cntr[res]++] = n; yading@10: } yading@10: INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits), yading@10: bits, 1, 1, codes, 2, 2, 132); yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Set up decoder with parameters from demuxer (extradata etc.). yading@10: */ yading@10: static av_cold int wmavoice_decode_init(AVCodecContext *ctx) yading@10: { yading@10: int n, flags, pitch_range, lsp16_flag; yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: yading@10: /** yading@10: * Extradata layout: yading@10: * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c), yading@10: * - byte 19-22: flags field (annoyingly in LE; see below for known yading@10: * values), yading@10: * - byte 23-46: variable bitmode tree (really just 17 * 3 bits, yading@10: * rest is 0). yading@10: */ yading@10: if (ctx->extradata_size != 46) { yading@10: av_log(ctx, AV_LOG_ERROR, yading@10: "Invalid extradata size %d (should be 46)\n", yading@10: ctx->extradata_size); yading@10: return -1; yading@10: } yading@10: flags = AV_RL32(ctx->extradata + 18); yading@10: s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align); yading@10: s->do_apf = flags & 0x1; yading@10: if (s->do_apf) { yading@10: ff_rdft_init(&s->rdft, 7, DFT_R2C); yading@10: ff_rdft_init(&s->irdft, 7, IDFT_C2R); yading@10: ff_dct_init(&s->dct, 6, DCT_I); yading@10: ff_dct_init(&s->dst, 6, DST_I); yading@10: yading@10: ff_sine_window_init(s->cos, 256); yading@10: memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0])); yading@10: for (n = 0; n < 255; n++) { yading@10: s->sin[n] = -s->sin[510 - n]; yading@10: s->cos[510 - n] = s->cos[n]; yading@10: } yading@10: } yading@10: s->denoise_strength = (flags >> 2) & 0xF; yading@10: if (s->denoise_strength >= 12) { yading@10: av_log(ctx, AV_LOG_ERROR, yading@10: "Invalid denoise filter strength %d (max=11)\n", yading@10: s->denoise_strength); yading@10: return -1; yading@10: } yading@10: s->denoise_tilt_corr = !!(flags & 0x40); yading@10: s->dc_level = (flags >> 7) & 0xF; yading@10: s->lsp_q_mode = !!(flags & 0x2000); yading@10: s->lsp_def_mode = !!(flags & 0x4000); yading@10: lsp16_flag = flags & 0x1000; yading@10: if (lsp16_flag) { yading@10: s->lsps = 16; yading@10: s->frame_lsp_bitsize = 34; yading@10: s->sframe_lsp_bitsize = 60; yading@10: } else { yading@10: s->lsps = 10; yading@10: s->frame_lsp_bitsize = 24; yading@10: s->sframe_lsp_bitsize = 48; yading@10: } yading@10: for (n = 0; n < s->lsps; n++) yading@10: s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); yading@10: yading@10: init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3); yading@10: if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) { yading@10: av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n"); yading@10: return -1; yading@10: } yading@10: yading@10: s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8; yading@10: s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8; yading@10: pitch_range = s->max_pitch_val - s->min_pitch_val; yading@10: if (pitch_range <= 0) { yading@10: av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n"); yading@10: return -1; yading@10: } yading@10: s->pitch_nbits = av_ceil_log2(pitch_range); yading@10: s->last_pitch_val = 40; yading@10: s->last_acb_type = ACB_TYPE_NONE; yading@10: s->history_nsamples = s->max_pitch_val + 8; yading@10: yading@10: if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) { yading@10: int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8, yading@10: max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8; yading@10: yading@10: av_log(ctx, AV_LOG_ERROR, yading@10: "Unsupported samplerate %d (min=%d, max=%d)\n", yading@10: ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz yading@10: yading@10: return -1; yading@10: } yading@10: yading@10: s->block_conv_table[0] = s->min_pitch_val; yading@10: s->block_conv_table[1] = (pitch_range * 25) >> 6; yading@10: s->block_conv_table[2] = (pitch_range * 44) >> 6; yading@10: s->block_conv_table[3] = s->max_pitch_val - 1; yading@10: s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF; yading@10: if (s->block_delta_pitch_hrange <= 0) { yading@10: av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n"); yading@10: return -1; yading@10: } yading@10: s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange); yading@10: s->block_pitch_range = s->block_conv_table[2] + yading@10: s->block_conv_table[3] + 1 + yading@10: 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val); yading@10: s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range); yading@10: yading@10: ctx->channels = 1; yading@10: ctx->channel_layout = AV_CH_LAYOUT_MONO; yading@10: ctx->sample_fmt = AV_SAMPLE_FMT_FLT; yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * @name Postfilter functions yading@10: * Postfilter functions (gain control, wiener denoise filter, DC filter, yading@10: * kalman smoothening, plus surrounding code to wrap it) yading@10: * @{ yading@10: */ yading@10: /** yading@10: * Adaptive gain control (as used in postfilter). yading@10: * yading@10: * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except yading@10: * that the energy here is calculated using sum(abs(...)), whereas the yading@10: * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)). yading@10: * yading@10: * @param out output buffer for filtered samples yading@10: * @param in input buffer containing the samples as they are after the yading@10: * postfilter steps so far yading@10: * @param speech_synth input buffer containing speech synth before postfilter yading@10: * @param size input buffer size yading@10: * @param alpha exponential filter factor yading@10: * @param gain_mem pointer to filter memory (single float) yading@10: */ yading@10: static void adaptive_gain_control(float *out, const float *in, yading@10: const float *speech_synth, yading@10: int size, float alpha, float *gain_mem) yading@10: { yading@10: int i; yading@10: float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor; yading@10: float mem = *gain_mem; yading@10: yading@10: for (i = 0; i < size; i++) { yading@10: speech_energy += fabsf(speech_synth[i]); yading@10: postfilter_energy += fabsf(in[i]); yading@10: } yading@10: gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy; yading@10: yading@10: for (i = 0; i < size; i++) { yading@10: mem = alpha * mem + gain_scale_factor; yading@10: out[i] = in[i] * mem; yading@10: } yading@10: yading@10: *gain_mem = mem; yading@10: } yading@10: yading@10: /** yading@10: * Kalman smoothing function. yading@10: * yading@10: * This function looks back pitch +/- 3 samples back into history to find yading@10: * the best fitting curve (that one giving the optimal gain of the two yading@10: * signals, i.e. the highest dot product between the two), and then yading@10: * uses that signal history to smoothen the output of the speech synthesis yading@10: * filter. yading@10: * yading@10: * @param s WMA Voice decoding context yading@10: * @param pitch pitch of the speech signal yading@10: * @param in input speech signal yading@10: * @param out output pointer for smoothened signal yading@10: * @param size input/output buffer size yading@10: * yading@10: * @returns -1 if no smoothening took place, e.g. because no optimal yading@10: * fit could be found, or 0 on success. yading@10: */ yading@10: static int kalman_smoothen(WMAVoiceContext *s, int pitch, yading@10: const float *in, float *out, int size) yading@10: { yading@10: int n; yading@10: float optimal_gain = 0, dot; yading@10: const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)], yading@10: *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)], yading@10: *best_hist_ptr = NULL; yading@10: yading@10: /* find best fitting point in history */ yading@10: do { yading@10: dot = avpriv_scalarproduct_float_c(in, ptr, size); yading@10: if (dot > optimal_gain) { yading@10: optimal_gain = dot; yading@10: best_hist_ptr = ptr; yading@10: } yading@10: } while (--ptr >= end); yading@10: yading@10: if (optimal_gain <= 0) yading@10: return -1; yading@10: dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size); yading@10: if (dot <= 0) // would be 1.0 yading@10: return -1; yading@10: yading@10: if (optimal_gain <= dot) { yading@10: dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000 yading@10: } else yading@10: dot = 0.625; yading@10: yading@10: /* actual smoothing */ yading@10: for (n = 0; n < size; n++) yading@10: out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Get the tilt factor of a formant filter from its transfer function yading@10: * @see #tilt_factor() in amrnbdec.c, which does essentially the same, yading@10: * but somehow (??) it does a speech synthesis filter in the yading@10: * middle, which is missing here yading@10: * yading@10: * @param lpcs LPC coefficients yading@10: * @param n_lpcs Size of LPC buffer yading@10: * @returns the tilt factor yading@10: */ yading@10: static float tilt_factor(const float *lpcs, int n_lpcs) yading@10: { yading@10: float rh0, rh1; yading@10: yading@10: rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs); yading@10: rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1); yading@10: yading@10: return rh1 / rh0; yading@10: } yading@10: yading@10: /** yading@10: * Derive denoise filter coefficients (in real domain) from the LPCs. yading@10: */ yading@10: static void calc_input_response(WMAVoiceContext *s, float *lpcs, yading@10: int fcb_type, float *coeffs, int remainder) yading@10: { yading@10: float last_coeff, min = 15.0, max = -15.0; yading@10: float irange, angle_mul, gain_mul, range, sq; yading@10: int n, idx; yading@10: yading@10: /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */ yading@10: s->rdft.rdft_calc(&s->rdft, lpcs); yading@10: #define log_range(var, assign) do { \ yading@10: float tmp = log10f(assign); var = tmp; \ yading@10: max = FFMAX(max, tmp); min = FFMIN(min, tmp); \ yading@10: } while (0) yading@10: log_range(last_coeff, lpcs[1] * lpcs[1]); yading@10: for (n = 1; n < 64; n++) yading@10: log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] + yading@10: lpcs[n * 2 + 1] * lpcs[n * 2 + 1]); yading@10: log_range(lpcs[0], lpcs[0] * lpcs[0]); yading@10: #undef log_range yading@10: range = max - min; yading@10: lpcs[64] = last_coeff; yading@10: yading@10: /* Now, use this spectrum to pick out these frequencies with higher yading@10: * (relative) power/energy (which we then take to be "not noise"), yading@10: * and set up a table (still in lpc[]) of (relative) gains per frequency. yading@10: * These frequencies will be maintained, while others ("noise") will be yading@10: * decreased in the filter output. */ yading@10: irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63] yading@10: gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) : yading@10: (5.0 / 14.7)); yading@10: angle_mul = gain_mul * (8.0 * M_LN10 / M_PI); yading@10: for (n = 0; n <= 64; n++) { yading@10: float pwr; yading@10: yading@10: idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1); yading@10: pwr = wmavoice_denoise_power_table[s->denoise_strength][idx]; yading@10: lpcs[n] = angle_mul * pwr; yading@10: yading@10: /* 70.57 =~ 1/log10(1.0331663) */ yading@10: idx = (pwr * gain_mul - 0.0295) * 70.570526123; yading@10: if (idx > 127) { // fallback if index falls outside table range yading@10: coeffs[n] = wmavoice_energy_table[127] * yading@10: powf(1.0331663, idx - 127); yading@10: } else yading@10: coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)]; yading@10: } yading@10: yading@10: /* calculate the Hilbert transform of the gains, which we do (since this yading@10: * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()). yading@10: * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the yading@10: * "moment" of the LPCs in this filter. */ yading@10: s->dct.dct_calc(&s->dct, lpcs); yading@10: s->dst.dct_calc(&s->dst, lpcs); yading@10: yading@10: /* Split out the coefficient indexes into phase/magnitude pairs */ yading@10: idx = 255 + av_clip(lpcs[64], -255, 255); yading@10: coeffs[0] = coeffs[0] * s->cos[idx]; yading@10: idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255); yading@10: last_coeff = coeffs[64] * s->cos[idx]; yading@10: for (n = 63;; n--) { yading@10: idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255); yading@10: coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; yading@10: coeffs[n * 2] = coeffs[n] * s->cos[idx]; yading@10: yading@10: if (!--n) break; yading@10: yading@10: idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255); yading@10: coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; yading@10: coeffs[n * 2] = coeffs[n] * s->cos[idx]; yading@10: } yading@10: coeffs[1] = last_coeff; yading@10: yading@10: /* move into real domain */ yading@10: s->irdft.rdft_calc(&s->irdft, coeffs); yading@10: yading@10: /* tilt correction and normalize scale */ yading@10: memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder)); yading@10: if (s->denoise_tilt_corr) { yading@10: float tilt_mem = 0; yading@10: yading@10: coeffs[remainder - 1] = 0; yading@10: ff_tilt_compensation(&tilt_mem, yading@10: -1.8 * tilt_factor(coeffs, remainder - 1), yading@10: coeffs, remainder); yading@10: } yading@10: sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs, yading@10: remainder)); yading@10: for (n = 0; n < remainder; n++) yading@10: coeffs[n] *= sq; yading@10: } yading@10: yading@10: /** yading@10: * This function applies a Wiener filter on the (noisy) speech signal as yading@10: * a means to denoise it. yading@10: * yading@10: * - take RDFT of LPCs to get the power spectrum of the noise + speech; yading@10: * - using this power spectrum, calculate (for each frequency) the Wiener yading@10: * filter gain, which depends on the frequency power and desired level yading@10: * of noise subtraction (when set too high, this leads to artifacts) yading@10: * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse yading@10: * of 4-8kHz); yading@10: * - by doing a phase shift, calculate the Hilbert transform of this array yading@10: * of per-frequency filter-gains to get the filtering coefficients; yading@10: * - smoothen/normalize/de-tilt these filter coefficients as desired; yading@10: * - take RDFT of noisy sound, apply the coefficients and take its IRDFT yading@10: * to get the denoised speech signal; yading@10: * - the leftover (i.e. output of the IRDFT on denoised speech data beyond yading@10: * the frame boundary) are saved and applied to subsequent frames by an yading@10: * overlap-add method (otherwise you get clicking-artifacts). yading@10: * yading@10: * @param s WMA Voice decoding context yading@10: * @param fcb_type Frame (codebook) type yading@10: * @param synth_pf input: the noisy speech signal, output: denoised speech yading@10: * data; should be 16-byte aligned (for ASM purposes) yading@10: * @param size size of the speech data yading@10: * @param lpcs LPCs used to synthesize this frame's speech data yading@10: */ yading@10: static void wiener_denoise(WMAVoiceContext *s, int fcb_type, yading@10: float *synth_pf, int size, yading@10: const float *lpcs) yading@10: { yading@10: int remainder, lim, n; yading@10: yading@10: if (fcb_type != FCB_TYPE_SILENCE) { yading@10: float *tilted_lpcs = s->tilted_lpcs_pf, yading@10: *coeffs = s->denoise_coeffs_pf, tilt_mem = 0; yading@10: yading@10: tilted_lpcs[0] = 1.0; yading@10: memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps); yading@10: memset(&tilted_lpcs[s->lsps + 1], 0, yading@10: sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1)); yading@10: ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps), yading@10: tilted_lpcs, s->lsps + 2); yading@10: yading@10: /* The IRDFT output (127 samples for 7-bit filter) beyond the frame yading@10: * size is applied to the next frame. All input beyond this is zero, yading@10: * and thus all output beyond this will go towards zero, hence we can yading@10: * limit to min(size-1, 127-size) as a performance consideration. */ yading@10: remainder = FFMIN(127 - size, size - 1); yading@10: calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder); yading@10: yading@10: /* apply coefficients (in frequency spectrum domain), i.e. complex yading@10: * number multiplication */ yading@10: memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size)); yading@10: s->rdft.rdft_calc(&s->rdft, synth_pf); yading@10: s->rdft.rdft_calc(&s->rdft, coeffs); yading@10: synth_pf[0] *= coeffs[0]; yading@10: synth_pf[1] *= coeffs[1]; yading@10: for (n = 1; n < 64; n++) { yading@10: float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1]; yading@10: synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1]; yading@10: synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1]; yading@10: } yading@10: s->irdft.rdft_calc(&s->irdft, synth_pf); yading@10: } yading@10: yading@10: /* merge filter output with the history of previous runs */ yading@10: if (s->denoise_filter_cache_size) { yading@10: lim = FFMIN(s->denoise_filter_cache_size, size); yading@10: for (n = 0; n < lim; n++) yading@10: synth_pf[n] += s->denoise_filter_cache[n]; yading@10: s->denoise_filter_cache_size -= lim; yading@10: memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size], yading@10: sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size); yading@10: } yading@10: yading@10: /* move remainder of filter output into a cache for future runs */ yading@10: if (fcb_type != FCB_TYPE_SILENCE) { yading@10: lim = FFMIN(remainder, s->denoise_filter_cache_size); yading@10: for (n = 0; n < lim; n++) yading@10: s->denoise_filter_cache[n] += synth_pf[size + n]; yading@10: if (lim < remainder) { yading@10: memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim], yading@10: sizeof(s->denoise_filter_cache[0]) * (remainder - lim)); yading@10: s->denoise_filter_cache_size = remainder; yading@10: } yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Averaging projection filter, the postfilter used in WMAVoice. yading@10: * yading@10: * This uses the following steps: yading@10: * - A zero-synthesis filter (generate excitation from synth signal) yading@10: * - Kalman smoothing on excitation, based on pitch yading@10: * - Re-synthesized smoothened output yading@10: * - Iterative Wiener denoise filter yading@10: * - Adaptive gain filter yading@10: * - DC filter yading@10: * yading@10: * @param s WMAVoice decoding context yading@10: * @param synth Speech synthesis output (before postfilter) yading@10: * @param samples Output buffer for filtered samples yading@10: * @param size Buffer size of synth & samples yading@10: * @param lpcs Generated LPCs used for speech synthesis yading@10: * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned) yading@10: * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses) yading@10: * @param pitch Pitch of the input signal yading@10: */ yading@10: static void postfilter(WMAVoiceContext *s, const float *synth, yading@10: float *samples, int size, yading@10: const float *lpcs, float *zero_exc_pf, yading@10: int fcb_type, int pitch) yading@10: { yading@10: float synth_filter_in_buf[MAX_FRAMESIZE / 2], yading@10: *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16], yading@10: *synth_filter_in = zero_exc_pf; yading@10: yading@10: av_assert0(size <= MAX_FRAMESIZE / 2); yading@10: yading@10: /* generate excitation from input signal */ yading@10: ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps); yading@10: yading@10: if (fcb_type >= FCB_TYPE_AW_PULSES && yading@10: !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size)) yading@10: synth_filter_in = synth_filter_in_buf; yading@10: yading@10: /* re-synthesize speech after smoothening, and keep history */ yading@10: ff_celp_lp_synthesis_filterf(synth_pf, lpcs, yading@10: synth_filter_in, size, s->lsps); yading@10: memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps], yading@10: sizeof(synth_pf[0]) * s->lsps); yading@10: yading@10: wiener_denoise(s, fcb_type, synth_pf, size, lpcs); yading@10: yading@10: adaptive_gain_control(samples, synth_pf, synth, size, 0.99, yading@10: &s->postfilter_agc); yading@10: yading@10: if (s->dc_level > 8) { yading@10: /* remove ultra-low frequency DC noise / highpass filter; yading@10: * coefficients are identical to those used in SIPR decoding, yading@10: * and very closely resemble those used in AMR-NB decoding. */ yading@10: ff_acelp_apply_order_2_transfer_function(samples, samples, yading@10: (const float[2]) { -1.99997, 1.0 }, yading@10: (const float[2]) { -1.9330735188, 0.93589198496 }, yading@10: 0.93980580475, s->dcf_mem, size); yading@10: } yading@10: } yading@10: /** yading@10: * @} yading@10: */ yading@10: yading@10: /** yading@10: * Dequantize LSPs yading@10: * @param lsps output pointer to the array that will hold the LSPs yading@10: * @param num number of LSPs to be dequantized yading@10: * @param values quantized values, contains n_stages values yading@10: * @param sizes range (i.e. max value) of each quantized value yading@10: * @param n_stages number of dequantization runs yading@10: * @param table dequantization table to be used yading@10: * @param mul_q LSF multiplier yading@10: * @param base_q base (lowest) LSF values yading@10: */ yading@10: static void dequant_lsps(double *lsps, int num, yading@10: const uint16_t *values, yading@10: const uint16_t *sizes, yading@10: int n_stages, const uint8_t *table, yading@10: const double *mul_q, yading@10: const double *base_q) yading@10: { yading@10: int n, m; yading@10: yading@10: memset(lsps, 0, num * sizeof(*lsps)); yading@10: for (n = 0; n < n_stages; n++) { yading@10: const uint8_t *t_off = &table[values[n] * num]; yading@10: double base = base_q[n], mul = mul_q[n]; yading@10: yading@10: for (m = 0; m < num; m++) yading@10: lsps[m] += base + mul * t_off[m]; yading@10: yading@10: table += sizes[n] * num; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * @name LSP dequantization routines yading@10: * LSP dequantization routines, for 10/16LSPs and independent/residual coding. yading@10: * @note we assume enough bits are available, caller should check. yading@10: * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits; yading@10: * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits. yading@10: * @{ yading@10: */ yading@10: /** yading@10: * Parse 10 independently-coded LSPs. yading@10: */ yading@10: static void dequant_lsp10i(GetBitContext *gb, double *lsps) yading@10: { yading@10: static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 }; yading@10: static const double mul_lsf[4] = { yading@10: 5.2187144800e-3, 1.4626986422e-3, yading@10: 9.6179549166e-4, 1.1325736225e-3 yading@10: }; yading@10: static const double base_lsf[4] = { yading@10: M_PI * -2.15522e-1, M_PI * -6.1646e-2, yading@10: M_PI * -3.3486e-2, M_PI * -5.7408e-2 yading@10: }; yading@10: uint16_t v[4]; yading@10: yading@10: v[0] = get_bits(gb, 8); yading@10: v[1] = get_bits(gb, 6); yading@10: v[2] = get_bits(gb, 5); yading@10: v[3] = get_bits(gb, 5); yading@10: yading@10: dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i, yading@10: mul_lsf, base_lsf); yading@10: } yading@10: yading@10: /** yading@10: * Parse 10 independently-coded LSPs, and then derive the tables to yading@10: * generate LSPs for the other frames from them (residual coding). yading@10: */ yading@10: static void dequant_lsp10r(GetBitContext *gb, yading@10: double *i_lsps, const double *old, yading@10: double *a1, double *a2, int q_mode) yading@10: { yading@10: static const uint16_t vec_sizes[3] = { 128, 64, 64 }; yading@10: static const double mul_lsf[3] = { yading@10: 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3 yading@10: }; yading@10: static const double base_lsf[3] = { yading@10: M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2 yading@10: }; yading@10: const float (*ipol_tab)[2][10] = q_mode ? yading@10: wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a; yading@10: uint16_t interpol, v[3]; yading@10: int n; yading@10: yading@10: dequant_lsp10i(gb, i_lsps); yading@10: yading@10: interpol = get_bits(gb, 5); yading@10: v[0] = get_bits(gb, 7); yading@10: v[1] = get_bits(gb, 6); yading@10: v[2] = get_bits(gb, 6); yading@10: yading@10: for (n = 0; n < 10; n++) { yading@10: double delta = old[n] - i_lsps[n]; yading@10: a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; yading@10: a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; yading@10: } yading@10: yading@10: dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r, yading@10: mul_lsf, base_lsf); yading@10: } yading@10: yading@10: /** yading@10: * Parse 16 independently-coded LSPs. yading@10: */ yading@10: static void dequant_lsp16i(GetBitContext *gb, double *lsps) yading@10: { yading@10: static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 }; yading@10: static const double mul_lsf[5] = { yading@10: 3.3439586280e-3, 6.9908173703e-4, yading@10: 3.3216608306e-3, 1.0334960326e-3, yading@10: 3.1899104283e-3 yading@10: }; yading@10: static const double base_lsf[5] = { yading@10: M_PI * -1.27576e-1, M_PI * -2.4292e-2, yading@10: M_PI * -1.28094e-1, M_PI * -3.2128e-2, yading@10: M_PI * -1.29816e-1 yading@10: }; yading@10: uint16_t v[5]; yading@10: yading@10: v[0] = get_bits(gb, 8); yading@10: v[1] = get_bits(gb, 6); yading@10: v[2] = get_bits(gb, 7); yading@10: v[3] = get_bits(gb, 6); yading@10: v[4] = get_bits(gb, 7); yading@10: yading@10: dequant_lsps( lsps, 5, v, vec_sizes, 2, yading@10: wmavoice_dq_lsp16i1, mul_lsf, base_lsf); yading@10: dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2, yading@10: wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]); yading@10: dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1, yading@10: wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]); yading@10: } yading@10: yading@10: /** yading@10: * Parse 16 independently-coded LSPs, and then derive the tables to yading@10: * generate LSPs for the other frames from them (residual coding). yading@10: */ yading@10: static void dequant_lsp16r(GetBitContext *gb, yading@10: double *i_lsps, const double *old, yading@10: double *a1, double *a2, int q_mode) yading@10: { yading@10: static const uint16_t vec_sizes[3] = { 128, 128, 128 }; yading@10: static const double mul_lsf[3] = { yading@10: 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3 yading@10: }; yading@10: static const double base_lsf[3] = { yading@10: M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2 yading@10: }; yading@10: const float (*ipol_tab)[2][16] = q_mode ? yading@10: wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a; yading@10: uint16_t interpol, v[3]; yading@10: int n; yading@10: yading@10: dequant_lsp16i(gb, i_lsps); yading@10: yading@10: interpol = get_bits(gb, 5); yading@10: v[0] = get_bits(gb, 7); yading@10: v[1] = get_bits(gb, 7); yading@10: v[2] = get_bits(gb, 7); yading@10: yading@10: for (n = 0; n < 16; n++) { yading@10: double delta = old[n] - i_lsps[n]; yading@10: a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; yading@10: a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; yading@10: } yading@10: yading@10: dequant_lsps( a2, 10, v, vec_sizes, 1, yading@10: wmavoice_dq_lsp16r1, mul_lsf, base_lsf); yading@10: dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1, yading@10: wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]); yading@10: dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1, yading@10: wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]); yading@10: } yading@10: yading@10: /** yading@10: * @} yading@10: * @name Pitch-adaptive window coding functions yading@10: * The next few functions are for pitch-adaptive window coding. yading@10: * @{ yading@10: */ yading@10: /** yading@10: * Parse the offset of the first pitch-adaptive window pulses, and yading@10: * the distribution of pulses between the two blocks in this frame. yading@10: * @param s WMA Voice decoding context private data yading@10: * @param gb bit I/O context yading@10: * @param pitch pitch for each block in this frame yading@10: */ yading@10: static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, yading@10: const int *pitch) yading@10: { yading@10: static const int16_t start_offset[94] = { yading@10: -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11, yading@10: 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26, yading@10: 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43, yading@10: 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, yading@10: 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91, yading@10: 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, yading@10: 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, yading@10: 141, 143, 145, 147, 149, 151, 153, 155, 157, 159 yading@10: }; yading@10: int bits, offset; yading@10: yading@10: /* position of pulse */ yading@10: s->aw_idx_is_ext = 0; yading@10: if ((bits = get_bits(gb, 6)) >= 54) { yading@10: s->aw_idx_is_ext = 1; yading@10: bits += (bits - 54) * 3 + get_bits(gb, 2); yading@10: } yading@10: yading@10: /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count yading@10: * the distribution of the pulses in each block contained in this frame. */ yading@10: s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16; yading@10: for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ; yading@10: s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0]; yading@10: s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2; yading@10: offset += s->aw_n_pulses[0] * pitch[0]; yading@10: s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1]; yading@10: s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2; yading@10: yading@10: /* if continuing from a position before the block, reset position to yading@10: * start of block (when corrected for the range over which it can be yading@10: * spread in aw_pulse_set1()). */ yading@10: if (start_offset[bits] < MAX_FRAMESIZE / 2) { yading@10: while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0) yading@10: s->aw_first_pulse_off[1] -= pitch[1]; yading@10: if (start_offset[bits] < 0) yading@10: while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0) yading@10: s->aw_first_pulse_off[0] -= pitch[0]; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Apply second set of pitch-adaptive window pulses. yading@10: * @param s WMA Voice decoding context private data yading@10: * @param gb bit I/O context yading@10: * @param block_idx block index in frame [0, 1] yading@10: * @param fcb structure containing fixed codebook vector info yading@10: */ yading@10: static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, yading@10: int block_idx, AMRFixed *fcb) yading@10: { yading@10: uint16_t use_mask_mem[9]; // only 5 are used, rest is padding yading@10: uint16_t *use_mask = use_mask_mem + 2; yading@10: /* in this function, idx is the index in the 80-bit (+ padding) use_mask yading@10: * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits yading@10: * of idx are the position of the bit within a particular item in the yading@10: * array (0 being the most significant bit, and 15 being the least yading@10: * significant bit), and the remainder (>> 4) is the index in the yading@10: * use_mask[]-array. This is faster and uses less memory than using a yading@10: * 80-byte/80-int array. */ yading@10: int pulse_off = s->aw_first_pulse_off[block_idx], yading@10: pulse_start, n, idx, range, aidx, start_off = 0; yading@10: yading@10: /* set offset of first pulse to within this block */ yading@10: if (s->aw_n_pulses[block_idx] > 0) yading@10: while (pulse_off + s->aw_pulse_range < 1) yading@10: pulse_off += fcb->pitch_lag; yading@10: yading@10: /* find range per pulse */ yading@10: if (s->aw_n_pulses[0] > 0) { yading@10: if (block_idx == 0) { yading@10: range = 32; yading@10: } else /* block_idx = 1 */ { yading@10: range = 8; yading@10: if (s->aw_n_pulses[block_idx] > 0) yading@10: pulse_off = s->aw_next_pulse_off_cache; yading@10: } yading@10: } else yading@10: range = 16; yading@10: pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0; yading@10: yading@10: /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly, yading@10: * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus yading@10: * we exclude that range from being pulsed again in this function. */ yading@10: memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0])); yading@10: memset( use_mask, -1, 5 * sizeof(use_mask[0])); yading@10: memset(&use_mask[5], 0, 2 * sizeof(use_mask[0])); yading@10: if (s->aw_n_pulses[block_idx] > 0) yading@10: for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) { yading@10: int excl_range = s->aw_pulse_range; // always 16 or 24 yading@10: uint16_t *use_mask_ptr = &use_mask[idx >> 4]; yading@10: int first_sh = 16 - (idx & 15); yading@10: *use_mask_ptr++ &= 0xFFFFu << first_sh; yading@10: excl_range -= first_sh; yading@10: if (excl_range >= 16) { yading@10: *use_mask_ptr++ = 0; yading@10: *use_mask_ptr &= 0xFFFF >> (excl_range - 16); yading@10: } else yading@10: *use_mask_ptr &= 0xFFFF >> excl_range; yading@10: } yading@10: yading@10: /* find the 'aidx'th offset that is not excluded */ yading@10: aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4); yading@10: for (n = 0; n <= aidx; pulse_start++) { yading@10: for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ; yading@10: if (idx >= MAX_FRAMESIZE / 2) { // find from zero yading@10: if (use_mask[0]) idx = 0x0F; yading@10: else if (use_mask[1]) idx = 0x1F; yading@10: else if (use_mask[2]) idx = 0x2F; yading@10: else if (use_mask[3]) idx = 0x3F; yading@10: else if (use_mask[4]) idx = 0x4F; yading@10: else return; yading@10: idx -= av_log2_16bit(use_mask[idx >> 4]); yading@10: } yading@10: if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) { yading@10: use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15)); yading@10: n++; yading@10: start_off = idx; yading@10: } yading@10: } yading@10: yading@10: fcb->x[fcb->n] = start_off; yading@10: fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0; yading@10: fcb->n++; yading@10: yading@10: /* set offset for next block, relative to start of that block */ yading@10: n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag; yading@10: s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0; yading@10: } yading@10: yading@10: /** yading@10: * Apply first set of pitch-adaptive window pulses. yading@10: * @param s WMA Voice decoding context private data yading@10: * @param gb bit I/O context yading@10: * @param block_idx block index in frame [0, 1] yading@10: * @param fcb storage location for fixed codebook pulse info yading@10: */ yading@10: static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, yading@10: int block_idx, AMRFixed *fcb) yading@10: { yading@10: int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx)); yading@10: float v; yading@10: yading@10: if (s->aw_n_pulses[block_idx] > 0) { yading@10: int n, v_mask, i_mask, sh, n_pulses; yading@10: yading@10: if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each yading@10: n_pulses = 3; yading@10: v_mask = 8; yading@10: i_mask = 7; yading@10: sh = 4; yading@10: } else { // 4 pulses, 1:sign + 2:index each yading@10: n_pulses = 4; yading@10: v_mask = 4; yading@10: i_mask = 3; yading@10: sh = 3; yading@10: } yading@10: yading@10: for (n = n_pulses - 1; n >= 0; n--, val >>= sh) { yading@10: fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0; yading@10: fcb->x[fcb->n] = (val & i_mask) * n_pulses + n + yading@10: s->aw_first_pulse_off[block_idx]; yading@10: while (fcb->x[fcb->n] < 0) yading@10: fcb->x[fcb->n] += fcb->pitch_lag; yading@10: if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2) yading@10: fcb->n++; yading@10: } yading@10: } else { yading@10: int num2 = (val & 0x1FF) >> 1, delta, idx; yading@10: yading@10: if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; } yading@10: else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; } yading@10: else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; } yading@10: else { delta = 7; idx = num2 + 1 - 3 * 75; } yading@10: v = (val & 0x200) ? -1.0 : 1.0; yading@10: yading@10: fcb->no_repeat_mask |= 3 << fcb->n; yading@10: fcb->x[fcb->n] = idx - delta; yading@10: fcb->y[fcb->n] = v; yading@10: fcb->x[fcb->n + 1] = idx; yading@10: fcb->y[fcb->n + 1] = (val & 1) ? -v : v; yading@10: fcb->n += 2; yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * @} yading@10: * yading@10: * Generate a random number from frame_cntr and block_idx, which will lief yading@10: * in the range [0, 1000 - block_size] (so it can be used as an index in a yading@10: * table of size 1000 of which you want to read block_size entries). yading@10: * yading@10: * @param frame_cntr current frame number yading@10: * @param block_num current block index yading@10: * @param block_size amount of entries we want to read from a table yading@10: * that has 1000 entries yading@10: * @return a (non-)random number in the [0, 1000 - block_size] range. yading@10: */ yading@10: static int pRNG(int frame_cntr, int block_num, int block_size) yading@10: { yading@10: /* array to simplify the calculation of z: yading@10: * y = (x % 9) * 5 + 6; yading@10: * z = (49995 * x) / y; yading@10: * Since y only has 9 values, we can remove the division by using a yading@10: * LUT and using FASTDIV-style divisions. For each of the 9 values yading@10: * of y, we can rewrite z as: yading@10: * z = x * (49995 / y) + x * ((49995 % y) / y) yading@10: * In this table, each col represents one possible value of y, the yading@10: * first number is 49995 / y, and the second is the FASTDIV variant yading@10: * of 49995 % y / y. */ yading@10: static const unsigned int div_tbl[9][2] = { yading@10: { 8332, 3 * 715827883U }, // y = 6 yading@10: { 4545, 0 * 390451573U }, // y = 11 yading@10: { 3124, 11 * 268435456U }, // y = 16 yading@10: { 2380, 15 * 204522253U }, // y = 21 yading@10: { 1922, 23 * 165191050U }, // y = 26 yading@10: { 1612, 23 * 138547333U }, // y = 31 yading@10: { 1388, 27 * 119304648U }, // y = 36 yading@10: { 1219, 16 * 104755300U }, // y = 41 yading@10: { 1086, 39 * 93368855U } // y = 46 yading@10: }; yading@10: unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr; yading@10: if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6, yading@10: // so this is effectively a modulo (%) yading@10: y = x - 9 * MULH(477218589, x); // x % 9 yading@10: z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1])); yading@10: // z = x * 49995 / (y * 5 + 6) yading@10: return z % (1000 - block_size); yading@10: } yading@10: yading@10: /** yading@10: * Parse hardcoded signal for a single block. yading@10: * @note see #synth_block(). yading@10: */ yading@10: static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, yading@10: int block_idx, int size, yading@10: const struct frame_type_desc *frame_desc, yading@10: float *excitation) yading@10: { yading@10: float gain; yading@10: int n, r_idx; yading@10: yading@10: av_assert0(size <= MAX_FRAMESIZE); yading@10: yading@10: /* Set the offset from which we start reading wmavoice_std_codebook */ yading@10: if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { yading@10: r_idx = pRNG(s->frame_cntr, block_idx, size); yading@10: gain = s->silence_gain; yading@10: } else /* FCB_TYPE_HARDCODED */ { yading@10: r_idx = get_bits(gb, 8); yading@10: gain = wmavoice_gain_universal[get_bits(gb, 6)]; yading@10: } yading@10: yading@10: /* Clear gain prediction parameters */ yading@10: memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err)); yading@10: yading@10: /* Apply gain to hardcoded codebook and use that as excitation signal */ yading@10: for (n = 0; n < size; n++) yading@10: excitation[n] = wmavoice_std_codebook[r_idx + n] * gain; yading@10: } yading@10: yading@10: /** yading@10: * Parse FCB/ACB signal for a single block. yading@10: * @note see #synth_block(). yading@10: */ yading@10: static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, yading@10: int block_idx, int size, yading@10: int block_pitch_sh2, yading@10: const struct frame_type_desc *frame_desc, yading@10: float *excitation) yading@10: { yading@10: static const float gain_coeff[6] = { yading@10: 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458 yading@10: }; yading@10: float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain; yading@10: int n, idx, gain_weight; yading@10: AMRFixed fcb; yading@10: yading@10: av_assert0(size <= MAX_FRAMESIZE / 2); yading@10: memset(pulses, 0, sizeof(*pulses) * size); yading@10: yading@10: fcb.pitch_lag = block_pitch_sh2 >> 2; yading@10: fcb.pitch_fac = 1.0; yading@10: fcb.no_repeat_mask = 0; yading@10: fcb.n = 0; yading@10: yading@10: /* For the other frame types, this is where we apply the innovation yading@10: * (fixed) codebook pulses of the speech signal. */ yading@10: if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { yading@10: aw_pulse_set1(s, gb, block_idx, &fcb); yading@10: aw_pulse_set2(s, gb, block_idx, &fcb); yading@10: } else /* FCB_TYPE_EXC_PULSES */ { yading@10: int offset_nbits = 5 - frame_desc->log_n_blocks; yading@10: yading@10: fcb.no_repeat_mask = -1; yading@10: /* similar to ff_decode_10_pulses_35bits(), but with single pulses yading@10: * (instead of double) for a subset of pulses */ yading@10: for (n = 0; n < 5; n++) { yading@10: float sign; yading@10: int pos1, pos2; yading@10: yading@10: sign = get_bits1(gb) ? 1.0 : -1.0; yading@10: pos1 = get_bits(gb, offset_nbits); yading@10: fcb.x[fcb.n] = n + 5 * pos1; yading@10: fcb.y[fcb.n++] = sign; yading@10: if (n < frame_desc->dbl_pulses) { yading@10: pos2 = get_bits(gb, offset_nbits); yading@10: fcb.x[fcb.n] = n + 5 * pos2; yading@10: fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign; yading@10: } yading@10: } yading@10: } yading@10: ff_set_fixed_vector(pulses, &fcb, 1.0, size); yading@10: yading@10: /* Calculate gain for adaptive & fixed codebook signal. yading@10: * see ff_amr_set_fixed_gain(). */ yading@10: idx = get_bits(gb, 7); yading@10: fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err, yading@10: gain_coeff, 6) - yading@10: 5.2409161640 + wmavoice_gain_codebook_fcb[idx]); yading@10: acb_gain = wmavoice_gain_codebook_acb[idx]; yading@10: pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx], yading@10: -2.9957322736 /* log(0.05) */, yading@10: 1.6094379124 /* log(5.0) */); yading@10: yading@10: gain_weight = 8 >> frame_desc->log_n_blocks; yading@10: memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err, yading@10: sizeof(*s->gain_pred_err) * (6 - gain_weight)); yading@10: for (n = 0; n < gain_weight; n++) yading@10: s->gain_pred_err[n] = pred_err; yading@10: yading@10: /* Calculation of adaptive codebook */ yading@10: if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { yading@10: int len; yading@10: for (n = 0; n < size; n += len) { yading@10: int next_idx_sh16; yading@10: int abs_idx = block_idx * size + n; yading@10: int pitch_sh16 = (s->last_pitch_val << 16) + yading@10: s->pitch_diff_sh16 * abs_idx; yading@10: int pitch = (pitch_sh16 + 0x6FFF) >> 16; yading@10: int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000; yading@10: idx = idx_sh16 >> 16; yading@10: if (s->pitch_diff_sh16) { yading@10: if (s->pitch_diff_sh16 > 0) { yading@10: next_idx_sh16 = (idx_sh16) &~ 0xFFFF; yading@10: } else yading@10: next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF; yading@10: len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8, yading@10: 1, size - n); yading@10: } else yading@10: len = size; yading@10: yading@10: ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch], yading@10: wmavoice_ipol1_coeffs, 17, yading@10: idx, 9, len); yading@10: } yading@10: } else /* ACB_TYPE_HAMMING */ { yading@10: int block_pitch = block_pitch_sh2 >> 2; yading@10: idx = block_pitch_sh2 & 3; yading@10: if (idx) { yading@10: ff_acelp_interpolatef(excitation, &excitation[-block_pitch], yading@10: wmavoice_ipol2_coeffs, 4, yading@10: idx, 8, size); yading@10: } else yading@10: av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch, yading@10: sizeof(float) * size); yading@10: } yading@10: yading@10: /* Interpolate ACB/FCB and use as excitation signal */ yading@10: ff_weighted_vector_sumf(excitation, excitation, pulses, yading@10: acb_gain, fcb_gain, size); yading@10: } yading@10: yading@10: /** yading@10: * Parse data in a single block. yading@10: * @note we assume enough bits are available, caller should check. yading@10: * yading@10: * @param s WMA Voice decoding context private data yading@10: * @param gb bit I/O context yading@10: * @param block_idx index of the to-be-read block yading@10: * @param size amount of samples to be read in this block yading@10: * @param block_pitch_sh2 pitch for this block << 2 yading@10: * @param lsps LSPs for (the end of) this frame yading@10: * @param prev_lsps LSPs for the last frame yading@10: * @param frame_desc frame type descriptor yading@10: * @param excitation target memory for the ACB+FCB interpolated signal yading@10: * @param synth target memory for the speech synthesis filter output yading@10: * @return 0 on success, <0 on error. yading@10: */ yading@10: static void synth_block(WMAVoiceContext *s, GetBitContext *gb, yading@10: int block_idx, int size, yading@10: int block_pitch_sh2, yading@10: const double *lsps, const double *prev_lsps, yading@10: const struct frame_type_desc *frame_desc, yading@10: float *excitation, float *synth) yading@10: { yading@10: double i_lsps[MAX_LSPS]; yading@10: float lpcs[MAX_LSPS]; yading@10: float fac; yading@10: int n; yading@10: yading@10: if (frame_desc->acb_type == ACB_TYPE_NONE) yading@10: synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation); yading@10: else yading@10: synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2, yading@10: frame_desc, excitation); yading@10: yading@10: /* convert interpolated LSPs to LPCs */ yading@10: fac = (block_idx + 0.5) / frame_desc->n_blocks; yading@10: for (n = 0; n < s->lsps; n++) // LSF -> LSP yading@10: i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n])); yading@10: ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); yading@10: yading@10: /* Speech synthesis */ yading@10: ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps); yading@10: } yading@10: yading@10: /** yading@10: * Synthesize output samples for a single frame. yading@10: * @note we assume enough bits are available, caller should check. yading@10: * yading@10: * @param ctx WMA Voice decoder context yading@10: * @param gb bit I/O context (s->gb or one for cross-packet superframes) yading@10: * @param frame_idx Frame number within superframe [0-2] yading@10: * @param samples pointer to output sample buffer, has space for at least 160 yading@10: * samples yading@10: * @param lsps LSP array yading@10: * @param prev_lsps array of previous frame's LSPs yading@10: * @param excitation target buffer for excitation signal yading@10: * @param synth target buffer for synthesized speech data yading@10: * @return 0 on success, <0 on error. yading@10: */ yading@10: static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, yading@10: float *samples, yading@10: const double *lsps, const double *prev_lsps, yading@10: float *excitation, float *synth) yading@10: { yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: int n, n_blocks_x2, log_n_blocks_x2, av_uninit(cur_pitch_val); yading@10: int pitch[MAX_BLOCKS], av_uninit(last_block_pitch); yading@10: yading@10: /* Parse frame type ("frame header"), see frame_descs */ yading@10: int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples; yading@10: yading@10: if (bd_idx < 0) { yading@10: av_log(ctx, AV_LOG_ERROR, yading@10: "Invalid frame type VLC code, skipping\n"); yading@10: return -1; yading@10: } yading@10: yading@10: block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks; yading@10: yading@10: /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */ yading@10: if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) { yading@10: /* Pitch is provided per frame, which is interpreted as the pitch of yading@10: * the last sample of the last block of this frame. We can interpolate yading@10: * the pitch of other blocks (and even pitch-per-sample) by gradually yading@10: * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */ yading@10: n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1; yading@10: log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1; yading@10: cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits); yading@10: cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1); yading@10: if (s->last_acb_type == ACB_TYPE_NONE || yading@10: 20 * abs(cur_pitch_val - s->last_pitch_val) > yading@10: (cur_pitch_val + s->last_pitch_val)) yading@10: s->last_pitch_val = cur_pitch_val; yading@10: yading@10: /* pitch per block */ yading@10: for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { yading@10: int fac = n * 2 + 1; yading@10: yading@10: pitch[n] = (MUL16(fac, cur_pitch_val) + yading@10: MUL16((n_blocks_x2 - fac), s->last_pitch_val) + yading@10: frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2; yading@10: } yading@10: yading@10: /* "pitch-diff-per-sample" for calculation of pitch per sample */ yading@10: s->pitch_diff_sh16 = yading@10: ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE; yading@10: } yading@10: yading@10: /* Global gain (if silence) and pitch-adaptive window coordinates */ yading@10: switch (frame_descs[bd_idx].fcb_type) { yading@10: case FCB_TYPE_SILENCE: yading@10: s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)]; yading@10: break; yading@10: case FCB_TYPE_AW_PULSES: yading@10: aw_parse_coords(s, gb, pitch); yading@10: break; yading@10: } yading@10: yading@10: for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { yading@10: int bl_pitch_sh2; yading@10: yading@10: /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */ yading@10: switch (frame_descs[bd_idx].acb_type) { yading@10: case ACB_TYPE_HAMMING: { yading@10: /* Pitch is given per block. Per-block pitches are encoded as an yading@10: * absolute value for the first block, and then delta values yading@10: * relative to this value) for all subsequent blocks. The scale of yading@10: * this pitch value is semi-logaritmic compared to its use in the yading@10: * decoder, so we convert it to normal scale also. */ yading@10: int block_pitch, yading@10: t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2, yading@10: t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1, yading@10: t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1; yading@10: yading@10: if (n == 0) { yading@10: block_pitch = get_bits(gb, s->block_pitch_nbits); yading@10: } else yading@10: block_pitch = last_block_pitch - s->block_delta_pitch_hrange + yading@10: get_bits(gb, s->block_delta_pitch_nbits); yading@10: /* Convert last_ so that any next delta is within _range */ yading@10: last_block_pitch = av_clip(block_pitch, yading@10: s->block_delta_pitch_hrange, yading@10: s->block_pitch_range - yading@10: s->block_delta_pitch_hrange); yading@10: yading@10: /* Convert semi-log-style scale back to normal scale */ yading@10: if (block_pitch < t1) { yading@10: bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch; yading@10: } else { yading@10: block_pitch -= t1; yading@10: if (block_pitch < t2) { yading@10: bl_pitch_sh2 = yading@10: (s->block_conv_table[1] << 2) + (block_pitch << 1); yading@10: } else { yading@10: block_pitch -= t2; yading@10: if (block_pitch < t3) { yading@10: bl_pitch_sh2 = yading@10: (s->block_conv_table[2] + block_pitch) << 2; yading@10: } else yading@10: bl_pitch_sh2 = s->block_conv_table[3] << 2; yading@10: } yading@10: } yading@10: pitch[n] = bl_pitch_sh2 >> 2; yading@10: break; yading@10: } yading@10: yading@10: case ACB_TYPE_ASYMMETRIC: { yading@10: bl_pitch_sh2 = pitch[n] << 2; yading@10: break; yading@10: } yading@10: yading@10: default: // ACB_TYPE_NONE has no pitch yading@10: bl_pitch_sh2 = 0; yading@10: break; yading@10: } yading@10: yading@10: synth_block(s, gb, n, block_nsamples, bl_pitch_sh2, yading@10: lsps, prev_lsps, &frame_descs[bd_idx], yading@10: &excitation[n * block_nsamples], yading@10: &synth[n * block_nsamples]); yading@10: } yading@10: yading@10: /* Averaging projection filter, if applicable. Else, just copy samples yading@10: * from synthesis buffer */ yading@10: if (s->do_apf) { yading@10: double i_lsps[MAX_LSPS]; yading@10: float lpcs[MAX_LSPS]; yading@10: yading@10: for (n = 0; n < s->lsps; n++) // LSF -> LSP yading@10: i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n])); yading@10: ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); yading@10: postfilter(s, synth, samples, 80, lpcs, yading@10: &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx], yading@10: frame_descs[bd_idx].fcb_type, pitch[0]); yading@10: yading@10: for (n = 0; n < s->lsps; n++) // LSF -> LSP yading@10: i_lsps[n] = cos(lsps[n]); yading@10: ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); yading@10: postfilter(s, &synth[80], &samples[80], 80, lpcs, yading@10: &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80], yading@10: frame_descs[bd_idx].fcb_type, pitch[0]); yading@10: } else yading@10: memcpy(samples, synth, 160 * sizeof(synth[0])); yading@10: yading@10: /* Cache values for next frame */ yading@10: s->frame_cntr++; yading@10: if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%) yading@10: s->last_acb_type = frame_descs[bd_idx].acb_type; yading@10: switch (frame_descs[bd_idx].acb_type) { yading@10: case ACB_TYPE_NONE: yading@10: s->last_pitch_val = 0; yading@10: break; yading@10: case ACB_TYPE_ASYMMETRIC: yading@10: s->last_pitch_val = cur_pitch_val; yading@10: break; yading@10: case ACB_TYPE_HAMMING: yading@10: s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1]; yading@10: break; yading@10: } yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Ensure minimum value for first item, maximum value for last value, yading@10: * proper spacing between each value and proper ordering. yading@10: * yading@10: * @param lsps array of LSPs yading@10: * @param num size of LSP array yading@10: * yading@10: * @note basically a double version of #ff_acelp_reorder_lsf(), might be yading@10: * useful to put in a generic location later on. Parts are also yading@10: * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(), yading@10: * which is in float. yading@10: */ yading@10: static void stabilize_lsps(double *lsps, int num) yading@10: { yading@10: int n, m, l; yading@10: yading@10: /* set minimum value for first, maximum value for last and minimum yading@10: * spacing between LSF values. yading@10: * Very similar to ff_set_min_dist_lsf(), but in double. */ yading@10: lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI); yading@10: for (n = 1; n < num; n++) yading@10: lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI); yading@10: lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI); yading@10: yading@10: /* reorder (looks like one-time / non-recursed bubblesort). yading@10: * Very similar to ff_sort_nearly_sorted_floats(), but in double. */ yading@10: for (n = 1; n < num; n++) { yading@10: if (lsps[n] < lsps[n - 1]) { yading@10: for (m = 1; m < num; m++) { yading@10: double tmp = lsps[m]; yading@10: for (l = m - 1; l >= 0; l--) { yading@10: if (lsps[l] <= tmp) break; yading@10: lsps[l + 1] = lsps[l]; yading@10: } yading@10: lsps[l + 1] = tmp; yading@10: } yading@10: break; yading@10: } yading@10: } yading@10: } yading@10: yading@10: /** yading@10: * Test if there's enough bits to read 1 superframe. yading@10: * yading@10: * @param orig_gb bit I/O context used for reading. This function yading@10: * does not modify the state of the bitreader; it yading@10: * only uses it to copy the current stream position yading@10: * @param s WMA Voice decoding context private data yading@10: * @return -1 if unsupported, 1 on not enough bits or 0 if OK. yading@10: */ yading@10: static int check_bits_for_superframe(GetBitContext *orig_gb, yading@10: WMAVoiceContext *s) yading@10: { yading@10: GetBitContext s_gb, *gb = &s_gb; yading@10: int n, need_bits, bd_idx; yading@10: const struct frame_type_desc *frame_desc; yading@10: yading@10: /* initialize a copy */ yading@10: init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits); yading@10: skip_bits_long(gb, get_bits_count(orig_gb)); yading@10: av_assert1(get_bits_left(gb) == get_bits_left(orig_gb)); yading@10: yading@10: /* superframe header */ yading@10: if (get_bits_left(gb) < 14) yading@10: return 1; yading@10: if (!get_bits1(gb)) yading@10: return -1; // WMAPro-in-WMAVoice superframe yading@10: if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe yading@10: if (s->has_residual_lsps) { // residual LSPs (for all frames) yading@10: if (get_bits_left(gb) < s->sframe_lsp_bitsize) yading@10: return 1; yading@10: skip_bits_long(gb, s->sframe_lsp_bitsize); yading@10: } yading@10: yading@10: /* frames */ yading@10: for (n = 0; n < MAX_FRAMES; n++) { yading@10: int aw_idx_is_ext = 0; yading@10: yading@10: if (!s->has_residual_lsps) { // independent LSPs (per-frame) yading@10: if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1; yading@10: skip_bits_long(gb, s->frame_lsp_bitsize); yading@10: } yading@10: bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)]; yading@10: if (bd_idx < 0) yading@10: return -1; // invalid frame type VLC code yading@10: frame_desc = &frame_descs[bd_idx]; yading@10: if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { yading@10: if (get_bits_left(gb) < s->pitch_nbits) yading@10: return 1; yading@10: skip_bits_long(gb, s->pitch_nbits); yading@10: } yading@10: if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { yading@10: skip_bits(gb, 8); yading@10: } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { yading@10: int tmp = get_bits(gb, 6); yading@10: if (tmp >= 0x36) { yading@10: skip_bits(gb, 2); yading@10: aw_idx_is_ext = 1; yading@10: } yading@10: } yading@10: yading@10: /* blocks */ yading@10: if (frame_desc->acb_type == ACB_TYPE_HAMMING) { yading@10: need_bits = s->block_pitch_nbits + yading@10: (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits; yading@10: } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { yading@10: need_bits = 2 * !aw_idx_is_ext; yading@10: } else yading@10: need_bits = 0; yading@10: need_bits += frame_desc->frame_size; yading@10: if (get_bits_left(gb) < need_bits) yading@10: return 1; yading@10: skip_bits_long(gb, need_bits); yading@10: } yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Synthesize output samples for a single superframe. If we have any data yading@10: * cached in s->sframe_cache, that will be used instead of whatever is loaded yading@10: * in s->gb. yading@10: * yading@10: * WMA Voice superframes contain 3 frames, each containing 160 audio samples, yading@10: * to give a total of 480 samples per frame. See #synth_frame() for frame yading@10: * parsing. In addition to 3 frames, superframes can also contain the LSPs yading@10: * (if these are globally specified for all frames (residually); they can yading@10: * also be specified individually per-frame. See the s->has_residual_lsps yading@10: * option), and can specify the number of samples encoded in this superframe yading@10: * (if less than 480), usually used to prevent blanks at track boundaries. yading@10: * yading@10: * @param ctx WMA Voice decoder context yading@10: * @return 0 on success, <0 on error or 1 if there was not enough data to yading@10: * fully parse the superframe yading@10: */ yading@10: static int synth_superframe(AVCodecContext *ctx, AVFrame *frame, yading@10: int *got_frame_ptr) yading@10: { yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: GetBitContext *gb = &s->gb, s_gb; yading@10: int n, res, n_samples = 480; yading@10: double lsps[MAX_FRAMES][MAX_LSPS]; yading@10: const double *mean_lsf = s->lsps == 16 ? yading@10: wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode]; yading@10: float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12]; yading@10: float synth[MAX_LSPS + MAX_SFRAMESIZE]; yading@10: float *samples; yading@10: yading@10: memcpy(synth, s->synth_history, yading@10: s->lsps * sizeof(*synth)); yading@10: memcpy(excitation, s->excitation_history, yading@10: s->history_nsamples * sizeof(*excitation)); yading@10: yading@10: if (s->sframe_cache_size > 0) { yading@10: gb = &s_gb; yading@10: init_get_bits(gb, s->sframe_cache, s->sframe_cache_size); yading@10: s->sframe_cache_size = 0; yading@10: } yading@10: yading@10: if ((res = check_bits_for_superframe(gb, s)) == 1) { yading@10: *got_frame_ptr = 0; yading@10: return 1; yading@10: } yading@10: yading@10: /* First bit is speech/music bit, it differentiates between WMAVoice yading@10: * speech samples (the actual codec) and WMAVoice music samples, which yading@10: * are really WMAPro-in-WMAVoice-superframes. I've never seen those in yading@10: * the wild yet. */ yading@10: if (!get_bits1(gb)) { yading@10: avpriv_request_sample(ctx, "WMAPro-in-WMAVoice"); yading@10: return AVERROR_PATCHWELCOME; yading@10: } yading@10: yading@10: /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */ yading@10: if (get_bits1(gb)) { yading@10: if ((n_samples = get_bits(gb, 12)) > 480) { yading@10: av_log(ctx, AV_LOG_ERROR, yading@10: "Superframe encodes >480 samples (%d), not allowed\n", yading@10: n_samples); yading@10: return -1; yading@10: } yading@10: } yading@10: /* Parse LSPs, if global for the superframe (can also be per-frame). */ yading@10: if (s->has_residual_lsps) { yading@10: double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2]; yading@10: yading@10: for (n = 0; n < s->lsps; n++) yading@10: prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n]; yading@10: yading@10: if (s->lsps == 10) { yading@10: dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); yading@10: } else /* s->lsps == 16 */ yading@10: dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); yading@10: yading@10: for (n = 0; n < s->lsps; n++) { yading@10: lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]); yading@10: lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]); yading@10: lsps[2][n] += mean_lsf[n]; yading@10: } yading@10: for (n = 0; n < 3; n++) yading@10: stabilize_lsps(lsps[n], s->lsps); yading@10: } yading@10: yading@10: /* get output buffer */ yading@10: frame->nb_samples = 480; yading@10: if ((res = ff_get_buffer(ctx, frame, 0)) < 0) yading@10: return res; yading@10: frame->nb_samples = n_samples; yading@10: samples = (float *)frame->data[0]; yading@10: yading@10: /* Parse frames, optionally preceded by per-frame (independent) LSPs. */ yading@10: for (n = 0; n < 3; n++) { yading@10: if (!s->has_residual_lsps) { yading@10: int m; yading@10: yading@10: if (s->lsps == 10) { yading@10: dequant_lsp10i(gb, lsps[n]); yading@10: } else /* s->lsps == 16 */ yading@10: dequant_lsp16i(gb, lsps[n]); yading@10: yading@10: for (m = 0; m < s->lsps; m++) yading@10: lsps[n][m] += mean_lsf[m]; yading@10: stabilize_lsps(lsps[n], s->lsps); yading@10: } yading@10: yading@10: if ((res = synth_frame(ctx, gb, n, yading@10: &samples[n * MAX_FRAMESIZE], yading@10: lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1], yading@10: &excitation[s->history_nsamples + n * MAX_FRAMESIZE], yading@10: &synth[s->lsps + n * MAX_FRAMESIZE]))) { yading@10: *got_frame_ptr = 0; yading@10: return res; yading@10: } yading@10: } yading@10: yading@10: /* Statistics? FIXME - we don't check for length, a slight overrun yading@10: * will be caught by internal buffer padding, and anything else yading@10: * will be skipped, not read. */ yading@10: if (get_bits1(gb)) { yading@10: res = get_bits(gb, 4); yading@10: skip_bits(gb, 10 * (res + 1)); yading@10: } yading@10: yading@10: *got_frame_ptr = 1; yading@10: yading@10: /* Update history */ yading@10: memcpy(s->prev_lsps, lsps[2], yading@10: s->lsps * sizeof(*s->prev_lsps)); yading@10: memcpy(s->synth_history, &synth[MAX_SFRAMESIZE], yading@10: s->lsps * sizeof(*synth)); yading@10: memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE], yading@10: s->history_nsamples * sizeof(*excitation)); yading@10: if (s->do_apf) yading@10: memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE], yading@10: s->history_nsamples * sizeof(*s->zero_exc_pf)); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Parse the packet header at the start of each packet (input data to this yading@10: * decoder). yading@10: * yading@10: * @param s WMA Voice decoding context private data yading@10: * @return 1 if not enough bits were available, or 0 on success. yading@10: */ yading@10: static int parse_packet_header(WMAVoiceContext *s) yading@10: { yading@10: GetBitContext *gb = &s->gb; yading@10: unsigned int res; yading@10: yading@10: if (get_bits_left(gb) < 11) yading@10: return 1; yading@10: skip_bits(gb, 4); // packet sequence number yading@10: s->has_residual_lsps = get_bits1(gb); yading@10: do { yading@10: res = get_bits(gb, 6); // number of superframes per packet yading@10: // (minus first one if there is spillover) yading@10: if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize) yading@10: return 1; yading@10: } while (res == 0x3F); yading@10: s->spillover_nbits = get_bits(gb, s->spillover_bitsize); yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: /** yading@10: * Copy (unaligned) bits from gb/data/size to pb. yading@10: * yading@10: * @param pb target buffer to copy bits into yading@10: * @param data source buffer to copy bits from yading@10: * @param size size of the source data, in bytes yading@10: * @param gb bit I/O context specifying the current position in the source. yading@10: * data. This function might use this to align the bit position to yading@10: * a whole-byte boundary before calling #avpriv_copy_bits() on aligned yading@10: * source data yading@10: * @param nbits the amount of bits to copy from source to target yading@10: * yading@10: * @note after calling this function, the current position in the input bit yading@10: * I/O context is undefined. yading@10: */ yading@10: static void copy_bits(PutBitContext *pb, yading@10: const uint8_t *data, int size, yading@10: GetBitContext *gb, int nbits) yading@10: { yading@10: int rmn_bytes, rmn_bits; yading@10: yading@10: rmn_bits = rmn_bytes = get_bits_left(gb); yading@10: if (rmn_bits < nbits) yading@10: return; yading@10: if (nbits > pb->size_in_bits - put_bits_count(pb)) yading@10: return; yading@10: rmn_bits &= 7; rmn_bytes >>= 3; yading@10: if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0) yading@10: put_bits(pb, rmn_bits, get_bits(gb, rmn_bits)); yading@10: avpriv_copy_bits(pb, data + size - rmn_bytes, yading@10: FFMIN(nbits - rmn_bits, rmn_bytes << 3)); yading@10: } yading@10: yading@10: /** yading@10: * Packet decoding: a packet is anything that the (ASF) demuxer contains, yading@10: * and we expect that the demuxer / application provides it to us as such yading@10: * (else you'll probably get garbage as output). Every packet has a size of yading@10: * ctx->block_align bytes, starts with a packet header (see yading@10: * #parse_packet_header()), and then a series of superframes. Superframe yading@10: * boundaries may exceed packets, i.e. superframes can split data over yading@10: * multiple (two) packets. yading@10: * yading@10: * For more information about frames, see #synth_superframe(). yading@10: */ yading@10: static int wmavoice_decode_packet(AVCodecContext *ctx, void *data, yading@10: int *got_frame_ptr, AVPacket *avpkt) yading@10: { yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: GetBitContext *gb = &s->gb; yading@10: int size, res, pos; yading@10: yading@10: /* Packets are sometimes a multiple of ctx->block_align, with a packet yading@10: * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer yading@10: * feeds us ASF packets, which may concatenate multiple "codec" packets yading@10: * in a single "muxer" packet, so we artificially emulate that by yading@10: * capping the packet size at ctx->block_align. */ yading@10: for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align); yading@10: if (!size) { yading@10: *got_frame_ptr = 0; yading@10: return 0; yading@10: } yading@10: init_get_bits(&s->gb, avpkt->data, size << 3); yading@10: yading@10: /* size == ctx->block_align is used to indicate whether we are dealing with yading@10: * a new packet or a packet of which we already read the packet header yading@10: * previously. */ yading@10: if (size == ctx->block_align) { // new packet header yading@10: if ((res = parse_packet_header(s)) < 0) yading@10: return res; yading@10: yading@10: /* If the packet header specifies a s->spillover_nbits, then we want yading@10: * to push out all data of the previous packet (+ spillover) before yading@10: * continuing to parse new superframes in the current packet. */ yading@10: if (s->spillover_nbits > 0) { yading@10: if (s->sframe_cache_size > 0) { yading@10: int cnt = get_bits_count(gb); yading@10: copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits); yading@10: flush_put_bits(&s->pb); yading@10: s->sframe_cache_size += s->spillover_nbits; yading@10: if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 && yading@10: *got_frame_ptr) { yading@10: cnt += s->spillover_nbits; yading@10: s->skip_bits_next = cnt & 7; yading@10: return cnt >> 3; yading@10: } else yading@10: skip_bits_long (gb, s->spillover_nbits - cnt + yading@10: get_bits_count(gb)); // resync yading@10: } else yading@10: skip_bits_long(gb, s->spillover_nbits); // resync yading@10: } yading@10: } else if (s->skip_bits_next) yading@10: skip_bits(gb, s->skip_bits_next); yading@10: yading@10: /* Try parsing superframes in current packet */ yading@10: s->sframe_cache_size = 0; yading@10: s->skip_bits_next = 0; yading@10: pos = get_bits_left(gb); yading@10: if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) { yading@10: return res; yading@10: } else if (*got_frame_ptr) { yading@10: int cnt = get_bits_count(gb); yading@10: s->skip_bits_next = cnt & 7; yading@10: return cnt >> 3; yading@10: } else if ((s->sframe_cache_size = pos) > 0) { yading@10: /* rewind bit reader to start of last (incomplete) superframe... */ yading@10: init_get_bits(gb, avpkt->data, size << 3); yading@10: skip_bits_long(gb, (size << 3) - pos); yading@10: av_assert1(get_bits_left(gb) == pos); yading@10: yading@10: /* ...and cache it for spillover in next packet */ yading@10: init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE); yading@10: copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size); yading@10: // FIXME bad - just copy bytes as whole and add use the yading@10: // skip_bits_next field yading@10: } yading@10: yading@10: return size; yading@10: } yading@10: yading@10: static av_cold int wmavoice_decode_end(AVCodecContext *ctx) yading@10: { yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: yading@10: if (s->do_apf) { yading@10: ff_rdft_end(&s->rdft); yading@10: ff_rdft_end(&s->irdft); yading@10: ff_dct_end(&s->dct); yading@10: ff_dct_end(&s->dst); yading@10: } yading@10: yading@10: return 0; yading@10: } yading@10: yading@10: static av_cold void wmavoice_flush(AVCodecContext *ctx) yading@10: { yading@10: WMAVoiceContext *s = ctx->priv_data; yading@10: int n; yading@10: yading@10: s->postfilter_agc = 0; yading@10: s->sframe_cache_size = 0; yading@10: s->skip_bits_next = 0; yading@10: for (n = 0; n < s->lsps; n++) yading@10: s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); yading@10: memset(s->excitation_history, 0, yading@10: sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY); yading@10: memset(s->synth_history, 0, yading@10: sizeof(*s->synth_history) * MAX_LSPS); yading@10: memset(s->gain_pred_err, 0, yading@10: sizeof(s->gain_pred_err)); yading@10: yading@10: if (s->do_apf) { yading@10: memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0, yading@10: sizeof(*s->synth_filter_out_buf) * s->lsps); yading@10: memset(s->dcf_mem, 0, yading@10: sizeof(*s->dcf_mem) * 2); yading@10: memset(s->zero_exc_pf, 0, yading@10: sizeof(*s->zero_exc_pf) * s->history_nsamples); yading@10: memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache)); yading@10: } yading@10: } yading@10: yading@10: AVCodec ff_wmavoice_decoder = { yading@10: .name = "wmavoice", yading@10: .type = AVMEDIA_TYPE_AUDIO, yading@10: .id = AV_CODEC_ID_WMAVOICE, yading@10: .priv_data_size = sizeof(WMAVoiceContext), yading@10: .init = wmavoice_decode_init, yading@10: .close = wmavoice_decode_end, yading@10: .decode = wmavoice_decode_packet, yading@10: .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1, yading@10: .flush = wmavoice_flush, yading@10: .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"), yading@10: };