/* * AMR narrowband decoder * Copyright (c) 2006-2007 Robert Swain * Copyright (c) 2009 Colin McQuillan * * This file is part of FFmpeg. * * FFmpeg is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public * License as published by the Free Software Foundation; either * version 2.1 of the License, or (at your option) any later version. * * FFmpeg is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * Lesser General Public License for more details. * * You should have received a copy of the GNU Lesser General Public * License along with FFmpeg; if not, write to the Free Software * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ /** * @file * AMR narrowband decoder * * This decoder uses floats for simplicity and so is not bit-exact. One * difference is that differences in phase can accumulate. The test sequences * in 3GPP TS 26.074 can still be useful. * * - Comparing this file's output to the output of the ref decoder gives a * PSNR of 30 to 80. Plotting the output samples shows a difference in * phase in some areas. * * - Comparing both decoders against their input, this decoder gives a similar * PSNR. If the test sequence homing frames are removed (this decoder does * not detect them), the PSNR is at least as good as the reference on 140 * out of 169 tests. */ #include #include #include #include #include #include #include "celp_filters.h" //#include "acelp_filters.h" #include "acelp_vectors.h" #include "amrnbdata.h" #define AMR_BLOCK_SIZE 160 ///< samples per frame #define AMR_SAMPLE_BOUND 32767.0f ///< threshold for synthesis overflow /* Useful macros. */ #define FFMAX(a,b) ((a) > (b) ? (a) : (b)) #define FFMIN(a,b) ((a) > (b) ? (b) : (a)) #define FFABS(a) ((a) >= 0 ? (a) : (-(a))) #define FF_ARRAY_ELEMS(a) (sizeof(a) / sizeof((a)[0])) #define DECLARE_ALIGNED(n,t,v) t __attribute__ ((aligned (n))) v #define MAX_LP_HALF_ORDER 8 #define MAX_LP_ORDER (2*MAX_LP_HALF_ORDER) #ifndef ff_lsp2polyf void ff_lsp2polyf(const float *lsp, float *f, int lp_half_order) { int i, j; f[0] = 1.0f; f[1] = -2 * lsp[0]; lsp -= 2; for(i=2; i<=lp_half_order; i++) { float val = -2 * lsp[2*i]; f[i] = val * f[i-1] + 2*f[i-2]; for(j=i-1; j>1; j--) f[j] += f[j-1] * val + f[j-2]; f[1] += val; } } #endif /* ff_lsp2polyf */ /** * Reconstruct LPC coefficients from the line spectral pair frequencies. * * @param lsp line spectral pairs in cosine domain * @param lpc linear predictive coding coefficients * @param lp_half_order half the number of the amount of LPCs to be * reconstructed, need to be smaller or equal to MAX_LP_HALF_ORDER * * @note buffers should have a minimux size of 2*lp_half_order elements. * * TIA/EIA/IS-733 2.4.3.3.5 */ void ff_acelp_lspf2lpc(const float *lsp, float *lpc, int lp_half_order) { float pa[MAX_LP_HALF_ORDER+1], qa[MAX_LP_HALF_ORDER+1]; float *lpc2 = lpc + (lp_half_order << 1) - 1; //av_assert2(lp_half_order <= MAX_LP_HALF_ORDER); ff_lsp2polyf(lsp, pa, lp_half_order); ff_lsp2polyf(lsp + 1, qa, lp_half_order); while (lp_half_order--) { float paf = pa[lp_half_order+1] + pa[lp_half_order]; float qaf = qa[lp_half_order+1] - qa[lp_half_order]; lpc [ lp_half_order] = 0.5f*(paf+qaf); lpc2[-lp_half_order] = 0.5f*(paf-qaf); } } const float ff_b60_sinc[61] = { 0.898529 , 0.865051 , 0.769257 , 0.624054 , 0.448639 , 0.265289 , 0.0959167 , -0.0412598 , -0.134338 , -0.178986 , -0.178528 , -0.142609 , -0.0849304 , -0.0205078 , 0.0369568 , 0.0773926 , 0.0955200 , 0.0912781 , 0.0689392 , 0.0357056 , 0. , -0.0305481 , -0.0504150 , -0.0570068 , -0.0508423 , -0.0350037 , -0.0141602 , 0.00665283, 0.0230713 , 0.0323486 , 0.0335388 , 0.0275879 , 0.0167847 , 0.00411987, -0.00747681, -0.0156860 , -0.0193481 , -0.0183716 , -0.0137634 , -0.00704956, 0. , 0.00582886 , 0.00939941, 0.0103760 , 0.00903320, 0.00604248, 0.00238037, -0.00109863 , -0.00366211, -0.00497437, -0.00503540, -0.00402832, -0.00241089, -0.000579834, 0.00103760, 0.00222778, 0.00277710, 0.00271606, 0.00213623, 0.00115967 , 0. }; void ff_acelp_interpolatef(float *out, const float *in, const float *filter_coeffs, int precision, int frac_pos, int filter_length, int length) { int n, i; for (n = 0; n < length; n++) { int idx = 0; float v = 0; for (i = 0; i < filter_length;) { v += in[n + i] * filter_coeffs[idx + frac_pos]; idx += precision; i++; v += in[n - i] * filter_coeffs[idx - frac_pos]; } out[n] = v; } } static inline int av_clip(int a, int amin, int amax) { if (a < amin) return amin; else if (a > amax) return amax; else return a; } static inline float av_clipf(float d0, float min, float max) { __asm__ __volatile__("maxss %1, %0\n" "minss %2, %0\n" :"+x"(d0) :"x"(min), "x"(max)); return d0; } /** * Scale from constructed speech to [-1,1] * * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but * upscales by two (section 6.2.2). * * Fundamentally, this scale is determined by energy_mean through * the fixed vector contribution to the excitation vector. */ #define AMR_SAMPLE_SCALE (2.0f / 32768.0f) /** Prediction factor for 12.2kbit/s mode */ #define PRED_FAC_MODE_12k2 0.65f #define LSF_R_FAC (8000.0f / 32768.0f) ///< LSF residual tables to Hertz #define MIN_LSF_SPACING (50.0488f / 8000.0f) ///< Ensures stability of LPC filter #define PITCH_DELAY_MIN 20 #define PITCH_DELAY_MAX 143 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode /** Initial energy in dB. Also used for bad frames (unimplemented). */ #define MIN_ENERGY -14.0f /** Maximum sharpening factor * * The specification says 0.8, which should be 13107, but the reference C code * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.) */ #define SHARP_MAX 0.79449462890625f /** Number of impulse response coefficients used for tilt factor */ #define AMR_TILT_RESPONSE 22 /** Tilt factor = 1st reflection coefficient * gamma_t */ #define AMR_TILT_GAMMA_T 0.8f typedef struct AVPacket { int64_t pts; uint8_t *data; uint64_t size; } AVPacket; typedef struct AMRContext { AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc) uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0 enum Mode cur_frame_mode; int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe float lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame float prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history float *excitation; ///< pointer to the current excitation vector in excitation_buf float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames) float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX] uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65 uint8_t hang_count; ///< the number of subframes since a hangover period started float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset" uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none uint8_t ir_filter_onset; ///< flag for impulse response filter strength float postfilter_mem[10]; ///< previous intermediate values in the formant filter float tilt_mem; ///< previous input to tilt compensation filter float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter CELPFContext celpf_ctx; float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples } AMRContext; int amrnb_decode_init(AMRContext *p) { int i; // p->excitation always points to the same position in p->excitation_buf p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1]; for (i = 0; i < LP_FILTER_ORDER; i++) { p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15); p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15); } for (i = 0; i < 4; i++) p->prediction_error[i] = MIN_ENERGY; ff_celp_filter_init(&p->celpf_ctx); return 0; } /** * Unpack an RFC4867 speech frame into the AMR frame mode and parameters. * * The order of speech bits is specified by 3GPP TS 26.101. * * @param p the context * @param buf pointer to the input buffer * @param buf_size size of the input buffer * * @return the frame mode */ static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf, int buf_size) { enum Mode mode; // Decode the first octet. mode = buf[0] >> 3 & 0x0F; // frame type p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit if (mode < MODE_DTX) { uint16_t *data = (uint16_t *)&p->frame; const uint8_t *order = amr_unpacking_bitmaps_per_mode[mode]; int field_size; memset(&p->frame, 0, sizeof(AMRNBFrame)); buf++; while ((field_size = *order++)) { int field = 0; int field_offset = *order++; while (field_size--) { int bit = *order++; field <<= 1; field |= buf[bit >> 3] >> (bit & 7) & 1; } data[field_offset] = field; } } return mode; } /// @defgroup amr_lpc_decoding AMR pitch LPC coefficient decoding functions /// @{ /** * Convert an lsf vector into an lsp vector. * * @param lsf input lsf vector * @param lsp output lsp vector */ // These are macros for evaluating polynomials using Estrin's method #define POLY2(x, c1, c0) ((x)*(c1)+(c0)) #define POLY3(x, c2, c1, c0) (((x)*(c2)+(c1))*(x)+(c0)) #define POLY4(x, x2, c3, c2, c1, c0) ((c0 + x*c1) + x2*(c2 + x*c3)) #define POLY5(x, x2, x4, c4, c3, c2, c1, c0) ((x4)*(c4)+POLY4(x, x2, c3, c2, c1, c0)) #define POLY6(x, x2, x4, c5, c4, c3, c2, c1, c0) ((x4)*POLY2(x, c5, c4)+POLY4(x, x2, c3, c2, c1, c0)) #define POLY7(x, x2, x4, c6, c5, c4, c3, c2, c1, c0) ((x4)*POLY3(x, c6, c5, c4)+POLY4(x, x2, c3, c2, c1, c0)) #define POLY8(x, x2, x4, c7, c6, c5, c4, c3, c2, c1, c0) ((x4)*POLY4(x, x2, c7, c6, c5, c4)+POLY4(x, x2, c3, c2, c1, c0)) static const float C1 = -5.0000000000e-01, C2 = 4.1666667908e-02, /* 0x3d2aaaab */ C3 = -1.3888889225e-03, /* 0xbab60b61 */ C4 = 2.4801587642e-05, /* 0x37d00d01 */ C5 = -2.7557314297e-07, /* 0xb493f27c */ C6 = 2.0875723372e-09, /* 0x310f74f6 */ C7 = -1.1359647598e-11; /* 0xad47d74e */ // Differs from libc cosf on [0, pi/2] at only the point 0x3fc460f8 (~1.5342093) by 0.0000001005828f static inline float k_cosf(float x) { float z,z2,z4,r; z = x*x; z2 = z*z; z4 = z2*z2; r = POLY6(z, z2, z4, C7, C6, C5, C4, C3, C2); r = z*r; return 1.0f + (z*r + C1*z); } static const float pi = 3.14159265358979323846f; typedef union _f32 { float f; uint32_t i; } _f32; float cosf_pi(float x) { unsigned int n = 0; float y; _f32 c; x = fabsf(x); /* argument reduction */ n = lrintf(x); y = (x - (float)n); /* y = |x| mod 1.0 */ c.f = k_cosf(pi*y); /* both gcc and clang keep stubbornly compiling this to a branch, so we need to handhold their hands. sigh. if (n&1) c = -c;*/ c.i ^= (n << 31); return c.f; } static void lsf2lsp(float *lsp, const float *lsf) { int i; for (i = 0; i < LP_FILTER_ORDER; i++) lsp[i] = cosf_pi(2.0f * lsf[i]); } float ff_dot_productf(const float* a, const float* b, int length) { float sum = 0; int i; for(i=0; i> 1; i++) memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset], 2 * sizeof(*lsf_r)); if (sign) { lsf_r[4] *= -1; lsf_r[5] *= -1; } if (update) memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(int16_t)); for (i = 0; i < LP_FILTER_ORDER; i++) lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0f) + lsf_no_r[i] * (1.0f / 8000.0f); //ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); for (i = 0; i < LP_FILTER_ORDER; i++) prev_lsf_q = lsf_q[i] = FFMAX(lsf_q[i], prev_lsf_q + MIN_LSF_SPACING); if (update) interpolate_lsf(p->lsf_q, lsf_q); lsf2lsp(lsp, lsf_q); } /** * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors. * * @param p pointer to the AMRContext */ static void lsf2lsp_5(AMRContext *p) { const uint16_t *lsf_param = p->frame.lsf; float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector const int16_t *lsf_quantizer[5]; int i; lsf_quantizer[0] = lsf_5_1[lsf_param[0]]; lsf_quantizer[1] = lsf_5_2[lsf_param[1]]; lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1]; lsf_quantizer[3] = lsf_5_4[lsf_param[3]]; lsf_quantizer[4] = lsf_5_5[lsf_param[4]]; for (i = 0; i < LP_FILTER_ORDER; i++) lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i]; lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0); lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1); // interpolate LSP vectors at subframes 1 and 3 ff_weighted_vector_sumf(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER); ff_weighted_vector_sumf(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER); } /** * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector. * * @param p pointer to the AMRContext */ static void lsf2lsp_3(AMRContext *p) { const uint16_t *lsf_param = p->frame.lsf; int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector float prev_lsf_q = 0.0; const int16_t *lsf_quantizer; int i, j; lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]]; memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r)); lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)]; memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r)); lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]]; memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r)); // calculate mean-removed LSF vector and add mean for (i = 0; i < LP_FILTER_ORDER; i++) lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0f) + lsf_3_mean[i] * (1.0f / 8000.0f); //ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); for (i = 0; i < LP_FILTER_ORDER; i++) prev_lsf_q = lsf_q[i] = FFMAX(lsf_q[i], prev_lsf_q + MIN_LSF_SPACING); // store data for computing the next frame's LSFs interpolate_lsf(p->lsf_q, lsf_q); memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); lsf2lsp(p->lsp[3], lsf_q); // interpolate LSP vectors at subframes 1, 2 and 3 for (i = 1; i <= 3; i++) for(j = 0; j < LP_FILTER_ORDER; j++) p->lsp[i-1][j] = p->prev_lsp_sub4[j] + (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25f * i; } /// @} /// @defgroup amr_pitch_vector_decoding AMR pitch vector decoding functions /// @{ /** * Like ff_decode_pitch_lag(), but with 1/6 resolution */ static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index, const int prev_lag_int, const int subframe) { if ((subframe == 0) || (subframe == 2)) { if (pitch_index < 463) { *lag_int = (pitch_index + 107) * 10923 >> 16; *lag_frac = pitch_index - *lag_int * 6 + 105; } else { *lag_int = pitch_index - 368; *lag_frac = 0; } } else { *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1; *lag_frac = pitch_index - *lag_int * 6 - 3; *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2, PITCH_DELAY_MAX - 9); } } static void decode_pitch_lag_1_4(int *lag_int, int *lag_frac, int pitch_index, const int prev_lag_int, const int subframe, int third_as_first) { /* Note n * 10923 >> 15 is floor(x/3) for 0 <= n <= 32767 */ if (subframe == 0 || (subframe == 2 && third_as_first)) { if (pitch_index < 197) pitch_index += 59; else pitch_index = 3 * pitch_index - 335; } else { int search_range_min = av_clip(prev_lag_int - 5, PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9); // decoding with 4-bit resolution if (pitch_index < 4) { // integer only precision for [search_range_min, search_range_min+3] pitch_index = 3 * (pitch_index + search_range_min) + 1; } else if (pitch_index < 12) { // 1/3 fractional precision for [search_range_min+3 1/3, search_range_min+5 2/3] pitch_index += 3 * search_range_min + 7; } else { // integer only precision for [search_range_min+6, search_range_min+9] pitch_index = 3 * (pitch_index + search_range_min - 6) + 1; } } *lag_int = pitch_index * 10923 >> 15; *lag_frac = pitch_index - 3 * *lag_int - 1; } static void decode_pitch_lag(int *lag_int, int *lag_frac, int pitch_index, const int prev_lag_int, const int subframe, int resolution) { int offs; /* Note n * 10923 >> 15 is floor(x/3) for 0 <= n <= 32767 */ if ((subframe == 0) || (subframe == 2)) { if (pitch_index < 197) pitch_index += 59; else pitch_index = 3 * pitch_index - 335; } else { // decoding with 5 or 6 bit resolution, 1/3 fractional precision pitch_index--; #if 0 if (resolution == 5) { pitch_index += 3 * av_clip(prev_lag_int - 10, PITCH_DELAY_MIN, PITCH_DELAY_MAX - 19); } else { pitch_index += 3 * av_clip(prev_lag_int - 5, PITCH_DELAY_MIN, PITCH_DELAY_MAX - 9); } #endif if (resolution == 5) { offs = 10; } else { offs = 5; } pitch_index += 3 * av_clip(prev_lag_int - offs, PITCH_DELAY_MIN, PITCH_DELAY_MAX - 2*offs + 1); } *lag_int = pitch_index * 10923 >> 15; *lag_frac = pitch_index - 3 * *lag_int - 1; } static void decode_pitch_vector(AMRContext *p, const AMRNBSubframe *amr_subframe, const int subframe) { int pitch_lag_int, pitch_lag_frac; enum Mode mode = p->cur_frame_mode; if (p->cur_frame_mode == MODE_12k2) { decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac, amr_subframe->p_lag, p->pitch_lag_int, subframe); } else if (p->cur_frame_mode <= MODE_6k7) { decode_pitch_lag_1_4(&pitch_lag_int, &pitch_lag_frac, amr_subframe->p_lag, p->pitch_lag_int, subframe, mode != MODE_4k75 && mode != MODE_5k15); } else { decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac, amr_subframe->p_lag, p->pitch_lag_int, subframe, (mode == MODE_7k95 ? 5 : 6)); } p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2); pitch_lag_int += pitch_lag_frac > 0; /* Calculate the pitch vector by interpolating the past excitation at the pitch lag using a b60 hamming windowed sinc function. */ ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int, ff_b60_sinc, 6, pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0), 10, AMR_SUBFRAME_SIZE); memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float)); } /// @} /// @defgroup amr_algebraic_code_book AMR algebraic code book (fixed) vector decoding functions /// @{ /** * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame. */ static void decode_10bit_pulse(int code, int pulse_position[8], int i1, int i2, int i3) { // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of // the 3 pulses and the upper 7 bits being coded in base 5 const uint8_t *positions = base_five_table[code >> 3]; pulse_position[i1] = (positions[2] << 1) + ( code & 1); pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1); pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1); } /** * Decode the algebraic codebook index to pulse positions and signs and * construct the algebraic codebook vector for MODE_10k2. * * @param fixed_index positions of the eight pulses * @param fixed_sparse pointer to the algebraic codebook vector */ static void decode_8_pulses_31bits(const int16_t *fixed_index, AMRFixed *fixed_sparse) { int pulse_position[8]; int i, temp; decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1); decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5); // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of // the 2 pulses and the upper 5 bits being coded in base 5 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5; pulse_position[3] = temp % 5; pulse_position[7] = temp / 5; if (pulse_position[7] & 1) pulse_position[3] = 4 - pulse_position[3]; pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1); pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1); fixed_sparse->no_repeat_mask = 0; fixed_sparse->n = 8; for (i = 0; i < 4; i++) { const int pos1 = (pulse_position[i] << 2) + i; const int pos2 = (pulse_position[i + 4] << 2) + i; const float sign = fixed_index[i] ? -1.0 : 1.0; fixed_sparse->x[i ] = pos1; fixed_sparse->x[i + 4] = pos2; fixed_sparse->y[i ] = sign; fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign; } } /** * Decode the algebraic codebook index to pulse positions and signs and * construct the algebraic codebook vector for MODE_12k2. * * @param fixed_index positions of the ten pulses * @param fixed_sparse pointer to the algebraic codebook vector */ static void decode_10_pulses_35bits(const int16_t *fixed_index, AMRFixed *fixed_sparse) { int i; fixed_sparse->no_repeat_mask = 0; fixed_sparse->n = 10; for (i = 0; i < 5; i++) { const int pos1 = 5*(fixed_index[2*i+1] & 7) + i; const int pos2 = 5*(fixed_index[2*i ] & 7) + i; const float sign = (fixed_index[2*i+1] & 0x08) ? -1.0f : 1.0f; fixed_sparse->x[2*i+1] = pos1; fixed_sparse->x[2*i ] = pos2; fixed_sparse->y[2*i+1] = sign; fixed_sparse->y[2*i ] = pos2 < pos1 ? -sign : sign; } } /** * Decode the algebraic codebook index to pulse positions and signs, * then construct the algebraic codebook vector. * * nb of pulses | bits encoding pulses * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13 * * @param fixed_sparse pointer to the algebraic codebook vector * @param pulses algebraic codebook indexes * @param mode mode of the current frame * @param subframe current subframe number */ static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses, const enum Mode mode, const int subframe) { //assert(MODE_4k75 <= mode && mode <= MODE_12k2); if (mode == MODE_12k2) { decode_10_pulses_35bits(pulses, fixed_sparse); } else if (mode == MODE_10k2) { decode_8_pulses_31bits(pulses, fixed_sparse); } else { int *pulse_position = fixed_sparse->x; int i, pulse_subset; const int fixed_index = pulses[0]; if (mode <= MODE_5k15) { pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1); pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset]; pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1]; fixed_sparse->n = 2; } else if (mode == MODE_5k9) { pulse_subset = ((fixed_index & 1) << 1) + 1; pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset; pulse_subset = (fixed_index >> 4) & 3; pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0); fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2; } else if (mode == MODE_6k7) { pulse_position[0] = (fixed_index & 7) * 5; pulse_subset = (fixed_index >> 2) & 2; pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1; pulse_subset = (fixed_index >> 6) & 2; pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2; fixed_sparse->n = 3; } else { // mode <= MODE_7k95 pulse_position[0] = 5*( fixed_index & 7); pulse_position[1] = 5*((fixed_index >> 3) & 7) + 1; pulse_position[2] = 5*((fixed_index >> 6) & 7) + 2; pulse_subset = (fixed_index >> 9) & 1; pulse_position[3] = 5*((fixed_index >> 10) & 7) + pulse_subset + 3; fixed_sparse->n = 4; } for (i = 0; i < fixed_sparse->n; i++) fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0; } } void ff_set_fixed_vector(float *out, const AMRFixed *in, float scale, int size) { int i; for (i=0; i < in->n; i++) { int x = in->x[i], repeats = !((in->no_repeat_mask >> i) & 1); float y = in->y[i] * scale; do { out[x] += y; y *= in->pitch_fac; x += in->pitch_lag; } while (x < size && repeats); } } void ff_clear_fixed_vector(float *out, const AMRFixed *in, int size) { int i; for (i=0; i < in->n; i++) { int x = in->x[i], repeats = !((in->no_repeat_mask >> i) & 1); do { out[x] = 0.0; x += in->pitch_lag; } while (x < size && repeats); } } /** * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2) * * @param p the context * @param subframe unpacked amr subframe * @param mode mode of the current frame * @param fixed_sparse sparse respresentation of the fixed vector */ static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode, AMRFixed *fixed_sparse) { // The spec suggests the current pitch gain is always used, but in other // modes the pitch and codebook gains are joinly quantized (sec 5.8.2) // so the codebook gain cannot depend on the quantized pitch gain. if (mode == MODE_12k2) p->beta = FFMIN(p->pitch_gain[4], 1.0f); fixed_sparse->pitch_lag = p->pitch_lag_int; fixed_sparse->pitch_fac = p->beta; // Save pitch sharpening factor for the next subframe // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from // the fact that the gains for two subframes are jointly quantized. if (mode != MODE_4k75 || subframe & 1) p->beta = av_clipf(p->pitch_gain[4], 0.0f, SHARP_MAX); } /// @} /// @defgroup amr_gain_decoding AMR gain decoding functions /// @{ /** * fixed gain smoothing * Note that where the spec specifies the "spectrum in the q domain" * in section 6.1.4, in fact frequencies should be used. * * @param p the context * @param lsf LSFs for the current subframe, in the range [0,1] * @param lsf_avg averaged LSFs * @param mode mode of the current frame * * @return fixed gain smoothed */ static float fixed_gain_smooth(AMRContext *p , const float *lsf, const float *lsf_avg, const enum Mode mode) { float diff = 0.0; int i; for (i = 0; i < LP_FILTER_ORDER; i++) diff += fabsf(lsf_avg[i] - lsf[i]) / lsf_avg[i]; // If diff is large for ten subframes, disable smoothing for a 40-subframe // hangover period. p->diff_count++; if (diff <= 0.65f) p->diff_count = 0; if (p->diff_count > 10) { p->hang_count = 0; p->diff_count--; // don't let diff_count overflow } if (p->hang_count < 40) { p->hang_count++; } else if (mode < MODE_7k4 || mode == MODE_10k2) { const float smoothing_factor = av_clipf(4.0f * diff - 1.6f, 0.0f, 1.0f); const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] + p->fixed_gain[2] + p->fixed_gain[3] + p->fixed_gain[4]) * 0.2f; return smoothing_factor * p->fixed_gain[4] + (1.0f - smoothing_factor) * fixed_gain_mean; } return p->fixed_gain[4]; } /** * Decode pitch gain and fixed gain factor (part of section 6.1.3). * * @param p the context * @param amr_subframe unpacked amr subframe * @param mode mode of the current frame * @param subframe current subframe number * @param fixed_gain_factor decoded gain correction factor */ static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe, const enum Mode mode, const int subframe, float *fixed_gain_factor) { if (mode == MODE_12k2 || mode == MODE_7k95) { p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ] * (1.0f / 16384.0f); *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain] * (1.0f / 2048.0f); } else { const uint16_t *gains; if (mode >= MODE_6k7) { gains = gains_high[amr_subframe->p_gain]; } else if (mode >= MODE_5k15) { gains = gains_low [amr_subframe->p_gain]; } else { // gain index is only coded in subframes 0,2 for MODE_4k75 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)]; } p->pitch_gain[4] = gains[0] * (1.0f / 16384.0f); *fixed_gain_factor = gains[1] * (1.0f / 4096.0f); } } /// @} /// @defgroup amr_pre_processing AMR pre-processing functions /// @{ /** * Circularly convolve a sparse fixed vector with a phase dispersion impulse * response filter (D.6.2 of G.729 and 6.1.5 of AMR). * * @param out vector with filter applied * @param in source vector * @param filter phase filter coefficients * * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] } */ static void apply_ir_filter(float *out, const AMRFixed *in, const float *filter) { float filter1[AMR_SUBFRAME_SIZE], //!< filters at pitch lag*1 and *2 filter2[AMR_SUBFRAME_SIZE]; int lag = in->pitch_lag; float fac = in->pitch_fac; int i; if (lag < AMR_SUBFRAME_SIZE) { ff_celp_circ_addf(filter1, filter, filter, lag, fac, AMR_SUBFRAME_SIZE); if (lag < AMR_SUBFRAME_SIZE >> 1) ff_celp_circ_addf(filter2, filter, filter1, lag, fac, AMR_SUBFRAME_SIZE); } memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE); for (i = 0; i < in->n; i++) { int x = in->x[i]; float y = in->y[i]; const float *filterp; if (x >= AMR_SUBFRAME_SIZE - lag) { filterp = filter; } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) { filterp = filter1; } else { filterp = filter2; } ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE); } } /** * Reduce fixed vector sparseness by smoothing with one of three IR filters. * Also know as "adaptive phase dispersion". * * This implements 3GPP TS 26.090 section 6.1(5). * * @param p the context * @param fixed_sparse algebraic codebook vector * @param fixed_vector unfiltered fixed vector * @param fixed_gain smoothed gain * @param out space for modified vector if necessary */ static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse, const float *fixed_vector, float fixed_gain, float *out) { int ir_filter_nr; if (p->pitch_gain[4] < 0.6f) { ir_filter_nr = 0; // strong filtering } else if (p->pitch_gain[4] < 0.9f) { ir_filter_nr = 1; // medium filtering } else ir_filter_nr = 2; // no filtering // detect 'onset' if (fixed_gain > 2.0f * p->prev_sparse_fixed_gain) { p->ir_filter_onset = 2; } else if (p->ir_filter_onset) p->ir_filter_onset--; if (!p->ir_filter_onset) { int i, count = 0; for (i = 0; i < 5; i++) if (p->pitch_gain[i] < 0.6f) count++; if (count > 2) ir_filter_nr = 0; if (ir_filter_nr > p->prev_ir_filter_nr + 1) ir_filter_nr--; } else if (ir_filter_nr < 2) ir_filter_nr++; // Disable filtering for very low level of fixed_gain. // Note this step is not specified in the technical description but is in // the reference source in the function Ph_disp. if (fixed_gain < 5.0f) ir_filter_nr = 2; if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2 && ir_filter_nr < 2) { apply_ir_filter(out, fixed_sparse, (p->cur_frame_mode == MODE_7k95 ? ir_filters_lookup_MODE_7k95 : ir_filters_lookup)[ir_filter_nr]); fixed_vector = out; } // update ir filter strength history p->prev_ir_filter_nr = ir_filter_nr; p->prev_sparse_fixed_gain = fixed_gain; return fixed_vector; } /// @} /// @defgroup amr_synthesis AMR synthesis functions /// @{ /** * Conduct 10th order linear predictive coding synthesis. * * @param p pointer to the AMRContext * @param lpc pointer to the LPC coefficients * @param fixed_gain fixed codebook gain for synthesis * @param fixed_vector algebraic codebook vector * @param samples pointer to the output speech samples * @param overflow 16-bit overflow flag */ static int synthesis(AMRContext *p, float *lpc, float fixed_gain, const float *fixed_vector, float *samples, uint8_t overflow) { int i; float scalefactor, excitation[AMR_SUBFRAME_SIZE]; // if an overflow has been detected, the pitch vector is scaled down by a // factor of 4 if (overflow) for (i = 0; i < AMR_SUBFRAME_SIZE; i++) p->pitch_vector[i] *= 0.25f; ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector, p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE); // emphasize pitch vector contribution if (p->pitch_gain[4] > 0.5f && !overflow) { float energy = ff_dot_productf(excitation, excitation, AMR_SUBFRAME_SIZE); float pitch_factor = p->pitch_gain[4] * (p->cur_frame_mode == MODE_12k2 ? 0.25f * FFMIN(p->pitch_gain[4], 1.0f) : 0.5f * FFMIN(p->pitch_gain[4], SHARP_MAX)); for (i = 0; i < AMR_SUBFRAME_SIZE; i++) excitation[i] += pitch_factor * p->pitch_vector[i]; scalefactor = ff_dot_productf(excitation, excitation, AMR_SUBFRAME_SIZE); if (scalefactor) scalefactor = sqrtf(energy / scalefactor); for (i = 0; i < AMR_SUBFRAME_SIZE; i++) excitation[i] = excitation[i] * scalefactor; } p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); // detect overflow for (i = 0; i < AMR_SUBFRAME_SIZE; i++) if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) { return 1; } return 0; } /// @} /// @defgroup amr_update AMR update functions /// @{ /** * Update buffers and history at the end of decoding a subframe. * * @param p pointer to the AMRContext */ static void update_state(AMRContext *p) { memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0])); memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE], (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float)); memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float)); memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float)); memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE], LP_FILTER_ORDER * sizeof(float)); } /// @} void ff_acelp_apply_order_2_pole_transfer_function(float *out, const float *in, const float pole_coeffs[2], float gain, float mem[2], int n) { int i; float tmp; for (i = 0; i < n; i++) { tmp = gain * in[i] - pole_coeffs[0] * mem[0] - pole_coeffs[1] * mem[1]; out[i] = tmp - mem[0] - mem[0] + mem[1]; mem[1] = mem[0]; mem[0] = tmp; } } void ff_tilt_compensation(float *mem, float tilt, float *samples, int size) { float new_tilt_mem = samples[size - 1]; int i; for (i = size - 1; i > 0; i--) samples[i] -= tilt * samples[i - 1]; samples[0] -= tilt * *mem; *mem = new_tilt_mem; } /// @defgroup amr_postproc AMR Post processing functions /// @{ /** * Get the tilt factor of a formant filter from its transfer function * * @param lpc_n LP_FILTER_ORDER coefficients of the numerator * @param lpc_d LP_FILTER_ORDER coefficients of the denominator */ static float tilt_factor(CELPFContext *c, float *lpc_n, float *lpc_d) { float rh0, rh1; // autocorrelation at lag 0 and 1 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 }; float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response hf[0] = 1.0; memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER); c->celp_lp_synthesis_filterf(hf, lpc_d, hf, AMR_TILT_RESPONSE, LP_FILTER_ORDER); rh0 = ff_dot_productf(hf, hf, AMR_TILT_RESPONSE); rh1 = ff_dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1); // The spec only specifies this check for 12.2 and 10.2 kbit/s // modes. But in the ref source the tilt is always non-negative. return rh1 >= 0.0f ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0f; } /** * Perform adaptive post-filtering to enhance the quality of the speech. * See section 6.2.1. * * @param p pointer to the AMRContext * @param lpc interpolated LP coefficients for this subframe * @param buf_out output of the filter */ static void postfilter(AMRContext *p, float *lpc, float *buf_out) { int i; float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter float gamma_n, gamma_d, g_n = 1.0f, g_d = 1.0f; // Formant filter factor table float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients float tilt_fac; if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) { gamma_n = 0.7f; gamma_d = 0.75f; } else { gamma_n = 0.55f; gamma_d = 0.7f; } for (i = 0; i < LP_FILTER_ORDER; i++) { g_n *= gamma_n; g_d *= gamma_d; lpc_n[i] = lpc[i] * g_n; lpc_d[i] = lpc[i] * g_d; } memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER); p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples, AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE, sizeof(float) * LP_FILTER_ORDER); p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n, pole_out + LP_FILTER_ORDER, AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); tilt_fac = tilt_factor(&p->celpf_ctx, lpc_n, lpc_d); ff_tilt_compensation(&p->tilt_mem, tilt_fac, buf_out, AMR_SUBFRAME_SIZE); } /// @} int amrnb_decode_frame(AMRContext *p, float *data, unsigned int *data_size, AVPacket *avpkt) { const uint8_t *buf = avpkt->data; int buf_size = avpkt->size; int i, subframe; float fixed_gain_factor, fixed_mean_energy; AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing float synth_fixed_gain; // the fixed gain that synthesis should use const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use p->cur_frame_mode = unpack_bitstream(p, buf, buf_size); if (p->cur_frame_mode == MODE_DTX) { //av_log_missing_feature("dtx mode", 1); return -1; } if (p->cur_frame_mode == MODE_12k2) { lsf2lsp_5(p); } else { lsf2lsp_3(p); } for (i = 0; i < 4; i++) ff_acelp_lspf2lpc(p->lsp[i], p->lpc[i], 5); for (subframe = 0; subframe < 4; subframe++) { const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe]; decode_pitch_vector(p, amr_subframe, subframe); decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses, p->cur_frame_mode, subframe); // The fixed gain (section 6.1.3) depends on the fixed vector // (section 6.1.2), but the fixed vector calculation uses // pitch sharpening based on the on the pitch gain (section 6.1.3). // So the correct order is: pitch gain, pitch sharpening, fixed gain. decode_gains(p, amr_subframe, p->cur_frame_mode, subframe, &fixed_gain_factor); pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse); ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0, AMR_SUBFRAME_SIZE); p->fixed_gain[4] = ff_amr_set_fixed_gain(fixed_gain_factor, /*fixed_mean_energy,*/ p->prediction_error, energy_mean[p->cur_frame_mode], energy_pred_fac); fixed_mean_energy = ff_dot_productf(p->fixed_vector, p->fixed_vector, AMR_SUBFRAME_SIZE)/AMR_SUBFRAME_SIZE; p->fixed_gain[4] = p->fixed_gain[4] / sqrtf(fixed_mean_energy); // The excitation feedback is calculated without any processing such // as fixed gain smoothing. This isn't mentioned in the specification. for (i = 0; i < AMR_SUBFRAME_SIZE; i++) p->excitation[i] *= p->pitch_gain[4]; ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4], AMR_SUBFRAME_SIZE); // In the ref decoder, excitation is stored with no fractional bits. // This step prevents buzz in silent periods. The ref encoder can // emit long sequences with pitch factor greater than one. This // creates unwanted feedback if the excitation vector is nonzero. // (e.g. test sequence T19_795.COD in 3GPP TS 26.074) for (i = 0; i < AMR_SUBFRAME_SIZE; i++) p->excitation[i] = truncf(p->excitation[i]); // Smooth fixed gain. // The specification is ambiguous, but in the reference source, the // smoothed value is NOT fed back into later fixed gain smoothing. synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe], p->lsf_avg, p->cur_frame_mode); synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector, synth_fixed_gain, spare_vector); if (synthesis(p, p->lpc[subframe], synth_fixed_gain, synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0)) // overflow detected -> rerun synthesis scaling pitch vector down // by a factor of 4, skipping adaptive gain control synthesis(p, p->lpc[subframe], synth_fixed_gain, synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1); postfilter(p, p->lpc[subframe], data + subframe * AMR_SUBFRAME_SIZE); // update buffers and history ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE); update_state(p); } ff_acelp_apply_order_2_pole_transfer_function(data, data, highpass_poles, highpass_gain * AMR_SAMPLE_SCALE, p->high_pass_mem, AMR_BLOCK_SIZE); /* Update averaged lsf vector (used for fixed gain smoothing). * * Note that lsf_avg should not incorporate the current frame's LSFs * for fixed_gain_smooth. * The specification has an incorrect formula: the reference decoder uses * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */ ff_weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3], 0.84, 0.16, LP_FILTER_ORDER); /* report how many samples we got */ *data_size = AMR_BLOCK_SIZE * sizeof(float); /* return the amount of bytes consumed if everything was OK */ return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC }