yading@10: /* yading@10: * Copyright (c) 2001, 2002 Fabrice Bellard 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: #include yading@10: yading@10: #include "libavutil/mem.h" yading@10: #include "dct32.h" yading@10: #include "mathops.h" yading@10: #include "mpegaudiodsp.h" yading@10: #include "mpegaudio.h" yading@10: yading@10: #if CONFIG_FLOAT yading@10: #define RENAME(n) n##_float yading@10: yading@10: static inline float round_sample(float *sum) yading@10: { yading@10: float sum1=*sum; yading@10: *sum = 0; yading@10: return sum1; yading@10: } yading@10: yading@10: #define MACS(rt, ra, rb) rt+=(ra)*(rb) yading@10: #define MULS(ra, rb) ((ra)*(rb)) yading@10: #define MULH3(x, y, s) ((s)*(y)*(x)) yading@10: #define MLSS(rt, ra, rb) rt-=(ra)*(rb) yading@10: #define MULLx(x, y, s) ((y)*(x)) yading@10: #define FIXHR(x) ((float)(x)) yading@10: #define FIXR(x) ((float)(x)) yading@10: #define SHR(a,b) ((a)*(1.0f/(1<<(b)))) yading@10: yading@10: #else yading@10: yading@10: #define RENAME(n) n##_fixed yading@10: #define OUT_SHIFT (WFRAC_BITS + FRAC_BITS - 15) yading@10: yading@10: static inline int round_sample(int64_t *sum) yading@10: { yading@10: int sum1; yading@10: sum1 = (int)((*sum) >> OUT_SHIFT); yading@10: *sum &= (1<>(b)) yading@10: # define FIXR(a) ((int)((a) * FRAC_ONE + 0.5)) yading@10: # define FIXHR(a) ((int)((a) * (1LL<<32) + 0.5)) yading@10: #endif yading@10: yading@10: /** Window for MDCT. Actually only the elements in [0,17] and yading@10: [MDCT_BUF_SIZE/2, MDCT_BUF_SIZE/2 + 17] are actually used. The rest yading@10: is just to preserve alignment for SIMD implementations. yading@10: */ yading@10: DECLARE_ALIGNED(16, INTFLOAT, RENAME(ff_mdct_win))[8][MDCT_BUF_SIZE]; yading@10: yading@10: DECLARE_ALIGNED(16, MPA_INT, RENAME(ff_mpa_synth_window))[512+256]; yading@10: yading@10: #define SUM8(op, sum, w, p) \ yading@10: { \ yading@10: op(sum, (w)[0 * 64], (p)[0 * 64]); \ yading@10: op(sum, (w)[1 * 64], (p)[1 * 64]); \ yading@10: op(sum, (w)[2 * 64], (p)[2 * 64]); \ yading@10: op(sum, (w)[3 * 64], (p)[3 * 64]); \ yading@10: op(sum, (w)[4 * 64], (p)[4 * 64]); \ yading@10: op(sum, (w)[5 * 64], (p)[5 * 64]); \ yading@10: op(sum, (w)[6 * 64], (p)[6 * 64]); \ yading@10: op(sum, (w)[7 * 64], (p)[7 * 64]); \ yading@10: } yading@10: yading@10: #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \ yading@10: { \ yading@10: INTFLOAT tmp;\ yading@10: tmp = p[0 * 64];\ yading@10: op1(sum1, (w1)[0 * 64], tmp);\ yading@10: op2(sum2, (w2)[0 * 64], tmp);\ yading@10: tmp = p[1 * 64];\ yading@10: op1(sum1, (w1)[1 * 64], tmp);\ yading@10: op2(sum2, (w2)[1 * 64], tmp);\ yading@10: tmp = p[2 * 64];\ yading@10: op1(sum1, (w1)[2 * 64], tmp);\ yading@10: op2(sum2, (w2)[2 * 64], tmp);\ yading@10: tmp = p[3 * 64];\ yading@10: op1(sum1, (w1)[3 * 64], tmp);\ yading@10: op2(sum2, (w2)[3 * 64], tmp);\ yading@10: tmp = p[4 * 64];\ yading@10: op1(sum1, (w1)[4 * 64], tmp);\ yading@10: op2(sum2, (w2)[4 * 64], tmp);\ yading@10: tmp = p[5 * 64];\ yading@10: op1(sum1, (w1)[5 * 64], tmp);\ yading@10: op2(sum2, (w2)[5 * 64], tmp);\ yading@10: tmp = p[6 * 64];\ yading@10: op1(sum1, (w1)[6 * 64], tmp);\ yading@10: op2(sum2, (w2)[6 * 64], tmp);\ yading@10: tmp = p[7 * 64];\ yading@10: op1(sum1, (w1)[7 * 64], tmp);\ yading@10: op2(sum2, (w2)[7 * 64], tmp);\ yading@10: } yading@10: yading@10: void RENAME(ff_mpadsp_apply_window)(MPA_INT *synth_buf, MPA_INT *window, yading@10: int *dither_state, OUT_INT *samples, yading@10: int incr) yading@10: { yading@10: register const MPA_INT *w, *w2, *p; yading@10: int j; yading@10: OUT_INT *samples2; yading@10: #if CONFIG_FLOAT yading@10: float sum, sum2; yading@10: #else yading@10: int64_t sum, sum2; yading@10: #endif yading@10: yading@10: /* copy to avoid wrap */ yading@10: memcpy(synth_buf + 512, synth_buf, 32 * sizeof(*synth_buf)); yading@10: yading@10: samples2 = samples + 31 * incr; yading@10: w = window; yading@10: w2 = window + 31; yading@10: yading@10: sum = *dither_state; yading@10: p = synth_buf + 16; yading@10: SUM8(MACS, sum, w, p); yading@10: p = synth_buf + 48; yading@10: SUM8(MLSS, sum, w + 32, p); yading@10: *samples = round_sample(&sum); yading@10: samples += incr; yading@10: w++; yading@10: yading@10: /* we calculate two samples at the same time to avoid one memory yading@10: access per two sample */ yading@10: for(j=1;j<16;j++) { yading@10: sum2 = 0; yading@10: p = synth_buf + 16 + j; yading@10: SUM8P2(sum, MACS, sum2, MLSS, w, w2, p); yading@10: p = synth_buf + 48 - j; yading@10: SUM8P2(sum, MLSS, sum2, MLSS, w + 32, w2 + 32, p); yading@10: yading@10: *samples = round_sample(&sum); yading@10: samples += incr; yading@10: sum += sum2; yading@10: *samples2 = round_sample(&sum); yading@10: samples2 -= incr; yading@10: w++; yading@10: w2--; yading@10: } yading@10: yading@10: p = synth_buf + 32; yading@10: SUM8(MLSS, sum, w + 32, p); yading@10: *samples = round_sample(&sum); yading@10: *dither_state= sum; yading@10: } yading@10: yading@10: /* 32 sub band synthesis filter. Input: 32 sub band samples, Output: yading@10: 32 samples. */ yading@10: void RENAME(ff_mpa_synth_filter)(MPADSPContext *s, MPA_INT *synth_buf_ptr, yading@10: int *synth_buf_offset, yading@10: MPA_INT *window, int *dither_state, yading@10: OUT_INT *samples, int incr, yading@10: MPA_INT *sb_samples) yading@10: { yading@10: MPA_INT *synth_buf; yading@10: int offset; yading@10: yading@10: offset = *synth_buf_offset; yading@10: synth_buf = synth_buf_ptr + offset; yading@10: yading@10: s->RENAME(dct32)(synth_buf, sb_samples); yading@10: s->RENAME(apply_window)(synth_buf, window, dither_state, samples, incr); yading@10: yading@10: offset = (offset - 32) & 511; yading@10: *synth_buf_offset = offset; yading@10: } yading@10: yading@10: av_cold void RENAME(ff_mpa_synth_init)(MPA_INT *window) yading@10: { yading@10: int i, j; yading@10: yading@10: /* max = 18760, max sum over all 16 coefs : 44736 */ yading@10: for(i=0;i<257;i++) { yading@10: INTFLOAT v; yading@10: v = ff_mpa_enwindow[i]; yading@10: #if CONFIG_FLOAT yading@10: v *= 1.0 / (1LL<<(16 + FRAC_BITS)); yading@10: #endif yading@10: window[i] = v; yading@10: if ((i & 63) != 0) yading@10: v = -v; yading@10: if (i != 0) yading@10: window[512 - i] = v; yading@10: } yading@10: yading@10: yading@10: // Needed for avoiding shuffles in ASM implementations yading@10: for(i=0; i < 8; i++) yading@10: for(j=0; j < 16; j++) yading@10: window[512+16*i+j] = window[64*i+32-j]; yading@10: yading@10: for(i=0; i < 8; i++) yading@10: for(j=0; j < 16; j++) yading@10: window[512+128+16*i+j] = window[64*i+48-j]; yading@10: } yading@10: yading@10: void RENAME(ff_init_mpadsp_tabs)(void) yading@10: { yading@10: int i, j; yading@10: /* compute mdct windows */ yading@10: for (i = 0; i < 36; i++) { yading@10: for (j = 0; j < 4; j++) { yading@10: double d; yading@10: yading@10: if (j == 2 && i % 3 != 1) yading@10: continue; yading@10: yading@10: d = sin(M_PI * (i + 0.5) / 36.0); yading@10: if (j == 1) { yading@10: if (i >= 30) d = 0; yading@10: else if (i >= 24) d = sin(M_PI * (i - 18 + 0.5) / 12.0); yading@10: else if (i >= 18) d = 1; yading@10: } else if (j == 3) { yading@10: if (i < 6) d = 0; yading@10: else if (i < 12) d = sin(M_PI * (i - 6 + 0.5) / 12.0); yading@10: else if (i < 18) d = 1; yading@10: } yading@10: //merge last stage of imdct into the window coefficients yading@10: d *= 0.5 * IMDCT_SCALAR / cos(M_PI * (2 * i + 19) / 72); yading@10: yading@10: if (j == 2) yading@10: RENAME(ff_mdct_win)[j][i/3] = FIXHR((d / (1<<5))); yading@10: else { yading@10: int idx = i < 18 ? i : i + (MDCT_BUF_SIZE/2 - 18); yading@10: RENAME(ff_mdct_win)[j][idx] = FIXHR((d / (1<<5))); yading@10: } yading@10: } yading@10: } yading@10: yading@10: /* NOTE: we do frequency inversion adter the MDCT by changing yading@10: the sign of the right window coefs */ yading@10: for (j = 0; j < 4; j++) { yading@10: for (i = 0; i < MDCT_BUF_SIZE; i += 2) { yading@10: RENAME(ff_mdct_win)[j + 4][i ] = RENAME(ff_mdct_win)[j][i ]; yading@10: RENAME(ff_mdct_win)[j + 4][i + 1] = -RENAME(ff_mdct_win)[j][i + 1]; yading@10: } yading@10: } yading@10: } yading@10: /* cos(pi*i/18) */ yading@10: #define C1 FIXHR(0.98480775301220805936/2) yading@10: #define C2 FIXHR(0.93969262078590838405/2) yading@10: #define C3 FIXHR(0.86602540378443864676/2) yading@10: #define C4 FIXHR(0.76604444311897803520/2) yading@10: #define C5 FIXHR(0.64278760968653932632/2) yading@10: #define C6 FIXHR(0.5/2) yading@10: #define C7 FIXHR(0.34202014332566873304/2) yading@10: #define C8 FIXHR(0.17364817766693034885/2) yading@10: yading@10: /* 0.5 / cos(pi*(2*i+1)/36) */ yading@10: static const INTFLOAT icos36[9] = { yading@10: FIXR(0.50190991877167369479), yading@10: FIXR(0.51763809020504152469), //0 yading@10: FIXR(0.55168895948124587824), yading@10: FIXR(0.61038729438072803416), yading@10: FIXR(0.70710678118654752439), //1 yading@10: FIXR(0.87172339781054900991), yading@10: FIXR(1.18310079157624925896), yading@10: FIXR(1.93185165257813657349), //2 yading@10: FIXR(5.73685662283492756461), yading@10: }; yading@10: yading@10: /* 0.5 / cos(pi*(2*i+1)/36) */ yading@10: static const INTFLOAT icos36h[9] = { yading@10: FIXHR(0.50190991877167369479/2), yading@10: FIXHR(0.51763809020504152469/2), //0 yading@10: FIXHR(0.55168895948124587824/2), yading@10: FIXHR(0.61038729438072803416/2), yading@10: FIXHR(0.70710678118654752439/2), //1 yading@10: FIXHR(0.87172339781054900991/2), yading@10: FIXHR(1.18310079157624925896/4), yading@10: FIXHR(1.93185165257813657349/4), //2 yading@10: // FIXHR(5.73685662283492756461), yading@10: }; yading@10: yading@10: /* using Lee like decomposition followed by hand coded 9 points DCT */ yading@10: static void imdct36(INTFLOAT *out, INTFLOAT *buf, INTFLOAT *in, INTFLOAT *win) yading@10: { yading@10: int i, j; yading@10: INTFLOAT t0, t1, t2, t3, s0, s1, s2, s3; yading@10: INTFLOAT tmp[18], *tmp1, *in1; yading@10: yading@10: for (i = 17; i >= 1; i--) yading@10: in[i] += in[i-1]; yading@10: for (i = 17; i >= 3; i -= 2) yading@10: in[i] += in[i-2]; yading@10: yading@10: for (j = 0; j < 2; j++) { yading@10: tmp1 = tmp + j; yading@10: in1 = in + j; yading@10: yading@10: t2 = in1[2*4] + in1[2*8] - in1[2*2]; yading@10: yading@10: t3 = in1[2*0] + SHR(in1[2*6],1); yading@10: t1 = in1[2*0] - in1[2*6]; yading@10: tmp1[ 6] = t1 - SHR(t2,1); yading@10: tmp1[16] = t1 + t2; yading@10: yading@10: t0 = MULH3(in1[2*2] + in1[2*4] , C2, 2); yading@10: t1 = MULH3(in1[2*4] - in1[2*8] , -2*C8, 1); yading@10: t2 = MULH3(in1[2*2] + in1[2*8] , -C4, 2); yading@10: yading@10: tmp1[10] = t3 - t0 - t2; yading@10: tmp1[ 2] = t3 + t0 + t1; yading@10: tmp1[14] = t3 + t2 - t1; yading@10: yading@10: tmp1[ 4] = MULH3(in1[2*5] + in1[2*7] - in1[2*1], -C3, 2); yading@10: t2 = MULH3(in1[2*1] + in1[2*5], C1, 2); yading@10: t3 = MULH3(in1[2*5] - in1[2*7], -2*C7, 1); yading@10: t0 = MULH3(in1[2*3], C3, 2); yading@10: yading@10: t1 = MULH3(in1[2*1] + in1[2*7], -C5, 2); yading@10: yading@10: tmp1[ 0] = t2 + t3 + t0; yading@10: tmp1[12] = t2 + t1 - t0; yading@10: tmp1[ 8] = t3 - t1 - t0; yading@10: } yading@10: yading@10: i = 0; yading@10: for (j = 0; j < 4; j++) { yading@10: t0 = tmp[i]; yading@10: t1 = tmp[i + 2]; yading@10: s0 = t1 + t0; yading@10: s2 = t1 - t0; yading@10: yading@10: t2 = tmp[i + 1]; yading@10: t3 = tmp[i + 3]; yading@10: s1 = MULH3(t3 + t2, icos36h[ j], 2); yading@10: s3 = MULLx(t3 - t2, icos36 [8 - j], FRAC_BITS); yading@10: yading@10: t0 = s0 + s1; yading@10: t1 = s0 - s1; yading@10: out[(9 + j) * SBLIMIT] = MULH3(t1, win[ 9 + j], 1) + buf[4*(9 + j)]; yading@10: out[(8 - j) * SBLIMIT] = MULH3(t1, win[ 8 - j], 1) + buf[4*(8 - j)]; yading@10: buf[4 * ( 9 + j )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + 9 + j], 1); yading@10: buf[4 * ( 8 - j )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + 8 - j], 1); yading@10: yading@10: t0 = s2 + s3; yading@10: t1 = s2 - s3; yading@10: out[(9 + 8 - j) * SBLIMIT] = MULH3(t1, win[ 9 + 8 - j], 1) + buf[4*(9 + 8 - j)]; yading@10: out[ j * SBLIMIT] = MULH3(t1, win[ j], 1) + buf[4*( j)]; yading@10: buf[4 * ( 9 + 8 - j )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + 9 + 8 - j], 1); yading@10: buf[4 * ( j )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + j], 1); yading@10: i += 4; yading@10: } yading@10: yading@10: s0 = tmp[16]; yading@10: s1 = MULH3(tmp[17], icos36h[4], 2); yading@10: t0 = s0 + s1; yading@10: t1 = s0 - s1; yading@10: out[(9 + 4) * SBLIMIT] = MULH3(t1, win[ 9 + 4], 1) + buf[4*(9 + 4)]; yading@10: out[(8 - 4) * SBLIMIT] = MULH3(t1, win[ 8 - 4], 1) + buf[4*(8 - 4)]; yading@10: buf[4 * ( 9 + 4 )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + 9 + 4], 1); yading@10: buf[4 * ( 8 - 4 )] = MULH3(t0, win[MDCT_BUF_SIZE/2 + 8 - 4], 1); yading@10: } yading@10: yading@10: void RENAME(ff_imdct36_blocks)(INTFLOAT *out, INTFLOAT *buf, INTFLOAT *in, yading@10: int count, int switch_point, int block_type) yading@10: { yading@10: int j; yading@10: for (j=0 ; j < count; j++) { yading@10: /* apply window & overlap with previous buffer */ yading@10: yading@10: /* select window */ yading@10: int win_idx = (switch_point && j < 2) ? 0 : block_type; yading@10: INTFLOAT *win = RENAME(ff_mdct_win)[win_idx + (4 & -(j & 1))]; yading@10: yading@10: imdct36(out, buf, in, win); yading@10: yading@10: in += 18; yading@10: buf += ((j&3) != 3 ? 1 : (72-3)); yading@10: out++; yading@10: } yading@10: } yading@10: