1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libopus/celt/mdct.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,311 @@
1.4 +/* Copyright (c) 2007-2008 CSIRO
1.5 + Copyright (c) 2007-2008 Xiph.Org Foundation
1.6 + Written by Jean-Marc Valin */
1.7 +/*
1.8 + Redistribution and use in source and binary forms, with or without
1.9 + modification, are permitted provided that the following conditions
1.10 + are met:
1.11 +
1.12 + - Redistributions of source code must retain the above copyright
1.13 + notice, this list of conditions and the following disclaimer.
1.14 +
1.15 + - Redistributions in binary form must reproduce the above copyright
1.16 + notice, this list of conditions and the following disclaimer in the
1.17 + documentation and/or other materials provided with the distribution.
1.18 +
1.19 + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
1.20 + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
1.21 + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
1.22 + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
1.23 + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
1.24 + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
1.25 + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
1.26 + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
1.27 + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
1.28 + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1.29 + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1.30 +*/
1.31 +
1.32 +/* This is a simple MDCT implementation that uses a N/4 complex FFT
1.33 + to do most of the work. It should be relatively straightforward to
1.34 + plug in pretty much and FFT here.
1.35 +
1.36 + This replaces the Vorbis FFT (and uses the exact same API), which
1.37 + was a bit too messy and that was ending up duplicating code
1.38 + (might as well use the same FFT everywhere).
1.39 +
1.40 + The algorithm is similar to (and inspired from) Fabrice Bellard's
1.41 + MDCT implementation in FFMPEG, but has differences in signs, ordering
1.42 + and scaling in many places.
1.43 +*/
1.44 +
1.45 +#ifndef SKIP_CONFIG_H
1.46 +#ifdef HAVE_CONFIG_H
1.47 +#include "config.h"
1.48 +#endif
1.49 +#endif
1.50 +
1.51 +#include "mdct.h"
1.52 +#include "kiss_fft.h"
1.53 +#include "_kiss_fft_guts.h"
1.54 +#include <math.h>
1.55 +#include "os_support.h"
1.56 +#include "mathops.h"
1.57 +#include "stack_alloc.h"
1.58 +
1.59 +#ifdef CUSTOM_MODES
1.60 +
1.61 +int clt_mdct_init(mdct_lookup *l,int N, int maxshift)
1.62 +{
1.63 + int i;
1.64 + int N4;
1.65 + kiss_twiddle_scalar *trig;
1.66 +#if defined(FIXED_POINT)
1.67 + int N2=N>>1;
1.68 +#endif
1.69 + l->n = N;
1.70 + N4 = N>>2;
1.71 + l->maxshift = maxshift;
1.72 + for (i=0;i<=maxshift;i++)
1.73 + {
1.74 + if (i==0)
1.75 + l->kfft[i] = opus_fft_alloc(N>>2>>i, 0, 0);
1.76 + else
1.77 + l->kfft[i] = opus_fft_alloc_twiddles(N>>2>>i, 0, 0, l->kfft[0]);
1.78 +#ifndef ENABLE_TI_DSPLIB55
1.79 + if (l->kfft[i]==NULL)
1.80 + return 0;
1.81 +#endif
1.82 + }
1.83 + l->trig = trig = (kiss_twiddle_scalar*)opus_alloc((N4+1)*sizeof(kiss_twiddle_scalar));
1.84 + if (l->trig==NULL)
1.85 + return 0;
1.86 + /* We have enough points that sine isn't necessary */
1.87 +#if defined(FIXED_POINT)
1.88 + for (i=0;i<=N4;i++)
1.89 + trig[i] = TRIG_UPSCALE*celt_cos_norm(DIV32(ADD32(SHL32(EXTEND32(i),17),N2),N));
1.90 +#else
1.91 + for (i=0;i<=N4;i++)
1.92 + trig[i] = (kiss_twiddle_scalar)cos(2*PI*i/N);
1.93 +#endif
1.94 + return 1;
1.95 +}
1.96 +
1.97 +void clt_mdct_clear(mdct_lookup *l)
1.98 +{
1.99 + int i;
1.100 + for (i=0;i<=l->maxshift;i++)
1.101 + opus_fft_free(l->kfft[i]);
1.102 + opus_free((kiss_twiddle_scalar*)l->trig);
1.103 +}
1.104 +
1.105 +#endif /* CUSTOM_MODES */
1.106 +
1.107 +/* Forward MDCT trashes the input array */
1.108 +void clt_mdct_forward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
1.109 + const opus_val16 *window, int overlap, int shift, int stride)
1.110 +{
1.111 + int i;
1.112 + int N, N2, N4;
1.113 + kiss_twiddle_scalar sine;
1.114 + VARDECL(kiss_fft_scalar, f);
1.115 + VARDECL(kiss_fft_scalar, f2);
1.116 + SAVE_STACK;
1.117 + N = l->n;
1.118 + N >>= shift;
1.119 + N2 = N>>1;
1.120 + N4 = N>>2;
1.121 + ALLOC(f, N2, kiss_fft_scalar);
1.122 + ALLOC(f2, N2, kiss_fft_scalar);
1.123 + /* sin(x) ~= x here */
1.124 +#ifdef FIXED_POINT
1.125 + sine = TRIG_UPSCALE*(QCONST16(0.7853981f, 15)+N2)/N;
1.126 +#else
1.127 + sine = (kiss_twiddle_scalar)2*PI*(.125f)/N;
1.128 +#endif
1.129 +
1.130 + /* Consider the input to be composed of four blocks: [a, b, c, d] */
1.131 + /* Window, shuffle, fold */
1.132 + {
1.133 + /* Temp pointers to make it really clear to the compiler what we're doing */
1.134 + const kiss_fft_scalar * OPUS_RESTRICT xp1 = in+(overlap>>1);
1.135 + const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+N2-1+(overlap>>1);
1.136 + kiss_fft_scalar * OPUS_RESTRICT yp = f;
1.137 + const opus_val16 * OPUS_RESTRICT wp1 = window+(overlap>>1);
1.138 + const opus_val16 * OPUS_RESTRICT wp2 = window+(overlap>>1)-1;
1.139 + for(i=0;i<((overlap+3)>>2);i++)
1.140 + {
1.141 + /* Real part arranged as -d-cR, Imag part arranged as -b+aR*/
1.142 + *yp++ = MULT16_32_Q15(*wp2, xp1[N2]) + MULT16_32_Q15(*wp1,*xp2);
1.143 + *yp++ = MULT16_32_Q15(*wp1, *xp1) - MULT16_32_Q15(*wp2, xp2[-N2]);
1.144 + xp1+=2;
1.145 + xp2-=2;
1.146 + wp1+=2;
1.147 + wp2-=2;
1.148 + }
1.149 + wp1 = window;
1.150 + wp2 = window+overlap-1;
1.151 + for(;i<N4-((overlap+3)>>2);i++)
1.152 + {
1.153 + /* Real part arranged as a-bR, Imag part arranged as -c-dR */
1.154 + *yp++ = *xp2;
1.155 + *yp++ = *xp1;
1.156 + xp1+=2;
1.157 + xp2-=2;
1.158 + }
1.159 + for(;i<N4;i++)
1.160 + {
1.161 + /* Real part arranged as a-bR, Imag part arranged as -c-dR */
1.162 + *yp++ = -MULT16_32_Q15(*wp1, xp1[-N2]) + MULT16_32_Q15(*wp2, *xp2);
1.163 + *yp++ = MULT16_32_Q15(*wp2, *xp1) + MULT16_32_Q15(*wp1, xp2[N2]);
1.164 + xp1+=2;
1.165 + xp2-=2;
1.166 + wp1+=2;
1.167 + wp2-=2;
1.168 + }
1.169 + }
1.170 + /* Pre-rotation */
1.171 + {
1.172 + kiss_fft_scalar * OPUS_RESTRICT yp = f;
1.173 + const kiss_twiddle_scalar *t = &l->trig[0];
1.174 + for(i=0;i<N4;i++)
1.175 + {
1.176 + kiss_fft_scalar re, im, yr, yi;
1.177 + re = yp[0];
1.178 + im = yp[1];
1.179 + yr = -S_MUL(re,t[i<<shift]) - S_MUL(im,t[(N4-i)<<shift]);
1.180 + yi = -S_MUL(im,t[i<<shift]) + S_MUL(re,t[(N4-i)<<shift]);
1.181 + /* works because the cos is nearly one */
1.182 + *yp++ = yr + S_MUL(yi,sine);
1.183 + *yp++ = yi - S_MUL(yr,sine);
1.184 + }
1.185 + }
1.186 +
1.187 + /* N/4 complex FFT, down-scales by 4/N */
1.188 + opus_fft(l->kfft[shift], (kiss_fft_cpx *)f, (kiss_fft_cpx *)f2);
1.189 +
1.190 + /* Post-rotate */
1.191 + {
1.192 + /* Temp pointers to make it really clear to the compiler what we're doing */
1.193 + const kiss_fft_scalar * OPUS_RESTRICT fp = f2;
1.194 + kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
1.195 + kiss_fft_scalar * OPUS_RESTRICT yp2 = out+stride*(N2-1);
1.196 + const kiss_twiddle_scalar *t = &l->trig[0];
1.197 + /* Temp pointers to make it really clear to the compiler what we're doing */
1.198 + for(i=0;i<N4;i++)
1.199 + {
1.200 + kiss_fft_scalar yr, yi;
1.201 + yr = S_MUL(fp[1],t[(N4-i)<<shift]) + S_MUL(fp[0],t[i<<shift]);
1.202 + yi = S_MUL(fp[0],t[(N4-i)<<shift]) - S_MUL(fp[1],t[i<<shift]);
1.203 + /* works because the cos is nearly one */
1.204 + *yp1 = yr - S_MUL(yi,sine);
1.205 + *yp2 = yi + S_MUL(yr,sine);;
1.206 + fp += 2;
1.207 + yp1 += 2*stride;
1.208 + yp2 -= 2*stride;
1.209 + }
1.210 + }
1.211 + RESTORE_STACK;
1.212 +}
1.213 +
1.214 +void clt_mdct_backward(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
1.215 + const opus_val16 * OPUS_RESTRICT window, int overlap, int shift, int stride)
1.216 +{
1.217 + int i;
1.218 + int N, N2, N4;
1.219 + kiss_twiddle_scalar sine;
1.220 + VARDECL(kiss_fft_scalar, f2);
1.221 + SAVE_STACK;
1.222 + N = l->n;
1.223 + N >>= shift;
1.224 + N2 = N>>1;
1.225 + N4 = N>>2;
1.226 + ALLOC(f2, N2, kiss_fft_scalar);
1.227 + /* sin(x) ~= x here */
1.228 +#ifdef FIXED_POINT
1.229 + sine = TRIG_UPSCALE*(QCONST16(0.7853981f, 15)+N2)/N;
1.230 +#else
1.231 + sine = (kiss_twiddle_scalar)2*PI*(.125f)/N;
1.232 +#endif
1.233 +
1.234 + /* Pre-rotate */
1.235 + {
1.236 + /* Temp pointers to make it really clear to the compiler what we're doing */
1.237 + const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
1.238 + const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
1.239 + kiss_fft_scalar * OPUS_RESTRICT yp = f2;
1.240 + const kiss_twiddle_scalar *t = &l->trig[0];
1.241 + for(i=0;i<N4;i++)
1.242 + {
1.243 + kiss_fft_scalar yr, yi;
1.244 + yr = -S_MUL(*xp2, t[i<<shift]) + S_MUL(*xp1,t[(N4-i)<<shift]);
1.245 + yi = -S_MUL(*xp2, t[(N4-i)<<shift]) - S_MUL(*xp1,t[i<<shift]);
1.246 + /* works because the cos is nearly one */
1.247 + *yp++ = yr - S_MUL(yi,sine);
1.248 + *yp++ = yi + S_MUL(yr,sine);
1.249 + xp1+=2*stride;
1.250 + xp2-=2*stride;
1.251 + }
1.252 + }
1.253 +
1.254 + /* Inverse N/4 complex FFT. This one should *not* downscale even in fixed-point */
1.255 + opus_ifft(l->kfft[shift], (kiss_fft_cpx *)f2, (kiss_fft_cpx *)(out+(overlap>>1)));
1.256 +
1.257 + /* Post-rotate and de-shuffle from both ends of the buffer at once to make
1.258 + it in-place. */
1.259 + {
1.260 + kiss_fft_scalar * OPUS_RESTRICT yp0 = out+(overlap>>1);
1.261 + kiss_fft_scalar * OPUS_RESTRICT yp1 = out+(overlap>>1)+N2-2;
1.262 + const kiss_twiddle_scalar *t = &l->trig[0];
1.263 + /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
1.264 + middle pair will be computed twice. */
1.265 + for(i=0;i<(N4+1)>>1;i++)
1.266 + {
1.267 + kiss_fft_scalar re, im, yr, yi;
1.268 + kiss_twiddle_scalar t0, t1;
1.269 + re = yp0[0];
1.270 + im = yp0[1];
1.271 + t0 = t[i<<shift];
1.272 + t1 = t[(N4-i)<<shift];
1.273 + /* We'd scale up by 2 here, but instead it's done when mixing the windows */
1.274 + yr = S_MUL(re,t0) - S_MUL(im,t1);
1.275 + yi = S_MUL(im,t0) + S_MUL(re,t1);
1.276 + re = yp1[0];
1.277 + im = yp1[1];
1.278 + /* works because the cos is nearly one */
1.279 + yp0[0] = -(yr - S_MUL(yi,sine));
1.280 + yp1[1] = yi + S_MUL(yr,sine);
1.281 +
1.282 + t0 = t[(N4-i-1)<<shift];
1.283 + t1 = t[(i+1)<<shift];
1.284 + /* We'd scale up by 2 here, but instead it's done when mixing the windows */
1.285 + yr = S_MUL(re,t0) - S_MUL(im,t1);
1.286 + yi = S_MUL(im,t0) + S_MUL(re,t1);
1.287 + /* works because the cos is nearly one */
1.288 + yp1[0] = -(yr - S_MUL(yi,sine));
1.289 + yp0[1] = yi + S_MUL(yr,sine);
1.290 + yp0 += 2;
1.291 + yp1 -= 2;
1.292 + }
1.293 + }
1.294 +
1.295 + /* Mirror on both sides for TDAC */
1.296 + {
1.297 + kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1;
1.298 + kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
1.299 + const opus_val16 * OPUS_RESTRICT wp1 = window;
1.300 + const opus_val16 * OPUS_RESTRICT wp2 = window+overlap-1;
1.301 +
1.302 + for(i = 0; i < overlap/2; i++)
1.303 + {
1.304 + kiss_fft_scalar x1, x2;
1.305 + x1 = *xp1;
1.306 + x2 = *yp1;
1.307 + *yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1);
1.308 + *xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1);
1.309 + wp1++;
1.310 + wp2--;
1.311 + }
1.312 + }
1.313 + RESTORE_STACK;
1.314 +}
