]> code.delx.au - pulseaudio/blob - src/modules/echo-cancel/adrian-aec.c
modules: Micro-optimisation for rewind_requested paths
[pulseaudio] / src / modules / echo-cancel / adrian-aec.c
1 /* aec.cpp
2 *
3 * Copyright (C) DFS Deutsche Flugsicherung (2004, 2005).
4 * All Rights Reserved.
5 *
6 * Acoustic Echo Cancellation NLMS-pw algorithm
7 *
8 * Version 0.3 filter created with www.dsptutor.freeuk.com
9 * Version 0.3.1 Allow change of stability parameter delta
10 * Version 0.4 Leaky Normalized LMS - pre whitening algorithm
11 */
12
13 #ifndef _GNU_SOURCE
14 #define _GNU_SOURCE
15 #endif
16
17 #include <math.h>
18 #include <string.h>
19 #include <stdint.h>
20
21 #include <pulse/xmalloc.h>
22
23 #include "adrian-aec.h"
24
25 #ifndef DISABLE_ORC
26 #include "adrian-aec-orc-gen.h"
27 #endif
28
29 #ifdef __SSE__
30 #include <xmmintrin.h>
31 #endif
32
33 /* Vector Dot Product */
34 static REAL dotp(REAL a[], REAL b[])
35 {
36 REAL sum0 = 0.0, sum1 = 0.0;
37 int j;
38
39 for (j = 0; j < NLMS_LEN; j += 2) {
40 // optimize: partial loop unrolling
41 sum0 += a[j] * b[j];
42 sum1 += a[j + 1] * b[j + 1];
43 }
44 return sum0 + sum1;
45 }
46
47 static REAL dotp_sse(REAL a[], REAL b[])
48 {
49 #ifdef __SSE__
50 /* This is taken from speex's inner product implementation */
51 int j;
52 REAL sum;
53 __m128 acc = _mm_setzero_ps();
54
55 for (j=0;j<NLMS_LEN;j+=8)
56 {
57 acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j), _mm_loadu_ps(b+j)));
58 acc = _mm_add_ps(acc, _mm_mul_ps(_mm_load_ps(a+j+4), _mm_loadu_ps(b+j+4)));
59 }
60 acc = _mm_add_ps(acc, _mm_movehl_ps(acc, acc));
61 acc = _mm_add_ss(acc, _mm_shuffle_ps(acc, acc, 0x55));
62 _mm_store_ss(&sum, acc);
63
64 return sum;
65 #else
66 return dotp(a, b);
67 #endif
68 }
69
70
71 AEC* AEC_init(int RATE, int have_vector)
72 {
73 AEC *a = pa_xnew(AEC, 1);
74 a->hangover = 0;
75 memset(a->x, 0, sizeof(a->x));
76 memset(a->xf, 0, sizeof(a->xf));
77 memset(a->w_arr, 0, sizeof(a->w_arr));
78 a->j = NLMS_EXT;
79 a->delta = 0.0f;
80 AEC_setambient(a, NoiseFloor);
81 a->dfast = a->dslow = M75dB_PCM;
82 a->xfast = a->xslow = M80dB_PCM;
83 a->gain = 1.0f;
84 a->Fx = IIR1_init(2000.0f/RATE);
85 a->Fe = IIR1_init(2000.0f/RATE);
86 a->cutoff = FIR_HP_300Hz_init();
87 a->acMic = IIR_HP_init();
88 a->acSpk = IIR_HP_init();
89
90 a->aes_y2 = M0dB;
91
92 a->fdwdisplay = -1;
93 a->dumpcnt = 0;
94 memset(a->ws, 0, sizeof(a->ws));
95
96 if (have_vector) {
97 /* Get a 16-byte aligned location */
98 a->w = (REAL *) (((uintptr_t) a->w_arr) + (((uintptr_t) a->w_arr) % 16));
99 a->dotp = dotp_sse;
100 } else {
101 /* We don't care about alignment, just use the array as-is */
102 a->w = a->w_arr;
103 a->dotp = dotp;
104 }
105
106 return a;
107 }
108
109 // Adrian soft decision DTD
110 // (Dual Average Near-End to Far-End signal Ratio DTD)
111 // This algorithm uses exponential smoothing with differnt
112 // ageing parameters to get fast and slow near-end and far-end
113 // signal averages. The ratio of NFRs term
114 // (dfast / xfast) / (dslow / xslow) is used to compute the stepsize
115 // A ratio value of 2.5 is mapped to stepsize 0, a ratio of 0 is
116 // mapped to 1.0 with a limited linear function.
117 static float AEC_dtd(AEC *a, REAL d, REAL x)
118 {
119 float ratio, stepsize;
120
121 // fast near-end and far-end average
122 a->dfast += ALPHAFAST * (fabsf(d) - a->dfast);
123 a->xfast += ALPHAFAST * (fabsf(x) - a->xfast);
124
125 // slow near-end and far-end average
126 a->dslow += ALPHASLOW * (fabsf(d) - a->dslow);
127 a->xslow += ALPHASLOW * (fabsf(x) - a->xslow);
128
129 if (a->xfast < M70dB_PCM) {
130 return 0.0; // no Spk signal
131 }
132
133 if (a->dfast < M70dB_PCM) {
134 return 0.0; // no Mic signal
135 }
136
137 // ratio of NFRs
138 ratio = (a->dfast * a->xslow) / (a->dslow * a->xfast);
139
140 // Linear interpolation with clamping at the limits
141 if (ratio < STEPX1)
142 stepsize = STEPY1;
143 else if (ratio > STEPX2)
144 stepsize = STEPY2;
145 else
146 stepsize = STEPY1 + (STEPY2 - STEPY1) * (ratio - STEPX1) / (STEPX2 - STEPX1);
147
148 return stepsize;
149 }
150
151
152 static void AEC_leaky(AEC *a)
153 // The xfast signal is used to charge the hangover timer to Thold.
154 // When hangover expires (no Spk signal for some time) the vector w
155 // is erased. This is my implementation of Leaky NLMS.
156 {
157 if (a->xfast >= M70dB_PCM) {
158 // vector w is valid for hangover Thold time
159 a->hangover = Thold;
160 } else {
161 if (a->hangover > 1) {
162 --(a->hangover);
163 } else if (1 == a->hangover) {
164 --(a->hangover);
165 // My Leaky NLMS is to erase vector w when hangover expires
166 memset(a->w, 0, sizeof(a->w));
167 }
168 }
169 }
170
171
172 #if 0
173 void AEC::openwdisplay() {
174 // open TCP connection to program wdisplay.tcl
175 fdwdisplay = socket_async("127.0.0.1", 50999);
176 };
177 #endif
178
179
180 static REAL AEC_nlms_pw(AEC *a, REAL d, REAL x_, float stepsize)
181 {
182 REAL e;
183 REAL ef;
184 a->x[a->j] = x_;
185 a->xf[a->j] = IIR1_highpass(a->Fx, x_); // pre-whitening of x
186
187 // calculate error value
188 // (mic signal - estimated mic signal from spk signal)
189 e = d;
190 if (a->hangover > 0) {
191 e -= a->dotp(a->w, a->x + a->j);
192 }
193 ef = IIR1_highpass(a->Fe, e); // pre-whitening of e
194
195 // optimize: iterative dotp(xf, xf)
196 a->dotp_xf_xf += (a->xf[a->j] * a->xf[a->j] - a->xf[a->j + NLMS_LEN - 1] * a->xf[a->j + NLMS_LEN - 1]);
197
198 if (stepsize > 0.0) {
199 // calculate variable step size
200 REAL mikro_ef = stepsize * ef / a->dotp_xf_xf;
201
202 #ifdef DISABLE_ORC
203 // update tap weights (filter learning)
204 int i;
205 for (i = 0; i < NLMS_LEN; i += 2) {
206 // optimize: partial loop unrolling
207 a->w[i] += mikro_ef * a->xf[i + a->j];
208 a->w[i + 1] += mikro_ef * a->xf[i + a->j + 1];
209 }
210 #else
211 update_tap_weights(a->w, &a->xf[a->j], mikro_ef, NLMS_LEN);
212 #endif
213 }
214
215 if (--(a->j) < 0) {
216 // optimize: decrease number of memory copies
217 a->j = NLMS_EXT;
218 memmove(a->x + a->j + 1, a->x, (NLMS_LEN - 1) * sizeof(REAL));
219 memmove(a->xf + a->j + 1, a->xf, (NLMS_LEN - 1) * sizeof(REAL));
220 }
221
222 // Saturation
223 if (e > MAXPCM) {
224 return MAXPCM;
225 } else if (e < -MAXPCM) {
226 return -MAXPCM;
227 } else {
228 return e;
229 }
230 }
231
232
233 int AEC_doAEC(AEC *a, int d_, int x_)
234 {
235 REAL d = (REAL) d_;
236 REAL x = (REAL) x_;
237
238 // Mic Highpass Filter - to remove DC
239 d = IIR_HP_highpass(a->acMic, d);
240
241 // Mic Highpass Filter - cut-off below 300Hz
242 d = FIR_HP_300Hz_highpass(a->cutoff, d);
243
244 // Amplify, for e.g. Soundcards with -6dB max. volume
245 d *= a->gain;
246
247 // Spk Highpass Filter - to remove DC
248 x = IIR_HP_highpass(a->acSpk, x);
249
250 // Double Talk Detector
251 a->stepsize = AEC_dtd(a, d, x);
252
253 // Leaky (ageing of vector w)
254 AEC_leaky(a);
255
256 // Acoustic Echo Cancellation
257 d = AEC_nlms_pw(a, d, x, a->stepsize);
258
259 #if 0
260 if (fdwdisplay >= 0) {
261 if (++dumpcnt >= (WIDEB*RATE/10)) {
262 // wdisplay creates 10 dumps per seconds = large CPU load!
263 dumpcnt = 0;
264 write(fdwdisplay, ws, DUMP_LEN*sizeof(float));
265 // we don't check return value. This is not production quality!!!
266 memset(ws, 0, sizeof(ws));
267 } else {
268 int i;
269 for (i = 0; i < DUMP_LEN; i += 2) {
270 // optimize: partial loop unrolling
271 ws[i] += w[i];
272 ws[i + 1] += w[i + 1];
273 }
274 }
275 }
276 #endif
277
278 return (int) d;
279 }