2 * Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
12 * The core AEC algorithm, SSE2 version of speed-critical functions.
15 #include "webrtc/modules/audio_processing/aec/aec_core.h"
17 #include <emmintrin.h>
19 #include <string.h> // memset
21 #include "webrtc/modules/audio_processing/aec/aec_core_internal.h"
22 #include "webrtc/modules/audio_processing/aec/aec_rdft.h"
24 __inline static float MulRe(float aRe, float aIm, float bRe, float bIm) {
25 return aRe * bRe - aIm * bIm;
28 __inline static float MulIm(float aRe, float aIm, float bRe, float bIm) {
29 return aRe * bIm + aIm * bRe;
32 static void FilterFarSSE2(AecCore* aec, float yf[2][PART_LEN1]) {
34 const int num_partitions = aec->num_partitions;
35 for (i = 0; i < num_partitions; i++) {
37 int xPos = (i + aec->xfBufBlockPos) * PART_LEN1;
38 int pos = i * PART_LEN1;
40 if (i + aec->xfBufBlockPos >= num_partitions) {
41 xPos -= num_partitions * (PART_LEN1);
44 // vectorized code (four at once)
45 for (j = 0; j + 3 < PART_LEN1; j += 4) {
46 const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
47 const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
48 const __m128 wfBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
49 const __m128 wfBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
50 const __m128 yf_re = _mm_loadu_ps(&yf[0][j]);
51 const __m128 yf_im = _mm_loadu_ps(&yf[1][j]);
52 const __m128 a = _mm_mul_ps(xfBuf_re, wfBuf_re);
53 const __m128 b = _mm_mul_ps(xfBuf_im, wfBuf_im);
54 const __m128 c = _mm_mul_ps(xfBuf_re, wfBuf_im);
55 const __m128 d = _mm_mul_ps(xfBuf_im, wfBuf_re);
56 const __m128 e = _mm_sub_ps(a, b);
57 const __m128 f = _mm_add_ps(c, d);
58 const __m128 g = _mm_add_ps(yf_re, e);
59 const __m128 h = _mm_add_ps(yf_im, f);
60 _mm_storeu_ps(&yf[0][j], g);
61 _mm_storeu_ps(&yf[1][j], h);
63 // scalar code for the remaining items.
64 for (; j < PART_LEN1; j++) {
65 yf[0][j] += MulRe(aec->xfBuf[0][xPos + j],
66 aec->xfBuf[1][xPos + j],
67 aec->wfBuf[0][pos + j],
68 aec->wfBuf[1][pos + j]);
69 yf[1][j] += MulIm(aec->xfBuf[0][xPos + j],
70 aec->xfBuf[1][xPos + j],
71 aec->wfBuf[0][pos + j],
72 aec->wfBuf[1][pos + j]);
77 static void ScaleErrorSignalSSE2(AecCore* aec, float ef[2][PART_LEN1]) {
78 const __m128 k1e_10f = _mm_set1_ps(1e-10f);
79 const __m128 kMu = aec->extended_filter_enabled ? _mm_set1_ps(kExtendedMu)
80 : _mm_set1_ps(aec->normal_mu);
81 const __m128 kThresh = aec->extended_filter_enabled
82 ? _mm_set1_ps(kExtendedErrorThreshold)
83 : _mm_set1_ps(aec->normal_error_threshold);
86 // vectorized code (four at once)
87 for (i = 0; i + 3 < PART_LEN1; i += 4) {
88 const __m128 xPow = _mm_loadu_ps(&aec->xPow[i]);
89 const __m128 ef_re_base = _mm_loadu_ps(&ef[0][i]);
90 const __m128 ef_im_base = _mm_loadu_ps(&ef[1][i]);
92 const __m128 xPowPlus = _mm_add_ps(xPow, k1e_10f);
93 __m128 ef_re = _mm_div_ps(ef_re_base, xPowPlus);
94 __m128 ef_im = _mm_div_ps(ef_im_base, xPowPlus);
95 const __m128 ef_re2 = _mm_mul_ps(ef_re, ef_re);
96 const __m128 ef_im2 = _mm_mul_ps(ef_im, ef_im);
97 const __m128 ef_sum2 = _mm_add_ps(ef_re2, ef_im2);
98 const __m128 absEf = _mm_sqrt_ps(ef_sum2);
99 const __m128 bigger = _mm_cmpgt_ps(absEf, kThresh);
100 __m128 absEfPlus = _mm_add_ps(absEf, k1e_10f);
101 const __m128 absEfInv = _mm_div_ps(kThresh, absEfPlus);
102 __m128 ef_re_if = _mm_mul_ps(ef_re, absEfInv);
103 __m128 ef_im_if = _mm_mul_ps(ef_im, absEfInv);
104 ef_re_if = _mm_and_ps(bigger, ef_re_if);
105 ef_im_if = _mm_and_ps(bigger, ef_im_if);
106 ef_re = _mm_andnot_ps(bigger, ef_re);
107 ef_im = _mm_andnot_ps(bigger, ef_im);
108 ef_re = _mm_or_ps(ef_re, ef_re_if);
109 ef_im = _mm_or_ps(ef_im, ef_im_if);
110 ef_re = _mm_mul_ps(ef_re, kMu);
111 ef_im = _mm_mul_ps(ef_im, kMu);
113 _mm_storeu_ps(&ef[0][i], ef_re);
114 _mm_storeu_ps(&ef[1][i], ef_im);
116 // scalar code for the remaining items.
119 aec->extended_filter_enabled ? kExtendedMu : aec->normal_mu;
120 const float error_threshold = aec->extended_filter_enabled
121 ? kExtendedErrorThreshold
122 : aec->normal_error_threshold;
123 for (; i < (PART_LEN1); i++) {
125 ef[0][i] /= (aec->xPow[i] + 1e-10f);
126 ef[1][i] /= (aec->xPow[i] + 1e-10f);
127 abs_ef = sqrtf(ef[0][i] * ef[0][i] + ef[1][i] * ef[1][i]);
129 if (abs_ef > error_threshold) {
130 abs_ef = error_threshold / (abs_ef + 1e-10f);
142 static void FilterAdaptationSSE2(AecCore* aec,
144 float ef[2][PART_LEN1]) {
146 const int num_partitions = aec->num_partitions;
147 for (i = 0; i < num_partitions; i++) {
148 int xPos = (i + aec->xfBufBlockPos) * (PART_LEN1);
149 int pos = i * PART_LEN1;
151 if (i + aec->xfBufBlockPos >= num_partitions) {
152 xPos -= num_partitions * PART_LEN1;
155 // Process the whole array...
156 for (j = 0; j < PART_LEN; j += 4) {
157 // Load xfBuf and ef.
158 const __m128 xfBuf_re = _mm_loadu_ps(&aec->xfBuf[0][xPos + j]);
159 const __m128 xfBuf_im = _mm_loadu_ps(&aec->xfBuf[1][xPos + j]);
160 const __m128 ef_re = _mm_loadu_ps(&ef[0][j]);
161 const __m128 ef_im = _mm_loadu_ps(&ef[1][j]);
162 // Calculate the product of conjugate(xfBuf) by ef.
163 // re(conjugate(a) * b) = aRe * bRe + aIm * bIm
164 // im(conjugate(a) * b)= aRe * bIm - aIm * bRe
165 const __m128 a = _mm_mul_ps(xfBuf_re, ef_re);
166 const __m128 b = _mm_mul_ps(xfBuf_im, ef_im);
167 const __m128 c = _mm_mul_ps(xfBuf_re, ef_im);
168 const __m128 d = _mm_mul_ps(xfBuf_im, ef_re);
169 const __m128 e = _mm_add_ps(a, b);
170 const __m128 f = _mm_sub_ps(c, d);
171 // Interleave real and imaginary parts.
172 const __m128 g = _mm_unpacklo_ps(e, f);
173 const __m128 h = _mm_unpackhi_ps(e, f);
175 _mm_storeu_ps(&fft[2 * j + 0], g);
176 _mm_storeu_ps(&fft[2 * j + 4], h);
178 // ... and fixup the first imaginary entry.
179 fft[1] = MulRe(aec->xfBuf[0][xPos + PART_LEN],
180 -aec->xfBuf[1][xPos + PART_LEN],
184 aec_rdft_inverse_128(fft);
185 memset(fft + PART_LEN, 0, sizeof(float) * PART_LEN);
189 float scale = 2.0f / PART_LEN2;
190 const __m128 scale_ps = _mm_load_ps1(&scale);
191 for (j = 0; j < PART_LEN; j += 4) {
192 const __m128 fft_ps = _mm_loadu_ps(&fft[j]);
193 const __m128 fft_scale = _mm_mul_ps(fft_ps, scale_ps);
194 _mm_storeu_ps(&fft[j], fft_scale);
197 aec_rdft_forward_128(fft);
200 float wt1 = aec->wfBuf[1][pos];
201 aec->wfBuf[0][pos + PART_LEN] += fft[1];
202 for (j = 0; j < PART_LEN; j += 4) {
203 __m128 wtBuf_re = _mm_loadu_ps(&aec->wfBuf[0][pos + j]);
204 __m128 wtBuf_im = _mm_loadu_ps(&aec->wfBuf[1][pos + j]);
205 const __m128 fft0 = _mm_loadu_ps(&fft[2 * j + 0]);
206 const __m128 fft4 = _mm_loadu_ps(&fft[2 * j + 4]);
207 const __m128 fft_re =
208 _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(2, 0, 2, 0));
209 const __m128 fft_im =
210 _mm_shuffle_ps(fft0, fft4, _MM_SHUFFLE(3, 1, 3, 1));
211 wtBuf_re = _mm_add_ps(wtBuf_re, fft_re);
212 wtBuf_im = _mm_add_ps(wtBuf_im, fft_im);
213 _mm_storeu_ps(&aec->wfBuf[0][pos + j], wtBuf_re);
214 _mm_storeu_ps(&aec->wfBuf[1][pos + j], wtBuf_im);
216 aec->wfBuf[1][pos] = wt1;
221 static __m128 mm_pow_ps(__m128 a, __m128 b) {
222 // a^b = exp2(b * log2(a))
223 // exp2(x) and log2(x) are calculated using polynomial approximations.
224 __m128 log2_a, b_log2_a, a_exp_b;
226 // Calculate log2(x), x = a.
228 // To calculate log2(x), we decompose x like this:
231 // y is in the [1.0, 2.0) range
233 // log2(x) = log2(y) + n
234 // n can be evaluated by playing with float representation.
235 // log2(y) in a small range can be approximated, this code uses an order
236 // five polynomial approximation. The coefficients have been
237 // estimated with the Remez algorithm and the resulting
238 // polynomial has a maximum relative error of 0.00086%.
241 // This is done by masking the exponent, shifting it into the top bit of
242 // the mantissa, putting eight into the biased exponent (to shift/
243 // compensate the fact that the exponent has been shifted in the top/
244 // fractional part and finally getting rid of the implicit leading one
245 // from the mantissa by substracting it out.
246 static const ALIGN16_BEG int float_exponent_mask[4] ALIGN16_END = {
247 0x7F800000, 0x7F800000, 0x7F800000, 0x7F800000};
248 static const ALIGN16_BEG int eight_biased_exponent[4] ALIGN16_END = {
249 0x43800000, 0x43800000, 0x43800000, 0x43800000};
250 static const ALIGN16_BEG int implicit_leading_one[4] ALIGN16_END = {
251 0x43BF8000, 0x43BF8000, 0x43BF8000, 0x43BF8000};
252 static const int shift_exponent_into_top_mantissa = 8;
253 const __m128 two_n = _mm_and_ps(a, *((__m128*)float_exponent_mask));
254 const __m128 n_1 = _mm_castsi128_ps(_mm_srli_epi32(
255 _mm_castps_si128(two_n), shift_exponent_into_top_mantissa));
256 const __m128 n_0 = _mm_or_ps(n_1, *((__m128*)eight_biased_exponent));
257 const __m128 n = _mm_sub_ps(n_0, *((__m128*)implicit_leading_one));
260 static const ALIGN16_BEG int mantissa_mask[4] ALIGN16_END = {
261 0x007FFFFF, 0x007FFFFF, 0x007FFFFF, 0x007FFFFF};
262 static const ALIGN16_BEG int zero_biased_exponent_is_one[4] ALIGN16_END = {
263 0x3F800000, 0x3F800000, 0x3F800000, 0x3F800000};
264 const __m128 mantissa = _mm_and_ps(a, *((__m128*)mantissa_mask));
266 _mm_or_ps(mantissa, *((__m128*)zero_biased_exponent_is_one));
268 // Approximate log2(y) ~= (y - 1) * pol5(y).
269 // pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
270 static const ALIGN16_BEG float ALIGN16_END C5[4] = {
271 -3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f, -3.4436006e-2f};
272 static const ALIGN16_BEG float ALIGN16_END
273 C4[4] = {3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f, 3.1821337e-1f};
274 static const ALIGN16_BEG float ALIGN16_END
275 C3[4] = {-1.2315303f, -1.2315303f, -1.2315303f, -1.2315303f};
276 static const ALIGN16_BEG float ALIGN16_END
277 C2[4] = {2.5988452f, 2.5988452f, 2.5988452f, 2.5988452f};
278 static const ALIGN16_BEG float ALIGN16_END
279 C1[4] = {-3.3241990f, -3.3241990f, -3.3241990f, -3.3241990f};
280 static const ALIGN16_BEG float ALIGN16_END
281 C0[4] = {3.1157899f, 3.1157899f, 3.1157899f, 3.1157899f};
282 const __m128 pol5_y_0 = _mm_mul_ps(y, *((__m128*)C5));
283 const __m128 pol5_y_1 = _mm_add_ps(pol5_y_0, *((__m128*)C4));
284 const __m128 pol5_y_2 = _mm_mul_ps(pol5_y_1, y);
285 const __m128 pol5_y_3 = _mm_add_ps(pol5_y_2, *((__m128*)C3));
286 const __m128 pol5_y_4 = _mm_mul_ps(pol5_y_3, y);
287 const __m128 pol5_y_5 = _mm_add_ps(pol5_y_4, *((__m128*)C2));
288 const __m128 pol5_y_6 = _mm_mul_ps(pol5_y_5, y);
289 const __m128 pol5_y_7 = _mm_add_ps(pol5_y_6, *((__m128*)C1));
290 const __m128 pol5_y_8 = _mm_mul_ps(pol5_y_7, y);
291 const __m128 pol5_y = _mm_add_ps(pol5_y_8, *((__m128*)C0));
292 const __m128 y_minus_one =
293 _mm_sub_ps(y, *((__m128*)zero_biased_exponent_is_one));
294 const __m128 log2_y = _mm_mul_ps(y_minus_one, pol5_y);
297 log2_a = _mm_add_ps(n, log2_y);
301 b_log2_a = _mm_mul_ps(b, log2_a);
303 // Calculate exp2(x), x = b * log2(a).
305 // To calculate 2^x, we decompose x like this:
307 // n is an integer, the value of x - 0.5 rounded down, therefore
308 // y is in the [0.5, 1.5) range
311 // 2^n can be evaluated by playing with float representation.
312 // 2^y in a small range can be approximated, this code uses an order two
313 // polynomial approximation. The coefficients have been estimated
314 // with the Remez algorithm and the resulting polynomial has a
315 // maximum relative error of 0.17%.
317 // To avoid over/underflow, we reduce the range of input to ]-127, 129].
318 static const ALIGN16_BEG float max_input[4] ALIGN16_END = {129.f, 129.f,
320 static const ALIGN16_BEG float min_input[4] ALIGN16_END = {
321 -126.99999f, -126.99999f, -126.99999f, -126.99999f};
322 const __m128 x_min = _mm_min_ps(b_log2_a, *((__m128*)max_input));
323 const __m128 x_max = _mm_max_ps(x_min, *((__m128*)min_input));
325 static const ALIGN16_BEG float half[4] ALIGN16_END = {0.5f, 0.5f,
327 const __m128 x_minus_half = _mm_sub_ps(x_max, *((__m128*)half));
328 const __m128i x_minus_half_floor = _mm_cvtps_epi32(x_minus_half);
330 static const ALIGN16_BEG int float_exponent_bias[4] ALIGN16_END = {
332 static const int float_exponent_shift = 23;
333 const __m128i two_n_exponent =
334 _mm_add_epi32(x_minus_half_floor, *((__m128i*)float_exponent_bias));
336 _mm_castsi128_ps(_mm_slli_epi32(two_n_exponent, float_exponent_shift));
338 const __m128 y = _mm_sub_ps(x_max, _mm_cvtepi32_ps(x_minus_half_floor));
339 // Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
340 static const ALIGN16_BEG float C2[4] ALIGN16_END = {
341 3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f, 3.3718944e-1f};
342 static const ALIGN16_BEG float C1[4] ALIGN16_END = {
343 6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f, 6.5763628e-1f};
344 static const ALIGN16_BEG float C0[4] ALIGN16_END = {1.0017247f, 1.0017247f,
345 1.0017247f, 1.0017247f};
346 const __m128 exp2_y_0 = _mm_mul_ps(y, *((__m128*)C2));
347 const __m128 exp2_y_1 = _mm_add_ps(exp2_y_0, *((__m128*)C1));
348 const __m128 exp2_y_2 = _mm_mul_ps(exp2_y_1, y);
349 const __m128 exp2_y = _mm_add_ps(exp2_y_2, *((__m128*)C0));
352 a_exp_b = _mm_mul_ps(exp2_y, two_n);
357 extern const float WebRtcAec_weightCurve[65];
358 extern const float WebRtcAec_overDriveCurve[65];
360 static void OverdriveAndSuppressSSE2(AecCore* aec,
361 float hNl[PART_LEN1],
363 float efw[2][PART_LEN1]) {
365 const __m128 vec_hNlFb = _mm_set1_ps(hNlFb);
366 const __m128 vec_one = _mm_set1_ps(1.0f);
367 const __m128 vec_minus_one = _mm_set1_ps(-1.0f);
368 const __m128 vec_overDriveSm = _mm_set1_ps(aec->overDriveSm);
369 // vectorized code (four at once)
370 for (i = 0; i + 3 < PART_LEN1; i += 4) {
372 __m128 vec_hNl = _mm_loadu_ps(&hNl[i]);
373 const __m128 vec_weightCurve = _mm_loadu_ps(&WebRtcAec_weightCurve[i]);
374 const __m128 bigger = _mm_cmpgt_ps(vec_hNl, vec_hNlFb);
375 const __m128 vec_weightCurve_hNlFb = _mm_mul_ps(vec_weightCurve, vec_hNlFb);
376 const __m128 vec_one_weightCurve = _mm_sub_ps(vec_one, vec_weightCurve);
377 const __m128 vec_one_weightCurve_hNl =
378 _mm_mul_ps(vec_one_weightCurve, vec_hNl);
379 const __m128 vec_if0 = _mm_andnot_ps(bigger, vec_hNl);
380 const __m128 vec_if1 = _mm_and_ps(
381 bigger, _mm_add_ps(vec_weightCurve_hNlFb, vec_one_weightCurve_hNl));
382 vec_hNl = _mm_or_ps(vec_if0, vec_if1);
385 const __m128 vec_overDriveCurve =
386 _mm_loadu_ps(&WebRtcAec_overDriveCurve[i]);
387 const __m128 vec_overDriveSm_overDriveCurve =
388 _mm_mul_ps(vec_overDriveSm, vec_overDriveCurve);
389 vec_hNl = mm_pow_ps(vec_hNl, vec_overDriveSm_overDriveCurve);
390 _mm_storeu_ps(&hNl[i], vec_hNl);
393 // Suppress error signal
395 __m128 vec_efw_re = _mm_loadu_ps(&efw[0][i]);
396 __m128 vec_efw_im = _mm_loadu_ps(&efw[1][i]);
397 vec_efw_re = _mm_mul_ps(vec_efw_re, vec_hNl);
398 vec_efw_im = _mm_mul_ps(vec_efw_im, vec_hNl);
400 // Ooura fft returns incorrect sign on imaginary component. It matters
401 // here because we are making an additive change with comfort noise.
402 vec_efw_im = _mm_mul_ps(vec_efw_im, vec_minus_one);
403 _mm_storeu_ps(&efw[0][i], vec_efw_re);
404 _mm_storeu_ps(&efw[1][i], vec_efw_im);
407 // scalar code for the remaining items.
408 for (; i < PART_LEN1; i++) {
410 if (hNl[i] > hNlFb) {
411 hNl[i] = WebRtcAec_weightCurve[i] * hNlFb +
412 (1 - WebRtcAec_weightCurve[i]) * hNl[i];
414 hNl[i] = powf(hNl[i], aec->overDriveSm * WebRtcAec_overDriveCurve[i]);
416 // Suppress error signal
420 // Ooura fft returns incorrect sign on imaginary component. It matters
421 // here because we are making an additive change with comfort noise.
426 void WebRtcAec_InitAec_SSE2(void) {
427 WebRtcAec_FilterFar = FilterFarSSE2;
428 WebRtcAec_ScaleErrorSignal = ScaleErrorSignalSSE2;
429 WebRtcAec_FilterAdaptation = FilterAdaptationSSE2;
430 WebRtcAec_OverdriveAndSuppress = OverdriveAndSuppressSSE2;