2 * Copyright (c) 2012 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.
11 #include "webrtc/modules/audio_processing/aecm/aecm_core.h"
16 #include "webrtc/common_audio/signal_processing/include/real_fft.h"
18 // TODO(kma): Re-write the corresponding assembly file, the offset
19 // generating script and makefile, to replace these C functions.
21 // Square root of Hanning window in Q14.
22 const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = {
24 399, 798, 1196, 1594, 1990, 2386, 2780, 3172,
25 3562, 3951, 4337, 4720, 5101, 5478, 5853, 6224,
26 6591, 6954, 7313, 7668, 8019, 8364, 8705, 9040,
27 9370, 9695, 10013, 10326, 10633, 10933, 11227, 11514,
28 11795, 12068, 12335, 12594, 12845, 13089, 13325, 13553,
29 13773, 13985, 14189, 14384, 14571, 14749, 14918, 15079,
30 15231, 15373, 15506, 15631, 15746, 15851, 15947, 16034,
31 16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384
34 // Square root of Hanning window in Q14, in reversed order.
35 static const ALIGN8_BEG int16_t kSqrtHanningReversed[] ALIGN8_END = {
36 16384, 16373, 16354, 16325, 16286, 16237, 16179, 16111,
37 16034, 15947, 15851, 15746, 15631, 15506, 15373, 15231,
38 15079, 14918, 14749, 14571, 14384, 14189, 13985, 13773,
39 13553, 13325, 13089, 12845, 12594, 12335, 12068, 11795,
40 11514, 11227, 10933, 10633, 10326, 10013, 9695, 9370,
41 9040, 8705, 8364, 8019, 7668, 7313, 6954, 6591,
42 6224, 5853, 5478, 5101, 4720, 4337, 3951, 3562,
43 3172, 2780, 2386, 1990, 1594, 1196, 798, 399
46 void WebRtcAecm_WindowAndFFTNeon(AecmCore_t* aecm,
48 const int16_t* time_signal,
49 complex16_t* freq_signal,
50 int time_signal_scaling) {
52 const int16_t* p_time_signal = time_signal;
53 const int16_t* p_time_signal_offset = &time_signal[PART_LEN];
54 const int16_t* p_hanning = WebRtcAecm_kSqrtHanning;
55 const int16_t* p_hanning_reversed = kSqrtHanningReversed;
57 int16_t* p_fft_offset = &fft[PART_LEN2];
59 assert((uintptr_t)p_time_signal % 8 == 0);
60 assert((uintptr_t)freq_signal % 32 == 0);
61 assert((uintptr_t)p_hanning % 8 == 0);
62 assert((uintptr_t)p_fft % 16 == 0);
66 "vmov.i16 d21, #0\n\t"
67 "vmov.i16 d27, #0\n\t"
69 :"r"(time_signal_scaling)
73 for (i = 0; i < PART_LEN; i += 4) {
75 "vld1.16 d0, [%[p_time_signal], :64]!\n\t"
76 "vld1.16 d22, [%[p_time_signal_offset], :64]!\n\t"
77 "vld1.16 d17, [%[p_hanning], :64]!\n\t"
78 "vld1.16 d23, [%[p_hanning_reversed], :64]!\n\t"
79 "vshl.s16 d18, d0, d16\n\t"
80 "vshl.s16 d22, d22, d16\n\t"
81 "vmull.s16 q9, d18, d17\n\t"
82 "vmull.s16 q12, d22, d23\n\t"
83 "vshrn.i32 d20, q9, #14\n\t"
84 "vshrn.i32 d26, q12, #14\n\t"
85 "vst2.16 {d20, d21}, [%[p_fft], :128]!\n\t"
86 "vst2.16 {d26, d27}, [%[p_fft_offset], :128]!\n\t"
87 :[p_time_signal]"+r"(p_time_signal),
88 [p_time_signal_offset]"+r"(p_time_signal_offset),
89 [p_hanning]"+r"(p_hanning),
90 [p_hanning_reversed]"+r"(p_hanning_reversed),
92 [p_fft_offset]"+r"(p_fft_offset)
94 :"d0", "d16", "d17", "d18", "d19", "d20", "d21",
95 "d22", "d23", "d24", "d25", "d26", "d27"
99 // Do forward FFT, then take only the first PART_LEN complex samples,
100 // and change signs of the imaginary parts.
101 WebRtcSpl_RealForwardFFT(aecm->real_fft, (int16_t*)fft,
102 (int16_t*)freq_signal);
104 for (i = 0; i < PART_LEN; i += 8) {
106 "vld2.16 {d20, d21, d22, d23}, [%[freq_signal], :256]\n\t"
107 "vneg.s16 d22, d22\n\t"
108 "vneg.s16 d23, d23\n\t"
109 "vst2.16 {d20, d21, d22, d23}, [%[freq_signal], :256]!\n\t"
110 :[freq_signal]"+r"(freq_signal)
112 : "d20", "d21", "d22", "d23"
117 void WebRtcAecm_InverseFFTAndWindowNeon(AecmCore_t* aecm,
121 const int16_t* nearendClean) {
124 assert((uintptr_t)efw % 32 == 0);
125 assert((uintptr_t)fft % 16 == 0);
126 assert((uintptr_t)output% 8 == 0);
127 assert((uintptr_t)WebRtcAecm_kSqrtHanning % 8 == 0);
128 assert((uintptr_t)kSqrtHanningReversed % 8 == 0);
129 assert((uintptr_t)(aecm->outBuf) % 8 == 0);
130 assert((uintptr_t)(aecm->xBuf) % 32 == 0);
131 assert((uintptr_t)(aecm->dBufNoisy) % 32 == 0);
132 assert((uintptr_t)(aecm->dBufClean) % 32 == 0);
135 complex16_t* p_efw = efw;
136 int16_t* p_fft = fft;
137 int16_t* p_fft_offset = &fft[PART_LEN4 - 6];
139 for (i = 0, j = 0; i < PART_LEN; i += 4, j += 8) {
140 // We overwrite two more elements in fft[], but it's ok.
142 "vld2.16 {q10}, [%[p_efw], :128]!\n\t"
144 "vneg.s16 d23, d23\n\t"
145 "vst2.16 {d22, d23}, [%[p_fft], :128]!\n\t"
146 "vrev64.16 q10, q10\n\t"
147 "vst2.16 {q10}, [%[p_fft_offset]], %[offset]\n\t"
150 [p_fft_offset]"+r"(p_fft_offset)
152 :"d20", "d21", "d22", "d23"
156 fft[PART_LEN2] = efw[PART_LEN].real;
157 fft[PART_LEN2 + 1] = -efw[PART_LEN].imag;
159 // Inverse FFT. Then take only the real values, and keep outCFFT
160 // to scale the samples.
161 outCFFT = WebRtcSpl_RealInverseFFT(aecm->real_fft, fft, (int16_t*)efw);
164 __asm __volatile("vdup.32 %q0, %1" : "=w"(tmp32x4_2) : "r"((int32_t)
165 (outCFFT - aecm->dfaCleanQDomain)));
166 for (i = 0; i < PART_LEN; i += 4) {
172 //efw[i].real = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
173 // efw[i].real, WebRtcAecm_kSqrtHanning[i], 14);
174 __asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&efw[i].real));
175 __asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_1) : "r"(&WebRtcAecm_kSqrtHanning[i]));
176 __asm __volatile("vmull.s16 %q0, %P1, %P2" : "=w"(tmp32x4_0) : "w"(tmp16x4_0), "w"(tmp16x4_1));
177 __asm __volatile("vrshr.s32 %q0, %q1, #14" : "=w"(tmp32x4_0) : "0"(tmp32x4_0));
179 //tmp32no1 = WEBRTC_SPL_SHIFT_W32((int32_t)efw[i].real,
180 // outCFFT - aecm->dfaCleanQDomain);
181 __asm __volatile("vshl.s32 %q0, %q1, %q2" : "=w"(tmp32x4_0) : "0"(tmp32x4_0), "w"(tmp32x4_2));
183 //efw[i].real = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX,
184 // tmp32no1 + aecm->outBuf[i], WEBRTC_SPL_WORD16_MIN);
185 // output[i] = efw[i].real;
186 __asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&aecm->outBuf[i]));
187 __asm __volatile("vmovl.s16 %q0, %P1" : "=w"(tmp32x4_1) : "w"(tmp16x4_0));
188 __asm __volatile("vadd.i32 %q0, %q1" : : "w"(tmp32x4_0), "w"(tmp32x4_1));
189 __asm __volatile("vqmovn.s32 %P0, %q1" : "=w"(tmp16x4_0) : "w"(tmp32x4_0));
190 __asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&efw[i].real));
191 __asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&output[i]));
193 // tmp32no1 = WEBRTC_SPL_MUL_16_16_RSFT(
194 // efw[PART_LEN + i].real, WebRtcAecm_kSqrtHanning[PART_LEN - i], 14);
195 __asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_0) : "r"(&efw[PART_LEN + i].real));
196 __asm __volatile("vld1.16 %P0, [%1, :64]" : "=w"(tmp16x4_1) : "r"(&kSqrtHanningReversed[i]));
197 __asm __volatile("vmull.s16 %q0, %P1, %P2" : "=w"(tmp32x4_0) : "w"(tmp16x4_0), "w"(tmp16x4_1));
198 __asm __volatile("vshr.s32 %q0, %q1, #14" : "=w"(tmp32x4_0) : "0"(tmp32x4_0));
200 // tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1, outCFFT - aecm->dfaCleanQDomain);
201 __asm __volatile("vshl.s32 %q0, %q1, %q2" : "=w"(tmp32x4_0) : "0"(tmp32x4_0), "w"(tmp32x4_2));
202 // aecm->outBuf[i] = (int16_t)WEBRTC_SPL_SAT(
203 // WEBRTC_SPL_WORD16_MAX, tmp32no1, WEBRTC_SPL_WORD16_MIN);
204 __asm __volatile("vqmovn.s32 %P0, %q1" : "=w"(tmp16x4_0) : "w"(tmp32x4_0));
205 __asm __volatile("vst1.16 %P0, [%1, :64]" : : "w"(tmp16x4_0), "r"(&aecm->outBuf[i]));
208 // Copy the current block to the old position (outBuf is shifted elsewhere).
209 for (i = 0; i < PART_LEN; i += 16) {
210 __asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
211 "r"(&aecm->xBuf[i + PART_LEN]) : "q10");
212 __asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : : "r"(&aecm->xBuf[i]): "q10");
214 for (i = 0; i < PART_LEN; i += 16) {
215 __asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
216 "r"(&aecm->dBufNoisy[i + PART_LEN]) : "q10");
217 __asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
218 "r"(&aecm->dBufNoisy[i]): "q10");
220 if (nearendClean != NULL) {
221 for (i = 0; i < PART_LEN; i += 16) {
222 __asm __volatile("vld1.16 {d20, d21, d22, d23}, [%0, :256]" : :
223 "r"(&aecm->dBufClean[i + PART_LEN]) : "q10");
224 __asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
225 "r"(&aecm->dBufClean[i]): "q10");
230 void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore_t* aecm,
231 const uint16_t* far_spectrum,
233 uint32_t* far_energy,
234 uint32_t* echo_energy_adapt,
235 uint32_t* echo_energy_stored) {
238 register uint32_t far_energy_r;
239 register uint32_t echo_energy_stored_r;
240 register uint32_t echo_energy_adapt_r;
242 assert((uintptr_t)echo_est % 32 == 0);
243 assert((uintptr_t)(aecm->channelStored) % 16 == 0);
244 assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0);
245 assert((uintptr_t)(aecm->channelStored) % 16 == 0);
246 assert((uintptr_t)(aecm->channelStored) % 16 == 0);
248 __asm __volatile("vmov.i32 q14, #0" : : : "q14"); // far_energy
249 __asm __volatile("vmov.i32 q8, #0" : : : "q8"); // echo_energy_stored
250 __asm __volatile("vmov.i32 q9, #0" : : : "q9"); // echo_energy_adapt
252 for (i = 0; i < PART_LEN - 7; i += 8) {
253 // far_energy += (uint32_t)(far_spectrum[i]);
254 __asm __volatile("vld1.16 {d26, d27}, [%0]" : : "r"(&far_spectrum[i]) : "q13");
255 __asm __volatile("vaddw.u16 q14, q14, d26" : : : "q14", "q13");
256 __asm __volatile("vaddw.u16 q14, q14, d27" : : : "q14", "q13");
258 // Get estimated echo energies for adaptive channel and stored channel.
259 // echoEst[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], far_spectrum[i]);
260 __asm __volatile("vld1.16 {d24, d25}, [%0, :128]" : : "r"(&aecm->channelStored[i]) : "q12");
261 __asm __volatile("vmull.u16 q10, d26, d24" : : : "q12", "q13", "q10");
262 __asm __volatile("vmull.u16 q11, d27, d25" : : : "q12", "q13", "q11");
263 __asm __volatile("vst1.32 {d20, d21, d22, d23}, [%0, :256]" : : "r"(&echo_est[i]):
266 // echo_energy_stored += (uint32_t)echoEst[i];
267 __asm __volatile("vadd.u32 q8, q10" : : : "q10", "q8");
268 __asm __volatile("vadd.u32 q8, q11" : : : "q11", "q8");
270 // echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i];
271 __asm __volatile("vld1.16 {d24, d25}, [%0, :128]" : : "r"(&aecm->channelAdapt16[i]) : "q12");
272 __asm __volatile("vmull.u16 q10, d26, d24" : : : "q12", "q13", "q10");
273 __asm __volatile("vmull.u16 q11, d27, d25" : : : "q12", "q13", "q11");
274 __asm __volatile("vadd.u32 q9, q10" : : : "q9", "q15");
275 __asm __volatile("vadd.u32 q9, q11" : : : "q9", "q11");
278 __asm __volatile("vadd.u32 d28, d29" : : : "q14");
279 __asm __volatile("vpadd.u32 d28, d28" : : : "q14");
280 __asm __volatile("vmov.32 %0, d28[0]" : "=r"(far_energy_r): : "q14");
282 __asm __volatile("vadd.u32 d18, d19" : : : "q9");
283 __asm __volatile("vpadd.u32 d18, d18" : : : "q9");
284 __asm __volatile("vmov.32 %0, d18[0]" : "=r"(echo_energy_adapt_r): : "q9");
286 __asm __volatile("vadd.u32 d16, d17" : : : "q8");
287 __asm __volatile("vpadd.u32 d16, d16" : : : "q8");
288 __asm __volatile("vmov.32 %0, d16[0]" : "=r"(echo_energy_stored_r): : "q8");
290 // Get estimated echo energies for adaptive channel and stored channel.
291 echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], far_spectrum[i]);
292 *echo_energy_stored = echo_energy_stored_r + (uint32_t)echo_est[i];
293 *far_energy = far_energy_r + (uint32_t)(far_spectrum[i]);
294 *echo_energy_adapt = echo_energy_adapt_r +
295 aecm->channelAdapt16[i] * far_spectrum[i];
298 void WebRtcAecm_StoreAdaptiveChannelNeon(AecmCore_t* aecm,
299 const uint16_t* far_spectrum,
303 assert((uintptr_t)echo_est % 32 == 0);
304 assert((uintptr_t)(aecm->channelStored) % 16 == 0);
305 assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0);
307 // During startup we store the channel every block.
308 // Recalculate echo estimate.
309 for (i = 0; i < PART_LEN - 7; i += 8) {
310 // aecm->channelStored[i] = acem->channelAdapt16[i];
311 // echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], far_spectrum[i]);
312 __asm __volatile("vld1.16 {d26, d27}, [%0]" : : "r"(&far_spectrum[i]) : "q13");
313 __asm __volatile("vld1.16 {d24, d25}, [%0, :128]" : : "r"(&aecm->channelAdapt16[i]) : "q12");
314 __asm __volatile("vst1.16 {d24, d25}, [%0, :128]" : : "r"(&aecm->channelStored[i]) : "q12");
315 __asm __volatile("vmull.u16 q10, d26, d24" : : : "q12", "q13", "q10");
316 __asm __volatile("vmull.u16 q11, d27, d25" : : : "q12", "q13", "q11");
317 __asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
318 "r"(&echo_est[i]) : "q10", "q11");
320 aecm->channelStored[i] = aecm->channelAdapt16[i];
321 echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], far_spectrum[i]);
324 void WebRtcAecm_ResetAdaptiveChannelNeon(AecmCore_t* aecm) {
327 assert((uintptr_t)(aecm->channelStored) % 16 == 0);
328 assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0);
329 assert((uintptr_t)(aecm->channelAdapt32) % 32 == 0);
331 for (i = 0; i < PART_LEN - 7; i += 8) {
332 // aecm->channelAdapt16[i] = aecm->channelStored[i];
333 // aecm->channelAdapt32[i] = WEBRTC_SPL_LSHIFT_W32((int32_t)
334 // aecm->channelStored[i], 16);
335 __asm __volatile("vld1.16 {d24, d25}, [%0, :128]" : :
336 "r"(&aecm->channelStored[i]) : "q12");
337 __asm __volatile("vst1.16 {d24, d25}, [%0, :128]" : :
338 "r"(&aecm->channelAdapt16[i]) : "q12");
339 __asm __volatile("vshll.s16 q10, d24, #16" : : : "q12", "q13", "q10");
340 __asm __volatile("vshll.s16 q11, d25, #16" : : : "q12", "q13", "q11");
341 __asm __volatile("vst1.16 {d20, d21, d22, d23}, [%0, :256]" : :
342 "r"(&aecm->channelAdapt32[i]): "q10", "q11");
344 aecm->channelAdapt16[i] = aecm->channelStored[i];
345 aecm->channelAdapt32[i] = WEBRTC_SPL_LSHIFT_W32(
346 (int32_t)aecm->channelStored[i], 16);