diff options
Diffstat (limited to 'third_party/webrtc/src/webrtc/modules/audio_processing/aecm')
8 files changed, 5223 insertions, 0 deletions
diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.c b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.c new file mode 100644 index 00000000..b801f07a --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.c @@ -0,0 +1,1233 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/aecm/aecm_core.h" + +#include <assert.h> +#include <stddef.h> +#include <stdlib.h> + +#include "webrtc/common_audio/ring_buffer.h" +#include "webrtc/common_audio/signal_processing/include/real_fft.h" +#include "webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h" +#include "webrtc/modules/audio_processing/utility/delay_estimator_wrapper.h" +#include "webrtc/system_wrappers/interface/compile_assert_c.h" +#include "webrtc/system_wrappers/interface/cpu_features_wrapper.h" +#include "webrtc/typedefs.h" + +#ifdef AEC_DEBUG +FILE *dfile; +FILE *testfile; +#endif + +const int16_t WebRtcAecm_kCosTable[] = { + 8192, 8190, 8187, 8180, 8172, 8160, 8147, 8130, 8112, + 8091, 8067, 8041, 8012, 7982, 7948, 7912, 7874, 7834, + 7791, 7745, 7697, 7647, 7595, 7540, 7483, 7424, 7362, + 7299, 7233, 7164, 7094, 7021, 6947, 6870, 6791, 6710, + 6627, 6542, 6455, 6366, 6275, 6182, 6087, 5991, 5892, + 5792, 5690, 5586, 5481, 5374, 5265, 5155, 5043, 4930, + 4815, 4698, 4580, 4461, 4341, 4219, 4096, 3971, 3845, + 3719, 3591, 3462, 3331, 3200, 3068, 2935, 2801, 2667, + 2531, 2395, 2258, 2120, 1981, 1842, 1703, 1563, 1422, + 1281, 1140, 998, 856, 713, 571, 428, 285, 142, + 0, -142, -285, -428, -571, -713, -856, -998, -1140, + -1281, -1422, -1563, -1703, -1842, -1981, -2120, -2258, -2395, + -2531, -2667, -2801, -2935, -3068, -3200, -3331, -3462, -3591, + -3719, -3845, -3971, -4095, -4219, -4341, -4461, -4580, -4698, + -4815, -4930, -5043, -5155, -5265, -5374, -5481, -5586, -5690, + -5792, -5892, -5991, -6087, -6182, -6275, -6366, -6455, -6542, + -6627, -6710, -6791, -6870, -6947, -7021, -7094, -7164, -7233, + -7299, -7362, -7424, -7483, -7540, -7595, -7647, -7697, -7745, + -7791, -7834, -7874, -7912, -7948, -7982, -8012, -8041, -8067, + -8091, -8112, -8130, -8147, -8160, -8172, -8180, -8187, -8190, + -8191, -8190, -8187, -8180, -8172, -8160, -8147, -8130, -8112, + -8091, -8067, -8041, -8012, -7982, -7948, -7912, -7874, -7834, + -7791, -7745, -7697, -7647, -7595, -7540, -7483, -7424, -7362, + -7299, -7233, -7164, -7094, -7021, -6947, -6870, -6791, -6710, + -6627, -6542, -6455, -6366, -6275, -6182, -6087, -5991, -5892, + -5792, -5690, -5586, -5481, -5374, -5265, -5155, -5043, -4930, + -4815, -4698, -4580, -4461, -4341, -4219, -4096, -3971, -3845, + -3719, -3591, -3462, -3331, -3200, -3068, -2935, -2801, -2667, + -2531, -2395, -2258, -2120, -1981, -1842, -1703, -1563, -1422, + -1281, -1140, -998, -856, -713, -571, -428, -285, -142, + 0, 142, 285, 428, 571, 713, 856, 998, 1140, + 1281, 1422, 1563, 1703, 1842, 1981, 2120, 2258, 2395, + 2531, 2667, 2801, 2935, 3068, 3200, 3331, 3462, 3591, + 3719, 3845, 3971, 4095, 4219, 4341, 4461, 4580, 4698, + 4815, 4930, 5043, 5155, 5265, 5374, 5481, 5586, 5690, + 5792, 5892, 5991, 6087, 6182, 6275, 6366, 6455, 6542, + 6627, 6710, 6791, 6870, 6947, 7021, 7094, 7164, 7233, + 7299, 7362, 7424, 7483, 7540, 7595, 7647, 7697, 7745, + 7791, 7834, 7874, 7912, 7948, 7982, 8012, 8041, 8067, + 8091, 8112, 8130, 8147, 8160, 8172, 8180, 8187, 8190 +}; + +const int16_t WebRtcAecm_kSinTable[] = { + 0, 142, 285, 428, 571, 713, 856, 998, + 1140, 1281, 1422, 1563, 1703, 1842, 1981, 2120, + 2258, 2395, 2531, 2667, 2801, 2935, 3068, 3200, + 3331, 3462, 3591, 3719, 3845, 3971, 4095, 4219, + 4341, 4461, 4580, 4698, 4815, 4930, 5043, 5155, + 5265, 5374, 5481, 5586, 5690, 5792, 5892, 5991, + 6087, 6182, 6275, 6366, 6455, 6542, 6627, 6710, + 6791, 6870, 6947, 7021, 7094, 7164, 7233, 7299, + 7362, 7424, 7483, 7540, 7595, 7647, 7697, 7745, + 7791, 7834, 7874, 7912, 7948, 7982, 8012, 8041, + 8067, 8091, 8112, 8130, 8147, 8160, 8172, 8180, + 8187, 8190, 8191, 8190, 8187, 8180, 8172, 8160, + 8147, 8130, 8112, 8091, 8067, 8041, 8012, 7982, + 7948, 7912, 7874, 7834, 7791, 7745, 7697, 7647, + 7595, 7540, 7483, 7424, 7362, 7299, 7233, 7164, + 7094, 7021, 6947, 6870, 6791, 6710, 6627, 6542, + 6455, 6366, 6275, 6182, 6087, 5991, 5892, 5792, + 5690, 5586, 5481, 5374, 5265, 5155, 5043, 4930, + 4815, 4698, 4580, 4461, 4341, 4219, 4096, 3971, + 3845, 3719, 3591, 3462, 3331, 3200, 3068, 2935, + 2801, 2667, 2531, 2395, 2258, 2120, 1981, 1842, + 1703, 1563, 1422, 1281, 1140, 998, 856, 713, + 571, 428, 285, 142, 0, -142, -285, -428, + -571, -713, -856, -998, -1140, -1281, -1422, -1563, + -1703, -1842, -1981, -2120, -2258, -2395, -2531, -2667, + -2801, -2935, -3068, -3200, -3331, -3462, -3591, -3719, + -3845, -3971, -4095, -4219, -4341, -4461, -4580, -4698, + -4815, -4930, -5043, -5155, -5265, -5374, -5481, -5586, + -5690, -5792, -5892, -5991, -6087, -6182, -6275, -6366, + -6455, -6542, -6627, -6710, -6791, -6870, -6947, -7021, + -7094, -7164, -7233, -7299, -7362, -7424, -7483, -7540, + -7595, -7647, -7697, -7745, -7791, -7834, -7874, -7912, + -7948, -7982, -8012, -8041, -8067, -8091, -8112, -8130, + -8147, -8160, -8172, -8180, -8187, -8190, -8191, -8190, + -8187, -8180, -8172, -8160, -8147, -8130, -8112, -8091, + -8067, -8041, -8012, -7982, -7948, -7912, -7874, -7834, + -7791, -7745, -7697, -7647, -7595, -7540, -7483, -7424, + -7362, -7299, -7233, -7164, -7094, -7021, -6947, -6870, + -6791, -6710, -6627, -6542, -6455, -6366, -6275, -6182, + -6087, -5991, -5892, -5792, -5690, -5586, -5481, -5374, + -5265, -5155, -5043, -4930, -4815, -4698, -4580, -4461, + -4341, -4219, -4096, -3971, -3845, -3719, -3591, -3462, + -3331, -3200, -3068, -2935, -2801, -2667, -2531, -2395, + -2258, -2120, -1981, -1842, -1703, -1563, -1422, -1281, + -1140, -998, -856, -713, -571, -428, -285, -142 +}; + +// Initialization table for echo channel in 8 kHz +static const int16_t kChannelStored8kHz[PART_LEN1] = { + 2040, 1815, 1590, 1498, 1405, 1395, 1385, 1418, + 1451, 1506, 1562, 1644, 1726, 1804, 1882, 1918, + 1953, 1982, 2010, 2025, 2040, 2034, 2027, 2021, + 2014, 1997, 1980, 1925, 1869, 1800, 1732, 1683, + 1635, 1604, 1572, 1545, 1517, 1481, 1444, 1405, + 1367, 1331, 1294, 1270, 1245, 1239, 1233, 1247, + 1260, 1282, 1303, 1338, 1373, 1407, 1441, 1470, + 1499, 1524, 1549, 1565, 1582, 1601, 1621, 1649, + 1676 +}; + +// Initialization table for echo channel in 16 kHz +static const int16_t kChannelStored16kHz[PART_LEN1] = { + 2040, 1590, 1405, 1385, 1451, 1562, 1726, 1882, + 1953, 2010, 2040, 2027, 2014, 1980, 1869, 1732, + 1635, 1572, 1517, 1444, 1367, 1294, 1245, 1233, + 1260, 1303, 1373, 1441, 1499, 1549, 1582, 1621, + 1676, 1741, 1802, 1861, 1921, 1983, 2040, 2102, + 2170, 2265, 2375, 2515, 2651, 2781, 2922, 3075, + 3253, 3471, 3738, 3976, 4151, 4258, 4308, 4288, + 4270, 4253, 4237, 4179, 4086, 3947, 3757, 3484, + 3153 +}; + +// Moves the pointer to the next entry and inserts |far_spectrum| and +// corresponding Q-domain in its buffer. +// +// Inputs: +// - self : Pointer to the delay estimation instance +// - far_spectrum : Pointer to the far end spectrum +// - far_q : Q-domain of far end spectrum +// +void WebRtcAecm_UpdateFarHistory(AecmCore* self, + uint16_t* far_spectrum, + int far_q) { + // Get new buffer position + self->far_history_pos++; + if (self->far_history_pos >= MAX_DELAY) { + self->far_history_pos = 0; + } + // Update Q-domain buffer + self->far_q_domains[self->far_history_pos] = far_q; + // Update far end spectrum buffer + memcpy(&(self->far_history[self->far_history_pos * PART_LEN1]), + far_spectrum, + sizeof(uint16_t) * PART_LEN1); +} + +// Returns a pointer to the far end spectrum aligned to current near end +// spectrum. The function WebRtc_DelayEstimatorProcessFix(...) should have been +// called before AlignedFarend(...). Otherwise, you get the pointer to the +// previous frame. The memory is only valid until the next call of +// WebRtc_DelayEstimatorProcessFix(...). +// +// Inputs: +// - self : Pointer to the AECM instance. +// - delay : Current delay estimate. +// +// Output: +// - far_q : The Q-domain of the aligned far end spectrum +// +// Return value: +// - far_spectrum : Pointer to the aligned far end spectrum +// NULL - Error +// +const uint16_t* WebRtcAecm_AlignedFarend(AecmCore* self, + int* far_q, + int delay) { + int buffer_position = 0; + assert(self != NULL); + buffer_position = self->far_history_pos - delay; + + // Check buffer position + if (buffer_position < 0) { + buffer_position += MAX_DELAY; + } + // Get Q-domain + *far_q = self->far_q_domains[buffer_position]; + // Return far end spectrum + return &(self->far_history[buffer_position * PART_LEN1]); +} + +// Declare function pointers. +CalcLinearEnergies WebRtcAecm_CalcLinearEnergies; +StoreAdaptiveChannel WebRtcAecm_StoreAdaptiveChannel; +ResetAdaptiveChannel WebRtcAecm_ResetAdaptiveChannel; + +AecmCore* WebRtcAecm_CreateCore() { + AecmCore* aecm = malloc(sizeof(AecmCore)); + + aecm->farFrameBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, + sizeof(int16_t)); + if (!aecm->farFrameBuf) + { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + + aecm->nearNoisyFrameBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, + sizeof(int16_t)); + if (!aecm->nearNoisyFrameBuf) + { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + + aecm->nearCleanFrameBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, + sizeof(int16_t)); + if (!aecm->nearCleanFrameBuf) + { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + + aecm->outFrameBuf = WebRtc_CreateBuffer(FRAME_LEN + PART_LEN, + sizeof(int16_t)); + if (!aecm->outFrameBuf) + { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + + aecm->delay_estimator_farend = WebRtc_CreateDelayEstimatorFarend(PART_LEN1, + MAX_DELAY); + if (aecm->delay_estimator_farend == NULL) { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + aecm->delay_estimator = + WebRtc_CreateDelayEstimator(aecm->delay_estimator_farend, 0); + if (aecm->delay_estimator == NULL) { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + // TODO(bjornv): Explicitly disable robust delay validation until no + // performance regression has been established. Then remove the line. + WebRtc_enable_robust_validation(aecm->delay_estimator, 0); + + aecm->real_fft = WebRtcSpl_CreateRealFFT(PART_LEN_SHIFT); + if (aecm->real_fft == NULL) { + WebRtcAecm_FreeCore(aecm); + return NULL; + } + + // Init some aecm pointers. 16 and 32 byte alignment is only necessary + // for Neon code currently. + aecm->xBuf = (int16_t*) (((uintptr_t)aecm->xBuf_buf + 31) & ~ 31); + aecm->dBufClean = (int16_t*) (((uintptr_t)aecm->dBufClean_buf + 31) & ~ 31); + aecm->dBufNoisy = (int16_t*) (((uintptr_t)aecm->dBufNoisy_buf + 31) & ~ 31); + aecm->outBuf = (int16_t*) (((uintptr_t)aecm->outBuf_buf + 15) & ~ 15); + aecm->channelStored = (int16_t*) (((uintptr_t) + aecm->channelStored_buf + 15) & ~ 15); + aecm->channelAdapt16 = (int16_t*) (((uintptr_t) + aecm->channelAdapt16_buf + 15) & ~ 15); + aecm->channelAdapt32 = (int32_t*) (((uintptr_t) + aecm->channelAdapt32_buf + 31) & ~ 31); + + return aecm; +} + +void WebRtcAecm_InitEchoPathCore(AecmCore* aecm, const int16_t* echo_path) { + int i = 0; + + // Reset the stored channel + memcpy(aecm->channelStored, echo_path, sizeof(int16_t) * PART_LEN1); + // Reset the adapted channels + memcpy(aecm->channelAdapt16, echo_path, sizeof(int16_t) * PART_LEN1); + for (i = 0; i < PART_LEN1; i++) + { + aecm->channelAdapt32[i] = (int32_t)aecm->channelAdapt16[i] << 16; + } + + // Reset channel storing variables + aecm->mseAdaptOld = 1000; + aecm->mseStoredOld = 1000; + aecm->mseThreshold = WEBRTC_SPL_WORD32_MAX; + aecm->mseChannelCount = 0; +} + +static void CalcLinearEnergiesC(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored) { + int i; + + // Get energy for the delayed far end signal and estimated + // echo using both stored and adapted channels. + for (i = 0; i < PART_LEN1; i++) + { + echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + far_spectrum[i]); + (*far_energy) += (uint32_t)(far_spectrum[i]); + *echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i]; + (*echo_energy_stored) += (uint32_t)echo_est[i]; + } +} + +static void StoreAdaptiveChannelC(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est) { + int i; + + // During startup we store the channel every block. + memcpy(aecm->channelStored, aecm->channelAdapt16, sizeof(int16_t) * PART_LEN1); + // Recalculate echo estimate + for (i = 0; i < PART_LEN; i += 4) + { + echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + far_spectrum[i]); + echo_est[i + 1] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 1], + far_spectrum[i + 1]); + echo_est[i + 2] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 2], + far_spectrum[i + 2]); + echo_est[i + 3] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 3], + far_spectrum[i + 3]); + } + echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + far_spectrum[i]); +} + +static void ResetAdaptiveChannelC(AecmCore* aecm) { + int i; + + // The stored channel has a significantly lower MSE than the adaptive one for + // two consecutive calculations. Reset the adaptive channel. + memcpy(aecm->channelAdapt16, aecm->channelStored, + sizeof(int16_t) * PART_LEN1); + // Restore the W32 channel + for (i = 0; i < PART_LEN; i += 4) + { + aecm->channelAdapt32[i] = (int32_t)aecm->channelStored[i] << 16; + aecm->channelAdapt32[i + 1] = (int32_t)aecm->channelStored[i + 1] << 16; + aecm->channelAdapt32[i + 2] = (int32_t)aecm->channelStored[i + 2] << 16; + aecm->channelAdapt32[i + 3] = (int32_t)aecm->channelStored[i + 3] << 16; + } + aecm->channelAdapt32[i] = (int32_t)aecm->channelStored[i] << 16; +} + +// Initialize function pointers for ARM Neon platform. +#if (defined WEBRTC_DETECT_NEON || defined WEBRTC_HAS_NEON) +static void WebRtcAecm_InitNeon(void) +{ + WebRtcAecm_StoreAdaptiveChannel = WebRtcAecm_StoreAdaptiveChannelNeon; + WebRtcAecm_ResetAdaptiveChannel = WebRtcAecm_ResetAdaptiveChannelNeon; + WebRtcAecm_CalcLinearEnergies = WebRtcAecm_CalcLinearEnergiesNeon; +} +#endif + +// Initialize function pointers for MIPS platform. +#if defined(MIPS32_LE) +static void WebRtcAecm_InitMips(void) +{ +#if defined(MIPS_DSP_R1_LE) + WebRtcAecm_StoreAdaptiveChannel = WebRtcAecm_StoreAdaptiveChannel_mips; + WebRtcAecm_ResetAdaptiveChannel = WebRtcAecm_ResetAdaptiveChannel_mips; +#endif + WebRtcAecm_CalcLinearEnergies = WebRtcAecm_CalcLinearEnergies_mips; +} +#endif + +// WebRtcAecm_InitCore(...) +// +// This function initializes the AECM instant created with WebRtcAecm_CreateCore(...) +// Input: +// - aecm : Pointer to the Echo Suppression instance +// - samplingFreq : Sampling Frequency +// +// Output: +// - aecm : Initialized instance +// +// Return value : 0 - Ok +// -1 - Error +// +int WebRtcAecm_InitCore(AecmCore* const aecm, int samplingFreq) { + int i = 0; + int32_t tmp32 = PART_LEN1 * PART_LEN1; + int16_t tmp16 = PART_LEN1; + + if (samplingFreq != 8000 && samplingFreq != 16000) + { + samplingFreq = 8000; + return -1; + } + // sanity check of sampling frequency + aecm->mult = (int16_t)samplingFreq / 8000; + + aecm->farBufWritePos = 0; + aecm->farBufReadPos = 0; + aecm->knownDelay = 0; + aecm->lastKnownDelay = 0; + + WebRtc_InitBuffer(aecm->farFrameBuf); + WebRtc_InitBuffer(aecm->nearNoisyFrameBuf); + WebRtc_InitBuffer(aecm->nearCleanFrameBuf); + WebRtc_InitBuffer(aecm->outFrameBuf); + + memset(aecm->xBuf_buf, 0, sizeof(aecm->xBuf_buf)); + memset(aecm->dBufClean_buf, 0, sizeof(aecm->dBufClean_buf)); + memset(aecm->dBufNoisy_buf, 0, sizeof(aecm->dBufNoisy_buf)); + memset(aecm->outBuf_buf, 0, sizeof(aecm->outBuf_buf)); + + aecm->seed = 666; + aecm->totCount = 0; + + if (WebRtc_InitDelayEstimatorFarend(aecm->delay_estimator_farend) != 0) { + return -1; + } + if (WebRtc_InitDelayEstimator(aecm->delay_estimator) != 0) { + return -1; + } + // Set far end histories to zero + memset(aecm->far_history, 0, sizeof(uint16_t) * PART_LEN1 * MAX_DELAY); + memset(aecm->far_q_domains, 0, sizeof(int) * MAX_DELAY); + aecm->far_history_pos = MAX_DELAY; + + aecm->nlpFlag = 1; + aecm->fixedDelay = -1; + + aecm->dfaCleanQDomain = 0; + aecm->dfaCleanQDomainOld = 0; + aecm->dfaNoisyQDomain = 0; + aecm->dfaNoisyQDomainOld = 0; + + memset(aecm->nearLogEnergy, 0, sizeof(aecm->nearLogEnergy)); + aecm->farLogEnergy = 0; + memset(aecm->echoAdaptLogEnergy, 0, sizeof(aecm->echoAdaptLogEnergy)); + memset(aecm->echoStoredLogEnergy, 0, sizeof(aecm->echoStoredLogEnergy)); + + // Initialize the echo channels with a stored shape. + if (samplingFreq == 8000) + { + WebRtcAecm_InitEchoPathCore(aecm, kChannelStored8kHz); + } + else + { + WebRtcAecm_InitEchoPathCore(aecm, kChannelStored16kHz); + } + + memset(aecm->echoFilt, 0, sizeof(aecm->echoFilt)); + memset(aecm->nearFilt, 0, sizeof(aecm->nearFilt)); + aecm->noiseEstCtr = 0; + + aecm->cngMode = AecmTrue; + + memset(aecm->noiseEstTooLowCtr, 0, sizeof(aecm->noiseEstTooLowCtr)); + memset(aecm->noiseEstTooHighCtr, 0, sizeof(aecm->noiseEstTooHighCtr)); + // Shape the initial noise level to an approximate pink noise. + for (i = 0; i < (PART_LEN1 >> 1) - 1; i++) + { + aecm->noiseEst[i] = (tmp32 << 8); + tmp16--; + tmp32 -= (int32_t)((tmp16 << 1) + 1); + } + for (; i < PART_LEN1; i++) + { + aecm->noiseEst[i] = (tmp32 << 8); + } + + aecm->farEnergyMin = WEBRTC_SPL_WORD16_MAX; + aecm->farEnergyMax = WEBRTC_SPL_WORD16_MIN; + aecm->farEnergyMaxMin = 0; + aecm->farEnergyVAD = FAR_ENERGY_MIN; // This prevents false speech detection at the + // beginning. + aecm->farEnergyMSE = 0; + aecm->currentVADValue = 0; + aecm->vadUpdateCount = 0; + aecm->firstVAD = 1; + + aecm->startupState = 0; + aecm->supGain = SUPGAIN_DEFAULT; + aecm->supGainOld = SUPGAIN_DEFAULT; + + aecm->supGainErrParamA = SUPGAIN_ERROR_PARAM_A; + aecm->supGainErrParamD = SUPGAIN_ERROR_PARAM_D; + aecm->supGainErrParamDiffAB = SUPGAIN_ERROR_PARAM_A - SUPGAIN_ERROR_PARAM_B; + aecm->supGainErrParamDiffBD = SUPGAIN_ERROR_PARAM_B - SUPGAIN_ERROR_PARAM_D; + + // Assert a preprocessor definition at compile-time. It's an assumption + // used in assembly code, so check the assembly files before any change. + COMPILE_ASSERT(PART_LEN % 16 == 0); + + // Initialize function pointers. + WebRtcAecm_CalcLinearEnergies = CalcLinearEnergiesC; + WebRtcAecm_StoreAdaptiveChannel = StoreAdaptiveChannelC; + WebRtcAecm_ResetAdaptiveChannel = ResetAdaptiveChannelC; + +#ifdef WEBRTC_DETECT_NEON + uint64_t features = WebRtc_GetCPUFeaturesARM(); + if ((features & kCPUFeatureNEON) != 0) + { + WebRtcAecm_InitNeon(); + } +#elif defined(WEBRTC_HAS_NEON) + WebRtcAecm_InitNeon(); +#endif + +#if defined(MIPS32_LE) + WebRtcAecm_InitMips(); +#endif + return 0; +} + +// TODO(bjornv): This function is currently not used. Add support for these +// parameters from a higher level +int WebRtcAecm_Control(AecmCore* aecm, int delay, int nlpFlag) { + aecm->nlpFlag = nlpFlag; + aecm->fixedDelay = delay; + + return 0; +} + +void WebRtcAecm_FreeCore(AecmCore* aecm) { + if (aecm == NULL) { + return; + } + + WebRtc_FreeBuffer(aecm->farFrameBuf); + WebRtc_FreeBuffer(aecm->nearNoisyFrameBuf); + WebRtc_FreeBuffer(aecm->nearCleanFrameBuf); + WebRtc_FreeBuffer(aecm->outFrameBuf); + + WebRtc_FreeDelayEstimator(aecm->delay_estimator); + WebRtc_FreeDelayEstimatorFarend(aecm->delay_estimator_farend); + WebRtcSpl_FreeRealFFT(aecm->real_fft); + + free(aecm); +} + +int WebRtcAecm_ProcessFrame(AecmCore* aecm, + const int16_t* farend, + const int16_t* nearendNoisy, + const int16_t* nearendClean, + int16_t* out) { + int16_t outBlock_buf[PART_LEN + 8]; // Align buffer to 8-byte boundary. + int16_t* outBlock = (int16_t*) (((uintptr_t) outBlock_buf + 15) & ~ 15); + + int16_t farFrame[FRAME_LEN]; + const int16_t* out_ptr = NULL; + int size = 0; + + // Buffer the current frame. + // Fetch an older one corresponding to the delay. + WebRtcAecm_BufferFarFrame(aecm, farend, FRAME_LEN); + WebRtcAecm_FetchFarFrame(aecm, farFrame, FRAME_LEN, aecm->knownDelay); + + // Buffer the synchronized far and near frames, + // to pass the smaller blocks individually. + WebRtc_WriteBuffer(aecm->farFrameBuf, farFrame, FRAME_LEN); + WebRtc_WriteBuffer(aecm->nearNoisyFrameBuf, nearendNoisy, FRAME_LEN); + if (nearendClean != NULL) + { + WebRtc_WriteBuffer(aecm->nearCleanFrameBuf, nearendClean, FRAME_LEN); + } + + // Process as many blocks as possible. + while (WebRtc_available_read(aecm->farFrameBuf) >= PART_LEN) + { + int16_t far_block[PART_LEN]; + const int16_t* far_block_ptr = NULL; + int16_t near_noisy_block[PART_LEN]; + const int16_t* near_noisy_block_ptr = NULL; + + WebRtc_ReadBuffer(aecm->farFrameBuf, (void**) &far_block_ptr, far_block, + PART_LEN); + WebRtc_ReadBuffer(aecm->nearNoisyFrameBuf, + (void**) &near_noisy_block_ptr, + near_noisy_block, + PART_LEN); + if (nearendClean != NULL) + { + int16_t near_clean_block[PART_LEN]; + const int16_t* near_clean_block_ptr = NULL; + + WebRtc_ReadBuffer(aecm->nearCleanFrameBuf, + (void**) &near_clean_block_ptr, + near_clean_block, + PART_LEN); + if (WebRtcAecm_ProcessBlock(aecm, + far_block_ptr, + near_noisy_block_ptr, + near_clean_block_ptr, + outBlock) == -1) + { + return -1; + } + } else + { + if (WebRtcAecm_ProcessBlock(aecm, + far_block_ptr, + near_noisy_block_ptr, + NULL, + outBlock) == -1) + { + return -1; + } + } + + WebRtc_WriteBuffer(aecm->outFrameBuf, outBlock, PART_LEN); + } + + // Stuff the out buffer if we have less than a frame to output. + // This should only happen for the first frame. + size = (int) WebRtc_available_read(aecm->outFrameBuf); + if (size < FRAME_LEN) + { + WebRtc_MoveReadPtr(aecm->outFrameBuf, size - FRAME_LEN); + } + + // Obtain an output frame. + WebRtc_ReadBuffer(aecm->outFrameBuf, (void**) &out_ptr, out, FRAME_LEN); + if (out_ptr != out) { + // ReadBuffer() hasn't copied to |out| in this case. + memcpy(out, out_ptr, FRAME_LEN * sizeof(int16_t)); + } + + return 0; +} + +// WebRtcAecm_AsymFilt(...) +// +// Performs asymmetric filtering. +// +// Inputs: +// - filtOld : Previous filtered value. +// - inVal : New input value. +// - stepSizePos : Step size when we have a positive contribution. +// - stepSizeNeg : Step size when we have a negative contribution. +// +// Output: +// +// Return: - Filtered value. +// +int16_t WebRtcAecm_AsymFilt(const int16_t filtOld, const int16_t inVal, + const int16_t stepSizePos, + const int16_t stepSizeNeg) +{ + int16_t retVal; + + if ((filtOld == WEBRTC_SPL_WORD16_MAX) | (filtOld == WEBRTC_SPL_WORD16_MIN)) + { + return inVal; + } + retVal = filtOld; + if (filtOld > inVal) + { + retVal -= (filtOld - inVal) >> stepSizeNeg; + } else + { + retVal += (inVal - filtOld) >> stepSizePos; + } + + return retVal; +} + +// ExtractFractionPart(a, zeros) +// +// returns the fraction part of |a|, with |zeros| number of leading zeros, as an +// int16_t scaled to Q8. There is no sanity check of |a| in the sense that the +// number of zeros match. +static int16_t ExtractFractionPart(uint32_t a, int zeros) { + return (int16_t)(((a << zeros) & 0x7FFFFFFF) >> 23); +} + +// Calculates and returns the log of |energy| in Q8. The input |energy| is +// supposed to be in Q(|q_domain|). +static int16_t LogOfEnergyInQ8(uint32_t energy, int q_domain) { + static const int16_t kLogLowValue = PART_LEN_SHIFT << 7; + int16_t log_energy_q8 = kLogLowValue; + if (energy > 0) { + int zeros = WebRtcSpl_NormU32(energy); + int16_t frac = ExtractFractionPart(energy, zeros); + // log2 of |energy| in Q8. + log_energy_q8 += ((31 - zeros) << 8) + frac - (q_domain << 8); + } + return log_energy_q8; +} + +// WebRtcAecm_CalcEnergies(...) +// +// This function calculates the log of energies for nearend, farend and estimated +// echoes. There is also an update of energy decision levels, i.e. internal VAD. +// +// +// @param aecm [i/o] Handle of the AECM instance. +// @param far_spectrum [in] Pointer to farend spectrum. +// @param far_q [in] Q-domain of farend spectrum. +// @param nearEner [in] Near end energy for current block in +// Q(aecm->dfaQDomain). +// @param echoEst [out] Estimated echo in Q(xfa_q+RESOLUTION_CHANNEL16). +// +void WebRtcAecm_CalcEnergies(AecmCore* aecm, + const uint16_t* far_spectrum, + const int16_t far_q, + const uint32_t nearEner, + int32_t* echoEst) { + // Local variables + uint32_t tmpAdapt = 0; + uint32_t tmpStored = 0; + uint32_t tmpFar = 0; + + int i; + + int16_t tmp16; + int16_t increase_max_shifts = 4; + int16_t decrease_max_shifts = 11; + int16_t increase_min_shifts = 11; + int16_t decrease_min_shifts = 3; + + // Get log of near end energy and store in buffer + + // Shift buffer + memmove(aecm->nearLogEnergy + 1, aecm->nearLogEnergy, + sizeof(int16_t) * (MAX_BUF_LEN - 1)); + + // Logarithm of integrated magnitude spectrum (nearEner) + aecm->nearLogEnergy[0] = LogOfEnergyInQ8(nearEner, aecm->dfaNoisyQDomain); + + WebRtcAecm_CalcLinearEnergies(aecm, far_spectrum, echoEst, &tmpFar, &tmpAdapt, &tmpStored); + + // Shift buffers + memmove(aecm->echoAdaptLogEnergy + 1, aecm->echoAdaptLogEnergy, + sizeof(int16_t) * (MAX_BUF_LEN - 1)); + memmove(aecm->echoStoredLogEnergy + 1, aecm->echoStoredLogEnergy, + sizeof(int16_t) * (MAX_BUF_LEN - 1)); + + // Logarithm of delayed far end energy + aecm->farLogEnergy = LogOfEnergyInQ8(tmpFar, far_q); + + // Logarithm of estimated echo energy through adapted channel + aecm->echoAdaptLogEnergy[0] = LogOfEnergyInQ8(tmpAdapt, + RESOLUTION_CHANNEL16 + far_q); + + // Logarithm of estimated echo energy through stored channel + aecm->echoStoredLogEnergy[0] = + LogOfEnergyInQ8(tmpStored, RESOLUTION_CHANNEL16 + far_q); + + // Update farend energy levels (min, max, vad, mse) + if (aecm->farLogEnergy > FAR_ENERGY_MIN) + { + if (aecm->startupState == 0) + { + increase_max_shifts = 2; + decrease_min_shifts = 2; + increase_min_shifts = 8; + } + + aecm->farEnergyMin = WebRtcAecm_AsymFilt(aecm->farEnergyMin, aecm->farLogEnergy, + increase_min_shifts, decrease_min_shifts); + aecm->farEnergyMax = WebRtcAecm_AsymFilt(aecm->farEnergyMax, aecm->farLogEnergy, + increase_max_shifts, decrease_max_shifts); + aecm->farEnergyMaxMin = (aecm->farEnergyMax - aecm->farEnergyMin); + + // Dynamic VAD region size + tmp16 = 2560 - aecm->farEnergyMin; + if (tmp16 > 0) + { + tmp16 = (int16_t)((tmp16 * FAR_ENERGY_VAD_REGION) >> 9); + } else + { + tmp16 = 0; + } + tmp16 += FAR_ENERGY_VAD_REGION; + + if ((aecm->startupState == 0) | (aecm->vadUpdateCount > 1024)) + { + // In startup phase or VAD update halted + aecm->farEnergyVAD = aecm->farEnergyMin + tmp16; + } else + { + if (aecm->farEnergyVAD > aecm->farLogEnergy) + { + aecm->farEnergyVAD += + (aecm->farLogEnergy + tmp16 - aecm->farEnergyVAD) >> 6; + aecm->vadUpdateCount = 0; + } else + { + aecm->vadUpdateCount++; + } + } + // Put MSE threshold higher than VAD + aecm->farEnergyMSE = aecm->farEnergyVAD + (1 << 8); + } + + // Update VAD variables + if (aecm->farLogEnergy > aecm->farEnergyVAD) + { + if ((aecm->startupState == 0) | (aecm->farEnergyMaxMin > FAR_ENERGY_DIFF)) + { + // We are in startup or have significant dynamics in input speech level + aecm->currentVADValue = 1; + } + } else + { + aecm->currentVADValue = 0; + } + if ((aecm->currentVADValue) && (aecm->firstVAD)) + { + aecm->firstVAD = 0; + if (aecm->echoAdaptLogEnergy[0] > aecm->nearLogEnergy[0]) + { + // The estimated echo has higher energy than the near end signal. + // This means that the initialization was too aggressive. Scale + // down by a factor 8 + for (i = 0; i < PART_LEN1; i++) + { + aecm->channelAdapt16[i] >>= 3; + } + // Compensate the adapted echo energy level accordingly. + aecm->echoAdaptLogEnergy[0] -= (3 << 8); + aecm->firstVAD = 1; + } + } +} + +// WebRtcAecm_CalcStepSize(...) +// +// This function calculates the step size used in channel estimation +// +// +// @param aecm [in] Handle of the AECM instance. +// @param mu [out] (Return value) Stepsize in log2(), i.e. number of shifts. +// +// +int16_t WebRtcAecm_CalcStepSize(AecmCore* const aecm) { + int32_t tmp32; + int16_t tmp16; + int16_t mu = MU_MAX; + + // Here we calculate the step size mu used in the + // following NLMS based Channel estimation algorithm + if (!aecm->currentVADValue) + { + // Far end energy level too low, no channel update + mu = 0; + } else if (aecm->startupState > 0) + { + if (aecm->farEnergyMin >= aecm->farEnergyMax) + { + mu = MU_MIN; + } else + { + tmp16 = (aecm->farLogEnergy - aecm->farEnergyMin); + tmp32 = tmp16 * MU_DIFF; + tmp32 = WebRtcSpl_DivW32W16(tmp32, aecm->farEnergyMaxMin); + mu = MU_MIN - 1 - (int16_t)(tmp32); + // The -1 is an alternative to rounding. This way we get a larger + // stepsize, so we in some sense compensate for truncation in NLMS + } + if (mu < MU_MAX) + { + mu = MU_MAX; // Equivalent with maximum step size of 2^-MU_MAX + } + } + + return mu; +} + +// WebRtcAecm_UpdateChannel(...) +// +// This function performs channel estimation. NLMS and decision on channel storage. +// +// +// @param aecm [i/o] Handle of the AECM instance. +// @param far_spectrum [in] Absolute value of the farend signal in Q(far_q) +// @param far_q [in] Q-domain of the farend signal +// @param dfa [in] Absolute value of the nearend signal (Q[aecm->dfaQDomain]) +// @param mu [in] NLMS step size. +// @param echoEst [i/o] Estimated echo in Q(far_q+RESOLUTION_CHANNEL16). +// +void WebRtcAecm_UpdateChannel(AecmCore* aecm, + const uint16_t* far_spectrum, + const int16_t far_q, + const uint16_t* const dfa, + const int16_t mu, + int32_t* echoEst) { + uint32_t tmpU32no1, tmpU32no2; + int32_t tmp32no1, tmp32no2; + int32_t mseStored; + int32_t mseAdapt; + + int i; + + int16_t zerosFar, zerosNum, zerosCh, zerosDfa; + int16_t shiftChFar, shiftNum, shift2ResChan; + int16_t tmp16no1; + int16_t xfaQ, dfaQ; + + // This is the channel estimation algorithm. It is base on NLMS but has a variable step + // length, which was calculated above. + if (mu) + { + for (i = 0; i < PART_LEN1; i++) + { + // Determine norm of channel and farend to make sure we don't get overflow in + // multiplication + zerosCh = WebRtcSpl_NormU32(aecm->channelAdapt32[i]); + zerosFar = WebRtcSpl_NormU32((uint32_t)far_spectrum[i]); + if (zerosCh + zerosFar > 31) + { + // Multiplication is safe + tmpU32no1 = WEBRTC_SPL_UMUL_32_16(aecm->channelAdapt32[i], + far_spectrum[i]); + shiftChFar = 0; + } else + { + // We need to shift down before multiplication + shiftChFar = 32 - zerosCh - zerosFar; + tmpU32no1 = (aecm->channelAdapt32[i] >> shiftChFar) * + far_spectrum[i]; + } + // Determine Q-domain of numerator + zerosNum = WebRtcSpl_NormU32(tmpU32no1); + if (dfa[i]) + { + zerosDfa = WebRtcSpl_NormU32((uint32_t)dfa[i]); + } else + { + zerosDfa = 32; + } + tmp16no1 = zerosDfa - 2 + aecm->dfaNoisyQDomain - + RESOLUTION_CHANNEL32 - far_q + shiftChFar; + if (zerosNum > tmp16no1 + 1) + { + xfaQ = tmp16no1; + dfaQ = zerosDfa - 2; + } else + { + xfaQ = zerosNum - 2; + dfaQ = RESOLUTION_CHANNEL32 + far_q - aecm->dfaNoisyQDomain - + shiftChFar + xfaQ; + } + // Add in the same Q-domain + tmpU32no1 = WEBRTC_SPL_SHIFT_W32(tmpU32no1, xfaQ); + tmpU32no2 = WEBRTC_SPL_SHIFT_W32((uint32_t)dfa[i], dfaQ); + tmp32no1 = (int32_t)tmpU32no2 - (int32_t)tmpU32no1; + zerosNum = WebRtcSpl_NormW32(tmp32no1); + if ((tmp32no1) && (far_spectrum[i] > (CHANNEL_VAD << far_q))) + { + // + // Update is needed + // + // This is what we would like to compute + // + // tmp32no1 = dfa[i] - (aecm->channelAdapt[i] * far_spectrum[i]) + // tmp32norm = (i + 1) + // aecm->channelAdapt[i] += (2^mu) * tmp32no1 + // / (tmp32norm * far_spectrum[i]) + // + + // Make sure we don't get overflow in multiplication. + if (zerosNum + zerosFar > 31) + { + if (tmp32no1 > 0) + { + tmp32no2 = (int32_t)WEBRTC_SPL_UMUL_32_16(tmp32no1, + far_spectrum[i]); + } else + { + tmp32no2 = -(int32_t)WEBRTC_SPL_UMUL_32_16(-tmp32no1, + far_spectrum[i]); + } + shiftNum = 0; + } else + { + shiftNum = 32 - (zerosNum + zerosFar); + if (tmp32no1 > 0) + { + tmp32no2 = (tmp32no1 >> shiftNum) * far_spectrum[i]; + } else + { + tmp32no2 = -((-tmp32no1 >> shiftNum) * far_spectrum[i]); + } + } + // Normalize with respect to frequency bin + tmp32no2 = WebRtcSpl_DivW32W16(tmp32no2, i + 1); + // Make sure we are in the right Q-domain + shift2ResChan = shiftNum + shiftChFar - xfaQ - mu - ((30 - zerosFar) << 1); + if (WebRtcSpl_NormW32(tmp32no2) < shift2ResChan) + { + tmp32no2 = WEBRTC_SPL_WORD32_MAX; + } else + { + tmp32no2 = WEBRTC_SPL_SHIFT_W32(tmp32no2, shift2ResChan); + } + aecm->channelAdapt32[i] = + WebRtcSpl_AddSatW32(aecm->channelAdapt32[i], tmp32no2); + if (aecm->channelAdapt32[i] < 0) + { + // We can never have negative channel gain + aecm->channelAdapt32[i] = 0; + } + aecm->channelAdapt16[i] = + (int16_t)(aecm->channelAdapt32[i] >> 16); + } + } + } + // END: Adaptive channel update + + // Determine if we should store or restore the channel + if ((aecm->startupState == 0) & (aecm->currentVADValue)) + { + // During startup we store the channel every block, + // and we recalculate echo estimate + WebRtcAecm_StoreAdaptiveChannel(aecm, far_spectrum, echoEst); + } else + { + if (aecm->farLogEnergy < aecm->farEnergyMSE) + { + aecm->mseChannelCount = 0; + } else + { + aecm->mseChannelCount++; + } + // Enough data for validation. Store channel if we can. + if (aecm->mseChannelCount >= (MIN_MSE_COUNT + 10)) + { + // We have enough data. + // Calculate MSE of "Adapt" and "Stored" versions. + // It is actually not MSE, but average absolute error. + mseStored = 0; + mseAdapt = 0; + for (i = 0; i < MIN_MSE_COUNT; i++) + { + tmp32no1 = ((int32_t)aecm->echoStoredLogEnergy[i] + - (int32_t)aecm->nearLogEnergy[i]); + tmp32no2 = WEBRTC_SPL_ABS_W32(tmp32no1); + mseStored += tmp32no2; + + tmp32no1 = ((int32_t)aecm->echoAdaptLogEnergy[i] + - (int32_t)aecm->nearLogEnergy[i]); + tmp32no2 = WEBRTC_SPL_ABS_W32(tmp32no1); + mseAdapt += tmp32no2; + } + if (((mseStored << MSE_RESOLUTION) < (MIN_MSE_DIFF * mseAdapt)) + & ((aecm->mseStoredOld << MSE_RESOLUTION) < (MIN_MSE_DIFF + * aecm->mseAdaptOld))) + { + // The stored channel has a significantly lower MSE than the adaptive one for + // two consecutive calculations. Reset the adaptive channel. + WebRtcAecm_ResetAdaptiveChannel(aecm); + } else if (((MIN_MSE_DIFF * mseStored) > (mseAdapt << MSE_RESOLUTION)) & (mseAdapt + < aecm->mseThreshold) & (aecm->mseAdaptOld < aecm->mseThreshold)) + { + // The adaptive channel has a significantly lower MSE than the stored one. + // The MSE for the adaptive channel has also been low for two consecutive + // calculations. Store the adaptive channel. + WebRtcAecm_StoreAdaptiveChannel(aecm, far_spectrum, echoEst); + + // Update threshold + if (aecm->mseThreshold == WEBRTC_SPL_WORD32_MAX) + { + aecm->mseThreshold = (mseAdapt + aecm->mseAdaptOld); + } else + { + int scaled_threshold = aecm->mseThreshold * 5 / 8; + aecm->mseThreshold += + ((mseAdapt - scaled_threshold) * 205) >> 8; + } + + } + + // Reset counter + aecm->mseChannelCount = 0; + + // Store the MSE values. + aecm->mseStoredOld = mseStored; + aecm->mseAdaptOld = mseAdapt; + } + } + // END: Determine if we should store or reset channel estimate. +} + +// CalcSuppressionGain(...) +// +// This function calculates the suppression gain that is used in the Wiener filter. +// +// +// @param aecm [i/n] Handle of the AECM instance. +// @param supGain [out] (Return value) Suppression gain with which to scale the noise +// level (Q14). +// +// +int16_t WebRtcAecm_CalcSuppressionGain(AecmCore* const aecm) { + int32_t tmp32no1; + + int16_t supGain = SUPGAIN_DEFAULT; + int16_t tmp16no1; + int16_t dE = 0; + + // Determine suppression gain used in the Wiener filter. The gain is based on a mix of far + // end energy and echo estimation error. + // Adjust for the far end signal level. A low signal level indicates no far end signal, + // hence we set the suppression gain to 0 + if (!aecm->currentVADValue) + { + supGain = 0; + } else + { + // Adjust for possible double talk. If we have large variations in estimation error we + // likely have double talk (or poor channel). + tmp16no1 = (aecm->nearLogEnergy[0] - aecm->echoStoredLogEnergy[0] - ENERGY_DEV_OFFSET); + dE = WEBRTC_SPL_ABS_W16(tmp16no1); + + if (dE < ENERGY_DEV_TOL) + { + // Likely no double talk. The better estimation, the more we can suppress signal. + // Update counters + if (dE < SUPGAIN_EPC_DT) + { + tmp32no1 = aecm->supGainErrParamDiffAB * dE; + tmp32no1 += (SUPGAIN_EPC_DT >> 1); + tmp16no1 = (int16_t)WebRtcSpl_DivW32W16(tmp32no1, SUPGAIN_EPC_DT); + supGain = aecm->supGainErrParamA - tmp16no1; + } else + { + tmp32no1 = aecm->supGainErrParamDiffBD * (ENERGY_DEV_TOL - dE); + tmp32no1 += ((ENERGY_DEV_TOL - SUPGAIN_EPC_DT) >> 1); + tmp16no1 = (int16_t)WebRtcSpl_DivW32W16(tmp32no1, (ENERGY_DEV_TOL + - SUPGAIN_EPC_DT)); + supGain = aecm->supGainErrParamD + tmp16no1; + } + } else + { + // Likely in double talk. Use default value + supGain = aecm->supGainErrParamD; + } + } + + if (supGain > aecm->supGainOld) + { + tmp16no1 = supGain; + } else + { + tmp16no1 = aecm->supGainOld; + } + aecm->supGainOld = supGain; + if (tmp16no1 < aecm->supGain) + { + aecm->supGain += (int16_t)((tmp16no1 - aecm->supGain) >> 4); + } else + { + aecm->supGain += (int16_t)((tmp16no1 - aecm->supGain) >> 4); + } + + // END: Update suppression gain + + return aecm->supGain; +} + +void WebRtcAecm_BufferFarFrame(AecmCore* const aecm, + const int16_t* const farend, + const int farLen) { + int writeLen = farLen, writePos = 0; + + // Check if the write position must be wrapped + while (aecm->farBufWritePos + writeLen > FAR_BUF_LEN) + { + // Write to remaining buffer space before wrapping + writeLen = FAR_BUF_LEN - aecm->farBufWritePos; + memcpy(aecm->farBuf + aecm->farBufWritePos, farend + writePos, + sizeof(int16_t) * writeLen); + aecm->farBufWritePos = 0; + writePos = writeLen; + writeLen = farLen - writeLen; + } + + memcpy(aecm->farBuf + aecm->farBufWritePos, farend + writePos, + sizeof(int16_t) * writeLen); + aecm->farBufWritePos += writeLen; +} + +void WebRtcAecm_FetchFarFrame(AecmCore* const aecm, + int16_t* const farend, + const int farLen, + const int knownDelay) { + int readLen = farLen; + int readPos = 0; + int delayChange = knownDelay - aecm->lastKnownDelay; + + aecm->farBufReadPos -= delayChange; + + // Check if delay forces a read position wrap + while (aecm->farBufReadPos < 0) + { + aecm->farBufReadPos += FAR_BUF_LEN; + } + while (aecm->farBufReadPos > FAR_BUF_LEN - 1) + { + aecm->farBufReadPos -= FAR_BUF_LEN; + } + + aecm->lastKnownDelay = knownDelay; + + // Check if read position must be wrapped + while (aecm->farBufReadPos + readLen > FAR_BUF_LEN) + { + + // Read from remaining buffer space before wrapping + readLen = FAR_BUF_LEN - aecm->farBufReadPos; + memcpy(farend + readPos, aecm->farBuf + aecm->farBufReadPos, + sizeof(int16_t) * readLen); + aecm->farBufReadPos = 0; + readPos = readLen; + readLen = farLen - readLen; + } + memcpy(farend + readPos, aecm->farBuf + aecm->farBufReadPos, + sizeof(int16_t) * readLen); + aecm->farBufReadPos += readLen; +} diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.h b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.h new file mode 100644 index 00000000..b52bb62d --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core.h @@ -0,0 +1,434 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +// Performs echo control (suppression) with fft routines in fixed-point. + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AECM_AECM_CORE_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_AECM_AECM_CORE_H_ + +#include "webrtc/common_audio/ring_buffer.h" +#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h" +#include "webrtc/modules/audio_processing/aecm/aecm_defines.h" +#include "webrtc/typedefs.h" + +#ifdef _MSC_VER // visual c++ +#define ALIGN8_BEG __declspec(align(8)) +#define ALIGN8_END +#else // gcc or icc +#define ALIGN8_BEG +#define ALIGN8_END __attribute__((aligned(8))) +#endif + +typedef struct { + int16_t real; + int16_t imag; +} ComplexInt16; + +typedef struct { + int farBufWritePos; + int farBufReadPos; + int knownDelay; + int lastKnownDelay; + int firstVAD; // Parameter to control poorly initialized channels + + RingBuffer* farFrameBuf; + RingBuffer* nearNoisyFrameBuf; + RingBuffer* nearCleanFrameBuf; + RingBuffer* outFrameBuf; + + int16_t farBuf[FAR_BUF_LEN]; + + int16_t mult; + uint32_t seed; + + // Delay estimation variables + void* delay_estimator_farend; + void* delay_estimator; + uint16_t currentDelay; + // Far end history variables + // TODO(bjornv): Replace |far_history| with ring_buffer. + uint16_t far_history[PART_LEN1 * MAX_DELAY]; + int far_history_pos; + int far_q_domains[MAX_DELAY]; + + int16_t nlpFlag; + int16_t fixedDelay; + + uint32_t totCount; + + int16_t dfaCleanQDomain; + int16_t dfaCleanQDomainOld; + int16_t dfaNoisyQDomain; + int16_t dfaNoisyQDomainOld; + + int16_t nearLogEnergy[MAX_BUF_LEN]; + int16_t farLogEnergy; + int16_t echoAdaptLogEnergy[MAX_BUF_LEN]; + int16_t echoStoredLogEnergy[MAX_BUF_LEN]; + + // The extra 16 or 32 bytes in the following buffers are for alignment based + // Neon code. + // It's designed this way since the current GCC compiler can't align a + // buffer in 16 or 32 byte boundaries properly. + int16_t channelStored_buf[PART_LEN1 + 8]; + int16_t channelAdapt16_buf[PART_LEN1 + 8]; + int32_t channelAdapt32_buf[PART_LEN1 + 8]; + int16_t xBuf_buf[PART_LEN2 + 16]; // farend + int16_t dBufClean_buf[PART_LEN2 + 16]; // nearend + int16_t dBufNoisy_buf[PART_LEN2 + 16]; // nearend + int16_t outBuf_buf[PART_LEN + 8]; + + // Pointers to the above buffers + int16_t *channelStored; + int16_t *channelAdapt16; + int32_t *channelAdapt32; + int16_t *xBuf; + int16_t *dBufClean; + int16_t *dBufNoisy; + int16_t *outBuf; + + int32_t echoFilt[PART_LEN1]; + int16_t nearFilt[PART_LEN1]; + int32_t noiseEst[PART_LEN1]; + int noiseEstTooLowCtr[PART_LEN1]; + int noiseEstTooHighCtr[PART_LEN1]; + int16_t noiseEstCtr; + int16_t cngMode; + + int32_t mseAdaptOld; + int32_t mseStoredOld; + int32_t mseThreshold; + + int16_t farEnergyMin; + int16_t farEnergyMax; + int16_t farEnergyMaxMin; + int16_t farEnergyVAD; + int16_t farEnergyMSE; + int currentVADValue; + int16_t vadUpdateCount; + + int16_t startupState; + int16_t mseChannelCount; + int16_t supGain; + int16_t supGainOld; + + int16_t supGainErrParamA; + int16_t supGainErrParamD; + int16_t supGainErrParamDiffAB; + int16_t supGainErrParamDiffBD; + + struct RealFFT* real_fft; + +#ifdef AEC_DEBUG + FILE *farFile; + FILE *nearFile; + FILE *outFile; +#endif +} AecmCore; + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_CreateCore() +// +// Allocates the memory needed by the AECM. The memory needs to be +// initialized separately using the WebRtcAecm_InitCore() function. +// Returns a pointer to the instance and a nullptr at failure. +AecmCore* WebRtcAecm_CreateCore(); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_InitCore(...) +// +// This function initializes the AECM instant created with +// WebRtcAecm_CreateCore() +// Input: +// - aecm : Pointer to the AECM instance +// - samplingFreq : Sampling Frequency +// +// Output: +// - aecm : Initialized instance +// +// Return value : 0 - Ok +// -1 - Error +// +int WebRtcAecm_InitCore(AecmCore* const aecm, int samplingFreq); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_FreeCore(...) +// +// This function releases the memory allocated by WebRtcAecm_CreateCore() +// Input: +// - aecm : Pointer to the AECM instance +// +void WebRtcAecm_FreeCore(AecmCore* aecm); + +int WebRtcAecm_Control(AecmCore* aecm, int delay, int nlpFlag); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_InitEchoPathCore(...) +// +// This function resets the echo channel adaptation with the specified channel. +// Input: +// - aecm : Pointer to the AECM instance +// - echo_path : Pointer to the data that should initialize the echo +// path +// +// Output: +// - aecm : Initialized instance +// +void WebRtcAecm_InitEchoPathCore(AecmCore* aecm, const int16_t* echo_path); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_ProcessFrame(...) +// +// This function processes frames and sends blocks to +// WebRtcAecm_ProcessBlock(...) +// +// Inputs: +// - aecm : Pointer to the AECM instance +// - farend : In buffer containing one frame of echo signal +// - nearendNoisy : In buffer containing one frame of nearend+echo signal +// without NS +// - nearendClean : In buffer containing one frame of nearend+echo signal +// with NS +// +// Output: +// - out : Out buffer, one frame of nearend signal : +// +// +int WebRtcAecm_ProcessFrame(AecmCore* aecm, + const int16_t* farend, + const int16_t* nearendNoisy, + const int16_t* nearendClean, + int16_t* out); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_ProcessBlock(...) +// +// This function is called for every block within one frame +// This function is called by WebRtcAecm_ProcessFrame(...) +// +// Inputs: +// - aecm : Pointer to the AECM instance +// - farend : In buffer containing one block of echo signal +// - nearendNoisy : In buffer containing one frame of nearend+echo signal +// without NS +// - nearendClean : In buffer containing one frame of nearend+echo signal +// with NS +// +// Output: +// - out : Out buffer, one block of nearend signal : +// +// +int WebRtcAecm_ProcessBlock(AecmCore* aecm, + const int16_t* farend, + const int16_t* nearendNoisy, + const int16_t* noisyClean, + int16_t* out); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_BufferFarFrame() +// +// Inserts a frame of data into farend buffer. +// +// Inputs: +// - aecm : Pointer to the AECM instance +// - farend : In buffer containing one frame of farend signal +// - farLen : Length of frame +// +void WebRtcAecm_BufferFarFrame(AecmCore* const aecm, + const int16_t* const farend, + const int farLen); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_FetchFarFrame() +// +// Read the farend buffer to account for known delay +// +// Inputs: +// - aecm : Pointer to the AECM instance +// - farend : In buffer containing one frame of farend signal +// - farLen : Length of frame +// - knownDelay : known delay +// +void WebRtcAecm_FetchFarFrame(AecmCore* const aecm, + int16_t* const farend, + const int farLen, + const int knownDelay); + +// All the functions below are intended to be private + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_UpdateFarHistory() +// +// Moves the pointer to the next entry and inserts |far_spectrum| and +// corresponding Q-domain in its buffer. +// +// Inputs: +// - self : Pointer to the delay estimation instance +// - far_spectrum : Pointer to the far end spectrum +// - far_q : Q-domain of far end spectrum +// +void WebRtcAecm_UpdateFarHistory(AecmCore* self, + uint16_t* far_spectrum, + int far_q); + +//////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_AlignedFarend() +// +// Returns a pointer to the far end spectrum aligned to current near end +// spectrum. The function WebRtc_DelayEstimatorProcessFix(...) should have been +// called before AlignedFarend(...). Otherwise, you get the pointer to the +// previous frame. The memory is only valid until the next call of +// WebRtc_DelayEstimatorProcessFix(...). +// +// Inputs: +// - self : Pointer to the AECM instance. +// - delay : Current delay estimate. +// +// Output: +// - far_q : The Q-domain of the aligned far end spectrum +// +// Return value: +// - far_spectrum : Pointer to the aligned far end spectrum +// NULL - Error +// +const uint16_t* WebRtcAecm_AlignedFarend(AecmCore* self, int* far_q, int delay); + +/////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_CalcSuppressionGain() +// +// This function calculates the suppression gain that is used in the +// Wiener filter. +// +// Inputs: +// - aecm : Pointer to the AECM instance. +// +// Return value: +// - supGain : Suppression gain with which to scale the noise +// level (Q14). +// +int16_t WebRtcAecm_CalcSuppressionGain(AecmCore* const aecm); + +/////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_CalcEnergies() +// +// This function calculates the log of energies for nearend, farend and +// estimated echoes. There is also an update of energy decision levels, +// i.e. internal VAD. +// +// Inputs: +// - aecm : Pointer to the AECM instance. +// - far_spectrum : Pointer to farend spectrum. +// - far_q : Q-domain of farend spectrum. +// - nearEner : Near end energy for current block in +// Q(aecm->dfaQDomain). +// +// Output: +// - echoEst : Estimated echo in Q(xfa_q+RESOLUTION_CHANNEL16). +// +void WebRtcAecm_CalcEnergies(AecmCore* aecm, + const uint16_t* far_spectrum, + const int16_t far_q, + const uint32_t nearEner, + int32_t* echoEst); + +/////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_CalcStepSize() +// +// This function calculates the step size used in channel estimation +// +// Inputs: +// - aecm : Pointer to the AECM instance. +// +// Return value: +// - mu : Stepsize in log2(), i.e. number of shifts. +// +int16_t WebRtcAecm_CalcStepSize(AecmCore* const aecm); + +/////////////////////////////////////////////////////////////////////////////// +// WebRtcAecm_UpdateChannel(...) +// +// This function performs channel estimation. +// NLMS and decision on channel storage. +// +// Inputs: +// - aecm : Pointer to the AECM instance. +// - far_spectrum : Absolute value of the farend signal in Q(far_q) +// - far_q : Q-domain of the farend signal +// - dfa : Absolute value of the nearend signal +// (Q[aecm->dfaQDomain]) +// - mu : NLMS step size. +// Input/Output: +// - echoEst : Estimated echo in Q(far_q+RESOLUTION_CHANNEL16). +// +void WebRtcAecm_UpdateChannel(AecmCore* aecm, + const uint16_t* far_spectrum, + const int16_t far_q, + const uint16_t* const dfa, + const int16_t mu, + int32_t* echoEst); + +extern const int16_t WebRtcAecm_kCosTable[]; +extern const int16_t WebRtcAecm_kSinTable[]; + +/////////////////////////////////////////////////////////////////////////////// +// Some function pointers, for internal functions shared by ARM NEON and +// generic C code. +// +typedef void (*CalcLinearEnergies)(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echoEst, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored); +extern CalcLinearEnergies WebRtcAecm_CalcLinearEnergies; + +typedef void (*StoreAdaptiveChannel)(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est); +extern StoreAdaptiveChannel WebRtcAecm_StoreAdaptiveChannel; + +typedef void (*ResetAdaptiveChannel)(AecmCore* aecm); +extern ResetAdaptiveChannel WebRtcAecm_ResetAdaptiveChannel; + +// For the above function pointers, functions for generic platforms are declared +// and defined as static in file aecm_core.c, while those for ARM Neon platforms +// are declared below and defined in file aecm_core_neon.c. +#if defined(WEBRTC_DETECT_NEON) || defined(WEBRTC_HAS_NEON) +void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored); + +void WebRtcAecm_StoreAdaptiveChannelNeon(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est); + +void WebRtcAecm_ResetAdaptiveChannelNeon(AecmCore* aecm); +#endif + +#if defined(MIPS32_LE) +void WebRtcAecm_CalcLinearEnergies_mips(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored); +#if defined(MIPS_DSP_R1_LE) +void WebRtcAecm_StoreAdaptiveChannel_mips(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est); + +void WebRtcAecm_ResetAdaptiveChannel_mips(AecmCore* aecm); +#endif +#endif + +#endif diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_c.c b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_c.c new file mode 100644 index 00000000..eb2bd918 --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_c.c @@ -0,0 +1,771 @@ +/* + * Copyright (c) 2013 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/aecm/aecm_core.h" + +#include <assert.h> +#include <stddef.h> +#include <stdlib.h> + +#include "webrtc/common_audio/ring_buffer.h" +#include "webrtc/common_audio/signal_processing/include/real_fft.h" +#include "webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h" +#include "webrtc/modules/audio_processing/utility/delay_estimator_wrapper.h" +#include "webrtc/system_wrappers/interface/compile_assert_c.h" +#include "webrtc/system_wrappers/interface/cpu_features_wrapper.h" +#include "webrtc/typedefs.h" + +// Square root of Hanning window in Q14. +#if defined(WEBRTC_DETECT_NEON) || defined(WEBRTC_HAS_NEON) +// Table is defined in an ARM assembly file. +extern const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END; +#else +static const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = { + 0, 399, 798, 1196, 1594, 1990, 2386, 2780, 3172, + 3562, 3951, 4337, 4720, 5101, 5478, 5853, 6224, + 6591, 6954, 7313, 7668, 8019, 8364, 8705, 9040, + 9370, 9695, 10013, 10326, 10633, 10933, 11227, 11514, + 11795, 12068, 12335, 12594, 12845, 13089, 13325, 13553, + 13773, 13985, 14189, 14384, 14571, 14749, 14918, 15079, + 15231, 15373, 15506, 15631, 15746, 15851, 15947, 16034, + 16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384 +}; +#endif + +#ifdef AECM_WITH_ABS_APPROX +//Q15 alpha = 0.99439986968132 const Factor for magnitude approximation +static const uint16_t kAlpha1 = 32584; +//Q15 beta = 0.12967166976970 const Factor for magnitude approximation +static const uint16_t kBeta1 = 4249; +//Q15 alpha = 0.94234827210087 const Factor for magnitude approximation +static const uint16_t kAlpha2 = 30879; +//Q15 beta = 0.33787806009150 const Factor for magnitude approximation +static const uint16_t kBeta2 = 11072; +//Q15 alpha = 0.82247698684306 const Factor for magnitude approximation +static const uint16_t kAlpha3 = 26951; +//Q15 beta = 0.57762063060713 const Factor for magnitude approximation +static const uint16_t kBeta3 = 18927; +#endif + +static const int16_t kNoiseEstQDomain = 15; +static const int16_t kNoiseEstIncCount = 5; + +static void ComfortNoise(AecmCore* aecm, + const uint16_t* dfa, + ComplexInt16* out, + const int16_t* lambda); + +static void WindowAndFFT(AecmCore* aecm, + int16_t* fft, + const int16_t* time_signal, + ComplexInt16* freq_signal, + int time_signal_scaling) { + int i = 0; + + // FFT of signal + for (i = 0; i < PART_LEN; i++) { + // Window time domain signal and insert into real part of + // transformation array |fft| + int16_t scaled_time_signal = time_signal[i] << time_signal_scaling; + fft[i] = (int16_t)((scaled_time_signal * WebRtcAecm_kSqrtHanning[i]) >> 14); + scaled_time_signal = time_signal[i + PART_LEN] << time_signal_scaling; + fft[PART_LEN + i] = (int16_t)(( + scaled_time_signal * WebRtcAecm_kSqrtHanning[PART_LEN - i]) >> 14); + } + + // Do forward FFT, then take only the first PART_LEN complex samples, + // and change signs of the imaginary parts. + WebRtcSpl_RealForwardFFT(aecm->real_fft, fft, (int16_t*)freq_signal); + for (i = 0; i < PART_LEN; i++) { + freq_signal[i].imag = -freq_signal[i].imag; + } +} + +static void InverseFFTAndWindow(AecmCore* aecm, + int16_t* fft, + ComplexInt16* efw, + int16_t* output, + const int16_t* nearendClean) { + int i, j, outCFFT; + int32_t tmp32no1; + // Reuse |efw| for the inverse FFT output after transferring + // the contents to |fft|. + int16_t* ifft_out = (int16_t*)efw; + + // Synthesis + for (i = 1, j = 2; i < PART_LEN; i += 1, j += 2) { + fft[j] = efw[i].real; + fft[j + 1] = -efw[i].imag; + } + fft[0] = efw[0].real; + fft[1] = -efw[0].imag; + + fft[PART_LEN2] = efw[PART_LEN].real; + fft[PART_LEN2 + 1] = -efw[PART_LEN].imag; + + // Inverse FFT. Keep outCFFT to scale the samples in the next block. + outCFFT = WebRtcSpl_RealInverseFFT(aecm->real_fft, fft, ifft_out); + for (i = 0; i < PART_LEN; i++) { + ifft_out[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( + ifft_out[i], WebRtcAecm_kSqrtHanning[i], 14); + tmp32no1 = WEBRTC_SPL_SHIFT_W32((int32_t)ifft_out[i], + outCFFT - aecm->dfaCleanQDomain); + output[i] = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, + tmp32no1 + aecm->outBuf[i], + WEBRTC_SPL_WORD16_MIN); + + tmp32no1 = (ifft_out[PART_LEN + i] * + WebRtcAecm_kSqrtHanning[PART_LEN - i]) >> 14; + tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1, + outCFFT - aecm->dfaCleanQDomain); + aecm->outBuf[i] = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, + tmp32no1, + WEBRTC_SPL_WORD16_MIN); + } + + // Copy the current block to the old position + // (aecm->outBuf is shifted elsewhere) + memcpy(aecm->xBuf, aecm->xBuf + PART_LEN, sizeof(int16_t) * PART_LEN); + memcpy(aecm->dBufNoisy, + aecm->dBufNoisy + PART_LEN, + sizeof(int16_t) * PART_LEN); + if (nearendClean != NULL) + { + memcpy(aecm->dBufClean, + aecm->dBufClean + PART_LEN, + sizeof(int16_t) * PART_LEN); + } +} + +// Transforms a time domain signal into the frequency domain, outputting the +// complex valued signal, absolute value and sum of absolute values. +// +// time_signal [in] Pointer to time domain signal +// freq_signal_real [out] Pointer to real part of frequency domain array +// freq_signal_imag [out] Pointer to imaginary part of frequency domain +// array +// freq_signal_abs [out] Pointer to absolute value of frequency domain +// array +// freq_signal_sum_abs [out] Pointer to the sum of all absolute values in +// the frequency domain array +// return value The Q-domain of current frequency values +// +static int TimeToFrequencyDomain(AecmCore* aecm, + const int16_t* time_signal, + ComplexInt16* freq_signal, + uint16_t* freq_signal_abs, + uint32_t* freq_signal_sum_abs) { + int i = 0; + int time_signal_scaling = 0; + + int32_t tmp32no1 = 0; + int32_t tmp32no2 = 0; + + // In fft_buf, +16 for 32-byte alignment. + int16_t fft_buf[PART_LEN4 + 16]; + int16_t *fft = (int16_t *) (((uintptr_t) fft_buf + 31) & ~31); + + int16_t tmp16no1; +#ifndef WEBRTC_ARCH_ARM_V7 + int16_t tmp16no2; +#endif +#ifdef AECM_WITH_ABS_APPROX + int16_t max_value = 0; + int16_t min_value = 0; + uint16_t alpha = 0; + uint16_t beta = 0; +#endif + +#ifdef AECM_DYNAMIC_Q + tmp16no1 = WebRtcSpl_MaxAbsValueW16(time_signal, PART_LEN2); + time_signal_scaling = WebRtcSpl_NormW16(tmp16no1); +#endif + + WindowAndFFT(aecm, fft, time_signal, freq_signal, time_signal_scaling); + + // Extract imaginary and real part, calculate the magnitude for + // all frequency bins + freq_signal[0].imag = 0; + freq_signal[PART_LEN].imag = 0; + freq_signal_abs[0] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[0].real); + freq_signal_abs[PART_LEN] = (uint16_t)WEBRTC_SPL_ABS_W16( + freq_signal[PART_LEN].real); + (*freq_signal_sum_abs) = (uint32_t)(freq_signal_abs[0]) + + (uint32_t)(freq_signal_abs[PART_LEN]); + + for (i = 1; i < PART_LEN; i++) + { + if (freq_signal[i].real == 0) + { + freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[i].imag); + } + else if (freq_signal[i].imag == 0) + { + freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[i].real); + } + else + { + // Approximation for magnitude of complex fft output + // magn = sqrt(real^2 + imag^2) + // magn ~= alpha * max(|imag|,|real|) + beta * min(|imag|,|real|) + // + // The parameters alpha and beta are stored in Q15 + +#ifdef AECM_WITH_ABS_APPROX + tmp16no1 = WEBRTC_SPL_ABS_W16(freq_signal[i].real); + tmp16no2 = WEBRTC_SPL_ABS_W16(freq_signal[i].imag); + + if(tmp16no1 > tmp16no2) + { + max_value = tmp16no1; + min_value = tmp16no2; + } else + { + max_value = tmp16no2; + min_value = tmp16no1; + } + + // Magnitude in Q(-6) + if ((max_value >> 2) > min_value) + { + alpha = kAlpha1; + beta = kBeta1; + } else if ((max_value >> 1) > min_value) + { + alpha = kAlpha2; + beta = kBeta2; + } else + { + alpha = kAlpha3; + beta = kBeta3; + } + tmp16no1 = (int16_t)((max_value * alpha) >> 15); + tmp16no2 = (int16_t)((min_value * beta) >> 15); + freq_signal_abs[i] = (uint16_t)tmp16no1 + (uint16_t)tmp16no2; +#else +#ifdef WEBRTC_ARCH_ARM_V7 + __asm __volatile( + "smulbb %[tmp32no1], %[real], %[real]\n\t" + "smlabb %[tmp32no2], %[imag], %[imag], %[tmp32no1]\n\t" + :[tmp32no1]"+&r"(tmp32no1), + [tmp32no2]"=r"(tmp32no2) + :[real]"r"(freq_signal[i].real), + [imag]"r"(freq_signal[i].imag) + ); +#else + tmp16no1 = WEBRTC_SPL_ABS_W16(freq_signal[i].real); + tmp16no2 = WEBRTC_SPL_ABS_W16(freq_signal[i].imag); + tmp32no1 = tmp16no1 * tmp16no1; + tmp32no2 = tmp16no2 * tmp16no2; + tmp32no2 = WebRtcSpl_AddSatW32(tmp32no1, tmp32no2); +#endif // WEBRTC_ARCH_ARM_V7 + tmp32no1 = WebRtcSpl_SqrtFloor(tmp32no2); + + freq_signal_abs[i] = (uint16_t)tmp32no1; +#endif // AECM_WITH_ABS_APPROX + } + (*freq_signal_sum_abs) += (uint32_t)freq_signal_abs[i]; + } + + return time_signal_scaling; +} + +int WebRtcAecm_ProcessBlock(AecmCore* aecm, + const int16_t* farend, + const int16_t* nearendNoisy, + const int16_t* nearendClean, + int16_t* output) { + int i; + + uint32_t xfaSum; + uint32_t dfaNoisySum; + uint32_t dfaCleanSum; + uint32_t echoEst32Gained; + uint32_t tmpU32; + + int32_t tmp32no1; + + uint16_t xfa[PART_LEN1]; + uint16_t dfaNoisy[PART_LEN1]; + uint16_t dfaClean[PART_LEN1]; + uint16_t* ptrDfaClean = dfaClean; + const uint16_t* far_spectrum_ptr = NULL; + + // 32 byte aligned buffers (with +8 or +16). + // TODO(kma): define fft with ComplexInt16. + int16_t fft_buf[PART_LEN4 + 2 + 16]; // +2 to make a loop safe. + int32_t echoEst32_buf[PART_LEN1 + 8]; + int32_t dfw_buf[PART_LEN2 + 8]; + int32_t efw_buf[PART_LEN2 + 8]; + + int16_t* fft = (int16_t*) (((uintptr_t) fft_buf + 31) & ~ 31); + int32_t* echoEst32 = (int32_t*) (((uintptr_t) echoEst32_buf + 31) & ~ 31); + ComplexInt16* dfw = (ComplexInt16*)(((uintptr_t)dfw_buf + 31) & ~31); + ComplexInt16* efw = (ComplexInt16*)(((uintptr_t)efw_buf + 31) & ~31); + + int16_t hnl[PART_LEN1]; + int16_t numPosCoef = 0; + int16_t nlpGain = ONE_Q14; + int delay; + int16_t tmp16no1; + int16_t tmp16no2; + int16_t mu; + int16_t supGain; + int16_t zeros32, zeros16; + int16_t zerosDBufNoisy, zerosDBufClean, zerosXBuf; + int far_q; + int16_t resolutionDiff, qDomainDiff, dfa_clean_q_domain_diff; + + const int kMinPrefBand = 4; + const int kMaxPrefBand = 24; + int32_t avgHnl32 = 0; + + // Determine startup state. There are three states: + // (0) the first CONV_LEN blocks + // (1) another CONV_LEN blocks + // (2) the rest + + if (aecm->startupState < 2) + { + aecm->startupState = (aecm->totCount >= CONV_LEN) + + (aecm->totCount >= CONV_LEN2); + } + // END: Determine startup state + + // Buffer near and far end signals + memcpy(aecm->xBuf + PART_LEN, farend, sizeof(int16_t) * PART_LEN); + memcpy(aecm->dBufNoisy + PART_LEN, nearendNoisy, sizeof(int16_t) * PART_LEN); + if (nearendClean != NULL) + { + memcpy(aecm->dBufClean + PART_LEN, + nearendClean, + sizeof(int16_t) * PART_LEN); + } + + // Transform far end signal from time domain to frequency domain. + far_q = TimeToFrequencyDomain(aecm, + aecm->xBuf, + dfw, + xfa, + &xfaSum); + + // Transform noisy near end signal from time domain to frequency domain. + zerosDBufNoisy = TimeToFrequencyDomain(aecm, + aecm->dBufNoisy, + dfw, + dfaNoisy, + &dfaNoisySum); + aecm->dfaNoisyQDomainOld = aecm->dfaNoisyQDomain; + aecm->dfaNoisyQDomain = (int16_t)zerosDBufNoisy; + + + if (nearendClean == NULL) + { + ptrDfaClean = dfaNoisy; + aecm->dfaCleanQDomainOld = aecm->dfaNoisyQDomainOld; + aecm->dfaCleanQDomain = aecm->dfaNoisyQDomain; + dfaCleanSum = dfaNoisySum; + } else + { + // Transform clean near end signal from time domain to frequency domain. + zerosDBufClean = TimeToFrequencyDomain(aecm, + aecm->dBufClean, + dfw, + dfaClean, + &dfaCleanSum); + aecm->dfaCleanQDomainOld = aecm->dfaCleanQDomain; + aecm->dfaCleanQDomain = (int16_t)zerosDBufClean; + } + + // Get the delay + // Save far-end history and estimate delay + WebRtcAecm_UpdateFarHistory(aecm, xfa, far_q); + if (WebRtc_AddFarSpectrumFix(aecm->delay_estimator_farend, + xfa, + PART_LEN1, + far_q) == -1) { + return -1; + } + delay = WebRtc_DelayEstimatorProcessFix(aecm->delay_estimator, + dfaNoisy, + PART_LEN1, + zerosDBufNoisy); + if (delay == -1) + { + return -1; + } + else if (delay == -2) + { + // If the delay is unknown, we assume zero. + // NOTE: this will have to be adjusted if we ever add lookahead. + delay = 0; + } + + if (aecm->fixedDelay >= 0) + { + // Use fixed delay + delay = aecm->fixedDelay; + } + + // Get aligned far end spectrum + far_spectrum_ptr = WebRtcAecm_AlignedFarend(aecm, &far_q, delay); + zerosXBuf = (int16_t) far_q; + if (far_spectrum_ptr == NULL) + { + return -1; + } + + // Calculate log(energy) and update energy threshold levels + WebRtcAecm_CalcEnergies(aecm, + far_spectrum_ptr, + zerosXBuf, + dfaNoisySum, + echoEst32); + + // Calculate stepsize + mu = WebRtcAecm_CalcStepSize(aecm); + + // Update counters + aecm->totCount++; + + // This is the channel estimation algorithm. + // It is base on NLMS but has a variable step length, + // which was calculated above. + WebRtcAecm_UpdateChannel(aecm, + far_spectrum_ptr, + zerosXBuf, + dfaNoisy, + mu, + echoEst32); + supGain = WebRtcAecm_CalcSuppressionGain(aecm); + + + // Calculate Wiener filter hnl[] + for (i = 0; i < PART_LEN1; i++) + { + // Far end signal through channel estimate in Q8 + // How much can we shift right to preserve resolution + tmp32no1 = echoEst32[i] - aecm->echoFilt[i]; + aecm->echoFilt[i] += (tmp32no1 * 50) >> 8; + + zeros32 = WebRtcSpl_NormW32(aecm->echoFilt[i]) + 1; + zeros16 = WebRtcSpl_NormW16(supGain) + 1; + if (zeros32 + zeros16 > 16) + { + // Multiplication is safe + // Result in + // Q(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN+ + // aecm->xfaQDomainBuf[diff]) + echoEst32Gained = WEBRTC_SPL_UMUL_32_16((uint32_t)aecm->echoFilt[i], + (uint16_t)supGain); + resolutionDiff = 14 - RESOLUTION_CHANNEL16 - RESOLUTION_SUPGAIN; + resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf); + } else + { + tmp16no1 = 17 - zeros32 - zeros16; + resolutionDiff = 14 + tmp16no1 - RESOLUTION_CHANNEL16 - + RESOLUTION_SUPGAIN; + resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf); + if (zeros32 > tmp16no1) + { + echoEst32Gained = WEBRTC_SPL_UMUL_32_16((uint32_t)aecm->echoFilt[i], + supGain >> tmp16no1); + } else + { + // Result in Q-(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN-16) + echoEst32Gained = (aecm->echoFilt[i] >> tmp16no1) * supGain; + } + } + + zeros16 = WebRtcSpl_NormW16(aecm->nearFilt[i]); + assert(zeros16 >= 0); // |zeros16| is a norm, hence non-negative. + dfa_clean_q_domain_diff = aecm->dfaCleanQDomain - aecm->dfaCleanQDomainOld; + if (zeros16 < dfa_clean_q_domain_diff && aecm->nearFilt[i]) { + tmp16no1 = aecm->nearFilt[i] << zeros16; + qDomainDiff = zeros16 - dfa_clean_q_domain_diff; + tmp16no2 = ptrDfaClean[i] >> -qDomainDiff; + } else { + tmp16no1 = dfa_clean_q_domain_diff < 0 + ? aecm->nearFilt[i] >> -dfa_clean_q_domain_diff + : aecm->nearFilt[i] << dfa_clean_q_domain_diff; + qDomainDiff = 0; + tmp16no2 = ptrDfaClean[i]; + } + tmp32no1 = (int32_t)(tmp16no2 - tmp16no1); + tmp16no2 = (int16_t)(tmp32no1 >> 4); + tmp16no2 += tmp16no1; + zeros16 = WebRtcSpl_NormW16(tmp16no2); + if ((tmp16no2) & (-qDomainDiff > zeros16)) { + aecm->nearFilt[i] = WEBRTC_SPL_WORD16_MAX; + } else { + aecm->nearFilt[i] = qDomainDiff < 0 ? tmp16no2 << -qDomainDiff + : tmp16no2 >> qDomainDiff; + } + + // Wiener filter coefficients, resulting hnl in Q14 + if (echoEst32Gained == 0) + { + hnl[i] = ONE_Q14; + } else if (aecm->nearFilt[i] == 0) + { + hnl[i] = 0; + } else + { + // Multiply the suppression gain + // Rounding + echoEst32Gained += (uint32_t)(aecm->nearFilt[i] >> 1); + tmpU32 = WebRtcSpl_DivU32U16(echoEst32Gained, + (uint16_t)aecm->nearFilt[i]); + + // Current resolution is + // Q-(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN- max(0,17-zeros16- zeros32)) + // Make sure we are in Q14 + tmp32no1 = (int32_t)WEBRTC_SPL_SHIFT_W32(tmpU32, resolutionDiff); + if (tmp32no1 > ONE_Q14) + { + hnl[i] = 0; + } else if (tmp32no1 < 0) + { + hnl[i] = ONE_Q14; + } else + { + // 1-echoEst/dfa + hnl[i] = ONE_Q14 - (int16_t)tmp32no1; + if (hnl[i] < 0) + { + hnl[i] = 0; + } + } + } + if (hnl[i]) + { + numPosCoef++; + } + } + // Only in wideband. Prevent the gain in upper band from being larger than + // in lower band. + if (aecm->mult == 2) + { + // TODO(bjornv): Investigate if the scaling of hnl[i] below can cause + // speech distortion in double-talk. + for (i = 0; i < PART_LEN1; i++) + { + hnl[i] = (int16_t)((hnl[i] * hnl[i]) >> 14); + } + + for (i = kMinPrefBand; i <= kMaxPrefBand; i++) + { + avgHnl32 += (int32_t)hnl[i]; + } + assert(kMaxPrefBand - kMinPrefBand + 1 > 0); + avgHnl32 /= (kMaxPrefBand - kMinPrefBand + 1); + + for (i = kMaxPrefBand; i < PART_LEN1; i++) + { + if (hnl[i] > (int16_t)avgHnl32) + { + hnl[i] = (int16_t)avgHnl32; + } + } + } + + // Calculate NLP gain, result is in Q14 + if (aecm->nlpFlag) + { + for (i = 0; i < PART_LEN1; i++) + { + // Truncate values close to zero and one. + if (hnl[i] > NLP_COMP_HIGH) + { + hnl[i] = ONE_Q14; + } else if (hnl[i] < NLP_COMP_LOW) + { + hnl[i] = 0; + } + + // Remove outliers + if (numPosCoef < 3) + { + nlpGain = 0; + } else + { + nlpGain = ONE_Q14; + } + + // NLP + if ((hnl[i] == ONE_Q14) && (nlpGain == ONE_Q14)) + { + hnl[i] = ONE_Q14; + } else + { + hnl[i] = (int16_t)((hnl[i] * nlpGain) >> 14); + } + + // multiply with Wiener coefficients + efw[i].real = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].real, + hnl[i], 14)); + efw[i].imag = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].imag, + hnl[i], 14)); + } + } + else + { + // multiply with Wiener coefficients + for (i = 0; i < PART_LEN1; i++) + { + efw[i].real = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].real, + hnl[i], 14)); + efw[i].imag = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].imag, + hnl[i], 14)); + } + } + + if (aecm->cngMode == AecmTrue) + { + ComfortNoise(aecm, ptrDfaClean, efw, hnl); + } + + InverseFFTAndWindow(aecm, fft, efw, output, nearendClean); + + return 0; +} + +static void ComfortNoise(AecmCore* aecm, + const uint16_t* dfa, + ComplexInt16* out, + const int16_t* lambda) { + int16_t i; + int16_t tmp16; + int32_t tmp32; + + int16_t randW16[PART_LEN]; + int16_t uReal[PART_LEN1]; + int16_t uImag[PART_LEN1]; + int32_t outLShift32; + int16_t noiseRShift16[PART_LEN1]; + + int16_t shiftFromNearToNoise = kNoiseEstQDomain - aecm->dfaCleanQDomain; + int16_t minTrackShift; + + assert(shiftFromNearToNoise >= 0); + assert(shiftFromNearToNoise < 16); + + if (aecm->noiseEstCtr < 100) + { + // Track the minimum more quickly initially. + aecm->noiseEstCtr++; + minTrackShift = 6; + } else + { + minTrackShift = 9; + } + + // Estimate noise power. + for (i = 0; i < PART_LEN1; i++) + { + // Shift to the noise domain. + tmp32 = (int32_t)dfa[i]; + outLShift32 = tmp32 << shiftFromNearToNoise; + + if (outLShift32 < aecm->noiseEst[i]) + { + // Reset "too low" counter + aecm->noiseEstTooLowCtr[i] = 0; + // Track the minimum. + if (aecm->noiseEst[i] < (1 << minTrackShift)) + { + // For small values, decrease noiseEst[i] every + // |kNoiseEstIncCount| block. The regular approach below can not + // go further down due to truncation. + aecm->noiseEstTooHighCtr[i]++; + if (aecm->noiseEstTooHighCtr[i] >= kNoiseEstIncCount) + { + aecm->noiseEst[i]--; + aecm->noiseEstTooHighCtr[i] = 0; // Reset the counter + } + } + else + { + aecm->noiseEst[i] -= ((aecm->noiseEst[i] - outLShift32) + >> minTrackShift); + } + } else + { + // Reset "too high" counter + aecm->noiseEstTooHighCtr[i] = 0; + // Ramp slowly upwards until we hit the minimum again. + if ((aecm->noiseEst[i] >> 19) > 0) + { + // Avoid overflow. + // Multiplication with 2049 will cause wrap around. Scale + // down first and then multiply + aecm->noiseEst[i] >>= 11; + aecm->noiseEst[i] *= 2049; + } + else if ((aecm->noiseEst[i] >> 11) > 0) + { + // Large enough for relative increase + aecm->noiseEst[i] *= 2049; + aecm->noiseEst[i] >>= 11; + } + else + { + // Make incremental increases based on size every + // |kNoiseEstIncCount| block + aecm->noiseEstTooLowCtr[i]++; + if (aecm->noiseEstTooLowCtr[i] >= kNoiseEstIncCount) + { + aecm->noiseEst[i] += (aecm->noiseEst[i] >> 9) + 1; + aecm->noiseEstTooLowCtr[i] = 0; // Reset counter + } + } + } + } + + for (i = 0; i < PART_LEN1; i++) + { + tmp32 = aecm->noiseEst[i] >> shiftFromNearToNoise; + if (tmp32 > 32767) + { + tmp32 = 32767; + aecm->noiseEst[i] = tmp32 << shiftFromNearToNoise; + } + noiseRShift16[i] = (int16_t)tmp32; + + tmp16 = ONE_Q14 - lambda[i]; + noiseRShift16[i] = (int16_t)((tmp16 * noiseRShift16[i]) >> 14); + } + + // Generate a uniform random array on [0 2^15-1]. + WebRtcSpl_RandUArray(randW16, PART_LEN, &aecm->seed); + + // Generate noise according to estimated energy. + uReal[0] = 0; // Reject LF noise. + uImag[0] = 0; + for (i = 1; i < PART_LEN1; i++) + { + // Get a random index for the cos and sin tables over [0 359]. + tmp16 = (int16_t)((359 * randW16[i - 1]) >> 15); + + // Tables are in Q13. + uReal[i] = (int16_t)((noiseRShift16[i] * WebRtcAecm_kCosTable[tmp16]) >> + 13); + uImag[i] = (int16_t)((-noiseRShift16[i] * WebRtcAecm_kSinTable[tmp16]) >> + 13); + } + uImag[PART_LEN] = 0; + + for (i = 0; i < PART_LEN1; i++) + { + out[i].real = WebRtcSpl_AddSatW16(out[i].real, uReal[i]); + out[i].imag = WebRtcSpl_AddSatW16(out[i].imag, uImag[i]); + } +} + diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_mips.c b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_mips.c new file mode 100644 index 00000000..3c2343a8 --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_mips.c @@ -0,0 +1,1566 @@ +/* + * Copyright (c) 2013 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/aecm/aecm_core.h" + +#include <assert.h> + +#include "webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h" +#include "webrtc/modules/audio_processing/utility/delay_estimator_wrapper.h" + +static const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = { + 0, 399, 798, 1196, 1594, 1990, 2386, 2780, 3172, + 3562, 3951, 4337, 4720, 5101, 5478, 5853, 6224, + 6591, 6954, 7313, 7668, 8019, 8364, 8705, 9040, + 9370, 9695, 10013, 10326, 10633, 10933, 11227, 11514, + 11795, 12068, 12335, 12594, 12845, 13089, 13325, 13553, + 13773, 13985, 14189, 14384, 14571, 14749, 14918, 15079, + 15231, 15373, 15506, 15631, 15746, 15851, 15947, 16034, + 16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384 +}; + +static const int16_t kNoiseEstQDomain = 15; +static const int16_t kNoiseEstIncCount = 5; + +static int16_t coefTable[] = { + 0, 4, 256, 260, 128, 132, 384, 388, + 64, 68, 320, 324, 192, 196, 448, 452, + 32, 36, 288, 292, 160, 164, 416, 420, + 96, 100, 352, 356, 224, 228, 480, 484, + 16, 20, 272, 276, 144, 148, 400, 404, + 80, 84, 336, 340, 208, 212, 464, 468, + 48, 52, 304, 308, 176, 180, 432, 436, + 112, 116, 368, 372, 240, 244, 496, 500, + 8, 12, 264, 268, 136, 140, 392, 396, + 72, 76, 328, 332, 200, 204, 456, 460, + 40, 44, 296, 300, 168, 172, 424, 428, + 104, 108, 360, 364, 232, 236, 488, 492, + 24, 28, 280, 284, 152, 156, 408, 412, + 88, 92, 344, 348, 216, 220, 472, 476, + 56, 60, 312, 316, 184, 188, 440, 444, + 120, 124, 376, 380, 248, 252, 504, 508 +}; + +static int16_t coefTable_ifft[] = { + 0, 512, 256, 508, 128, 252, 384, 380, + 64, 124, 320, 444, 192, 188, 448, 316, + 32, 60, 288, 476, 160, 220, 416, 348, + 96, 92, 352, 412, 224, 156, 480, 284, + 16, 28, 272, 492, 144, 236, 400, 364, + 80, 108, 336, 428, 208, 172, 464, 300, + 48, 44, 304, 460, 176, 204, 432, 332, + 112, 76, 368, 396, 240, 140, 496, 268, + 8, 12, 264, 500, 136, 244, 392, 372, + 72, 116, 328, 436, 200, 180, 456, 308, + 40, 52, 296, 468, 168, 212, 424, 340, + 104, 84, 360, 404, 232, 148, 488, 276, + 24, 20, 280, 484, 152, 228, 408, 356, + 88, 100, 344, 420, 216, 164, 472, 292, + 56, 36, 312, 452, 184, 196, 440, 324, + 120, 68, 376, 388, 248, 132, 504, 260 +}; + +static void ComfortNoise(AecmCore* aecm, + const uint16_t* dfa, + ComplexInt16* out, + const int16_t* lambda); + +static void WindowAndFFT(AecmCore* aecm, + int16_t* fft, + const int16_t* time_signal, + ComplexInt16* freq_signal, + int time_signal_scaling) { + int i, j; + int32_t tmp1, tmp2, tmp3, tmp4; + int16_t* pfrfi; + ComplexInt16* pfreq_signal; + int16_t f_coef, s_coef; + int32_t load_ptr, store_ptr1, store_ptr2, shift, shift1; + int32_t hann, hann1, coefs; + + memset(fft, 0, sizeof(int16_t) * PART_LEN4); + + // FFT of signal + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[shift], %[time_signal_scaling], -14 \n\t" + "addiu %[i], $zero, 64 \n\t" + "addiu %[load_ptr], %[time_signal], 0 \n\t" + "addiu %[hann], %[hanning], 0 \n\t" + "addiu %[hann1], %[hanning], 128 \n\t" + "addiu %[coefs], %[coefTable], 0 \n\t" + "bltz %[shift], 2f \n\t" + " negu %[shift1], %[shift] \n\t" + "1: \n\t" + "lh %[tmp1], 0(%[load_ptr]) \n\t" + "lh %[tmp2], 0(%[hann]) \n\t" + "lh %[tmp3], 128(%[load_ptr]) \n\t" + "lh %[tmp4], 0(%[hann1]) \n\t" + "addiu %[i], %[i], -1 \n\t" + "mul %[tmp1], %[tmp1], %[tmp2] \n\t" + "mul %[tmp3], %[tmp3], %[tmp4] \n\t" + "lh %[f_coef], 0(%[coefs]) \n\t" + "lh %[s_coef], 2(%[coefs]) \n\t" + "addiu %[load_ptr], %[load_ptr], 2 \n\t" + "addiu %[hann], %[hann], 2 \n\t" + "addiu %[hann1], %[hann1], -2 \n\t" + "addu %[store_ptr1], %[fft], %[f_coef] \n\t" + "addu %[store_ptr2], %[fft], %[s_coef] \n\t" + "sllv %[tmp1], %[tmp1], %[shift] \n\t" + "sllv %[tmp3], %[tmp3], %[shift] \n\t" + "sh %[tmp1], 0(%[store_ptr1]) \n\t" + "sh %[tmp3], 0(%[store_ptr2]) \n\t" + "bgtz %[i], 1b \n\t" + " addiu %[coefs], %[coefs], 4 \n\t" + "b 3f \n\t" + " nop \n\t" + "2: \n\t" + "lh %[tmp1], 0(%[load_ptr]) \n\t" + "lh %[tmp2], 0(%[hann]) \n\t" + "lh %[tmp3], 128(%[load_ptr]) \n\t" + "lh %[tmp4], 0(%[hann1]) \n\t" + "addiu %[i], %[i], -1 \n\t" + "mul %[tmp1], %[tmp1], %[tmp2] \n\t" + "mul %[tmp3], %[tmp3], %[tmp4] \n\t" + "lh %[f_coef], 0(%[coefs]) \n\t" + "lh %[s_coef], 2(%[coefs]) \n\t" + "addiu %[load_ptr], %[load_ptr], 2 \n\t" + "addiu %[hann], %[hann], 2 \n\t" + "addiu %[hann1], %[hann1], -2 \n\t" + "addu %[store_ptr1], %[fft], %[f_coef] \n\t" + "addu %[store_ptr2], %[fft], %[s_coef] \n\t" + "srav %[tmp1], %[tmp1], %[shift1] \n\t" + "srav %[tmp3], %[tmp3], %[shift1] \n\t" + "sh %[tmp1], 0(%[store_ptr1]) \n\t" + "sh %[tmp3], 0(%[store_ptr2]) \n\t" + "bgtz %[i], 2b \n\t" + " addiu %[coefs], %[coefs], 4 \n\t" + "3: \n\t" + ".set pop \n\t" + : [load_ptr] "=&r" (load_ptr), [shift] "=&r" (shift), [hann] "=&r" (hann), + [hann1] "=&r" (hann1), [shift1] "=&r" (shift1), [coefs] "=&r" (coefs), + [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2), [tmp3] "=&r" (tmp3), + [tmp4] "=&r" (tmp4), [i] "=&r" (i), [f_coef] "=&r" (f_coef), + [s_coef] "=&r" (s_coef), [store_ptr1] "=&r" (store_ptr1), + [store_ptr2] "=&r" (store_ptr2) + : [time_signal] "r" (time_signal), [coefTable] "r" (coefTable), + [time_signal_scaling] "r" (time_signal_scaling), + [hanning] "r" (WebRtcAecm_kSqrtHanning), [fft] "r" (fft) + : "memory", "hi", "lo" + ); + + WebRtcSpl_ComplexFFT(fft, PART_LEN_SHIFT, 1); + pfrfi = fft; + pfreq_signal = freq_signal; + + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[j], $zero, 128 \n\t" + "1: \n\t" + "lh %[tmp1], 0(%[pfrfi]) \n\t" + "lh %[tmp2], 2(%[pfrfi]) \n\t" + "lh %[tmp3], 4(%[pfrfi]) \n\t" + "lh %[tmp4], 6(%[pfrfi]) \n\t" + "subu %[tmp2], $zero, %[tmp2] \n\t" + "sh %[tmp1], 0(%[pfreq_signal]) \n\t" + "sh %[tmp2], 2(%[pfreq_signal]) \n\t" + "subu %[tmp4], $zero, %[tmp4] \n\t" + "sh %[tmp3], 4(%[pfreq_signal]) \n\t" + "sh %[tmp4], 6(%[pfreq_signal]) \n\t" + "lh %[tmp1], 8(%[pfrfi]) \n\t" + "lh %[tmp2], 10(%[pfrfi]) \n\t" + "lh %[tmp3], 12(%[pfrfi]) \n\t" + "lh %[tmp4], 14(%[pfrfi]) \n\t" + "addiu %[j], %[j], -8 \n\t" + "subu %[tmp2], $zero, %[tmp2] \n\t" + "sh %[tmp1], 8(%[pfreq_signal]) \n\t" + "sh %[tmp2], 10(%[pfreq_signal]) \n\t" + "subu %[tmp4], $zero, %[tmp4] \n\t" + "sh %[tmp3], 12(%[pfreq_signal]) \n\t" + "sh %[tmp4], 14(%[pfreq_signal]) \n\t" + "addiu %[pfreq_signal], %[pfreq_signal], 16 \n\t" + "bgtz %[j], 1b \n\t" + " addiu %[pfrfi], %[pfrfi], 16 \n\t" + ".set pop \n\t" + : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2), [tmp3] "=&r" (tmp3), + [j] "=&r" (j), [pfrfi] "+r" (pfrfi), [pfreq_signal] "+r" (pfreq_signal), + [tmp4] "=&r" (tmp4) + : + : "memory" + ); +} + +static void InverseFFTAndWindow(AecmCore* aecm, + int16_t* fft, + ComplexInt16* efw, + int16_t* output, + const int16_t* nearendClean) { + int i, outCFFT; + int32_t tmp1, tmp2, tmp3, tmp4, tmp_re, tmp_im; + int16_t* pcoefTable_ifft = coefTable_ifft; + int16_t* pfft = fft; + int16_t* ppfft = fft; + ComplexInt16* pefw = efw; + int32_t out_aecm; + int16_t* paecm_buf = aecm->outBuf; + const int16_t* p_kSqrtHanning = WebRtcAecm_kSqrtHanning; + const int16_t* pp_kSqrtHanning = &WebRtcAecm_kSqrtHanning[PART_LEN]; + int16_t* output1 = output; + + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[i], $zero, 64 \n\t" + "1: \n\t" + "lh %[tmp1], 0(%[pcoefTable_ifft]) \n\t" + "lh %[tmp2], 2(%[pcoefTable_ifft]) \n\t" + "lh %[tmp_re], 0(%[pefw]) \n\t" + "lh %[tmp_im], 2(%[pefw]) \n\t" + "addu %[pfft], %[fft], %[tmp2] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "addu %[pfft], %[fft], %[tmp1] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "subu %[tmp_im], $zero, %[tmp_im] \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "lh %[tmp1], 4(%[pcoefTable_ifft]) \n\t" + "lh %[tmp2], 6(%[pcoefTable_ifft]) \n\t" + "lh %[tmp_re], 4(%[pefw]) \n\t" + "lh %[tmp_im], 6(%[pefw]) \n\t" + "addu %[pfft], %[fft], %[tmp2] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "addu %[pfft], %[fft], %[tmp1] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "subu %[tmp_im], $zero, %[tmp_im] \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "lh %[tmp1], 8(%[pcoefTable_ifft]) \n\t" + "lh %[tmp2], 10(%[pcoefTable_ifft]) \n\t" + "lh %[tmp_re], 8(%[pefw]) \n\t" + "lh %[tmp_im], 10(%[pefw]) \n\t" + "addu %[pfft], %[fft], %[tmp2] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "addu %[pfft], %[fft], %[tmp1] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "subu %[tmp_im], $zero, %[tmp_im] \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "lh %[tmp1], 12(%[pcoefTable_ifft]) \n\t" + "lh %[tmp2], 14(%[pcoefTable_ifft]) \n\t" + "lh %[tmp_re], 12(%[pefw]) \n\t" + "lh %[tmp_im], 14(%[pefw]) \n\t" + "addu %[pfft], %[fft], %[tmp2] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "addu %[pfft], %[fft], %[tmp1] \n\t" + "sh %[tmp_re], 0(%[pfft]) \n\t" + "subu %[tmp_im], $zero, %[tmp_im] \n\t" + "sh %[tmp_im], 2(%[pfft]) \n\t" + "addiu %[pcoefTable_ifft], %[pcoefTable_ifft], 16 \n\t" + "addiu %[i], %[i], -4 \n\t" + "bgtz %[i], 1b \n\t" + " addiu %[pefw], %[pefw], 16 \n\t" + ".set pop \n\t" + : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2), [pfft] "+r" (pfft), + [i] "=&r" (i), [tmp_re] "=&r" (tmp_re), [tmp_im] "=&r" (tmp_im), + [pefw] "+r" (pefw), [pcoefTable_ifft] "+r" (pcoefTable_ifft), + [fft] "+r" (fft) + : + : "memory" + ); + + fft[2] = efw[PART_LEN].real; + fft[3] = -efw[PART_LEN].imag; + + outCFFT = WebRtcSpl_ComplexIFFT(fft, PART_LEN_SHIFT, 1); + pfft = fft; + + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[i], $zero, 128 \n\t" + "1: \n\t" + "lh %[tmp1], 0(%[ppfft]) \n\t" + "lh %[tmp2], 4(%[ppfft]) \n\t" + "lh %[tmp3], 8(%[ppfft]) \n\t" + "lh %[tmp4], 12(%[ppfft]) \n\t" + "addiu %[i], %[i], -4 \n\t" + "sh %[tmp1], 0(%[pfft]) \n\t" + "sh %[tmp2], 2(%[pfft]) \n\t" + "sh %[tmp3], 4(%[pfft]) \n\t" + "sh %[tmp4], 6(%[pfft]) \n\t" + "addiu %[ppfft], %[ppfft], 16 \n\t" + "bgtz %[i], 1b \n\t" + " addiu %[pfft], %[pfft], 8 \n\t" + ".set pop \n\t" + : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2), [pfft] "+r" (pfft), + [i] "=&r" (i), [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4), + [ppfft] "+r" (ppfft) + : + : "memory" + ); + + pfft = fft; + out_aecm = (int32_t)(outCFFT - aecm->dfaCleanQDomain); + + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[i], $zero, 64 \n\t" + "11: \n\t" + "lh %[tmp1], 0(%[pfft]) \n\t" + "lh %[tmp2], 0(%[p_kSqrtHanning]) \n\t" + "addiu %[i], %[i], -2 \n\t" + "mul %[tmp1], %[tmp1], %[tmp2] \n\t" + "lh %[tmp3], 2(%[pfft]) \n\t" + "lh %[tmp4], 2(%[p_kSqrtHanning]) \n\t" + "mul %[tmp3], %[tmp3], %[tmp4] \n\t" + "addiu %[tmp1], %[tmp1], 8192 \n\t" + "sra %[tmp1], %[tmp1], 14 \n\t" + "addiu %[tmp3], %[tmp3], 8192 \n\t" + "sra %[tmp3], %[tmp3], 14 \n\t" + "bgez %[out_aecm], 1f \n\t" + " negu %[tmp2], %[out_aecm] \n\t" + "srav %[tmp1], %[tmp1], %[tmp2] \n\t" + "b 2f \n\t" + " srav %[tmp3], %[tmp3], %[tmp2] \n\t" + "1: \n\t" + "sllv %[tmp1], %[tmp1], %[out_aecm] \n\t" + "sllv %[tmp3], %[tmp3], %[out_aecm] \n\t" + "2: \n\t" + "lh %[tmp4], 0(%[paecm_buf]) \n\t" + "lh %[tmp2], 2(%[paecm_buf]) \n\t" + "addu %[tmp3], %[tmp3], %[tmp2] \n\t" + "addu %[tmp1], %[tmp1], %[tmp4] \n\t" +#if defined(MIPS_DSP_R1_LE) + "shll_s.w %[tmp1], %[tmp1], 16 \n\t" + "sra %[tmp1], %[tmp1], 16 \n\t" + "shll_s.w %[tmp3], %[tmp3], 16 \n\t" + "sra %[tmp3], %[tmp3], 16 \n\t" +#else // #if defined(MIPS_DSP_R1_LE) + "sra %[tmp4], %[tmp1], 31 \n\t" + "sra %[tmp2], %[tmp1], 15 \n\t" + "beq %[tmp4], %[tmp2], 3f \n\t" + " ori %[tmp2], $zero, 0x7fff \n\t" + "xor %[tmp1], %[tmp2], %[tmp4] \n\t" + "3: \n\t" + "sra %[tmp2], %[tmp3], 31 \n\t" + "sra %[tmp4], %[tmp3], 15 \n\t" + "beq %[tmp2], %[tmp4], 4f \n\t" + " ori %[tmp4], $zero, 0x7fff \n\t" + "xor %[tmp3], %[tmp4], %[tmp2] \n\t" + "4: \n\t" +#endif // #if defined(MIPS_DSP_R1_LE) + "sh %[tmp1], 0(%[pfft]) \n\t" + "sh %[tmp1], 0(%[output1]) \n\t" + "sh %[tmp3], 2(%[pfft]) \n\t" + "sh %[tmp3], 2(%[output1]) \n\t" + "lh %[tmp1], 128(%[pfft]) \n\t" + "lh %[tmp2], 0(%[pp_kSqrtHanning]) \n\t" + "mul %[tmp1], %[tmp1], %[tmp2] \n\t" + "lh %[tmp3], 130(%[pfft]) \n\t" + "lh %[tmp4], -2(%[pp_kSqrtHanning]) \n\t" + "mul %[tmp3], %[tmp3], %[tmp4] \n\t" + "sra %[tmp1], %[tmp1], 14 \n\t" + "sra %[tmp3], %[tmp3], 14 \n\t" + "bgez %[out_aecm], 5f \n\t" + " negu %[tmp2], %[out_aecm] \n\t" + "srav %[tmp3], %[tmp3], %[tmp2] \n\t" + "b 6f \n\t" + " srav %[tmp1], %[tmp1], %[tmp2] \n\t" + "5: \n\t" + "sllv %[tmp1], %[tmp1], %[out_aecm] \n\t" + "sllv %[tmp3], %[tmp3], %[out_aecm] \n\t" + "6: \n\t" +#if defined(MIPS_DSP_R1_LE) + "shll_s.w %[tmp1], %[tmp1], 16 \n\t" + "sra %[tmp1], %[tmp1], 16 \n\t" + "shll_s.w %[tmp3], %[tmp3], 16 \n\t" + "sra %[tmp3], %[tmp3], 16 \n\t" +#else // #if defined(MIPS_DSP_R1_LE) + "sra %[tmp4], %[tmp1], 31 \n\t" + "sra %[tmp2], %[tmp1], 15 \n\t" + "beq %[tmp4], %[tmp2], 7f \n\t" + " ori %[tmp2], $zero, 0x7fff \n\t" + "xor %[tmp1], %[tmp2], %[tmp4] \n\t" + "7: \n\t" + "sra %[tmp2], %[tmp3], 31 \n\t" + "sra %[tmp4], %[tmp3], 15 \n\t" + "beq %[tmp2], %[tmp4], 8f \n\t" + " ori %[tmp4], $zero, 0x7fff \n\t" + "xor %[tmp3], %[tmp4], %[tmp2] \n\t" + "8: \n\t" +#endif // #if defined(MIPS_DSP_R1_LE) + "sh %[tmp1], 0(%[paecm_buf]) \n\t" + "sh %[tmp3], 2(%[paecm_buf]) \n\t" + "addiu %[output1], %[output1], 4 \n\t" + "addiu %[paecm_buf], %[paecm_buf], 4 \n\t" + "addiu %[pfft], %[pfft], 4 \n\t" + "addiu %[p_kSqrtHanning], %[p_kSqrtHanning], 4 \n\t" + "bgtz %[i], 11b \n\t" + " addiu %[pp_kSqrtHanning], %[pp_kSqrtHanning], -4 \n\t" + ".set pop \n\t" + : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2), [pfft] "+r" (pfft), + [output1] "+r" (output1), [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4), + [paecm_buf] "+r" (paecm_buf), [i] "=&r" (i), + [pp_kSqrtHanning] "+r" (pp_kSqrtHanning), + [p_kSqrtHanning] "+r" (p_kSqrtHanning) + : [out_aecm] "r" (out_aecm), + [WebRtcAecm_kSqrtHanning] "r" (WebRtcAecm_kSqrtHanning) + : "hi", "lo","memory" + ); + + // Copy the current block to the old position + // (aecm->outBuf is shifted elsewhere) + memcpy(aecm->xBuf, aecm->xBuf + PART_LEN, sizeof(int16_t) * PART_LEN); + memcpy(aecm->dBufNoisy, + aecm->dBufNoisy + PART_LEN, + sizeof(int16_t) * PART_LEN); + if (nearendClean != NULL) { + memcpy(aecm->dBufClean, + aecm->dBufClean + PART_LEN, + sizeof(int16_t) * PART_LEN); + } +} + +void WebRtcAecm_CalcLinearEnergies_mips(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored) { + int i; + uint32_t par1 = (*far_energy); + uint32_t par2 = (*echo_energy_adapt); + uint32_t par3 = (*echo_energy_stored); + int16_t* ch_stored_p = &(aecm->channelStored[0]); + int16_t* ch_adapt_p = &(aecm->channelAdapt16[0]); + uint16_t* spectrum_p = (uint16_t*)(&(far_spectrum[0])); + int32_t* echo_p = &(echo_est[0]); + int32_t temp0, stored0, echo0, adept0, spectrum0; + int32_t stored1, adept1, spectrum1, echo1, temp1; + + // Get energy for the delayed far end signal and estimated + // echo using both stored and adapted channels. + for (i = 0; i < PART_LEN; i+= 4) { + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "lh %[stored0], 0(%[ch_stored_p]) \n\t" + "lhu %[adept0], 0(%[ch_adapt_p]) \n\t" + "lhu %[spectrum0], 0(%[spectrum_p]) \n\t" + "lh %[stored1], 2(%[ch_stored_p]) \n\t" + "lhu %[adept1], 2(%[ch_adapt_p]) \n\t" + "lhu %[spectrum1], 2(%[spectrum_p]) \n\t" + "mul %[echo0], %[stored0], %[spectrum0] \n\t" + "mul %[temp0], %[adept0], %[spectrum0] \n\t" + "mul %[echo1], %[stored1], %[spectrum1] \n\t" + "mul %[temp1], %[adept1], %[spectrum1] \n\t" + "addu %[par1], %[par1], %[spectrum0] \n\t" + "addu %[par1], %[par1], %[spectrum1] \n\t" + "addiu %[echo_p], %[echo_p], 16 \n\t" + "addu %[par3], %[par3], %[echo0] \n\t" + "addu %[par2], %[par2], %[temp0] \n\t" + "addu %[par3], %[par3], %[echo1] \n\t" + "addu %[par2], %[par2], %[temp1] \n\t" + "usw %[echo0], -16(%[echo_p]) \n\t" + "usw %[echo1], -12(%[echo_p]) \n\t" + "lh %[stored0], 4(%[ch_stored_p]) \n\t" + "lhu %[adept0], 4(%[ch_adapt_p]) \n\t" + "lhu %[spectrum0], 4(%[spectrum_p]) \n\t" + "lh %[stored1], 6(%[ch_stored_p]) \n\t" + "lhu %[adept1], 6(%[ch_adapt_p]) \n\t" + "lhu %[spectrum1], 6(%[spectrum_p]) \n\t" + "mul %[echo0], %[stored0], %[spectrum0] \n\t" + "mul %[temp0], %[adept0], %[spectrum0] \n\t" + "mul %[echo1], %[stored1], %[spectrum1] \n\t" + "mul %[temp1], %[adept1], %[spectrum1] \n\t" + "addu %[par1], %[par1], %[spectrum0] \n\t" + "addu %[par1], %[par1], %[spectrum1] \n\t" + "addiu %[ch_stored_p], %[ch_stored_p], 8 \n\t" + "addiu %[ch_adapt_p], %[ch_adapt_p], 8 \n\t" + "addiu %[spectrum_p], %[spectrum_p], 8 \n\t" + "addu %[par3], %[par3], %[echo0] \n\t" + "addu %[par2], %[par2], %[temp0] \n\t" + "addu %[par3], %[par3], %[echo1] \n\t" + "addu %[par2], %[par2], %[temp1] \n\t" + "usw %[echo0], -8(%[echo_p]) \n\t" + "usw %[echo1], -4(%[echo_p]) \n\t" + ".set pop \n\t" + : [temp0] "=&r" (temp0), [stored0] "=&r" (stored0), + [adept0] "=&r" (adept0), [spectrum0] "=&r" (spectrum0), + [echo0] "=&r" (echo0), [echo_p] "+r" (echo_p), [par3] "+r" (par3), + [par1] "+r" (par1), [par2] "+r" (par2), [stored1] "=&r" (stored1), + [adept1] "=&r" (adept1), [echo1] "=&r" (echo1), + [spectrum1] "=&r" (spectrum1), [temp1] "=&r" (temp1), + [ch_stored_p] "+r" (ch_stored_p), [ch_adapt_p] "+r" (ch_adapt_p), + [spectrum_p] "+r" (spectrum_p) + : + : "hi", "lo", "memory" + ); + } + + echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN], + far_spectrum[PART_LEN]); + par1 += (uint32_t)(far_spectrum[PART_LEN]); + par2 += aecm->channelAdapt16[PART_LEN] * far_spectrum[PART_LEN]; + par3 += (uint32_t)echo_est[PART_LEN]; + + (*far_energy) = par1; + (*echo_energy_adapt) = par2; + (*echo_energy_stored) = par3; +} + +#if defined(MIPS_DSP_R1_LE) +void WebRtcAecm_StoreAdaptiveChannel_mips(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est) { + int i; + int16_t* temp1; + uint16_t* temp8; + int32_t temp0, temp2, temp3, temp4, temp5, temp6; + int32_t* temp7 = &(echo_est[0]); + temp1 = &(aecm->channelStored[0]); + temp8 = (uint16_t*)(&far_spectrum[0]); + + // During startup we store the channel every block. + memcpy(aecm->channelStored, aecm->channelAdapt16, + sizeof(int16_t) * PART_LEN1); + // Recalculate echo estimate + for (i = 0; i < PART_LEN; i += 4) { + __asm __volatile ( + "ulw %[temp0], 0(%[temp8]) \n\t" + "ulw %[temp2], 0(%[temp1]) \n\t" + "ulw %[temp4], 4(%[temp8]) \n\t" + "ulw %[temp5], 4(%[temp1]) \n\t" + "muleq_s.w.phl %[temp3], %[temp2], %[temp0] \n\t" + "muleq_s.w.phr %[temp0], %[temp2], %[temp0] \n\t" + "muleq_s.w.phl %[temp6], %[temp5], %[temp4] \n\t" + "muleq_s.w.phr %[temp4], %[temp5], %[temp4] \n\t" + "addiu %[temp7], %[temp7], 16 \n\t" + "addiu %[temp1], %[temp1], 8 \n\t" + "addiu %[temp8], %[temp8], 8 \n\t" + "sra %[temp3], %[temp3], 1 \n\t" + "sra %[temp0], %[temp0], 1 \n\t" + "sra %[temp6], %[temp6], 1 \n\t" + "sra %[temp4], %[temp4], 1 \n\t" + "usw %[temp3], -12(%[temp7]) \n\t" + "usw %[temp0], -16(%[temp7]) \n\t" + "usw %[temp6], -4(%[temp7]) \n\t" + "usw %[temp4], -8(%[temp7]) \n\t" + : [temp0] "=&r" (temp0), [temp2] "=&r" (temp2), [temp3] "=&r" (temp3), + [temp4] "=&r" (temp4), [temp5] "=&r" (temp5), [temp6] "=&r" (temp6), + [temp1] "+r" (temp1), [temp8] "+r" (temp8), [temp7] "+r" (temp7) + : + : "hi", "lo", "memory" + ); + } + echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + far_spectrum[i]); +} + +void WebRtcAecm_ResetAdaptiveChannel_mips(AecmCore* aecm) { + int i; + int32_t* temp3; + int16_t* temp0; + int32_t temp1, temp2, temp4, temp5; + + temp0 = &(aecm->channelStored[0]); + temp3 = &(aecm->channelAdapt32[0]); + + // The stored channel has a significantly lower MSE than the adaptive one for + // two consecutive calculations. Reset the adaptive channel. + memcpy(aecm->channelAdapt16, + aecm->channelStored, + sizeof(int16_t) * PART_LEN1); + + // Restore the W32 channel + for (i = 0; i < PART_LEN; i += 4) { + __asm __volatile ( + "ulw %[temp1], 0(%[temp0]) \n\t" + "ulw %[temp4], 4(%[temp0]) \n\t" + "preceq.w.phl %[temp2], %[temp1] \n\t" + "preceq.w.phr %[temp1], %[temp1] \n\t" + "preceq.w.phl %[temp5], %[temp4] \n\t" + "preceq.w.phr %[temp4], %[temp4] \n\t" + "addiu %[temp0], %[temp0], 8 \n\t" + "usw %[temp2], 4(%[temp3]) \n\t" + "usw %[temp1], 0(%[temp3]) \n\t" + "usw %[temp5], 12(%[temp3]) \n\t" + "usw %[temp4], 8(%[temp3]) \n\t" + "addiu %[temp3], %[temp3], 16 \n\t" + : [temp1] "=&r" (temp1), [temp2] "=&r" (temp2), + [temp4] "=&r" (temp4), [temp5] "=&r" (temp5), + [temp3] "+r" (temp3), [temp0] "+r" (temp0) + : + : "memory" + ); + } + + aecm->channelAdapt32[i] = (int32_t)aecm->channelStored[i] << 16; +} +#endif // #if defined(MIPS_DSP_R1_LE) + +// Transforms a time domain signal into the frequency domain, outputting the +// complex valued signal, absolute value and sum of absolute values. +// +// time_signal [in] Pointer to time domain signal +// freq_signal_real [out] Pointer to real part of frequency domain array +// freq_signal_imag [out] Pointer to imaginary part of frequency domain +// array +// freq_signal_abs [out] Pointer to absolute value of frequency domain +// array +// freq_signal_sum_abs [out] Pointer to the sum of all absolute values in +// the frequency domain array +// return value The Q-domain of current frequency values +// +static int TimeToFrequencyDomain(AecmCore* aecm, + const int16_t* time_signal, + ComplexInt16* freq_signal, + uint16_t* freq_signal_abs, + uint32_t* freq_signal_sum_abs) { + int i = 0; + int time_signal_scaling = 0; + + // In fft_buf, +16 for 32-byte alignment. + int16_t fft_buf[PART_LEN4 + 16]; + int16_t *fft = (int16_t *) (((uintptr_t) fft_buf + 31) & ~31); + + int16_t tmp16no1; +#if !defined(MIPS_DSP_R2_LE) + int32_t tmp32no1; + int32_t tmp32no2; + int16_t tmp16no2; +#else + int32_t tmp32no10, tmp32no11, tmp32no12, tmp32no13; + int32_t tmp32no20, tmp32no21, tmp32no22, tmp32no23; + int16_t* freqp; + uint16_t* freqabsp; + uint32_t freqt0, freqt1, freqt2, freqt3; + uint32_t freqs; +#endif + +#ifdef AECM_DYNAMIC_Q + tmp16no1 = WebRtcSpl_MaxAbsValueW16(time_signal, PART_LEN2); + time_signal_scaling = WebRtcSpl_NormW16(tmp16no1); +#endif + + WindowAndFFT(aecm, fft, time_signal, freq_signal, time_signal_scaling); + + // Extract imaginary and real part, + // calculate the magnitude for all frequency bins + freq_signal[0].imag = 0; + freq_signal[PART_LEN].imag = 0; + freq_signal[PART_LEN].real = fft[PART_LEN2]; + freq_signal_abs[0] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[0].real); + freq_signal_abs[PART_LEN] = (uint16_t)WEBRTC_SPL_ABS_W16( + freq_signal[PART_LEN].real); + (*freq_signal_sum_abs) = (uint32_t)(freq_signal_abs[0]) + + (uint32_t)(freq_signal_abs[PART_LEN]); + +#if !defined(MIPS_DSP_R2_LE) + for (i = 1; i < PART_LEN; i++) { + if (freq_signal[i].real == 0) + { + freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16( + freq_signal[i].imag); + } + else if (freq_signal[i].imag == 0) + { + freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16( + freq_signal[i].real); + } + else + { + // Approximation for magnitude of complex fft output + // magn = sqrt(real^2 + imag^2) + // magn ~= alpha * max(|imag|,|real|) + beta * min(|imag|,|real|) + // + // The parameters alpha and beta are stored in Q15 + tmp16no1 = WEBRTC_SPL_ABS_W16(freq_signal[i].real); + tmp16no2 = WEBRTC_SPL_ABS_W16(freq_signal[i].imag); + tmp32no1 = tmp16no1 * tmp16no1; + tmp32no2 = tmp16no2 * tmp16no2; + tmp32no2 = WebRtcSpl_AddSatW32(tmp32no1, tmp32no2); + tmp32no1 = WebRtcSpl_SqrtFloor(tmp32no2); + + freq_signal_abs[i] = (uint16_t)tmp32no1; + } + (*freq_signal_sum_abs) += (uint32_t)freq_signal_abs[i]; + } +#else // #if !defined(MIPS_DSP_R2_LE) + freqs = (uint32_t)(freq_signal_abs[0]) + + (uint32_t)(freq_signal_abs[PART_LEN]); + freqp = &(freq_signal[1].real); + + __asm __volatile ( + "lw %[freqt0], 0(%[freqp]) \n\t" + "lw %[freqt1], 4(%[freqp]) \n\t" + "lw %[freqt2], 8(%[freqp]) \n\t" + "mult $ac0, $zero, $zero \n\t" + "mult $ac1, $zero, $zero \n\t" + "mult $ac2, $zero, $zero \n\t" + "dpaq_s.w.ph $ac0, %[freqt0], %[freqt0] \n\t" + "dpaq_s.w.ph $ac1, %[freqt1], %[freqt1] \n\t" + "dpaq_s.w.ph $ac2, %[freqt2], %[freqt2] \n\t" + "addiu %[freqp], %[freqp], 12 \n\t" + "extr.w %[tmp32no20], $ac0, 1 \n\t" + "extr.w %[tmp32no21], $ac1, 1 \n\t" + "extr.w %[tmp32no22], $ac2, 1 \n\t" + : [freqt0] "=&r" (freqt0), [freqt1] "=&r" (freqt1), + [freqt2] "=&r" (freqt2), [freqp] "+r" (freqp), + [tmp32no20] "=r" (tmp32no20), [tmp32no21] "=r" (tmp32no21), + [tmp32no22] "=r" (tmp32no22) + : + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", "$ac2hi", "$ac2lo" + ); + + tmp32no10 = WebRtcSpl_SqrtFloor(tmp32no20); + tmp32no11 = WebRtcSpl_SqrtFloor(tmp32no21); + tmp32no12 = WebRtcSpl_SqrtFloor(tmp32no22); + freq_signal_abs[1] = (uint16_t)tmp32no10; + freq_signal_abs[2] = (uint16_t)tmp32no11; + freq_signal_abs[3] = (uint16_t)tmp32no12; + freqs += (uint32_t)tmp32no10; + freqs += (uint32_t)tmp32no11; + freqs += (uint32_t)tmp32no12; + freqabsp = &(freq_signal_abs[4]); + for (i = 4; i < PART_LEN; i+=4) + { + __asm __volatile ( + "ulw %[freqt0], 0(%[freqp]) \n\t" + "ulw %[freqt1], 4(%[freqp]) \n\t" + "ulw %[freqt2], 8(%[freqp]) \n\t" + "ulw %[freqt3], 12(%[freqp]) \n\t" + "mult $ac0, $zero, $zero \n\t" + "mult $ac1, $zero, $zero \n\t" + "mult $ac2, $zero, $zero \n\t" + "mult $ac3, $zero, $zero \n\t" + "dpaq_s.w.ph $ac0, %[freqt0], %[freqt0] \n\t" + "dpaq_s.w.ph $ac1, %[freqt1], %[freqt1] \n\t" + "dpaq_s.w.ph $ac2, %[freqt2], %[freqt2] \n\t" + "dpaq_s.w.ph $ac3, %[freqt3], %[freqt3] \n\t" + "addiu %[freqp], %[freqp], 16 \n\t" + "addiu %[freqabsp], %[freqabsp], 8 \n\t" + "extr.w %[tmp32no20], $ac0, 1 \n\t" + "extr.w %[tmp32no21], $ac1, 1 \n\t" + "extr.w %[tmp32no22], $ac2, 1 \n\t" + "extr.w %[tmp32no23], $ac3, 1 \n\t" + : [freqt0] "=&r" (freqt0), [freqt1] "=&r" (freqt1), + [freqt2] "=&r" (freqt2), [freqt3] "=&r" (freqt3), + [tmp32no20] "=r" (tmp32no20), [tmp32no21] "=r" (tmp32no21), + [tmp32no22] "=r" (tmp32no22), [tmp32no23] "=r" (tmp32no23), + [freqabsp] "+r" (freqabsp), [freqp] "+r" (freqp) + : + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + + tmp32no10 = WebRtcSpl_SqrtFloor(tmp32no20); + tmp32no11 = WebRtcSpl_SqrtFloor(tmp32no21); + tmp32no12 = WebRtcSpl_SqrtFloor(tmp32no22); + tmp32no13 = WebRtcSpl_SqrtFloor(tmp32no23); + + __asm __volatile ( + "sh %[tmp32no10], -8(%[freqabsp]) \n\t" + "sh %[tmp32no11], -6(%[freqabsp]) \n\t" + "sh %[tmp32no12], -4(%[freqabsp]) \n\t" + "sh %[tmp32no13], -2(%[freqabsp]) \n\t" + "addu %[freqs], %[freqs], %[tmp32no10] \n\t" + "addu %[freqs], %[freqs], %[tmp32no11] \n\t" + "addu %[freqs], %[freqs], %[tmp32no12] \n\t" + "addu %[freqs], %[freqs], %[tmp32no13] \n\t" + : [freqs] "+r" (freqs) + : [tmp32no10] "r" (tmp32no10), [tmp32no11] "r" (tmp32no11), + [tmp32no12] "r" (tmp32no12), [tmp32no13] "r" (tmp32no13), + [freqabsp] "r" (freqabsp) + : "memory" + ); + } + + (*freq_signal_sum_abs) = freqs; +#endif + + return time_signal_scaling; +} + +int WebRtcAecm_ProcessBlock(AecmCore* aecm, + const int16_t* farend, + const int16_t* nearendNoisy, + const int16_t* nearendClean, + int16_t* output) { + int i; + uint32_t xfaSum; + uint32_t dfaNoisySum; + uint32_t dfaCleanSum; + uint32_t echoEst32Gained; + uint32_t tmpU32; + int32_t tmp32no1; + + uint16_t xfa[PART_LEN1]; + uint16_t dfaNoisy[PART_LEN1]; + uint16_t dfaClean[PART_LEN1]; + uint16_t* ptrDfaClean = dfaClean; + const uint16_t* far_spectrum_ptr = NULL; + + // 32 byte aligned buffers (with +8 or +16). + int16_t fft_buf[PART_LEN4 + 2 + 16]; // +2 to make a loop safe. + int32_t echoEst32_buf[PART_LEN1 + 8]; + int32_t dfw_buf[PART_LEN2 + 8]; + int32_t efw_buf[PART_LEN2 + 8]; + + int16_t* fft = (int16_t*)(((uint32_t)fft_buf + 31) & ~ 31); + int32_t* echoEst32 = (int32_t*)(((uint32_t)echoEst32_buf + 31) & ~ 31); + ComplexInt16* dfw = (ComplexInt16*)(((uint32_t)dfw_buf + 31) & ~31); + ComplexInt16* efw = (ComplexInt16*)(((uint32_t)efw_buf + 31) & ~31); + + int16_t hnl[PART_LEN1]; + int16_t numPosCoef = 0; + int delay; + int16_t tmp16no1; + int16_t tmp16no2; + int16_t mu; + int16_t supGain; + int16_t zeros32, zeros16; + int16_t zerosDBufNoisy, zerosDBufClean, zerosXBuf; + int far_q; + int16_t resolutionDiff, qDomainDiff, dfa_clean_q_domain_diff; + + const int kMinPrefBand = 4; + const int kMaxPrefBand = 24; + int32_t avgHnl32 = 0; + + int32_t temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + int16_t* ptr; + int16_t* ptr1; + int16_t* er_ptr; + int16_t* dr_ptr; + + ptr = &hnl[0]; + ptr1 = &hnl[0]; + er_ptr = &efw[0].real; + dr_ptr = &dfw[0].real; + + // Determine startup state. There are three states: + // (0) the first CONV_LEN blocks + // (1) another CONV_LEN blocks + // (2) the rest + + if (aecm->startupState < 2) { + aecm->startupState = (aecm->totCount >= CONV_LEN) + + (aecm->totCount >= CONV_LEN2); + } + // END: Determine startup state + + // Buffer near and far end signals + memcpy(aecm->xBuf + PART_LEN, farend, sizeof(int16_t) * PART_LEN); + memcpy(aecm->dBufNoisy + PART_LEN, + nearendNoisy, + sizeof(int16_t) * PART_LEN); + if (nearendClean != NULL) { + memcpy(aecm->dBufClean + PART_LEN, + nearendClean, + sizeof(int16_t) * PART_LEN); + } + + // Transform far end signal from time domain to frequency domain. + far_q = TimeToFrequencyDomain(aecm, + aecm->xBuf, + dfw, + xfa, + &xfaSum); + + // Transform noisy near end signal from time domain to frequency domain. + zerosDBufNoisy = TimeToFrequencyDomain(aecm, + aecm->dBufNoisy, + dfw, + dfaNoisy, + &dfaNoisySum); + aecm->dfaNoisyQDomainOld = aecm->dfaNoisyQDomain; + aecm->dfaNoisyQDomain = (int16_t)zerosDBufNoisy; + + if (nearendClean == NULL) { + ptrDfaClean = dfaNoisy; + aecm->dfaCleanQDomainOld = aecm->dfaNoisyQDomainOld; + aecm->dfaCleanQDomain = aecm->dfaNoisyQDomain; + dfaCleanSum = dfaNoisySum; + } else { + // Transform clean near end signal from time domain to frequency domain. + zerosDBufClean = TimeToFrequencyDomain(aecm, + aecm->dBufClean, + dfw, + dfaClean, + &dfaCleanSum); + aecm->dfaCleanQDomainOld = aecm->dfaCleanQDomain; + aecm->dfaCleanQDomain = (int16_t)zerosDBufClean; + } + + // Get the delay + // Save far-end history and estimate delay + WebRtcAecm_UpdateFarHistory(aecm, xfa, far_q); + + if (WebRtc_AddFarSpectrumFix(aecm->delay_estimator_farend, xfa, PART_LEN1, + far_q) == -1) { + return -1; + } + delay = WebRtc_DelayEstimatorProcessFix(aecm->delay_estimator, + dfaNoisy, + PART_LEN1, + zerosDBufNoisy); + if (delay == -1) { + return -1; + } + else if (delay == -2) { + // If the delay is unknown, we assume zero. + // NOTE: this will have to be adjusted if we ever add lookahead. + delay = 0; + } + + if (aecm->fixedDelay >= 0) { + // Use fixed delay + delay = aecm->fixedDelay; + } + + // Get aligned far end spectrum + far_spectrum_ptr = WebRtcAecm_AlignedFarend(aecm, &far_q, delay); + zerosXBuf = (int16_t) far_q; + + if (far_spectrum_ptr == NULL) { + return -1; + } + + // Calculate log(energy) and update energy threshold levels + WebRtcAecm_CalcEnergies(aecm, + far_spectrum_ptr, + zerosXBuf, + dfaNoisySum, + echoEst32); + // Calculate stepsize + mu = WebRtcAecm_CalcStepSize(aecm); + + // Update counters + aecm->totCount++; + + // This is the channel estimation algorithm. + // It is base on NLMS but has a variable step length, + // which was calculated above. + WebRtcAecm_UpdateChannel(aecm, + far_spectrum_ptr, + zerosXBuf, + dfaNoisy, + mu, + echoEst32); + + supGain = WebRtcAecm_CalcSuppressionGain(aecm); + + // Calculate Wiener filter hnl[] + for (i = 0; i < PART_LEN1; i++) { + // Far end signal through channel estimate in Q8 + // How much can we shift right to preserve resolution + tmp32no1 = echoEst32[i] - aecm->echoFilt[i]; + aecm->echoFilt[i] += (tmp32no1 * 50) >> 8; + + zeros32 = WebRtcSpl_NormW32(aecm->echoFilt[i]) + 1; + zeros16 = WebRtcSpl_NormW16(supGain) + 1; + if (zeros32 + zeros16 > 16) { + // Multiplication is safe + // Result in + // Q(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN+aecm->xfaQDomainBuf[diff]) + echoEst32Gained = WEBRTC_SPL_UMUL_32_16((uint32_t)aecm->echoFilt[i], + (uint16_t)supGain); + resolutionDiff = 14 - RESOLUTION_CHANNEL16 - RESOLUTION_SUPGAIN; + resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf); + } else { + tmp16no1 = 17 - zeros32 - zeros16; + resolutionDiff = 14 + tmp16no1 - RESOLUTION_CHANNEL16 - + RESOLUTION_SUPGAIN; + resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf); + if (zeros32 > tmp16no1) { + echoEst32Gained = WEBRTC_SPL_UMUL_32_16( + (uint32_t)aecm->echoFilt[i], + supGain >> tmp16no1); + } else { + // Result in Q-(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN-16) + echoEst32Gained = (aecm->echoFilt[i] >> tmp16no1) * supGain; + } + } + + zeros16 = WebRtcSpl_NormW16(aecm->nearFilt[i]); + assert(zeros16 >= 0); // |zeros16| is a norm, hence non-negative. + dfa_clean_q_domain_diff = aecm->dfaCleanQDomain - aecm->dfaCleanQDomainOld; + if (zeros16 < dfa_clean_q_domain_diff && aecm->nearFilt[i]) { + tmp16no1 = aecm->nearFilt[i] << zeros16; + qDomainDiff = zeros16 - dfa_clean_q_domain_diff; + tmp16no2 = ptrDfaClean[i] >> -qDomainDiff; + } else { + tmp16no1 = dfa_clean_q_domain_diff < 0 + ? aecm->nearFilt[i] >> -dfa_clean_q_domain_diff + : aecm->nearFilt[i] << dfa_clean_q_domain_diff; + qDomainDiff = 0; + tmp16no2 = ptrDfaClean[i]; + } + + tmp32no1 = (int32_t)(tmp16no2 - tmp16no1); + tmp16no2 = (int16_t)(tmp32no1 >> 4); + tmp16no2 += tmp16no1; + zeros16 = WebRtcSpl_NormW16(tmp16no2); + if ((tmp16no2) & (-qDomainDiff > zeros16)) { + aecm->nearFilt[i] = WEBRTC_SPL_WORD16_MAX; + } else { + aecm->nearFilt[i] = qDomainDiff < 0 ? tmp16no2 << -qDomainDiff + : tmp16no2 >> qDomainDiff; + } + + // Wiener filter coefficients, resulting hnl in Q14 + if (echoEst32Gained == 0) { + hnl[i] = ONE_Q14; + numPosCoef++; + } else if (aecm->nearFilt[i] == 0) { + hnl[i] = 0; + } else { + // Multiply the suppression gain + // Rounding + echoEst32Gained += (uint32_t)(aecm->nearFilt[i] >> 1); + tmpU32 = WebRtcSpl_DivU32U16(echoEst32Gained, + (uint16_t)aecm->nearFilt[i]); + + // Current resolution is + // Q-(RESOLUTION_CHANNEL + RESOLUTION_SUPGAIN + // - max(0, 17 - zeros16 - zeros32)) + // Make sure we are in Q14 + tmp32no1 = (int32_t)WEBRTC_SPL_SHIFT_W32(tmpU32, resolutionDiff); + if (tmp32no1 > ONE_Q14) { + hnl[i] = 0; + } else if (tmp32no1 < 0) { + hnl[i] = ONE_Q14; + numPosCoef++; + } else { + // 1-echoEst/dfa + hnl[i] = ONE_Q14 - (int16_t)tmp32no1; + if (hnl[i] <= 0) { + hnl[i] = 0; + } else { + numPosCoef++; + } + } + } + } + + // Only in wideband. Prevent the gain in upper band from being larger than + // in lower band. + if (aecm->mult == 2) { + // TODO(bjornv): Investigate if the scaling of hnl[i] below can cause + // speech distortion in double-talk. + for (i = 0; i < (PART_LEN1 >> 3); i++) { + __asm __volatile ( + "lh %[temp1], 0(%[ptr1]) \n\t" + "lh %[temp2], 2(%[ptr1]) \n\t" + "lh %[temp3], 4(%[ptr1]) \n\t" + "lh %[temp4], 6(%[ptr1]) \n\t" + "lh %[temp5], 8(%[ptr1]) \n\t" + "lh %[temp6], 10(%[ptr1]) \n\t" + "lh %[temp7], 12(%[ptr1]) \n\t" + "lh %[temp8], 14(%[ptr1]) \n\t" + "mul %[temp1], %[temp1], %[temp1] \n\t" + "mul %[temp2], %[temp2], %[temp2] \n\t" + "mul %[temp3], %[temp3], %[temp3] \n\t" + "mul %[temp4], %[temp4], %[temp4] \n\t" + "mul %[temp5], %[temp5], %[temp5] \n\t" + "mul %[temp6], %[temp6], %[temp6] \n\t" + "mul %[temp7], %[temp7], %[temp7] \n\t" + "mul %[temp8], %[temp8], %[temp8] \n\t" + "sra %[temp1], %[temp1], 14 \n\t" + "sra %[temp2], %[temp2], 14 \n\t" + "sra %[temp3], %[temp3], 14 \n\t" + "sra %[temp4], %[temp4], 14 \n\t" + "sra %[temp5], %[temp5], 14 \n\t" + "sra %[temp6], %[temp6], 14 \n\t" + "sra %[temp7], %[temp7], 14 \n\t" + "sra %[temp8], %[temp8], 14 \n\t" + "sh %[temp1], 0(%[ptr1]) \n\t" + "sh %[temp2], 2(%[ptr1]) \n\t" + "sh %[temp3], 4(%[ptr1]) \n\t" + "sh %[temp4], 6(%[ptr1]) \n\t" + "sh %[temp5], 8(%[ptr1]) \n\t" + "sh %[temp6], 10(%[ptr1]) \n\t" + "sh %[temp7], 12(%[ptr1]) \n\t" + "sh %[temp8], 14(%[ptr1]) \n\t" + "addiu %[ptr1], %[ptr1], 16 \n\t" + : [temp1] "=&r" (temp1), [temp2] "=&r" (temp2), [temp3] "=&r" (temp3), + [temp4] "=&r" (temp4), [temp5] "=&r" (temp5), [temp6] "=&r" (temp6), + [temp7] "=&r" (temp7), [temp8] "=&r" (temp8), [ptr1] "+r" (ptr1) + : + : "memory", "hi", "lo" + ); + } + for(i = 0; i < (PART_LEN1 & 7); i++) { + __asm __volatile ( + "lh %[temp1], 0(%[ptr1]) \n\t" + "mul %[temp1], %[temp1], %[temp1] \n\t" + "sra %[temp1], %[temp1], 14 \n\t" + "sh %[temp1], 0(%[ptr1]) \n\t" + "addiu %[ptr1], %[ptr1], 2 \n\t" + : [temp1] "=&r" (temp1), [ptr1] "+r" (ptr1) + : + : "memory", "hi", "lo" + ); + } + + for (i = kMinPrefBand; i <= kMaxPrefBand; i++) { + avgHnl32 += (int32_t)hnl[i]; + } + + assert(kMaxPrefBand - kMinPrefBand + 1 > 0); + avgHnl32 /= (kMaxPrefBand - kMinPrefBand + 1); + + for (i = kMaxPrefBand; i < PART_LEN1; i++) { + if (hnl[i] > (int16_t)avgHnl32) { + hnl[i] = (int16_t)avgHnl32; + } + } + } + + // Calculate NLP gain, result is in Q14 + if (aecm->nlpFlag) { + if (numPosCoef < 3) { + for (i = 0; i < PART_LEN1; i++) { + efw[i].real = 0; + efw[i].imag = 0; + hnl[i] = 0; + } + } else { + for (i = 0; i < PART_LEN1; i++) { +#if defined(MIPS_DSP_R1_LE) + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "lh %[temp1], 0(%[ptr]) \n\t" + "lh %[temp2], 0(%[dr_ptr]) \n\t" + "slti %[temp4], %[temp1], 0x4001 \n\t" + "beqz %[temp4], 3f \n\t" + " lh %[temp3], 2(%[dr_ptr]) \n\t" + "slti %[temp5], %[temp1], 3277 \n\t" + "bnez %[temp5], 2f \n\t" + " addiu %[dr_ptr], %[dr_ptr], 4 \n\t" + "mul %[temp2], %[temp2], %[temp1] \n\t" + "mul %[temp3], %[temp3], %[temp1] \n\t" + "shra_r.w %[temp2], %[temp2], 14 \n\t" + "shra_r.w %[temp3], %[temp3], 14 \n\t" + "b 4f \n\t" + " nop \n\t" + "2: \n\t" + "addu %[temp1], $zero, $zero \n\t" + "addu %[temp2], $zero, $zero \n\t" + "addu %[temp3], $zero, $zero \n\t" + "b 1f \n\t" + " nop \n\t" + "3: \n\t" + "addiu %[temp1], $0, 0x4000 \n\t" + "1: \n\t" + "sh %[temp1], 0(%[ptr]) \n\t" + "4: \n\t" + "sh %[temp2], 0(%[er_ptr]) \n\t" + "sh %[temp3], 2(%[er_ptr]) \n\t" + "addiu %[ptr], %[ptr], 2 \n\t" + "addiu %[er_ptr], %[er_ptr], 4 \n\t" + ".set pop \n\t" + : [temp1] "=&r" (temp1), [temp2] "=&r" (temp2), [temp3] "=&r" (temp3), + [temp4] "=&r" (temp4), [temp5] "=&r" (temp5), [ptr] "+r" (ptr), + [er_ptr] "+r" (er_ptr), [dr_ptr] "+r" (dr_ptr) + : + : "memory", "hi", "lo" + ); +#else + __asm __volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "lh %[temp1], 0(%[ptr]) \n\t" + "lh %[temp2], 0(%[dr_ptr]) \n\t" + "slti %[temp4], %[temp1], 0x4001 \n\t" + "beqz %[temp4], 3f \n\t" + " lh %[temp3], 2(%[dr_ptr]) \n\t" + "slti %[temp5], %[temp1], 3277 \n\t" + "bnez %[temp5], 2f \n\t" + " addiu %[dr_ptr], %[dr_ptr], 4 \n\t" + "mul %[temp2], %[temp2], %[temp1] \n\t" + "mul %[temp3], %[temp3], %[temp1] \n\t" + "addiu %[temp2], %[temp2], 0x2000 \n\t" + "addiu %[temp3], %[temp3], 0x2000 \n\t" + "sra %[temp2], %[temp2], 14 \n\t" + "sra %[temp3], %[temp3], 14 \n\t" + "b 4f \n\t" + " nop \n\t" + "2: \n\t" + "addu %[temp1], $zero, $zero \n\t" + "addu %[temp2], $zero, $zero \n\t" + "addu %[temp3], $zero, $zero \n\t" + "b 1f \n\t" + " nop \n\t" + "3: \n\t" + "addiu %[temp1], $0, 0x4000 \n\t" + "1: \n\t" + "sh %[temp1], 0(%[ptr]) \n\t" + "4: \n\t" + "sh %[temp2], 0(%[er_ptr]) \n\t" + "sh %[temp3], 2(%[er_ptr]) \n\t" + "addiu %[ptr], %[ptr], 2 \n\t" + "addiu %[er_ptr], %[er_ptr], 4 \n\t" + ".set pop \n\t" + : [temp1] "=&r" (temp1), [temp2] "=&r" (temp2), [temp3] "=&r" (temp3), + [temp4] "=&r" (temp4), [temp5] "=&r" (temp5), [ptr] "+r" (ptr), + [er_ptr] "+r" (er_ptr), [dr_ptr] "+r" (dr_ptr) + : + : "memory", "hi", "lo" + ); +#endif + } + } + } + else { + // multiply with Wiener coefficients + for (i = 0; i < PART_LEN1; i++) { + efw[i].real = (int16_t) + (WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].real, + hnl[i], + 14)); + efw[i].imag = (int16_t) + (WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].imag, + hnl[i], + 14)); + } + } + + if (aecm->cngMode == AecmTrue) { + ComfortNoise(aecm, ptrDfaClean, efw, hnl); + } + + InverseFFTAndWindow(aecm, fft, efw, output, nearendClean); + + return 0; +} + +// Generate comfort noise and add to output signal. +static void ComfortNoise(AecmCore* aecm, + const uint16_t* dfa, + ComplexInt16* out, + const int16_t* lambda) { + int16_t i; + int16_t tmp16, tmp161, tmp162, tmp163, nrsh1, nrsh2; + int32_t tmp32, tmp321, tnoise, tnoise1; + int32_t tmp322, tmp323, *tmp1; + int16_t* dfap; + int16_t* lambdap; + const int32_t c2049 = 2049; + const int32_t c359 = 359; + const int32_t c114 = ONE_Q14; + + int16_t randW16[PART_LEN]; + int16_t uReal[PART_LEN1]; + int16_t uImag[PART_LEN1]; + int32_t outLShift32; + + int16_t shiftFromNearToNoise = kNoiseEstQDomain - aecm->dfaCleanQDomain; + int16_t minTrackShift = 9; + + assert(shiftFromNearToNoise >= 0); + assert(shiftFromNearToNoise < 16); + + if (aecm->noiseEstCtr < 100) { + // Track the minimum more quickly initially. + aecm->noiseEstCtr++; + minTrackShift = 6; + } + + // Generate a uniform random array on [0 2^15-1]. + WebRtcSpl_RandUArray(randW16, PART_LEN, &aecm->seed); + int16_t* randW16p = (int16_t*)randW16; +#if defined (MIPS_DSP_R1_LE) + int16_t* kCosTablep = (int16_t*)WebRtcAecm_kCosTable; + int16_t* kSinTablep = (int16_t*)WebRtcAecm_kSinTable; +#endif // #if defined(MIPS_DSP_R1_LE) + tmp1 = (int32_t*)aecm->noiseEst + 1; + dfap = (int16_t*)dfa + 1; + lambdap = (int16_t*)lambda + 1; + // Estimate noise power. + for (i = 1; i < PART_LEN1; i+=2) { + // Shift to the noise domain. + __asm __volatile ( + "lh %[tmp32], 0(%[dfap]) \n\t" + "lw %[tnoise], 0(%[tmp1]) \n\t" + "sllv %[outLShift32], %[tmp32], %[shiftFromNearToNoise] \n\t" + : [tmp32] "=&r" (tmp32), [outLShift32] "=r" (outLShift32), + [tnoise] "=&r" (tnoise) + : [tmp1] "r" (tmp1), [dfap] "r" (dfap), + [shiftFromNearToNoise] "r" (shiftFromNearToNoise) + : "memory" + ); + + if (outLShift32 < tnoise) { + // Reset "too low" counter + aecm->noiseEstTooLowCtr[i] = 0; + // Track the minimum. + if (tnoise < (1 << minTrackShift)) { + // For small values, decrease noiseEst[i] every + // |kNoiseEstIncCount| block. The regular approach below can not + // go further down due to truncation. + aecm->noiseEstTooHighCtr[i]++; + if (aecm->noiseEstTooHighCtr[i] >= kNoiseEstIncCount) { + tnoise--; + aecm->noiseEstTooHighCtr[i] = 0; // Reset the counter + } + } else { + __asm __volatile ( + "subu %[tmp32], %[tnoise], %[outLShift32] \n\t" + "srav %[tmp32], %[tmp32], %[minTrackShift] \n\t" + "subu %[tnoise], %[tnoise], %[tmp32] \n\t" + : [tmp32] "=&r" (tmp32), [tnoise] "+r" (tnoise) + : [outLShift32] "r" (outLShift32), [minTrackShift] "r" (minTrackShift) + ); + } + } else { + // Reset "too high" counter + aecm->noiseEstTooHighCtr[i] = 0; + // Ramp slowly upwards until we hit the minimum again. + if ((tnoise >> 19) <= 0) { + if ((tnoise >> 11) > 0) { + // Large enough for relative increase + __asm __volatile ( + "mul %[tnoise], %[tnoise], %[c2049] \n\t" + "sra %[tnoise], %[tnoise], 11 \n\t" + : [tnoise] "+r" (tnoise) + : [c2049] "r" (c2049) + : "hi", "lo" + ); + } else { + // Make incremental increases based on size every + // |kNoiseEstIncCount| block + aecm->noiseEstTooLowCtr[i]++; + if (aecm->noiseEstTooLowCtr[i] >= kNoiseEstIncCount) { + __asm __volatile ( + "sra %[tmp32], %[tnoise], 9 \n\t" + "addi %[tnoise], %[tnoise], 1 \n\t" + "addu %[tnoise], %[tnoise], %[tmp32] \n\t" + : [tnoise] "+r" (tnoise), [tmp32] "=&r" (tmp32) + : + ); + aecm->noiseEstTooLowCtr[i] = 0; // Reset counter + } + } + } else { + // Avoid overflow. + // Multiplication with 2049 will cause wrap around. Scale + // down first and then multiply + __asm __volatile ( + "sra %[tnoise], %[tnoise], 11 \n\t" + "mul %[tnoise], %[tnoise], %[c2049] \n\t" + : [tnoise] "+r" (tnoise) + : [c2049] "r" (c2049) + : "hi", "lo" + ); + } + } + + // Shift to the noise domain. + __asm __volatile ( + "lh %[tmp32], 2(%[dfap]) \n\t" + "lw %[tnoise1], 4(%[tmp1]) \n\t" + "addiu %[dfap], %[dfap], 4 \n\t" + "sllv %[outLShift32], %[tmp32], %[shiftFromNearToNoise] \n\t" + : [tmp32] "=&r" (tmp32), [dfap] "+r" (dfap), + [outLShift32] "=r" (outLShift32), [tnoise1] "=&r" (tnoise1) + : [tmp1] "r" (tmp1), [shiftFromNearToNoise] "r" (shiftFromNearToNoise) + : "memory" + ); + + if (outLShift32 < tnoise1) { + // Reset "too low" counter + aecm->noiseEstTooLowCtr[i + 1] = 0; + // Track the minimum. + if (tnoise1 < (1 << minTrackShift)) { + // For small values, decrease noiseEst[i] every + // |kNoiseEstIncCount| block. The regular approach below can not + // go further down due to truncation. + aecm->noiseEstTooHighCtr[i + 1]++; + if (aecm->noiseEstTooHighCtr[i + 1] >= kNoiseEstIncCount) { + tnoise1--; + aecm->noiseEstTooHighCtr[i + 1] = 0; // Reset the counter + } + } else { + __asm __volatile ( + "subu %[tmp32], %[tnoise1], %[outLShift32] \n\t" + "srav %[tmp32], %[tmp32], %[minTrackShift] \n\t" + "subu %[tnoise1], %[tnoise1], %[tmp32] \n\t" + : [tmp32] "=&r" (tmp32), [tnoise1] "+r" (tnoise1) + : [outLShift32] "r" (outLShift32), [minTrackShift] "r" (minTrackShift) + ); + } + } else { + // Reset "too high" counter + aecm->noiseEstTooHighCtr[i + 1] = 0; + // Ramp slowly upwards until we hit the minimum again. + if ((tnoise1 >> 19) <= 0) { + if ((tnoise1 >> 11) > 0) { + // Large enough for relative increase + __asm __volatile ( + "mul %[tnoise1], %[tnoise1], %[c2049] \n\t" + "sra %[tnoise1], %[tnoise1], 11 \n\t" + : [tnoise1] "+r" (tnoise1) + : [c2049] "r" (c2049) + : "hi", "lo" + ); + } else { + // Make incremental increases based on size every + // |kNoiseEstIncCount| block + aecm->noiseEstTooLowCtr[i + 1]++; + if (aecm->noiseEstTooLowCtr[i + 1] >= kNoiseEstIncCount) { + __asm __volatile ( + "sra %[tmp32], %[tnoise1], 9 \n\t" + "addi %[tnoise1], %[tnoise1], 1 \n\t" + "addu %[tnoise1], %[tnoise1], %[tmp32] \n\t" + : [tnoise1] "+r" (tnoise1), [tmp32] "=&r" (tmp32) + : + ); + aecm->noiseEstTooLowCtr[i + 1] = 0; // Reset counter + } + } + } else { + // Avoid overflow. + // Multiplication with 2049 will cause wrap around. Scale + // down first and then multiply + __asm __volatile ( + "sra %[tnoise1], %[tnoise1], 11 \n\t" + "mul %[tnoise1], %[tnoise1], %[c2049] \n\t" + : [tnoise1] "+r" (tnoise1) + : [c2049] "r" (c2049) + : "hi", "lo" + ); + } + } + + __asm __volatile ( + "lh %[tmp16], 0(%[lambdap]) \n\t" + "lh %[tmp161], 2(%[lambdap]) \n\t" + "sw %[tnoise], 0(%[tmp1]) \n\t" + "sw %[tnoise1], 4(%[tmp1]) \n\t" + "subu %[tmp16], %[c114], %[tmp16] \n\t" + "subu %[tmp161], %[c114], %[tmp161] \n\t" + "srav %[tmp32], %[tnoise], %[shiftFromNearToNoise] \n\t" + "srav %[tmp321], %[tnoise1], %[shiftFromNearToNoise] \n\t" + "addiu %[lambdap], %[lambdap], 4 \n\t" + "addiu %[tmp1], %[tmp1], 8 \n\t" + : [tmp16] "=&r" (tmp16), [tmp161] "=&r" (tmp161), [tmp1] "+r" (tmp1), + [tmp32] "=&r" (tmp32), [tmp321] "=&r" (tmp321), [lambdap] "+r" (lambdap) + : [tnoise] "r" (tnoise), [tnoise1] "r" (tnoise1), [c114] "r" (c114), + [shiftFromNearToNoise] "r" (shiftFromNearToNoise) + : "memory" + ); + + if (tmp32 > 32767) { + tmp32 = 32767; + aecm->noiseEst[i] = tmp32 << shiftFromNearToNoise; + } + if (tmp321 > 32767) { + tmp321 = 32767; + aecm->noiseEst[i+1] = tmp321 << shiftFromNearToNoise; + } + + __asm __volatile ( + "mul %[tmp32], %[tmp32], %[tmp16] \n\t" + "mul %[tmp321], %[tmp321], %[tmp161] \n\t" + "sra %[nrsh1], %[tmp32], 14 \n\t" + "sra %[nrsh2], %[tmp321], 14 \n\t" + : [nrsh1] "=&r" (nrsh1), [nrsh2] "=r" (nrsh2) + : [tmp16] "r" (tmp16), [tmp161] "r" (tmp161), [tmp32] "r" (tmp32), + [tmp321] "r" (tmp321) + : "memory", "hi", "lo" + ); + + __asm __volatile ( + "lh %[tmp32], 0(%[randW16p]) \n\t" + "lh %[tmp321], 2(%[randW16p]) \n\t" + "addiu %[randW16p], %[randW16p], 4 \n\t" + "mul %[tmp32], %[tmp32], %[c359] \n\t" + "mul %[tmp321], %[tmp321], %[c359] \n\t" + "sra %[tmp16], %[tmp32], 15 \n\t" + "sra %[tmp161], %[tmp321], 15 \n\t" + : [randW16p] "+r" (randW16p), [tmp32] "=&r" (tmp32), + [tmp16] "=r" (tmp16), [tmp161] "=r" (tmp161), [tmp321] "=&r" (tmp321) + : [c359] "r" (c359) + : "memory", "hi", "lo" + ); + +#if !defined(MIPS_DSP_R1_LE) + tmp32 = WebRtcAecm_kCosTable[tmp16]; + tmp321 = WebRtcAecm_kSinTable[tmp16]; + tmp322 = WebRtcAecm_kCosTable[tmp161]; + tmp323 = WebRtcAecm_kSinTable[tmp161]; +#else + __asm __volatile ( + "sll %[tmp16], %[tmp16], 1 \n\t" + "sll %[tmp161], %[tmp161], 1 \n\t" + "lhx %[tmp32], %[tmp16](%[kCosTablep]) \n\t" + "lhx %[tmp321], %[tmp16](%[kSinTablep]) \n\t" + "lhx %[tmp322], %[tmp161](%[kCosTablep]) \n\t" + "lhx %[tmp323], %[tmp161](%[kSinTablep]) \n\t" + : [tmp32] "=&r" (tmp32), [tmp321] "=&r" (tmp321), + [tmp322] "=&r" (tmp322), [tmp323] "=&r" (tmp323) + : [kCosTablep] "r" (kCosTablep), [tmp16] "r" (tmp16), + [tmp161] "r" (tmp161), [kSinTablep] "r" (kSinTablep) + : "memory" + ); +#endif + __asm __volatile ( + "mul %[tmp32], %[tmp32], %[nrsh1] \n\t" + "negu %[tmp162], %[nrsh1] \n\t" + "mul %[tmp322], %[tmp322], %[nrsh2] \n\t" + "negu %[tmp163], %[nrsh2] \n\t" + "sra %[tmp32], %[tmp32], 13 \n\t" + "mul %[tmp321], %[tmp321], %[tmp162] \n\t" + "sra %[tmp322], %[tmp322], 13 \n\t" + "mul %[tmp323], %[tmp323], %[tmp163] \n\t" + "sra %[tmp321], %[tmp321], 13 \n\t" + "sra %[tmp323], %[tmp323], 13 \n\t" + : [tmp32] "+r" (tmp32), [tmp321] "+r" (tmp321), [tmp162] "=&r" (tmp162), + [tmp322] "+r" (tmp322), [tmp323] "+r" (tmp323), [tmp163] "=&r" (tmp163) + : [nrsh1] "r" (nrsh1), [nrsh2] "r" (nrsh2) + : "hi", "lo" + ); + // Tables are in Q13. + uReal[i] = (int16_t)tmp32; + uImag[i] = (int16_t)tmp321; + uReal[i + 1] = (int16_t)tmp322; + uImag[i + 1] = (int16_t)tmp323; + } + + int32_t tt, sgn; + tt = out[0].real; + sgn = ((int)tt) >> 31; + out[0].real = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); + tt = out[0].imag; + sgn = ((int)tt) >> 31; + out[0].imag = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); + for (i = 1; i < PART_LEN; i++) { + tt = out[i].real + uReal[i]; + sgn = ((int)tt) >> 31; + out[i].real = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); + tt = out[i].imag + uImag[i]; + sgn = ((int)tt) >> 31; + out[i].imag = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); + } + tt = out[PART_LEN].real + uReal[PART_LEN]; + sgn = ((int)tt) >> 31; + out[PART_LEN].real = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); + tt = out[PART_LEN].imag; + sgn = ((int)tt) >> 31; + out[PART_LEN].imag = sgn == (int16_t)(tt >> 15) ? (int16_t)tt : (16384 ^ sgn); +} + diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_neon.c b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_neon.c new file mode 100644 index 00000000..1751fcf7 --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_core_neon.c @@ -0,0 +1,212 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/aecm/aecm_core.h" + +#include <arm_neon.h> +#include <assert.h> + +#include "webrtc/common_audio/signal_processing/include/real_fft.h" + +// TODO(kma): Re-write the corresponding assembly file, the offset +// generating script and makefile, to replace these C functions. + +// Square root of Hanning window in Q14. +const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = { + 0, + 399, 798, 1196, 1594, 1990, 2386, 2780, 3172, + 3562, 3951, 4337, 4720, 5101, 5478, 5853, 6224, + 6591, 6954, 7313, 7668, 8019, 8364, 8705, 9040, + 9370, 9695, 10013, 10326, 10633, 10933, 11227, 11514, + 11795, 12068, 12335, 12594, 12845, 13089, 13325, 13553, + 13773, 13985, 14189, 14384, 14571, 14749, 14918, 15079, + 15231, 15373, 15506, 15631, 15746, 15851, 15947, 16034, + 16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384 +}; + +static inline void AddLanes(uint32_t* ptr, uint32x4_t v) { +#if defined(WEBRTC_ARCH_ARM64) + *(ptr) = vaddvq_u32(v); +#else + uint32x2_t tmp_v; + tmp_v = vadd_u32(vget_low_u32(v), vget_high_u32(v)); + tmp_v = vpadd_u32(tmp_v, tmp_v); + *(ptr) = vget_lane_u32(tmp_v, 0); +#endif +} + +void WebRtcAecm_CalcLinearEnergiesNeon(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est, + uint32_t* far_energy, + uint32_t* echo_energy_adapt, + uint32_t* echo_energy_stored) { + int16_t* start_stored_p = aecm->channelStored; + int16_t* start_adapt_p = aecm->channelAdapt16; + int32_t* echo_est_p = echo_est; + const int16_t* end_stored_p = aecm->channelStored + PART_LEN; + const uint16_t* far_spectrum_p = far_spectrum; + int16x8_t store_v, adapt_v; + uint16x8_t spectrum_v; + uint32x4_t echo_est_v_low, echo_est_v_high; + uint32x4_t far_energy_v, echo_stored_v, echo_adapt_v; + + far_energy_v = vdupq_n_u32(0); + echo_adapt_v = vdupq_n_u32(0); + echo_stored_v = vdupq_n_u32(0); + + // Get energy for the delayed far end signal and estimated + // echo using both stored and adapted channels. + // The C code: + // for (i = 0; i < PART_LEN1; i++) { + // echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + // far_spectrum[i]); + // (*far_energy) += (uint32_t)(far_spectrum[i]); + // *echo_energy_adapt += aecm->channelAdapt16[i] * far_spectrum[i]; + // (*echo_energy_stored) += (uint32_t)echo_est[i]; + // } + while (start_stored_p < end_stored_p) { + spectrum_v = vld1q_u16(far_spectrum_p); + adapt_v = vld1q_s16(start_adapt_p); + store_v = vld1q_s16(start_stored_p); + + far_energy_v = vaddw_u16(far_energy_v, vget_low_u16(spectrum_v)); + far_energy_v = vaddw_u16(far_energy_v, vget_high_u16(spectrum_v)); + + echo_est_v_low = vmull_u16(vreinterpret_u16_s16(vget_low_s16(store_v)), + vget_low_u16(spectrum_v)); + echo_est_v_high = vmull_u16(vreinterpret_u16_s16(vget_high_s16(store_v)), + vget_high_u16(spectrum_v)); + vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low)); + vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high)); + + echo_stored_v = vaddq_u32(echo_est_v_low, echo_stored_v); + echo_stored_v = vaddq_u32(echo_est_v_high, echo_stored_v); + + echo_adapt_v = vmlal_u16(echo_adapt_v, + vreinterpret_u16_s16(vget_low_s16(adapt_v)), + vget_low_u16(spectrum_v)); + echo_adapt_v = vmlal_u16(echo_adapt_v, + vreinterpret_u16_s16(vget_high_s16(adapt_v)), + vget_high_u16(spectrum_v)); + + start_stored_p += 8; + start_adapt_p += 8; + far_spectrum_p += 8; + echo_est_p += 8; + } + + AddLanes(far_energy, far_energy_v); + AddLanes(echo_energy_stored, echo_stored_v); + AddLanes(echo_energy_adapt, echo_adapt_v); + + echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN], + far_spectrum[PART_LEN]); + *echo_energy_stored += (uint32_t)echo_est[PART_LEN]; + *far_energy += (uint32_t)far_spectrum[PART_LEN]; + *echo_energy_adapt += aecm->channelAdapt16[PART_LEN] * far_spectrum[PART_LEN]; +} + +void WebRtcAecm_StoreAdaptiveChannelNeon(AecmCore* aecm, + const uint16_t* far_spectrum, + int32_t* echo_est) { + assert((uintptr_t)echo_est % 32 == 0); + assert((uintptr_t)(aecm->channelStored) % 16 == 0); + assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0); + + // This is C code of following optimized code. + // During startup we store the channel every block. + // memcpy(aecm->channelStored, + // aecm->channelAdapt16, + // sizeof(int16_t) * PART_LEN1); + // Recalculate echo estimate + // for (i = 0; i < PART_LEN; i += 4) { + // echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + // far_spectrum[i]); + // echo_est[i + 1] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 1], + // far_spectrum[i + 1]); + // echo_est[i + 2] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 2], + // far_spectrum[i + 2]); + // echo_est[i + 3] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i + 3], + // far_spectrum[i + 3]); + // } + // echo_est[i] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[i], + // far_spectrum[i]); + const uint16_t* far_spectrum_p = far_spectrum; + int16_t* start_adapt_p = aecm->channelAdapt16; + int16_t* start_stored_p = aecm->channelStored; + const int16_t* end_stored_p = aecm->channelStored + PART_LEN; + int32_t* echo_est_p = echo_est; + + uint16x8_t far_spectrum_v; + int16x8_t adapt_v; + uint32x4_t echo_est_v_low, echo_est_v_high; + + while (start_stored_p < end_stored_p) { + far_spectrum_v = vld1q_u16(far_spectrum_p); + adapt_v = vld1q_s16(start_adapt_p); + + vst1q_s16(start_stored_p, adapt_v); + + echo_est_v_low = vmull_u16(vget_low_u16(far_spectrum_v), + vget_low_u16(vreinterpretq_u16_s16(adapt_v))); + echo_est_v_high = vmull_u16(vget_high_u16(far_spectrum_v), + vget_high_u16(vreinterpretq_u16_s16(adapt_v))); + + vst1q_s32(echo_est_p, vreinterpretq_s32_u32(echo_est_v_low)); + vst1q_s32(echo_est_p + 4, vreinterpretq_s32_u32(echo_est_v_high)); + + far_spectrum_p += 8; + start_adapt_p += 8; + start_stored_p += 8; + echo_est_p += 8; + } + aecm->channelStored[PART_LEN] = aecm->channelAdapt16[PART_LEN]; + echo_est[PART_LEN] = WEBRTC_SPL_MUL_16_U16(aecm->channelStored[PART_LEN], + far_spectrum[PART_LEN]); +} + +void WebRtcAecm_ResetAdaptiveChannelNeon(AecmCore* aecm) { + assert((uintptr_t)(aecm->channelStored) % 16 == 0); + assert((uintptr_t)(aecm->channelAdapt16) % 16 == 0); + assert((uintptr_t)(aecm->channelAdapt32) % 32 == 0); + + // The C code of following optimized code. + // for (i = 0; i < PART_LEN1; i++) { + // aecm->channelAdapt16[i] = aecm->channelStored[i]; + // aecm->channelAdapt32[i] = WEBRTC_SPL_LSHIFT_W32( + // (int32_t)aecm->channelStored[i], 16); + // } + + int16_t* start_stored_p = aecm->channelStored; + int16_t* start_adapt16_p = aecm->channelAdapt16; + int32_t* start_adapt32_p = aecm->channelAdapt32; + const int16_t* end_stored_p = start_stored_p + PART_LEN; + + int16x8_t stored_v; + int32x4_t adapt32_v_low, adapt32_v_high; + + while (start_stored_p < end_stored_p) { + stored_v = vld1q_s16(start_stored_p); + vst1q_s16(start_adapt16_p, stored_v); + + adapt32_v_low = vshll_n_s16(vget_low_s16(stored_v), 16); + adapt32_v_high = vshll_n_s16(vget_high_s16(stored_v), 16); + + vst1q_s32(start_adapt32_p, adapt32_v_low); + vst1q_s32(start_adapt32_p + 4, adapt32_v_high); + + start_stored_p += 8; + start_adapt16_p += 8; + start_adapt32_p += 8; + } + aecm->channelAdapt16[PART_LEN] = aecm->channelStored[PART_LEN]; + aecm->channelAdapt32[PART_LEN] = (int32_t)aecm->channelStored[PART_LEN] << 16; +} diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_defines.h b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_defines.h new file mode 100644 index 00000000..6d63990b --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/aecm_defines.h @@ -0,0 +1,87 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AECM_AECM_DEFINES_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_AECM_AECM_DEFINES_H_ + +#define AECM_DYNAMIC_Q /* Turn on/off dynamic Q-domain. */ + +/* Algorithm parameters */ +#define FRAME_LEN 80 /* Total frame length, 10 ms. */ + +#define PART_LEN 64 /* Length of partition. */ +#define PART_LEN_SHIFT 7 /* Length of (PART_LEN * 2) in base 2. */ + +#define PART_LEN1 (PART_LEN + 1) /* Unique fft coefficients. */ +#define PART_LEN2 (PART_LEN << 1) /* Length of partition * 2. */ +#define PART_LEN4 (PART_LEN << 2) /* Length of partition * 4. */ +#define FAR_BUF_LEN PART_LEN4 /* Length of buffers. */ +#define MAX_DELAY 100 + +/* Counter parameters */ +#define CONV_LEN 512 /* Convergence length used at startup. */ +#define CONV_LEN2 (CONV_LEN << 1) /* Used at startup. */ + +/* Energy parameters */ +#define MAX_BUF_LEN 64 /* History length of energy signals. */ +#define FAR_ENERGY_MIN 1025 /* Lowest Far energy level: At least 2 */ + /* in energy. */ +#define FAR_ENERGY_DIFF 929 /* Allowed difference between max */ + /* and min. */ +#define ENERGY_DEV_OFFSET 0 /* The energy error offset in Q8. */ +#define ENERGY_DEV_TOL 400 /* The energy estimation tolerance (Q8). */ +#define FAR_ENERGY_VAD_REGION 230 /* Far VAD tolerance region. */ + +/* Stepsize parameters */ +#define MU_MIN 10 /* Min stepsize 2^-MU_MIN (far end energy */ + /* dependent). */ +#define MU_MAX 1 /* Max stepsize 2^-MU_MAX (far end energy */ + /* dependent). */ +#define MU_DIFF 9 /* MU_MIN - MU_MAX */ + +/* Channel parameters */ +#define MIN_MSE_COUNT 20 /* Min number of consecutive blocks with enough */ + /* far end energy to compare channel estimates. */ +#define MIN_MSE_DIFF 29 /* The ratio between adapted and stored channel to */ + /* accept a new storage (0.8 in Q-MSE_RESOLUTION). */ +#define MSE_RESOLUTION 5 /* MSE parameter resolution. */ +#define RESOLUTION_CHANNEL16 12 /* W16 Channel in Q-RESOLUTION_CHANNEL16. */ +#define RESOLUTION_CHANNEL32 28 /* W32 Channel in Q-RESOLUTION_CHANNEL. */ +#define CHANNEL_VAD 16 /* Minimum energy in frequency band */ + /* to update channel. */ + +/* Suppression gain parameters: SUPGAIN parameters in Q-(RESOLUTION_SUPGAIN). */ +#define RESOLUTION_SUPGAIN 8 /* Channel in Q-(RESOLUTION_SUPGAIN). */ +#define SUPGAIN_DEFAULT (1 << RESOLUTION_SUPGAIN) /* Default. */ +#define SUPGAIN_ERROR_PARAM_A 3072 /* Estimation error parameter */ + /* (Maximum gain) (8 in Q8). */ +#define SUPGAIN_ERROR_PARAM_B 1536 /* Estimation error parameter */ + /* (Gain before going down). */ +#define SUPGAIN_ERROR_PARAM_D SUPGAIN_DEFAULT /* Estimation error parameter */ + /* (Should be the same as Default) (1 in Q8). */ +#define SUPGAIN_EPC_DT 200 /* SUPGAIN_ERROR_PARAM_C * ENERGY_DEV_TOL */ + +/* Defines for "check delay estimation" */ +#define CORR_WIDTH 31 /* Number of samples to correlate over. */ +#define CORR_MAX 16 /* Maximum correlation offset. */ +#define CORR_MAX_BUF 63 +#define CORR_DEV 4 +#define CORR_MAX_LEVEL 20 +#define CORR_MAX_LOW 4 +#define CORR_BUF_LEN (CORR_MAX << 1) + 1 +/* Note that CORR_WIDTH + 2*CORR_MAX <= MAX_BUF_LEN. */ + +#define ONE_Q14 (1 << 14) + +/* NLP defines */ +#define NLP_COMP_LOW 3277 /* 0.2 in Q14 */ +#define NLP_COMP_HIGH ONE_Q14 /* 1 in Q14 */ + +#endif diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/echo_control_mobile.c b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/echo_control_mobile.c new file mode 100644 index 00000000..83781e97 --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/echo_control_mobile.c @@ -0,0 +1,702 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h" + +#ifdef AEC_DEBUG +#include <stdio.h> +#endif +#include <stdlib.h> + +#include "webrtc/common_audio/ring_buffer.h" +#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h" +#include "webrtc/modules/audio_processing/aecm/aecm_core.h" + +#define BUF_SIZE_FRAMES 50 // buffer size (frames) +// Maximum length of resampled signal. Must be an integer multiple of frames +// (ceil(1/(1 + MIN_SKEW)*2) + 1)*FRAME_LEN +// The factor of 2 handles wb, and the + 1 is as a safety margin +#define MAX_RESAMP_LEN (5 * FRAME_LEN) + +static const size_t kBufSizeSamp = BUF_SIZE_FRAMES * FRAME_LEN; // buffer size (samples) +static const int kSampMsNb = 8; // samples per ms in nb +// Target suppression levels for nlp modes +// log{0.001, 0.00001, 0.00000001} +static const int kInitCheck = 42; + +typedef struct +{ + int sampFreq; + int scSampFreq; + short bufSizeStart; + int knownDelay; + + // Stores the last frame added to the farend buffer + short farendOld[2][FRAME_LEN]; + short initFlag; // indicates if AEC has been initialized + + // Variables used for averaging far end buffer size + short counter; + short sum; + short firstVal; + short checkBufSizeCtr; + + // Variables used for delay shifts + short msInSndCardBuf; + short filtDelay; + int timeForDelayChange; + int ECstartup; + int checkBuffSize; + int delayChange; + short lastDelayDiff; + + int16_t echoMode; + +#ifdef AEC_DEBUG + FILE *bufFile; + FILE *delayFile; + FILE *preCompFile; + FILE *postCompFile; +#endif // AEC_DEBUG + // Structures + RingBuffer *farendBuf; + + int lastError; + + AecmCore* aecmCore; +} AecMobile; + +// Estimates delay to set the position of the farend buffer read pointer +// (controlled by knownDelay) +static int WebRtcAecm_EstBufDelay(AecMobile* aecmInst, short msInSndCardBuf); + +// Stuffs the farend buffer if the estimated delay is too large +static int WebRtcAecm_DelayComp(AecMobile* aecmInst); + +void* WebRtcAecm_Create() { + AecMobile* aecm = malloc(sizeof(AecMobile)); + + WebRtcSpl_Init(); + + aecm->aecmCore = WebRtcAecm_CreateCore(); + if (!aecm->aecmCore) { + WebRtcAecm_Free(aecm); + return NULL; + } + + aecm->farendBuf = WebRtc_CreateBuffer(kBufSizeSamp, + sizeof(int16_t)); + if (!aecm->farendBuf) + { + WebRtcAecm_Free(aecm); + return NULL; + } + + aecm->initFlag = 0; + aecm->lastError = 0; + +#ifdef AEC_DEBUG + aecm->aecmCore->farFile = fopen("aecFar.pcm","wb"); + aecm->aecmCore->nearFile = fopen("aecNear.pcm","wb"); + aecm->aecmCore->outFile = fopen("aecOut.pcm","wb"); + //aecm->aecmCore->outLpFile = fopen("aecOutLp.pcm","wb"); + + aecm->bufFile = fopen("aecBuf.dat", "wb"); + aecm->delayFile = fopen("aecDelay.dat", "wb"); + aecm->preCompFile = fopen("preComp.pcm", "wb"); + aecm->postCompFile = fopen("postComp.pcm", "wb"); +#endif // AEC_DEBUG + return aecm; +} + +void WebRtcAecm_Free(void* aecmInst) { + AecMobile* aecm = aecmInst; + + if (aecm == NULL) { + return; + } + +#ifdef AEC_DEBUG + fclose(aecm->aecmCore->farFile); + fclose(aecm->aecmCore->nearFile); + fclose(aecm->aecmCore->outFile); + //fclose(aecm->aecmCore->outLpFile); + + fclose(aecm->bufFile); + fclose(aecm->delayFile); + fclose(aecm->preCompFile); + fclose(aecm->postCompFile); +#endif // AEC_DEBUG + WebRtcAecm_FreeCore(aecm->aecmCore); + WebRtc_FreeBuffer(aecm->farendBuf); + free(aecm); +} + +int32_t WebRtcAecm_Init(void *aecmInst, int32_t sampFreq) +{ + AecMobile* aecm = aecmInst; + AecmConfig aecConfig; + + if (aecm == NULL) + { + return -1; + } + + if (sampFreq != 8000 && sampFreq != 16000) + { + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + aecm->sampFreq = sampFreq; + + // Initialize AECM core + if (WebRtcAecm_InitCore(aecm->aecmCore, aecm->sampFreq) == -1) + { + aecm->lastError = AECM_UNSPECIFIED_ERROR; + return -1; + } + + // Initialize farend buffer + WebRtc_InitBuffer(aecm->farendBuf); + + aecm->initFlag = kInitCheck; // indicates that initialization has been done + + aecm->delayChange = 1; + + aecm->sum = 0; + aecm->counter = 0; + aecm->checkBuffSize = 1; + aecm->firstVal = 0; + + aecm->ECstartup = 1; + aecm->bufSizeStart = 0; + aecm->checkBufSizeCtr = 0; + aecm->filtDelay = 0; + aecm->timeForDelayChange = 0; + aecm->knownDelay = 0; + aecm->lastDelayDiff = 0; + + memset(&aecm->farendOld[0][0], 0, 160); + + // Default settings. + aecConfig.cngMode = AecmTrue; + aecConfig.echoMode = 3; + + if (WebRtcAecm_set_config(aecm, aecConfig) == -1) + { + aecm->lastError = AECM_UNSPECIFIED_ERROR; + return -1; + } + + return 0; +} + +int32_t WebRtcAecm_BufferFarend(void *aecmInst, const int16_t *farend, + size_t nrOfSamples) +{ + AecMobile* aecm = aecmInst; + int32_t retVal = 0; + + if (aecm == NULL) + { + return -1; + } + + if (farend == NULL) + { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + if (nrOfSamples != 80 && nrOfSamples != 160) + { + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + + // TODO: Is this really a good idea? + if (!aecm->ECstartup) + { + WebRtcAecm_DelayComp(aecm); + } + + WebRtc_WriteBuffer(aecm->farendBuf, farend, nrOfSamples); + + return retVal; +} + +int32_t WebRtcAecm_Process(void *aecmInst, const int16_t *nearendNoisy, + const int16_t *nearendClean, int16_t *out, + size_t nrOfSamples, int16_t msInSndCardBuf) +{ + AecMobile* aecm = aecmInst; + int32_t retVal = 0; + size_t i; + short nmbrOfFilledBuffers; + size_t nBlocks10ms; + size_t nFrames; +#ifdef AEC_DEBUG + short msInAECBuf; +#endif + + if (aecm == NULL) + { + return -1; + } + + if (nearendNoisy == NULL) + { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + + if (out == NULL) + { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + if (nrOfSamples != 80 && nrOfSamples != 160) + { + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + + if (msInSndCardBuf < 0) + { + msInSndCardBuf = 0; + aecm->lastError = AECM_BAD_PARAMETER_WARNING; + retVal = -1; + } else if (msInSndCardBuf > 500) + { + msInSndCardBuf = 500; + aecm->lastError = AECM_BAD_PARAMETER_WARNING; + retVal = -1; + } + msInSndCardBuf += 10; + aecm->msInSndCardBuf = msInSndCardBuf; + + nFrames = nrOfSamples / FRAME_LEN; + nBlocks10ms = nFrames / aecm->aecmCore->mult; + + if (aecm->ECstartup) + { + if (nearendClean == NULL) + { + if (out != nearendNoisy) + { + memcpy(out, nearendNoisy, sizeof(short) * nrOfSamples); + } + } else if (out != nearendClean) + { + memcpy(out, nearendClean, sizeof(short) * nrOfSamples); + } + + nmbrOfFilledBuffers = + (short) WebRtc_available_read(aecm->farendBuf) / FRAME_LEN; + // The AECM is in the start up mode + // AECM is disabled until the soundcard buffer and farend buffers are OK + + // Mechanism to ensure that the soundcard buffer is reasonably stable. + if (aecm->checkBuffSize) + { + aecm->checkBufSizeCtr++; + // Before we fill up the far end buffer we require the amount of data on the + // sound card to be stable (+/-8 ms) compared to the first value. This + // comparison is made during the following 4 consecutive frames. If it seems + // to be stable then we start to fill up the far end buffer. + + if (aecm->counter == 0) + { + aecm->firstVal = aecm->msInSndCardBuf; + aecm->sum = 0; + } + + if (abs(aecm->firstVal - aecm->msInSndCardBuf) + < WEBRTC_SPL_MAX(0.2 * aecm->msInSndCardBuf, kSampMsNb)) + { + aecm->sum += aecm->msInSndCardBuf; + aecm->counter++; + } else + { + aecm->counter = 0; + } + + if (aecm->counter * nBlocks10ms >= 6) + { + // The farend buffer size is determined in blocks of 80 samples + // Use 75% of the average value of the soundcard buffer + aecm->bufSizeStart + = WEBRTC_SPL_MIN((3 * aecm->sum + * aecm->aecmCore->mult) / (aecm->counter * 40), BUF_SIZE_FRAMES); + // buffersize has now been determined + aecm->checkBuffSize = 0; + } + + if (aecm->checkBufSizeCtr * nBlocks10ms > 50) + { + // for really bad sound cards, don't disable echocanceller for more than 0.5 sec + aecm->bufSizeStart = WEBRTC_SPL_MIN((3 * aecm->msInSndCardBuf + * aecm->aecmCore->mult) / 40, BUF_SIZE_FRAMES); + aecm->checkBuffSize = 0; + } + } + + // if checkBuffSize changed in the if-statement above + if (!aecm->checkBuffSize) + { + // soundcard buffer is now reasonably stable + // When the far end buffer is filled with approximately the same amount of + // data as the amount on the sound card we end the start up phase and start + // to cancel echoes. + + if (nmbrOfFilledBuffers == aecm->bufSizeStart) + { + aecm->ECstartup = 0; // Enable the AECM + } else if (nmbrOfFilledBuffers > aecm->bufSizeStart) + { + WebRtc_MoveReadPtr(aecm->farendBuf, + (int) WebRtc_available_read(aecm->farendBuf) + - (int) aecm->bufSizeStart * FRAME_LEN); + aecm->ECstartup = 0; + } + } + + } else + { + // AECM is enabled + + // Note only 1 block supported for nb and 2 blocks for wb + for (i = 0; i < nFrames; i++) + { + int16_t farend[FRAME_LEN]; + const int16_t* farend_ptr = NULL; + + nmbrOfFilledBuffers = + (short) WebRtc_available_read(aecm->farendBuf) / FRAME_LEN; + + // Check that there is data in the far end buffer + if (nmbrOfFilledBuffers > 0) + { + // Get the next 80 samples from the farend buffer + WebRtc_ReadBuffer(aecm->farendBuf, (void**) &farend_ptr, farend, + FRAME_LEN); + + // Always store the last frame for use when we run out of data + memcpy(&(aecm->farendOld[i][0]), farend_ptr, + FRAME_LEN * sizeof(short)); + } else + { + // We have no data so we use the last played frame + memcpy(farend, &(aecm->farendOld[i][0]), FRAME_LEN * sizeof(short)); + farend_ptr = farend; + } + + // Call buffer delay estimator when all data is extracted, + // i,e. i = 0 for NB and i = 1 for WB + if ((i == 0 && aecm->sampFreq == 8000) || (i == 1 && aecm->sampFreq == 16000)) + { + WebRtcAecm_EstBufDelay(aecm, aecm->msInSndCardBuf); + } + + // Call the AECM + /*WebRtcAecm_ProcessFrame(aecm->aecmCore, farend, &nearend[FRAME_LEN * i], + &out[FRAME_LEN * i], aecm->knownDelay);*/ + if (WebRtcAecm_ProcessFrame(aecm->aecmCore, + farend_ptr, + &nearendNoisy[FRAME_LEN * i], + (nearendClean + ? &nearendClean[FRAME_LEN * i] + : NULL), + &out[FRAME_LEN * i]) == -1) + return -1; + } + } + +#ifdef AEC_DEBUG + msInAECBuf = (short) WebRtc_available_read(aecm->farendBuf) / + (kSampMsNb * aecm->aecmCore->mult); + fwrite(&msInAECBuf, 2, 1, aecm->bufFile); + fwrite(&(aecm->knownDelay), sizeof(aecm->knownDelay), 1, aecm->delayFile); +#endif + + return retVal; +} + +int32_t WebRtcAecm_set_config(void *aecmInst, AecmConfig config) +{ + AecMobile* aecm = aecmInst; + + if (aecm == NULL) + { + return -1; + } + + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + if (config.cngMode != AecmFalse && config.cngMode != AecmTrue) + { + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + aecm->aecmCore->cngMode = config.cngMode; + + if (config.echoMode < 0 || config.echoMode > 4) + { + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + aecm->echoMode = config.echoMode; + + if (aecm->echoMode == 0) + { + aecm->aecmCore->supGain = SUPGAIN_DEFAULT >> 3; + aecm->aecmCore->supGainOld = SUPGAIN_DEFAULT >> 3; + aecm->aecmCore->supGainErrParamA = SUPGAIN_ERROR_PARAM_A >> 3; + aecm->aecmCore->supGainErrParamD = SUPGAIN_ERROR_PARAM_D >> 3; + aecm->aecmCore->supGainErrParamDiffAB = (SUPGAIN_ERROR_PARAM_A >> 3) + - (SUPGAIN_ERROR_PARAM_B >> 3); + aecm->aecmCore->supGainErrParamDiffBD = (SUPGAIN_ERROR_PARAM_B >> 3) + - (SUPGAIN_ERROR_PARAM_D >> 3); + } else if (aecm->echoMode == 1) + { + aecm->aecmCore->supGain = SUPGAIN_DEFAULT >> 2; + aecm->aecmCore->supGainOld = SUPGAIN_DEFAULT >> 2; + aecm->aecmCore->supGainErrParamA = SUPGAIN_ERROR_PARAM_A >> 2; + aecm->aecmCore->supGainErrParamD = SUPGAIN_ERROR_PARAM_D >> 2; + aecm->aecmCore->supGainErrParamDiffAB = (SUPGAIN_ERROR_PARAM_A >> 2) + - (SUPGAIN_ERROR_PARAM_B >> 2); + aecm->aecmCore->supGainErrParamDiffBD = (SUPGAIN_ERROR_PARAM_B >> 2) + - (SUPGAIN_ERROR_PARAM_D >> 2); + } else if (aecm->echoMode == 2) + { + aecm->aecmCore->supGain = SUPGAIN_DEFAULT >> 1; + aecm->aecmCore->supGainOld = SUPGAIN_DEFAULT >> 1; + aecm->aecmCore->supGainErrParamA = SUPGAIN_ERROR_PARAM_A >> 1; + aecm->aecmCore->supGainErrParamD = SUPGAIN_ERROR_PARAM_D >> 1; + aecm->aecmCore->supGainErrParamDiffAB = (SUPGAIN_ERROR_PARAM_A >> 1) + - (SUPGAIN_ERROR_PARAM_B >> 1); + aecm->aecmCore->supGainErrParamDiffBD = (SUPGAIN_ERROR_PARAM_B >> 1) + - (SUPGAIN_ERROR_PARAM_D >> 1); + } else if (aecm->echoMode == 3) + { + aecm->aecmCore->supGain = SUPGAIN_DEFAULT; + aecm->aecmCore->supGainOld = SUPGAIN_DEFAULT; + aecm->aecmCore->supGainErrParamA = SUPGAIN_ERROR_PARAM_A; + aecm->aecmCore->supGainErrParamD = SUPGAIN_ERROR_PARAM_D; + aecm->aecmCore->supGainErrParamDiffAB = SUPGAIN_ERROR_PARAM_A - SUPGAIN_ERROR_PARAM_B; + aecm->aecmCore->supGainErrParamDiffBD = SUPGAIN_ERROR_PARAM_B - SUPGAIN_ERROR_PARAM_D; + } else if (aecm->echoMode == 4) + { + aecm->aecmCore->supGain = SUPGAIN_DEFAULT << 1; + aecm->aecmCore->supGainOld = SUPGAIN_DEFAULT << 1; + aecm->aecmCore->supGainErrParamA = SUPGAIN_ERROR_PARAM_A << 1; + aecm->aecmCore->supGainErrParamD = SUPGAIN_ERROR_PARAM_D << 1; + aecm->aecmCore->supGainErrParamDiffAB = (SUPGAIN_ERROR_PARAM_A << 1) + - (SUPGAIN_ERROR_PARAM_B << 1); + aecm->aecmCore->supGainErrParamDiffBD = (SUPGAIN_ERROR_PARAM_B << 1) + - (SUPGAIN_ERROR_PARAM_D << 1); + } + + return 0; +} + +int32_t WebRtcAecm_get_config(void *aecmInst, AecmConfig *config) +{ + AecMobile* aecm = aecmInst; + + if (aecm == NULL) + { + return -1; + } + + if (config == NULL) + { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + config->cngMode = aecm->aecmCore->cngMode; + config->echoMode = aecm->echoMode; + + return 0; +} + +int32_t WebRtcAecm_InitEchoPath(void* aecmInst, + const void* echo_path, + size_t size_bytes) +{ + AecMobile* aecm = aecmInst; + const int16_t* echo_path_ptr = echo_path; + + if (aecmInst == NULL) { + return -1; + } + if (echo_path == NULL) { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + if (size_bytes != WebRtcAecm_echo_path_size_bytes()) + { + // Input channel size does not match the size of AECM + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + WebRtcAecm_InitEchoPathCore(aecm->aecmCore, echo_path_ptr); + + return 0; +} + +int32_t WebRtcAecm_GetEchoPath(void* aecmInst, + void* echo_path, + size_t size_bytes) +{ + AecMobile* aecm = aecmInst; + int16_t* echo_path_ptr = echo_path; + + if (aecmInst == NULL) { + return -1; + } + if (echo_path == NULL) { + aecm->lastError = AECM_NULL_POINTER_ERROR; + return -1; + } + if (size_bytes != WebRtcAecm_echo_path_size_bytes()) + { + // Input channel size does not match the size of AECM + aecm->lastError = AECM_BAD_PARAMETER_ERROR; + return -1; + } + if (aecm->initFlag != kInitCheck) + { + aecm->lastError = AECM_UNINITIALIZED_ERROR; + return -1; + } + + memcpy(echo_path_ptr, aecm->aecmCore->channelStored, size_bytes); + return 0; +} + +size_t WebRtcAecm_echo_path_size_bytes() +{ + return (PART_LEN1 * sizeof(int16_t)); +} + +int32_t WebRtcAecm_get_error_code(void *aecmInst) +{ + AecMobile* aecm = aecmInst; + + if (aecm == NULL) + { + return -1; + } + + return aecm->lastError; +} + +static int WebRtcAecm_EstBufDelay(AecMobile* aecm, short msInSndCardBuf) { + short delayNew, nSampSndCard; + short nSampFar = (short) WebRtc_available_read(aecm->farendBuf); + short diff; + + nSampSndCard = msInSndCardBuf * kSampMsNb * aecm->aecmCore->mult; + + delayNew = nSampSndCard - nSampFar; + + if (delayNew < FRAME_LEN) + { + WebRtc_MoveReadPtr(aecm->farendBuf, FRAME_LEN); + delayNew += FRAME_LEN; + } + + aecm->filtDelay = WEBRTC_SPL_MAX(0, (8 * aecm->filtDelay + 2 * delayNew) / 10); + + diff = aecm->filtDelay - aecm->knownDelay; + if (diff > 224) + { + if (aecm->lastDelayDiff < 96) + { + aecm->timeForDelayChange = 0; + } else + { + aecm->timeForDelayChange++; + } + } else if (diff < 96 && aecm->knownDelay > 0) + { + if (aecm->lastDelayDiff > 224) + { + aecm->timeForDelayChange = 0; + } else + { + aecm->timeForDelayChange++; + } + } else + { + aecm->timeForDelayChange = 0; + } + aecm->lastDelayDiff = diff; + + if (aecm->timeForDelayChange > 25) + { + aecm->knownDelay = WEBRTC_SPL_MAX((int)aecm->filtDelay - 160, 0); + } + return 0; +} + +static int WebRtcAecm_DelayComp(AecMobile* aecm) { + int nSampFar = (int) WebRtc_available_read(aecm->farendBuf); + int nSampSndCard, delayNew, nSampAdd; + const int maxStuffSamp = 10 * FRAME_LEN; + + nSampSndCard = aecm->msInSndCardBuf * kSampMsNb * aecm->aecmCore->mult; + delayNew = nSampSndCard - nSampFar; + + if (delayNew > FAR_BUF_LEN - FRAME_LEN * aecm->aecmCore->mult) + { + // The difference of the buffer sizes is larger than the maximum + // allowed known delay. Compensate by stuffing the buffer. + nSampAdd = (int)(WEBRTC_SPL_MAX(((nSampSndCard >> 1) - nSampFar), + FRAME_LEN)); + nSampAdd = WEBRTC_SPL_MIN(nSampAdd, maxStuffSamp); + + WebRtc_MoveReadPtr(aecm->farendBuf, -nSampAdd); + aecm->delayChange = 1; // the delay needs to be updated + } + + return 0; +} diff --git a/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h new file mode 100644 index 00000000..7ae15c2a --- /dev/null +++ b/third_party/webrtc/src/webrtc/modules/audio_processing/aecm/include/echo_control_mobile.h @@ -0,0 +1,218 @@ +/* + * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_AECM_INCLUDE_ECHO_CONTROL_MOBILE_H_ +#define WEBRTC_MODULES_AUDIO_PROCESSING_AECM_INCLUDE_ECHO_CONTROL_MOBILE_H_ + +#include <stdlib.h> + +#include "webrtc/typedefs.h" + +enum { + AecmFalse = 0, + AecmTrue +}; + +// Errors +#define AECM_UNSPECIFIED_ERROR 12000 +#define AECM_UNSUPPORTED_FUNCTION_ERROR 12001 +#define AECM_UNINITIALIZED_ERROR 12002 +#define AECM_NULL_POINTER_ERROR 12003 +#define AECM_BAD_PARAMETER_ERROR 12004 + +// Warnings +#define AECM_BAD_PARAMETER_WARNING 12100 + +typedef struct { + int16_t cngMode; // AECM_FALSE, AECM_TRUE (default) + int16_t echoMode; // 0, 1, 2, 3 (default), 4 +} AecmConfig; + +#ifdef __cplusplus +extern "C" { +#endif + +/* + * Allocates the memory needed by the AECM. The memory needs to be + * initialized separately using the WebRtcAecm_Init() function. + * Returns a pointer to the instance and a nullptr at failure. + */ +void* WebRtcAecm_Create(); + +/* + * This function releases the memory allocated by WebRtcAecm_Create() + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + */ +void WebRtcAecm_Free(void* aecmInst); + +/* + * Initializes an AECM instance. + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * int32_t sampFreq Sampling frequency of data + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_Init(void* aecmInst, int32_t sampFreq); + +/* + * Inserts an 80 or 160 sample block of data into the farend buffer. + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * int16_t* farend In buffer containing one frame of + * farend signal + * int16_t nrOfSamples Number of samples in farend buffer + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_BufferFarend(void* aecmInst, + const int16_t* farend, + size_t nrOfSamples); + +/* + * Runs the AECM on an 80 or 160 sample blocks of data. + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * int16_t* nearendNoisy In buffer containing one frame of + * reference nearend+echo signal. If + * noise reduction is active, provide + * the noisy signal here. + * int16_t* nearendClean In buffer containing one frame of + * nearend+echo signal. If noise + * reduction is active, provide the + * clean signal here. Otherwise pass a + * NULL pointer. + * int16_t nrOfSamples Number of samples in nearend buffer + * int16_t msInSndCardBuf Delay estimate for sound card and + * system buffers + * + * Outputs Description + * ------------------------------------------------------------------- + * int16_t* out Out buffer, one frame of processed nearend + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_Process(void* aecmInst, + const int16_t* nearendNoisy, + const int16_t* nearendClean, + int16_t* out, + size_t nrOfSamples, + int16_t msInSndCardBuf); + +/* + * This function enables the user to set certain parameters on-the-fly + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * AecmConfig config Config instance that contains all + * properties to be set + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_set_config(void* aecmInst, AecmConfig config); + +/* + * This function enables the user to set certain parameters on-the-fly + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * + * Outputs Description + * ------------------------------------------------------------------- + * AecmConfig* config Pointer to the config instance that + * all properties will be written to + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_get_config(void *aecmInst, AecmConfig *config); + +/* + * This function enables the user to set the echo path on-the-fly. + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * void* echo_path Pointer to the echo path to be set + * size_t size_bytes Size in bytes of the echo path + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_InitEchoPath(void* aecmInst, + const void* echo_path, + size_t size_bytes); + +/* + * This function enables the user to get the currently used echo path + * on-the-fly + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * void* echo_path Pointer to echo path + * size_t size_bytes Size in bytes of the echo path + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 0: OK + * -1: error + */ +int32_t WebRtcAecm_GetEchoPath(void* aecmInst, + void* echo_path, + size_t size_bytes); + +/* + * This function enables the user to get the echo path size in bytes + * + * Outputs Description + * ------------------------------------------------------------------- + * size_t return Size in bytes + */ +size_t WebRtcAecm_echo_path_size_bytes(); + +/* + * Gets the last error code. + * + * Inputs Description + * ------------------------------------------------------------------- + * void* aecmInst Pointer to the AECM instance + * + * Outputs Description + * ------------------------------------------------------------------- + * int32_t return 11000-11100: error code + */ +int32_t WebRtcAecm_get_error_code(void *aecmInst); + +#ifdef __cplusplus +} +#endif +#endif // WEBRTC_MODULES_AUDIO_PROCESSING_AECM_INCLUDE_ECHO_CONTROL_MOBILE_H_ |