adaptive_fir_filter.cc | Explore in Territory

/*
 *  Copyright (c) 2017 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 "modules/audio_processing/aec3/adaptive_fir_filter.h"

// Defines WEBRTC_ARCH_X86_FAMILY, used below.
#include "rtc_base/system/arch.h"

#if defined(WEBRTC_HAS_NEON)
#include <arm_neon.h>
#endif
#if defined(WEBRTC_ARCH_X86_FAMILY)
#include <emmintrin.h>
#endif
#include <math.h>

#include <algorithm>
#include <functional>

#include "modules/audio_processing/aec3/fft_data.h"
#include "rtc_base/checks.h"

namespace webrtc {

namespace aec3 {

// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse(
    size_t num_partitions,
    const std::vector<std::vector<FftData>>& H,
    std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) { … }

#if defined(WEBRTC_HAS_NEON)
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse_Neon(
    size_t num_partitions,
    const std::vector<std::vector<FftData>>& H,
    std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) {
  for (auto& H2_ch : *H2) {
    H2_ch.fill(0.f);
  }

  const size_t num_render_channels = H[0].size();
  RTC_DCHECK_EQ(H.size(), H2->capacity());
  for (size_t p = 0; p < num_partitions; ++p) {
    RTC_DCHECK_EQ(kFftLengthBy2Plus1, (*H2)[p].size());
    auto& H2_p = (*H2)[p];
    for (size_t ch = 0; ch < num_render_channels; ++ch) {
      const FftData& H_p_ch = H[p][ch];
      for (size_t j = 0; j < kFftLengthBy2; j += 4) {
        const float32x4_t re = vld1q_f32(&H_p_ch.re[j]);
        const float32x4_t im = vld1q_f32(&H_p_ch.im[j]);
        float32x4_t H2_new = vmulq_f32(re, re);
        H2_new = vmlaq_f32(H2_new, im, im);
        float32x4_t H2_p_j = vld1q_f32(&H2_p[j]);
        H2_p_j = vmaxq_f32(H2_p_j, H2_new);
        vst1q_f32(&H2_p[j], H2_p_j);
      }
      float H2_new = H_p_ch.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] +
                     H_p_ch.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2];
      H2_p[kFftLengthBy2] = std::max(H2_p[kFftLengthBy2], H2_new);
    }
  }
}
#endif

#if defined(WEBRTC_ARCH_X86_FAMILY)
// Computes and stores the frequency response of the filter.
void ComputeFrequencyResponse_Sse2(
    size_t num_partitions,
    const std::vector<std::vector<FftData>>& H,
    std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) { … }
#endif

// Adapts the filter partitions as H(t+1)=H(t)+G(t)*conj(X(t)).
void AdaptPartitions(const RenderBuffer& render_buffer,
                     const FftData& G,
                     size_t num_partitions,
                     std::vector<std::vector<FftData>>* H) { … }

#if defined(WEBRTC_HAS_NEON)
// Adapts the filter partitions. (Neon variant)
void AdaptPartitions_Neon(const RenderBuffer& render_buffer,
                          const FftData& G,
                          size_t num_partitions,
                          std::vector<std::vector<FftData>>* H) {
  rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
      render_buffer.GetFftBuffer();
  const size_t num_render_channels = render_buffer_data[0].size();
  const size_t lim1 = std::min(
      render_buffer_data.size() - render_buffer.Position(), num_partitions);
  const size_t lim2 = num_partitions;
  constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;

  size_t X_partition = render_buffer.Position();
  size_t limit = lim1;
  size_t p = 0;
  do {
    for (; p < limit; ++p, ++X_partition) {
      for (size_t ch = 0; ch < num_render_channels; ++ch) {
        FftData& H_p_ch = (*H)[p][ch];
        const FftData& X = render_buffer_data[X_partition][ch];
        for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
          const float32x4_t G_re = vld1q_f32(&G.re[k]);
          const float32x4_t G_im = vld1q_f32(&G.im[k]);
          const float32x4_t X_re = vld1q_f32(&X.re[k]);
          const float32x4_t X_im = vld1q_f32(&X.im[k]);
          const float32x4_t H_re = vld1q_f32(&H_p_ch.re[k]);
          const float32x4_t H_im = vld1q_f32(&H_p_ch.im[k]);
          const float32x4_t a = vmulq_f32(X_re, G_re);
          const float32x4_t e = vmlaq_f32(a, X_im, G_im);
          const float32x4_t c = vmulq_f32(X_re, G_im);
          const float32x4_t f = vmlsq_f32(c, X_im, G_re);
          const float32x4_t g = vaddq_f32(H_re, e);
          const float32x4_t h = vaddq_f32(H_im, f);
          vst1q_f32(&H_p_ch.re[k], g);
          vst1q_f32(&H_p_ch.im[k], h);
        }
      }
    }

    X_partition = 0;
    limit = lim2;
  } while (p < lim2);

  X_partition = render_buffer.Position();
  limit = lim1;
  p = 0;
  do {
    for (; p < limit; ++p, ++X_partition) {
      for (size_t ch = 0; ch < num_render_channels; ++ch) {
        FftData& H_p_ch = (*H)[p][ch];
        const FftData& X = render_buffer_data[X_partition][ch];

        H_p_ch.re[kFftLengthBy2] += X.re[kFftLengthBy2] * G.re[kFftLengthBy2] +
                                    X.im[kFftLengthBy2] * G.im[kFftLengthBy2];
        H_p_ch.im[kFftLengthBy2] += X.re[kFftLengthBy2] * G.im[kFftLengthBy2] -
                                    X.im[kFftLengthBy2] * G.re[kFftLengthBy2];
      }
    }
    X_partition = 0;
    limit = lim2;
  } while (p < lim2);
}
#endif

#if defined(WEBRTC_ARCH_X86_FAMILY)
// Adapts the filter partitions. (SSE2 variant)
void AdaptPartitions_Sse2(const RenderBuffer& render_buffer,
                          const FftData& G,
                          size_t num_partitions,
                          std::vector<std::vector<FftData>>* H) { … }
#endif

// Produces the filter output.
void ApplyFilter(const RenderBuffer& render_buffer,
                 size_t num_partitions,
                 const std::vector<std::vector<FftData>>& H,
                 FftData* S) { … }

#if defined(WEBRTC_HAS_NEON)
// Produces the filter output (Neon variant).
void ApplyFilter_Neon(const RenderBuffer& render_buffer,
                      size_t num_partitions,
                      const std::vector<std::vector<FftData>>& H,
                      FftData* S) {
  // const RenderBuffer& render_buffer,
  //                     rtc::ArrayView<const FftData> H,
  //                     FftData* S) {
  RTC_DCHECK_GE(H.size(), H.size() - 1);
  S->Clear();

  rtc::ArrayView<const std::vector<FftData>> render_buffer_data =
      render_buffer.GetFftBuffer();
  const size_t num_render_channels = render_buffer_data[0].size();
  const size_t lim1 = std::min(
      render_buffer_data.size() - render_buffer.Position(), num_partitions);
  const size_t lim2 = num_partitions;
  constexpr size_t kNumFourBinBands = kFftLengthBy2 / 4;

  size_t X_partition = render_buffer.Position();
  size_t p = 0;
  size_t limit = lim1;
  do {
    for (; p < limit; ++p, ++X_partition) {
      for (size_t ch = 0; ch < num_render_channels; ++ch) {
        const FftData& H_p_ch = H[p][ch];
        const FftData& X = render_buffer_data[X_partition][ch];
        for (size_t k = 0, n = 0; n < kNumFourBinBands; ++n, k += 4) {
          const float32x4_t X_re = vld1q_f32(&X.re[k]);
          const float32x4_t X_im = vld1q_f32(&X.im[k]);
          const float32x4_t H_re = vld1q_f32(&H_p_ch.re[k]);
          const float32x4_t H_im = vld1q_f32(&H_p_ch.im[k]);
          const float32x4_t S_re = vld1q_f32(&S->re[k]);
          const float32x4_t S_im = vld1q_f32(&S->im[k]);
          const float32x4_t a = vmulq_f32(X_re, H_re);
          const float32x4_t e = vmlsq_f32(a, X_im, H_im);
          const float32x4_t c = vmulq_f32(X_re, H_im);
          const float32x4_t f = vmlaq_f32(c, X_im, H_re);
          const float32x4_t g = vaddq_f32(S_re, e);
          const float32x4_t h = vaddq_f32(S_im, f);
          vst1q_f32(&S->re[k], g);
          vst1q_f32(&S->im[k], h);
        }
      }
    }
    limit = lim2;
    X_partition = 0;
  } while (p < lim2);

  X_partition = render_buffer.Position();
  p = 0;
  limit = lim1;
  do {
    for (; p < limit; ++p, ++X_partition) {
      for (size_t ch = 0; ch < num_render_channels; ++ch) {
        const FftData& H_p_ch = H[p][ch];
        const FftData& X = render_buffer_data[X_partition][ch];
        S->re[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2] -
                                X.im[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2];
        S->im[kFftLengthBy2] += X.re[kFftLengthBy2] * H_p_ch.im[kFftLengthBy2] +
                                X.im[kFftLengthBy2] * H_p_ch.re[kFftLengthBy2];
      }
    }
    limit = lim2;
    X_partition = 0;
  } while (p < lim2);
}
#endif

#if defined(WEBRTC_ARCH_X86_FAMILY)
// Produces the filter output (SSE2 variant).
void ApplyFilter_Sse2(const RenderBuffer& render_buffer,
                      size_t num_partitions,
                      const std::vector<std::vector<FftData>>& H,
                      FftData* S) { … }
#endif

}  // namespace aec3

namespace {

// Ensures that the newly added filter partitions after a size increase are set
// to zero.
void ZeroFilter(size_t old_size,
                size_t new_size,
                std::vector<std::vector<FftData>>* H) { … }

}  // namespace

AdaptiveFirFilter::AdaptiveFirFilter(size_t max_size_partitions,
                                     size_t initial_size_partitions,
                                     size_t size_change_duration_blocks,
                                     size_t num_render_channels,
                                     Aec3Optimization optimization,
                                     ApmDataDumper* data_dumper)
    : … { … }

AdaptiveFirFilter::~AdaptiveFirFilter() = default;

void AdaptiveFirFilter::HandleEchoPathChange() { … }

void AdaptiveFirFilter::SetSizePartitions(size_t size, bool immediate_effect) { … }

void AdaptiveFirFilter::UpdateSize() { … }

void AdaptiveFirFilter::Filter(const RenderBuffer& render_buffer,
                               FftData* S) const { … }

void AdaptiveFirFilter::Adapt(const RenderBuffer& render_buffer,
                              const FftData& G) { … }

void AdaptiveFirFilter::Adapt(const RenderBuffer& render_buffer,
                              const FftData& G,
                              std::vector<float>* impulse_response) { … }

void AdaptiveFirFilter::ComputeFrequencyResponse(
    std::vector<std::array<float, kFftLengthBy2Plus1>>* H2) const { … }

void AdaptiveFirFilter::AdaptAndUpdateSize(const RenderBuffer& render_buffer,
                                           const FftData& G) { … }

// Constrains the partition of the frequency domain filter to be limited in
// time via setting the relevant time-domain coefficients to zero and updates
// the corresponding values in an externally stored impulse response estimate.
void AdaptiveFirFilter::ConstrainAndUpdateImpulseResponse(
    std::vector<float>* impulse_response) { … }

// Constrains the a partiton of the frequency domain filter to be limited in
// time via setting the relevant time-domain coefficients to zero.
void AdaptiveFirFilter::Constrain() { … }

void AdaptiveFirFilter::ScaleFilter(float factor) { … }

// Set the filter coefficients.
void AdaptiveFirFilter::SetFilter(size_t num_partitions,
                                  const std::vector<std::vector<FftData>>& H) { … }

}  // namespace webrtc
chromium/third_party/webrtc/modules/audio_processing/aec3/adaptive_fir_filter.cc
