1/*
2 *  Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include "webrtc/common_audio/fir_filter_sse.h"
12
13#include <assert.h>
14#include <string.h>
15#include <xmmintrin.h>
16
17#include "webrtc/system_wrappers/interface/aligned_malloc.h"
18
19namespace webrtc {
20
21FIRFilterSSE2::FIRFilterSSE2(const float* coefficients,
22                             size_t coefficients_length,
23                             size_t max_input_length)
24    :  // Closest higher multiple of four.
25      coefficients_length_((coefficients_length + 3) & ~0x03),
26      state_length_(coefficients_length_ - 1),
27      coefficients_(static_cast<float*>(
28          AlignedMalloc(sizeof(float) * coefficients_length_, 16))),
29      state_(static_cast<float*>(
30          AlignedMalloc(sizeof(float) * (max_input_length + state_length_),
31                        16))) {
32  // Add zeros at the end of the coefficients.
33  size_t padding = coefficients_length_ - coefficients_length;
34  memset(coefficients_.get(), 0, padding * sizeof(coefficients_[0]));
35  // The coefficients are reversed to compensate for the order in which the
36  // input samples are acquired (most recent last).
37  for (size_t i = 0; i < coefficients_length; ++i) {
38    coefficients_[i + padding] = coefficients[coefficients_length - i - 1];
39  }
40  memset(state_.get(),
41         0,
42         (max_input_length + state_length_) * sizeof(state_[0]));
43}
44
45void FIRFilterSSE2::Filter(const float* in, size_t length, float* out) {
46  assert(length > 0);
47
48  memcpy(&state_[state_length_], in, length * sizeof(*in));
49
50  // Convolves the input signal |in| with the filter kernel |coefficients_|
51  // taking into account the previous state.
52  for (size_t i = 0; i < length; ++i) {
53    float* in_ptr = &state_[i];
54    float* coef_ptr = coefficients_.get();
55
56    __m128 m_sum = _mm_setzero_ps();
57    __m128 m_in;
58
59    // Depending on if the pointer is aligned with 16 bytes or not it is loaded
60    // differently.
61    if (reinterpret_cast<uintptr_t>(in_ptr) & 0x0F) {
62      for (size_t j = 0; j < coefficients_length_; j += 4) {
63        m_in = _mm_loadu_ps(in_ptr + j);
64        m_sum = _mm_add_ps(m_sum, _mm_mul_ps(m_in, _mm_load_ps(coef_ptr + j)));
65      }
66    } else {
67      for (size_t j = 0; j < coefficients_length_; j += 4) {
68        m_in = _mm_load_ps(in_ptr + j);
69        m_sum = _mm_add_ps(m_sum, _mm_mul_ps(m_in, _mm_load_ps(coef_ptr + j)));
70      }
71    }
72    m_sum = _mm_add_ps(_mm_movehl_ps(m_sum, m_sum), m_sum);
73    _mm_store_ss(out + i, _mm_add_ss(m_sum, _mm_shuffle_ps(m_sum, m_sum, 1)));
74  }
75
76  // Update current state.
77  memmove(state_.get(), &state_[length], state_length_ * sizeof(state_[0]));
78}
79
80}  // namespace webrtc
81