1 /*
2 * Copyright (c) 2020 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 <immintrin.h>
12 #include <stddef.h>
13 #include <stdint.h>
14 #include <xmmintrin.h>
15
16 #include "common_audio/resampler/sinc_resampler.h"
17
18 namespace webrtc {
19
Convolve_AVX2(const float * input_ptr,const float * k1,const float * k2,double kernel_interpolation_factor)20 float SincResampler::Convolve_AVX2(const float* input_ptr,
21 const float* k1,
22 const float* k2,
23 double kernel_interpolation_factor) {
24 __m256 m_input;
25 __m256 m_sums1 = _mm256_setzero_ps();
26 __m256 m_sums2 = _mm256_setzero_ps();
27
28 // Based on |input_ptr| alignment, we need to use loadu or load. Unrolling
29 // these loops has not been tested or benchmarked.
30 bool aligned_input = (reinterpret_cast<uintptr_t>(input_ptr) & 0x1F) == 0;
31 if (!aligned_input) {
32 for (size_t i = 0; i < kKernelSize; i += 8) {
33 m_input = _mm256_loadu_ps(input_ptr + i);
34 m_sums1 = _mm256_fmadd_ps(m_input, _mm256_load_ps(k1 + i), m_sums1);
35 m_sums2 = _mm256_fmadd_ps(m_input, _mm256_load_ps(k2 + i), m_sums2);
36 }
37 } else {
38 for (size_t i = 0; i < kKernelSize; i += 8) {
39 m_input = _mm256_load_ps(input_ptr + i);
40 m_sums1 = _mm256_fmadd_ps(m_input, _mm256_load_ps(k1 + i), m_sums1);
41 m_sums2 = _mm256_fmadd_ps(m_input, _mm256_load_ps(k2 + i), m_sums2);
42 }
43 }
44
45 // Linearly interpolate the two "convolutions".
46 __m128 m128_sums1 = _mm_add_ps(_mm256_extractf128_ps(m_sums1, 0),
47 _mm256_extractf128_ps(m_sums1, 1));
48 __m128 m128_sums2 = _mm_add_ps(_mm256_extractf128_ps(m_sums2, 0),
49 _mm256_extractf128_ps(m_sums2, 1));
50 m128_sums1 = _mm_mul_ps(
51 m128_sums1,
52 _mm_set_ps1(static_cast<float>(1.0 - kernel_interpolation_factor)));
53 m128_sums2 = _mm_mul_ps(
54 m128_sums2, _mm_set_ps1(static_cast<float>(kernel_interpolation_factor)));
55 m128_sums1 = _mm_add_ps(m128_sums1, m128_sums2);
56
57 // Sum components together.
58 float result;
59 m128_sums2 = _mm_add_ps(_mm_movehl_ps(m128_sums1, m128_sums1), m128_sums1);
60 _mm_store_ss(&result, _mm_add_ss(m128_sums2,
61 _mm_shuffle_ps(m128_sums2, m128_sums2, 1)));
62
63 return result;
64 }
65
66 } // namespace webrtc
67