1 /*
2  *  Copyright (c) 2013 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 // Modified from the Chromium original:
12 // src/media/base/simd/sinc_resampler_sse.cc
13 
14 #include <stddef.h>
15 #include <stdint.h>
16 #include <xmmintrin.h>
17 
18 #include "common_audio/resampler/sinc_resampler.h"
19 
20 namespace webrtc {
21 
Convolve_SSE(const float * input_ptr,const float * k1,const float * k2,double kernel_interpolation_factor)22 float SincResampler::Convolve_SSE(const float* input_ptr,
23                                   const float* k1,
24                                   const float* k2,
25                                   double kernel_interpolation_factor) {
26   __m128 m_input;
27   __m128 m_sums1 = _mm_setzero_ps();
28   __m128 m_sums2 = _mm_setzero_ps();
29 
30   // Based on |input_ptr| alignment, we need to use loadu or load.  Unrolling
31   // these loops hurt performance in local testing.
32   if (reinterpret_cast<uintptr_t>(input_ptr) & 0x0F) {
33     for (size_t i = 0; i < kKernelSize; i += 4) {
34       m_input = _mm_loadu_ps(input_ptr + i);
35       m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
36       m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
37     }
38   } else {
39     for (size_t i = 0; i < kKernelSize; i += 4) {
40       m_input = _mm_load_ps(input_ptr + i);
41       m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
42       m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
43     }
44   }
45 
46   // Linearly interpolate the two "convolutions".
47   m_sums1 = _mm_mul_ps(
48       m_sums1,
49       _mm_set_ps1(static_cast<float>(1.0 - kernel_interpolation_factor)));
50   m_sums2 = _mm_mul_ps(
51       m_sums2, _mm_set_ps1(static_cast<float>(kernel_interpolation_factor)));
52   m_sums1 = _mm_add_ps(m_sums1, m_sums2);
53 
54   // Sum components together.
55   float result;
56   m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
57   _mm_store_ss(&result,
58                _mm_add_ss(m_sums2, _mm_shuffle_ps(m_sums2, m_sums2, 1)));
59 
60   return result;
61 }
62 
63 }  // namespace webrtc
64