1 // Copyright 2013 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4
5 #include "media/base/sinc_resampler.h"
6
7 #include <xmmintrin.h>
8
9 namespace media {
10
Convolve_SSE(const float * input_ptr,const float * k1,const float * k2,double kernel_interpolation_factor)11 float SincResampler::Convolve_SSE(const float* input_ptr, const float* k1,
12 const float* k2,
13 double kernel_interpolation_factor) {
14 __m128 m_input;
15 __m128 m_sums1 = _mm_setzero_ps();
16 __m128 m_sums2 = _mm_setzero_ps();
17
18 // Based on |input_ptr| alignment, we need to use loadu or load. Unrolling
19 // these loops hurt performance in local testing.
20 if (reinterpret_cast<uintptr_t>(input_ptr) & 0x0F) {
21 for (int i = 0; i < kKernelSize; i += 4) {
22 m_input = _mm_loadu_ps(input_ptr + i);
23 m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
24 m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
25 }
26 } else {
27 for (int i = 0; i < kKernelSize; i += 4) {
28 m_input = _mm_load_ps(input_ptr + i);
29 m_sums1 = _mm_add_ps(m_sums1, _mm_mul_ps(m_input, _mm_load_ps(k1 + i)));
30 m_sums2 = _mm_add_ps(m_sums2, _mm_mul_ps(m_input, _mm_load_ps(k2 + i)));
31 }
32 }
33
34 // Linearly interpolate the two "convolutions".
35 m_sums1 = _mm_mul_ps(m_sums1, _mm_set_ps1(1.0 - kernel_interpolation_factor));
36 m_sums2 = _mm_mul_ps(m_sums2, _mm_set_ps1(kernel_interpolation_factor));
37 m_sums1 = _mm_add_ps(m_sums1, m_sums2);
38
39 // Sum components together.
40 float result;
41 m_sums2 = _mm_add_ps(_mm_movehl_ps(m_sums1, m_sums1), m_sums1);
42 _mm_store_ss(&result, _mm_add_ss(m_sums2, _mm_shuffle_ps(
43 m_sums2, m_sums2, 1)));
44
45 return result;
46 }
47
48 } // namespace media
49