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 "modules/audio_processing/agc2/rnn_vad/vector_math.h"
12
13 #include <immintrin.h>
14
15 #include "api/array_view.h"
16 #include "rtc_base/checks.h"
17 #include "rtc_base/numerics/safe_conversions.h"
18
19 namespace webrtc {
20 namespace rnn_vad {
21
DotProductAvx2(rtc::ArrayView<const float> x,rtc::ArrayView<const float> y) const22 float VectorMath::DotProductAvx2(rtc::ArrayView<const float> x,
23 rtc::ArrayView<const float> y) const {
24 RTC_DCHECK(cpu_features_.avx2);
25 RTC_DCHECK_EQ(x.size(), y.size());
26 __m256 accumulator = _mm256_setzero_ps();
27 constexpr int kBlockSizeLog2 = 3;
28 constexpr int kBlockSize = 1 << kBlockSizeLog2;
29 const int incomplete_block_index = (x.size() >> kBlockSizeLog2)
30 << kBlockSizeLog2;
31 for (int i = 0; i < incomplete_block_index; i += kBlockSize) {
32 RTC_DCHECK_LE(i + kBlockSize, x.size());
33 const __m256 x_i = _mm256_loadu_ps(&x[i]);
34 const __m256 y_i = _mm256_loadu_ps(&y[i]);
35 accumulator = _mm256_fmadd_ps(x_i, y_i, accumulator);
36 }
37 // Reduce `accumulator` by addition.
38 __m128 high = _mm256_extractf128_ps(accumulator, 1);
39 __m128 low = _mm256_extractf128_ps(accumulator, 0);
40 low = _mm_add_ps(high, low);
41 high = _mm_movehl_ps(high, low);
42 low = _mm_add_ps(high, low);
43 high = _mm_shuffle_ps(low, low, 1);
44 low = _mm_add_ss(high, low);
45 float dot_product = _mm_cvtss_f32(low);
46 // Add the result for the last block if incomplete.
47 for (int i = incomplete_block_index; i < rtc::dchecked_cast<int>(x.size());
48 ++i) {
49 dot_product += x[i] * y[i];
50 }
51 return dot_product;
52 }
53
54 } // namespace rnn_vad
55 } // namespace webrtc
56