1 ///////////////////////////////////////////////////////////////////////
2 // File: dotproductavx.cpp
3 // Description: Architecture-specific dot-product function.
4 // Author: Ray Smith
5 //
6 // (C) Copyright 2015, Google Inc.
7 // Licensed under the Apache License, Version 2.0 (the "License");
8 // you may not use this file except in compliance with the License.
9 // You may obtain a copy of the License at
10 // http://www.apache.org/licenses/LICENSE-2.0
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS,
13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 // See the License for the specific language governing permissions and
15 // limitations under the License.
16 ///////////////////////////////////////////////////////////////////////
17
18 #if !defined(__AVX__)
19 # if defined(__i686__) || defined(__x86_64__)
20 # error Implementation only for AVX capable architectures
21 # endif
22 #else
23
24 # include <immintrin.h>
25 # include <cstdint>
26 # include "dotproduct.h"
27
28 namespace tesseract {
29
30 // Computes and returns the dot product of the n-vectors u and v.
31 // Uses Intel AVX intrinsics to access the SIMD instruction set.
32 #if defined(FAST_FLOAT)
DotProductAVX(const float * u,const float * v,int n)33 float DotProductAVX(const float *u, const float *v, int n) {
34 const unsigned quot = n / 8;
35 const unsigned rem = n % 8;
36 __m256 t0 = _mm256_setzero_ps();
37 for (unsigned k = 0; k < quot; k++) {
38 __m256 f0 = _mm256_loadu_ps(u);
39 __m256 f1 = _mm256_loadu_ps(v);
40 f0 = _mm256_mul_ps(f0, f1);
41 t0 = _mm256_add_ps(t0, f0);
42 u += 8;
43 v += 8;
44 }
45 alignas(32) float tmp[8];
46 _mm256_store_ps(tmp, t0);
47 float result = tmp[0] + tmp[1] + tmp[2] + tmp[3] + tmp[4] + tmp[5] + tmp[6] + tmp[7];
48 for (unsigned k = 0; k < rem; k++) {
49 result += *u++ * *v++;
50 }
51 return result;
52 }
53 #else
54 double DotProductAVX(const double *u, const double *v, int n) {
55 const unsigned quot = n / 8;
56 const unsigned rem = n % 8;
57 __m256d t0 = _mm256_setzero_pd();
58 __m256d t1 = _mm256_setzero_pd();
59 for (unsigned k = 0; k < quot; k++) {
60 __m256d f0 = _mm256_loadu_pd(u);
61 __m256d f1 = _mm256_loadu_pd(v);
62 f0 = _mm256_mul_pd(f0, f1);
63 t0 = _mm256_add_pd(t0, f0);
64 u += 4;
65 v += 4;
66 __m256d f2 = _mm256_loadu_pd(u);
67 __m256d f3 = _mm256_loadu_pd(v);
68 f2 = _mm256_mul_pd(f2, f3);
69 t1 = _mm256_add_pd(t1, f2);
70 u += 4;
71 v += 4;
72 }
73 t0 = _mm256_hadd_pd(t0, t1);
74 alignas(32) double tmp[4];
75 _mm256_store_pd(tmp, t0);
76 double result = tmp[0] + tmp[1] + tmp[2] + tmp[3];
77 for (unsigned k = 0; k < rem; k++) {
78 result += *u++ * *v++;
79 }
80 return result;
81 }
82 #endif
83
84 } // namespace tesseract.
85
86 #endif
87