tesseract  5.0.0
intsimdmatrixsse.cpp
Go to the documentation of this file.
1 // File: intsindmatrixsse.cpp
3 // Description: SSE implementation of 8-bit int SIMD matrix multiply.
4 // Author: Ray Smith
5 //
6 // (C) Copyright 2017, 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.
17 
18 #if !defined(__SSE4_1__)
19 # if defined(__i686__) || defined(__x86_64__)
20 # error Implementation only for SSE 4.1 capable architectures
21 # endif
22 #else
23 
24 # include "intsimdmatrix.h"
25 
26 # include <emmintrin.h>
27 # include <smmintrin.h>
28 # include <cstdint>
29 
30 namespace tesseract {
31 
32 // Computes and returns the dot product of the n-vectors u and v.
33 // Uses Intel SSE intrinsics to access the SIMD instruction set.
34 static int32_t IntDotProductSSE(const int8_t *u, const int8_t *v, int n) {
35  int max_offset = n - 8;
36  int offset = 0;
37  // Accumulate a set of 4 32-bit sums in sum, by loading 8 pairs of 8-bit
38  // values, extending to 16 bit, multiplying to make 32 bit results.
39  int32_t result = 0;
40  if (offset <= max_offset) {
41  offset = 8;
42  __m128i packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(u));
43  __m128i packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(v));
44  __m128i sum = _mm_cvtepi8_epi16(packed1);
45  packed2 = _mm_cvtepi8_epi16(packed2);
46  // The magic _mm_add_epi16 is perfect here. It multiplies 8 pairs of 16 bit
47  // ints to make 32 bit results, which are then horizontally added in pairs
48  // to make 4 32 bit results that still fit in a 128 bit register.
49  sum = _mm_madd_epi16(sum, packed2);
50  while (offset <= max_offset) {
51  packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(u + offset));
52  packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i *>(v + offset));
53  offset += 8;
54  packed1 = _mm_cvtepi8_epi16(packed1);
55  packed2 = _mm_cvtepi8_epi16(packed2);
56  packed1 = _mm_madd_epi16(packed1, packed2);
57  sum = _mm_add_epi32(sum, packed1);
58  }
59  // Sum the 4 packed 32 bit sums and extract the low result.
60  sum = _mm_hadd_epi32(sum, sum);
61  sum = _mm_hadd_epi32(sum, sum);
62  result = _mm_cvtsi128_si32(sum);
63  }
64  while (offset < n) {
65  result += u[offset] * v[offset];
66  ++offset;
67  }
68  return result;
69 }
70 
71 // Computes part of matrix.vector v = Wu. Computes 1 result.
72 static void PartialMatrixDotVector1(const int8_t *wi, const TFloat *scales, const int8_t *u,
73  int num_in, TFloat *v) {
74  TFloat total = IntDotProductSSE(u, wi, num_in);
75  // Add in the bias and correct for integer values.
76  *v = (total + wi[num_in] * INT8_MAX) * *scales;
77 }
78 
79 static void matrixDotVector(int dim1, int dim2, const int8_t *wi, const TFloat *scales,
80  const int8_t *u, TFloat *v) {
81  const int num_out = dim1;
82  const int num_in = dim2 - 1;
83  int output = 0;
84 
85  for (; output < num_out; output++) {
86  PartialMatrixDotVector1(wi, scales, u, num_in, v);
87  wi += dim2;
88  scales++;
89  v++;
90  }
91 }
92 
93 const IntSimdMatrix IntSimdMatrix::intSimdMatrixSSE = {
94  matrixDotVector,
95  // Number of 32 bit outputs held in each register.
96  1,
97  // Maximum number of registers that we will use to hold outputs.
98  1,
99  // Number of 8 bit inputs in the inputs register.
100  1,
101  // Number of inputs in each weight group.
102  1
103 };
104 
105 } // namespace tesseract.
106 
107 #endif
double TFloat
Definition: tesstypes.h:39
static const IntSimdMatrix intSimdMatrixSSE