1 // Tencent is pleased to support the open source community by making ncnn available.
2 //
3 // Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
4 //
5 // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6 // in compliance with the License. You may obtain a copy of the License at
7 //
8 // https://opensource.org/licenses/BSD-3-Clause
9 //
10 // Unless required by applicable law or agreed to in writing, software distributed
11 // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
13 // specific language governing permissions and limitations under the License.
14
convolution_pack8to1_int8_sse(const Mat & bottom_blob,Mat & top_blob,const Mat & weight_data_int8,int kernel_w,int kernel_h,int dilation_w,int dilation_h,int stride_w,int stride_h,const Option & opt)15 static void convolution_pack8to1_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data_int8, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, const Option& opt)
16 {
17 int w = bottom_blob.w;
18 int channels = bottom_blob.c;
19
20 int outw = top_blob.w;
21 int outh = top_blob.h;
22 int outch = top_blob.c;
23
24 const int maxk = kernel_w * kernel_h;
25
26 // kernel offsets
27 std::vector<int> _space_ofs(maxk);
28 int* space_ofs = &_space_ofs[0];
29 {
30 int p1 = 0;
31 int p2 = 0;
32 int gap = w * dilation_h - kernel_w * dilation_w;
33 for (int i = 0; i < kernel_h; i++)
34 {
35 for (int j = 0; j < kernel_w; j++)
36 {
37 space_ofs[p1] = p2;
38 p1++;
39 p2 += dilation_w;
40 }
41 p2 += gap;
42 }
43 }
44
45 // num_output
46 #pragma omp parallel for num_threads(opt.num_threads)
47 for (int p = 0; p < outch; p++)
48 {
49 int* outptr = top_blob.channel(p);
50
51 for (int i = 0; i < outh; i++)
52 {
53 for (int j = 0; j < outw; j++)
54 {
55 int sum = 0;
56
57 const signed char* kptr = weight_data_int8.channel(p);
58
59 // channels
60 for (int q = 0; q < channels; q++)
61 {
62 const Mat m = bottom_blob.channel(q);
63 const signed char* sptr = m.row<const signed char>(i * stride_h) + j * stride_w * 8;
64
65 for (int k = 0; k < maxk; k++)
66 {
67 // TODO use _mm_cvtepi8_epi16 on sse4.1
68 __m128i _val = _mm_loadl_epi64((const __m128i*)(sptr + space_ofs[k] * 8));
69 _val = _mm_unpacklo_epi8(_val, _mm_cmpgt_epi8(_mm_setzero_si128(), _val));
70
71 __m128i _w = _mm_loadl_epi64((const __m128i*)kptr);
72 _w = _mm_unpacklo_epi8(_w, _mm_cmpgt_epi8(_mm_setzero_si128(), _w));
73
74 __m128i _sl = _mm_mullo_epi16(_val, _w);
75 __m128i _sh = _mm_mulhi_epi16(_val, _w);
76 __m128i _s0 = _mm_unpacklo_epi16(_sl, _sh);
77 __m128i _s1 = _mm_unpackhi_epi16(_sl, _sh);
78
79 __m128i _s4 = _mm_add_epi32(_s0, _s1);
80
81 // TODO use _mm_hadd_epi32 on ssse3
82 int s4[4];
83 _mm_storeu_si128((__m128i*)s4, _s4);
84 sum += s4[0] + s4[1] + s4[2] + s4[3]; // dot
85
86 kptr += 8;
87 }
88 }
89
90 outptr[j] = sum;
91 }
92
93 outptr += outw;
94 }
95 }
96 }
97