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 
conv1x1s1_sgemm_pack8to4_int8_sse(const Mat & bottom_blob,Mat & top_blob,const Mat & kernel,const Option & opt)15 static void conv1x1s1_sgemm_pack8to4_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
16 {
17     int w = bottom_blob.w;
18     int h = bottom_blob.h;
19     const int size = w * h;
20 
21     Mat bottom_im2col = bottom_blob;
22     bottom_im2col.w = size;
23     bottom_im2col.h = 1;
24 
25     im2col_sgemm_pack8to4_int8_sse(bottom_im2col, top_blob, kernel, opt);
26 }
27 
conv1x1s2_pack8to4_int8_sse(const Mat & bottom_blob,Mat & top_blob,const Mat & kernel,const Option & opt)28 static void conv1x1s2_pack8to4_int8_sse(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
29 {
30     int w = bottom_blob.w;
31     int channels = bottom_blob.c;
32     size_t elemsize = bottom_blob.elemsize;
33     int elempack = bottom_blob.elempack;
34 
35     int outw = top_blob.w;
36     int outh = top_blob.h;
37 
38     const int tailstep = w - 2 * outw + w;
39 
40     Mat bottom_blob_shrinked;
41     bottom_blob_shrinked.create(outw, outh, channels, elemsize, elempack, opt.workspace_allocator);
42 
43     #pragma omp parallel for num_threads(opt.num_threads)
44     for (int p = 0; p < channels; p++)
45     {
46         const int64_t* r0 = bottom_blob.channel(p);
47         int64_t* outptr = bottom_blob_shrinked.channel(p);
48 
49         for (int i = 0; i < outh; i++)
50         {
51             int j = 0;
52             for (; j < outw; j++)
53             {
54                 outptr[0] = r0[0];
55 
56                 r0 += 2;
57                 outptr += 1;
58             }
59 
60             r0 += tailstep;
61         }
62     }
63 
64     conv1x1s1_sgemm_pack8to4_int8_sse(bottom_blob_shrinked, top_blob, kernel, opt);
65 }
66