1 /*
2  * Copyright (c) 2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "IndirectConv2dAddressPrecalculation.h"
25 
26 #include "arm_compute/core/Types.h"
27 
28 #include "tests/validation/Helpers.h"
29 
30 namespace arm_compute
31 {
32 namespace test
33 {
34 namespace validation
35 {
36 namespace reference
37 {
indirect_conv2d_addr_precalculation(const TensorShape & shape_conv_src,const TensorShape & shape_conv_wei,const TensorShape & shape_conv_dst,const TensorShape & shape_dst,const PadStrideInfo & conv_info)38 SimpleTensor<int32_t> indirect_conv2d_addr_precalculation(const TensorShape &shape_conv_src, const TensorShape &shape_conv_wei, const TensorShape &shape_conv_dst, const TensorShape &shape_dst,
39                                                           const PadStrideInfo &conv_info)
40 {
41     SimpleTensor<int32_t> out{ shape_dst, DataType::S32 };
42 
43     constexpr unsigned int width_idx = 1;
44     constexpr unsigned int heigh_idx = 2;
45 
46     const int src_conv_width  = static_cast<int32_t>(shape_conv_src[width_idx]); // NHWC
47     const int src_conv_height = static_cast<int32_t>(shape_conv_src[heigh_idx]); // NHWC
48     const int dst_conv_width  = static_cast<int32_t>(shape_conv_dst[width_idx]); // NHWC
49     const int wei_conv_width  = static_cast<int32_t>(shape_conv_wei[width_idx]); // NHWC
50     const int wei_conv_height = static_cast<int32_t>(shape_conv_wei[heigh_idx]); // NHWC
51     const int dst_width       = static_cast<int32_t>(shape_dst[0]);
52     const int dst_height      = static_cast<int32_t>(shape_dst[1]);
53     const int dst_batch       = static_cast<int32_t>(shape_dst[2]);
54     const int ks              = wei_conv_width * wei_conv_height;
55     const int stride_x        = static_cast<int32_t>(conv_info.stride().first);
56     const int stride_y        = static_cast<int32_t>(conv_info.stride().second);
57     const int pad_left        = static_cast<int32_t>(conv_info.pad_left());
58     const int pad_top         = static_cast<int32_t>(conv_info.pad_top());
59 
60     const int m0 = dst_width / ks;
61 
62     for(int z = 0; z < dst_batch; ++z)
63     {
64         for(int y = 0; y < dst_height; ++y)
65         {
66             const int mout = y * m0;
67             for(int ki = 0; ki < ks; ++ki)
68             {
69                 const int xk = ki % wei_conv_width;
70                 const int yk = ki / wei_conv_width;
71                 for(int mi = 0; mi < m0; ++mi)
72                 {
73                     int xi = ((mout + mi) % dst_conv_width) * stride_x;
74                     int yi = ((mout + mi) / dst_conv_width) * stride_y;
75                     xi -= pad_left;
76                     yi -= pad_top;
77                     const int x_s = xi + xk;
78                     const int y_s = yi + yk;
79                     int       my  = x_s + y_s * src_conv_width;
80                     my            = my + z * src_conv_width * src_conv_height;
81                     my            = x_s >= 0 ? my : -1;
82                     my            = x_s < src_conv_width ? my : -1;
83                     my            = y_s >= 0 ? my : -1;
84                     my            = y_s < src_conv_height ? my : -1;
85 
86                     const unsigned int addr_out = mi + ki * m0 + y * (dst_width) + z * (dst_width * dst_height);
87                     out[addr_out]               = my;
88                 }
89             }
90         }
91     }
92 
93     return out;
94 }
95 } // namespace reference
96 } // namespace validation
97 } // namespace test
98 } // namespace arm_compute