xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/convolver.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2020 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 #pragma once
25 
26 #include "convolution_parameters.hpp"
27 
28 #include <algorithm>
29 #include <cstddef>
30 #include <tuple>
31 #include <vector>
32 
33 namespace arm_gemm {
34 
35 // Class to assist with convolution calculations.
36 //
37 // This is framed as a hierarchy of objects:
38 //
39 //  - Top level object which depends only on convolution parameters.  This sets up std::vectors for the padding and
40 //    kernel offset arrays.  From this you can request:
41 //
42 //  - Mid level object (e.g. instantiated at start of 'ConvolutionInterleave').  This holds specifics about the
43 //    input tensor, and the desired column range.  Calculations specific to this can be done once when this is set
44 //    up.  From this you can request:
45 //
46 //  - Low level object (instantiated for each range of rows).  This contains methods to actually populate a row
47 //    pointer array.
48 
49 
50 template<typename T>
51 class convolver {
52 private:
53     const ConvolutionParameters  m_params;
54 
55     // Vector of padding data
56     const std::vector<T>         m_pad_row;
57 
58     // X/Y offsets for each kernel position
59     std::vector<int>             m_kernel_y;
60     std::vector<int>             m_kernel_x;
61 
62     class column_handler {
63     private:
64         const convolver<T>          &m_parent;
65 
66         // Base/stride of input image
67         const T * const              m_input_base;
68         const size_t                 m_input_stride;
69 
70         // Starting kernel point and channel offset within that point
71         const unsigned int           m_start_pos;
72         const unsigned int           m_start_offset;
73 
74         // Total length to process, rounded length of each input channel block.
75         const unsigned int           m_length;
76         const unsigned int           m_rounded_stringlen;
77 
78         class row_handler {
79         private:
80             const convolver<T>          &m_convolver;
81             const column_handler        &m_parent;
82 
83             // These variables track progress through the current block of rows
84             unsigned int                 m_start_output_y=0;
85             unsigned int                 m_start_output_x=0;
86 
87             unsigned int                 m_length_remaining=0;
88             unsigned int                 m_current_pos=0;
89 
90             unsigned int                 m_active_height=0;
91 
92         public:
row_handler(const column_handler & parent,unsigned int start_row,unsigned int active_height)93             row_handler(const column_handler &parent, unsigned int start_row, unsigned int active_height) :
94                 m_convolver(parent.m_parent),
95                 m_parent(parent),
96                 m_start_output_y(start_row / m_convolver.m_params.output_width),
97                 m_start_output_x(start_row % m_convolver.m_params.output_width),
98                 m_length_remaining(m_parent.m_length),
99                 m_current_pos(m_parent.m_start_pos),
100                 m_active_height(active_height) { }
101 
finished() const102             bool finished() const {
103                 return (m_length_remaining == 0);
104             }
105 
next_block(const T ** const row_ptr)106             std::tuple<unsigned int, unsigned int> next_block(const T ** const row_ptr) {
107                 if (finished()) {
108                     return std::make_tuple(0, 0);
109                 }
110 
111                 // "in_width" in the amount of data that will be read in (copied)
112                 // "out_width" is the total amount of data that will be produced (including padding)
113                 unsigned int offset = (m_current_pos == m_parent.m_start_pos) ? m_parent.m_start_offset : 0;
114                 unsigned int in_width = std::min(m_length_remaining, static_cast<unsigned int>(m_convolver.m_params.input_channels) - offset);
115                 unsigned int out_width = std::min(m_length_remaining, m_parent.m_rounded_stringlen - offset);
116 
117                 unsigned int output_y = m_start_output_y;
118                 unsigned int output_x = m_start_output_x;
119 
120                 for (unsigned int row=0; row<m_active_height; row++) {
121                     int input_y = (output_y * m_convolver.m_params.output_stride_h) + m_convolver.m_kernel_y[m_current_pos];
122                     int input_x = (output_x * m_convolver.m_params.output_stride_w) + m_convolver.m_kernel_x[m_current_pos];
123 
124                     // Out-of-bounds points will read the padding data,
125                     // otherwise find the correct address in the input image.
126                     if (input_y < 0 || input_y >= m_convolver.m_params.input_height || input_x < 0 || input_x >= m_convolver.m_params.input_width) {
127                         row_ptr[row] = m_convolver.m_pad_row.data();
128                     } else {
129                         row_ptr[row] = m_parent.m_input_base + ((input_y * m_convolver.m_params.input_width) + input_x) * m_parent.m_input_stride;
130                     }
131 
132                     output_x++;
133                     if (output_x == m_convolver.m_params.output_width) {
134                         output_y++;
135                         output_x=0;
136                     }
137                 }
138 
139                 m_current_pos++;
140                 m_length_remaining-=out_width;
141 
142                 return std::make_tuple(in_width, offset);
143             }
144         }; // end of "row handler" class
145 
146     public:
column_handler(const convolver<T> & parent,const T * input_base,size_t input_stride,unsigned int k_start,unsigned int k_end,unsigned int rounded_stringlen)147         column_handler(const convolver<T> &parent, const T *input_base, size_t input_stride,
148                        unsigned int k_start, unsigned int k_end, unsigned int rounded_stringlen)
149                      : m_parent(parent), m_input_base(input_base), m_input_stride(input_stride),
150                        m_start_pos(k_start / rounded_stringlen),
151                        m_start_offset(k_start % rounded_stringlen),
152                        m_length(k_end - k_start),
153                        m_rounded_stringlen(rounded_stringlen) { }
154 
process_rows(unsigned int start_row,unsigned int active_height) const155         row_handler process_rows(unsigned int start_row, unsigned int active_height) const {
156             return row_handler(*this, start_row, active_height);
157         }
158     }; // end of "column handler" class
159 
160 public:
convolver(ConvolutionParameters params)161     convolver(ConvolutionParameters params) :
162         m_params (params), m_pad_row(params.input_channels, static_cast<T>(params.padding_value)),
163         m_kernel_y(params.kernel_width * params.kernel_height, 0),
164         m_kernel_x(params.kernel_width * params.kernel_height, 0) {
165 
166         // Kernel points are addressed across, then down (assumed weight layout is WHIO)
167         for (unsigned int ky=0; ky<params.kernel_height; ky++) {
168             for (unsigned int kx=0; kx<params.kernel_width; kx++) {
169                 unsigned int n = (ky * params.kernel_width) + kx;
170                 m_kernel_y[n] = ky - params.padding_top;
171                 m_kernel_x[n] = kx - params.padding_left;
172             }
173         }
174     }
175 
process_columns(const T * input_base,size_t input_stride,unsigned int k_start,unsigned int k_end,unsigned int rounded_stringlen) const176     column_handler process_columns(const T *input_base, size_t input_stride,
177                                    unsigned int k_start, unsigned int k_end, unsigned int rounded_stringlen) const {
178         return column_handler(*this, input_base, input_stride, k_start, k_end, rounded_stringlen);
179     }
180 };
181 
182 } // namespace arm_gemm
183