1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
|
/*
* Copyright (c) 2019 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#pragma once
#include <deque>
#include <functional>
#include <memory>
#include "depthwise.hpp"
namespace depthwise
{
template <
unsigned int OutputTileRows, unsigned int OutputTileCols,
unsigned int KernelRows, unsigned int KernelCols,
unsigned int StrideRows, unsigned int StrideCols,
typename TIn, typename TBias, typename TOut
>
class DilatedDepthwiseConvolution : public IDepthwiseConvolution
{
public:
/** Create a new dilated depthwise convolution engine.
*/
DilatedDepthwiseConvolution(
int n_batches, int n_input_rows, int n_input_cols, int n_channels,
int dilation_factor,
nck::ActivationFunction activation,
unsigned int padding_top,
unsigned int padding_left,
unsigned int padding_bottom,
unsigned int padding_right
);
/** Create a new dilated depthwise convolution engine.
*/
DilatedDepthwiseConvolution(
int n_batches, int n_input_rows, int n_input_cols, int n_channels,
int dilation_factor, int n_output_rows, int n_output_cols,
nck::ActivationFunction activation,
unsigned int padding_top,
unsigned int padding_left,
unsigned int padding_bottom,
unsigned int padding_right
);
// Cannot copy or move a DilatedDepthwiseConvolution.
DilatedDepthwiseConvolution(DilatedDepthwiseConvolution&) = delete;
DilatedDepthwiseConvolution operator=(DilatedDepthwiseConvolution&) = delete;
/* Set input tensor and stride. */
void set_input(const void *inptr) override;
void set_input(const void *inptr, int column_stride) override;
void set_input(const void *inptr, int row_stride, int column_stride) override;
void set_input(const void *inptr, int batch_stride, int row_stride, int column_stride) override;
/* Set output tensor and stride. */
void set_output(void *outptr) override;
void set_output(void *outptr, int column_stride) override;
void set_output(void *outptr, int row_stride, int column_stride) override;
void set_output(void *outptr, int batch_stride, int row_stride, int column_stride) override;
static int get_output_size(
int dim_size,
unsigned int padding_before,
unsigned int padding_after,
int dilation_factor
);
int output_size(
int dim_size, unsigned int padding_before, unsigned int padding_after
) const override;
/* Weights and biases are re-ordered to improve memory access patterns. Use
* these methods to determine the size of the re-pack buffer and to set the
* address (and implicitly reorder the weights and biases into) the buffer.
*/
size_t get_packed_params_size(void) const override;
void set_packed_params_buffer(void *) override;
void pack_params(const void *weights, const void *biases=nullptr) const override;
void pack_params(void *buffer, const void *weights, const void *biases=nullptr) const override;
void pack_params(
void *buffer,
const void* weights,
unsigned int weight_row_stride,
unsigned int weight_col_stride,
const void *biases=nullptr
) const override;
/* Working space is used to pad tensors on the fly. Before running any
* inference check the amount of space required, allocate and provide a
* pointer to the convolution engine.
*/
size_t get_working_space_size(unsigned int nthreads=1) const override;
void set_working_space(void *) override;
unsigned int get_window(void) const override;
void run(unsigned int start, unsigned int stop, unsigned int threadid=0) override;
protected:
/** Protected constructor which also accepts a function to construct a new
* subconvolution
*/
DilatedDepthwiseConvolution(
int n_batches, int n_input_rows, int n_input_cols, int n_channels,
int dilation_factor, int n_output_rows, int n_output_cols,
nck::ActivationFunction activation,
unsigned int padding_top,
unsigned int padding_left,
unsigned int padding_bottom,
unsigned int padding_right,
std::function<IDepthwiseConvolution *(int, int, int, int, int, int, nck::ActivationFunction, unsigned int, unsigned int, unsigned int, unsigned int)> subconvfn
);
const int _dilation_factor;
const int _n_input_rows, _n_input_cols, _n_channels;
const int _padding_top, _padding_left;
const int _n_output_rows, _n_output_cols;
/* Dilated depthwise convolution is performed through repeated calls to
* non-dilated convolutions. If the dilation factor is $n$, then we perform
* $(n + 1)^2$ depthwise convolutions.
*/
using BaseDepthwise = DepthwiseConvolution<
OutputTileRows, OutputTileCols,
KernelRows, KernelCols,
StrideRows, StrideCols,
TIn, TBias, TOut
>;
std::deque<std::deque<std::unique_ptr<IDepthwiseConvolution>>> _convs;
};
} // namespace depthwise
|
