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
|
/*
* Copyright (c) 2017 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 <utility>
#include "batched_blocked_gemm.hpp"
#include "winograd_gemm.hpp"
/** Example of how to construct an ACL-like interface.
*
* Use `get_weight_storage_size`, `get_input_storage_size` and
* `get_output_storage_size` to allocate memory for the convolution engine.
* Then create a `WinogradConvolutionLayer`.
*
* Initialise the weights using `weights_transform.run(...)`.
*
* For each inference:
* 1. Transform the inputs to the Winograd domain using `input_transform.run(...)`
* 2. Perform a number of GEMMs using `gemms.run(...)`
* 3. Transform the output to the spatial domain using `output_transform.run(...)`
*/
template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols,
typename TIn, typename TOut>
class WinogradConvolutionLayer
{
private:
const KernelShape _kernel_shape;
const Tensor4DShape _input_shape;
const PaddingType _padding;
const Tensor4DShape _output_shape;
const int _n_output_rows, _n_output_cols;
const int _kernel_matrix_stride, _kernel_matrix_row_stride;
const int _input_matrix_stride, _input_matrix_row_stride;
const int _output_matrix_stride, _output_matrix_row_stride;
const int _tile_rows, _tile_cols;
const int _m, _k, _n;
public:
using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols>;
using WeightsTransform = typename WinogradBase::template WeightsTransform<TIn>;
using InputTransform = typename WinogradBase::template InputTransform<TIn>;
using WinogradConv = typename WinogradBase::template Convolution<TOut, TIn>;
using MultiGEMM = winograd::BatchedBlockedGemm<WinogradConv::M_BLOCK, WinogradConv::N_BLOCK, TIn, TOut>;
using OutputTransform = typename WinogradBase::template OutputTransform<TOut>;
/* Public member variables. */
WeightsTransform weights_transform; /** Operator to transform weights to Winograd domain. */
InputTransform input_transform; /** Operator to transform input to Winograd domain. */
MultiGEMM gemms; /** Operator to perform multiple GEMMs. */
OutputTransform output_transform; /** Operator to transform output from Winograd domain. */
/** Determine how much memory (in units of TIn) to allocate for the
* transformed weights.
*/
static unsigned int get_weight_storage_size(
const int n_output_channels, /** Number of output feature maps. */
const int n_input_channels /** Number of input feature maps. */
);
/** Determine how much memory (in units of TIn) to allocate for the
* transformed input.
*/
static unsigned int get_input_storage_size(
const int n_batches, /** Number of batches in the input tensor. */
const int n_channels, /** Number of feature maps in the input tensor. */
const int n_rows, /** Number of rows in each feature map. */
const int n_cols, /** Number of columns in each feature map. */
const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
);
/** Determine how much memory (in units of TOut) to allocate for the
* (Winograd domain) output.
*/
static unsigned int get_output_storage_size(
const int n_batches, /** Number of batches in the output tensor. */
const int n_rows, /** Number of rows in each feature map of the input tensor. */
const int n_cols, /** Number of columns in each feature map of the input tensor. */
const int n_output_channels, /** Number of feature maps in the output tensor. */
const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
);
/** Get the shape (rows, cols) of a feature map of the output tensor. */
static std::pair<int, int> get_output_feature_map_shape(
const int n_input_rows, /** Number of rows in the input feature map. */
const int n_input_cols, /** Number of columns in the input feature map. */
const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */
);
/** Create a new Winograd convolution layer.
*/
WinogradConvolutionLayer(
const int n_batches, /** Number of batches in the input and output tensors. */
const int n_input_channels, /** Number of feature maps in a batch of the input tensor. */
const int n_input_rows, /** Number of rows in a feature map of the input tensor. */
const int n_input_cols, /** Number of columns in a feature map of the input tensor. */
const int n_output_channels, /** Number of feature maps in the output tensor. */
const bool same_padding, /** Use "SAME" padding, otherwise use "VALID". */
const TIn* const weights, /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */
TIn* const weights_storage, /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */
const TIn* const input, /** Pointer to NHWC ordered input tensor, in the spatial domain. */
TIn* const winograd_input, /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */
TOut* const output, /** Pointer to NHWC ordered output tensor, in the spatial domain. */
TOut* const winograd_output /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */
);
};
|
