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
|
/*
* Copyright (c) 2017-2018 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.
*/
#include "arm_compute/runtime/NEON/functions/NELocallyConnectedLayer.h"
#include "arm_compute/core/PixelValue.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include <cmath>
#include <tuple>
using namespace arm_compute;
NELocallyConnectedLayer::NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
_is_first_run(false)
{
}
void NELocallyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
ARM_COMPUTE_ERROR_ON(weights->info()->dimension(2) != input->info()->dimension(2));
ARM_COMPUTE_ERROR_ON(!conv_info.padding_is_symmetric());
if(biases != nullptr)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 2);
}
bool _has_bias = (biases != nullptr);
_is_first_run = true;
// Get parameters for conv_info
unsigned int stride_x = 0;
unsigned int stride_y = 0;
unsigned int pad_x = 0;
unsigned int pad_y = 0;
std::tie(stride_x, stride_y) = conv_info.stride();
std::tie(pad_x, pad_y) = conv_info.pad();
const unsigned int kernel_width = weights->info()->dimension(0);
const unsigned int kernel_height = weights->info()->dimension(1);
// Get convolved dimensions
unsigned int conv_w = 0;
unsigned int conv_h = 0;
std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
conv_info);
ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one");
// Create tensor to store the reshaped weights
const size_t mat_weights_cols = weights->info()->dimension(3);
const size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
const size_t mat_weights_num = weights->info()->dimension(4);
const TensorShape shape_wr(mat_weights_cols, mat_weights_rows, mat_weights_num);
_weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type()));
// Create tensor to store im2col reshaped inputs
const size_t mat_input_cols = mat_weights_rows;
const size_t mat_input_rows = conv_w * conv_h;
TensorShape shape_im2col = input->info()->tensor_shape();
shape_im2col.set(0, mat_input_cols);
shape_im2col.set(1, mat_input_rows);
shape_im2col.set(2, 1);
_input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
// Create locally connected layer output tensor
TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
shape_gemm.set(0, mat_weights_cols);
shape_gemm.set(1, mat_input_rows);
_gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
// Manage intermediate buffers
_memory_group.manage(&_input_im2col_reshaped);
_memory_group.manage(&_gemm_output);
// Configure kernels
_input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
_weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
_mm_kernel.configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output);
_output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
// Allocate intermediate tensors
_weights_reshaped.allocator()->allocate();
_input_im2col_reshaped.allocator()->allocate();
_gemm_output.allocator()->allocate();
}
void NELocallyConnectedLayer::run()
{
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
_is_first_run = false;
NEScheduler::get().schedule(&_weights_reshape_kernel, 3);
}
_memory_group.acquire();
// Run input reshaping
NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY);
// Runs GEMM on reshaped matrices
NEScheduler::get().schedule(&_mm_kernel, Window::DimX);
// Reshape output matrix
NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);
_memory_group.release();
}
|
