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
|
/*
* 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/NEDirectConvolutionLayer.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;
NEDirectConvolutionLayer::NEDirectConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _output_stage_kernel(), _conv_kernel(), _input_border_handler(), _activationlayer_function(), _accumulator(), _has_bias(false), _is_fixed_point(false),
_is_activationlayer_enabled(false)
{
}
void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info)
{
// Free accumulator
if(_accumulator.buffer() != nullptr)
{
_accumulator.allocator()->free();
}
// Check if bias should be added in the convolution result
_has_bias = (bias != nullptr);
// Allocate the intermediate accumulator tensor in case of fixed point input
_is_fixed_point = is_data_type_fixed_point(input->info()->data_type());
if(_is_fixed_point)
{
const DataType promoted_dt = (input->info()->data_type() == DataType::QS8) ? DataType::QS16 : DataType::QS32;
_accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, promoted_dt, output->info()->fixed_point_position()));
_memory_group.manage(&_accumulator);
_conv_kernel.configure(input, weights, &_accumulator, conv_info);
// When no bias is provided, we need to downscale the accumulator tensor
_output_stage_kernel.configure(&_accumulator, bias, output);
_accumulator.allocator()->allocate();
}
else
{
_conv_kernel.configure(input, weights, output, conv_info);
if(_has_bias)
{
_output_stage_kernel.configure(output, bias);
}
}
// Add zero padding XY
_input_border_handler.configure(input, _conv_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<float>(0.f)));
//Configure Activation Layer
_is_activationlayer_enabled = act_info.enabled();
if(_is_activationlayer_enabled)
{
_activationlayer_function.configure(output, nullptr, act_info);
}
}
Status NEDirectConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &conv_info,
const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
DataType data_type = output->data_type();
if(is_data_type_fixed_point(data_type))
{
// Promote data type in case of fixed point
data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32);
}
TensorInfo accumulator(output->clone()->set_is_resizable(true).reset_padding().set_data_type(data_type));
// Validate Convolution kernel
ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerKernel::validate(input, weights, &accumulator, conv_info));
if(bias != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, bias);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != weights->dimension(3),
"Biases size and number of input feature maps should match");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Biases should be one dimensional");
}
// Validate bias kernel
ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerOutputStageKernel::validate(&accumulator, bias, output));
if(act_info.enabled())
{
ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info));
}
return Status{};
}
void NEDirectConvolutionLayer::run()
{
NEScheduler::get().schedule(&_input_border_handler, Window::DimZ);
_memory_group.acquire();
NEScheduler::get().schedule(&_conv_kernel, Window::DimZ);
if(_has_bias || _is_fixed_point)
{
NEScheduler::get().schedule(&_output_stage_kernel, Window::DimY);
}
if(_is_activationlayer_enabled)
{
_activationlayer_function.run();
}
_memory_group.release();
}
|
