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
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
|
/*
* 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/CL/functions/CLDepthwiseConvolutionLayer.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "arm_compute/core/PixelValue.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
#include "support/ToolchainSupport.h"
using namespace arm_compute;
using namespace arm_compute::misc;
using namespace arm_compute::misc::shape_calculator;
CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3()
: _kernel(), _border_handler()
{
}
void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, ActivationLayerInfo act_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
_kernel.set_target(CLScheduler::get().target());
_kernel.configure(input, weights, biases, output, conv_info, act_info);
// Configure border handler
PixelValue &&zero_value(0.f);
if(is_data_type_quantized_asymmetric(input->info()->data_type()))
{
zero_value = PixelValue(static_cast<uint8_t>(input->info()->quantization_info().offset));
}
_border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, zero_value);
}
void CLDepthwiseConvolutionLayer3x3::run()
{
CLScheduler::get().enqueue(_border_handler);
CLScheduler::get().enqueue(_kernel);
}
CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
: _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
_weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_first_run(true), _is_quantized(false), _original_weights(nullptr)
{
}
void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2));
const size_t weights_w = weights->info()->dimension(0);
const size_t weights_h = weights->info()->dimension(1);
const size_t weights_z = weights->info()->dimension(2);
_is_first_run = true;
_original_weights = weights;
_is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
bool append_bias = (biases != nullptr) && !_is_quantized;
const GPUTarget gpu_target = CLScheduler::get().target();
// Calculate output shape
TensorShape dwc_output_shape = shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info);
// Output width and height
const unsigned int conv_w = dwc_output_shape.x();
const unsigned int conv_h = dwc_output_shape.y();
// Set up intermediate tensors
const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
const size_t conv_size = conv_w * conv_h;
// Im2Col configuration
TensorShape shape_im2col = input->info()->tensor_shape();
shape_im2col.set(0, patch_size);
shape_im2col.set(1, conv_size);
shape_im2col.set(2, weights_z);
_input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
_im2col_kernel.set_target(gpu_target);
_im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias);
// Weights reshape configuration
const TensorShape shape_weights_reshape(patch_size, weights_z);
_weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
_weights_reshape_kernel.configure(weights, &_weights_reshaped, append_bias ? biases : nullptr);
// GEMV configuration
DataType v2mm_dt = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
TensorShape shape_v2mm_out = input->info()->tensor_shape();
shape_v2mm_out.set(0, conv_size * weights_z);
shape_v2mm_out.set(1, 1);
shape_v2mm_out.set(2, 1);
_v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
_v2mm_kernel.set_target(gpu_target);
_v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
_output_reshaped.allocator()->init(_v2mm_output.info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(dwc_output_shape));
_vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
// Output staged configuration
if(_is_quantized)
{
const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale;
int output_multiplier, output_shift;
quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
_output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output_quant_info.offset);
_output_reshaped.allocator()->allocate();
}
// Fill borders on inputs
PixelValue zero_in(static_cast<int32_t>(0));
PixelValue zero_w(static_cast<int32_t>(0));
if(_is_quantized)
{
zero_in = PixelValue(static_cast<int32_t>(input->info()->quantization_info().offset));
zero_w = PixelValue(static_cast<int32_t>(weights->info()->quantization_info().offset));
}
BorderSize border_size = _v2mm_kernel.border_size();
_v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);
border_size.bottom = 0;
_v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, zero_w);
// Allocate intermediate tensors
_input_reshaped.allocator()->allocate();
_weights_reshaped.allocator()->allocate();
_v2mm_output.allocator()->allocate();
}
void CLDepthwiseConvolutionLayer::run()
{
// Run weights reshaping (Runs once for every configure)
if(_is_first_run)
{
ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
CLScheduler::get().enqueue(_weights_reshape_kernel);
CLScheduler::get().enqueue(_v2mm_weights_fill_border);
_is_first_run = false;
// Mark original weights tensor as unused
_original_weights->mark_as_unused();
}
CLScheduler::get().enqueue(_im2col_kernel);
CLScheduler::get().enqueue(_v2mm_input_fill_border);
CLScheduler::get().enqueue(_v2mm_kernel);
CLScheduler::get().enqueue(_vector_to_tensor_kernel);
if(_is_quantized)
{
CLScheduler::get().enqueue(_output_stage_kernel);
}
}
|
