1 /*
2 * Copyright (c) 2019 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
25
26 #include "arm_compute/core/Utils.h"
27 #include "arm_compute/graph/Graph.h"
28 #include "arm_compute/graph/INodeVisitor.h"
29 #include "arm_compute/graph/Utils.h"
30
31 namespace arm_compute
32 {
33 namespace graph
34 {
FusedDepthwiseConvolutionBatchNormalizationNode(float epsilon,PadStrideInfo info,unsigned int depth_multiplier,DepthwiseConvolutionMethod method,ActivationLayerInfo fused_activation)35 FusedDepthwiseConvolutionBatchNormalizationNode::FusedDepthwiseConvolutionBatchNormalizationNode(float epsilon,
36 PadStrideInfo info,
37 unsigned int depth_multiplier,
38 DepthwiseConvolutionMethod method,
39 ActivationLayerInfo fused_activation)
40 : _epsilon(epsilon), _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _fused_activation(fused_activation)
41 {
42 _input_edges.resize(7, EmptyEdgeID);
43 _outputs.resize(1, NullTensorID);
44 }
45
set_depthwise_convolution_method(DepthwiseConvolutionMethod method)46 void FusedDepthwiseConvolutionBatchNormalizationNode::set_depthwise_convolution_method(DepthwiseConvolutionMethod method)
47 {
48 _method = method;
49 }
50
depthwise_convolution_method() const51 DepthwiseConvolutionMethod FusedDepthwiseConvolutionBatchNormalizationNode::depthwise_convolution_method() const
52 {
53 return _method;
54 }
55
epsilon() const56 float FusedDepthwiseConvolutionBatchNormalizationNode::epsilon() const
57 {
58 return _epsilon;
59 }
60
convolution_info() const61 PadStrideInfo FusedDepthwiseConvolutionBatchNormalizationNode::convolution_info() const
62 {
63 return _info;
64 }
65
depth_multiplier() const66 unsigned int FusedDepthwiseConvolutionBatchNormalizationNode::depth_multiplier() const
67 {
68 return _depth_multiplier;
69 }
70
fused_activation() const71 ActivationLayerInfo FusedDepthwiseConvolutionBatchNormalizationNode::fused_activation() const
72 {
73 return _fused_activation;
74 }
75
set_fused_activation(ActivationLayerInfo fused_activation)76 void FusedDepthwiseConvolutionBatchNormalizationNode::set_fused_activation(ActivationLayerInfo fused_activation)
77 {
78 _fused_activation = fused_activation;
79 }
80
compute_output_descriptor(const TensorDescriptor & input_descriptor,const TensorDescriptor & weights_descriptor,const PadStrideInfo & info,int depth_multiplier)81 TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
82 const TensorDescriptor &weights_descriptor,
83 const PadStrideInfo &info,
84 int depth_multiplier)
85 {
86 unsigned int output_width = 0;
87 unsigned int output_height = 0;
88
89 const unsigned int input_width = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH);
90 const unsigned int input_height = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT);
91 const unsigned int input_channels = get_dimension_size(input_descriptor, DataLayoutDimension::CHANNEL);
92 const unsigned int kernel_width = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH);
93 const unsigned int kernel_height = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT);
94
95 std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
96
97 TensorDescriptor output_descriptor = input_descriptor;
98 output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::WIDTH), output_width);
99 output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::HEIGHT), output_height);
100 output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::CHANNEL), input_channels * depth_multiplier);
101
102 return output_descriptor;
103 }
104
forward_descriptors()105 bool FusedDepthwiseConvolutionBatchNormalizationNode::forward_descriptors()
106 {
107 if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (output_id(0) != NullTensorID))
108 {
109 Tensor *dst = output(0);
110 ARM_COMPUTE_ERROR_ON(dst == nullptr);
111 dst->desc() = configure_output(0);
112 return true;
113 }
114 return false;
115 }
116
configure_output(size_t idx) const117 TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::configure_output(size_t idx) const
118 {
119 ARM_COMPUTE_UNUSED(idx);
120 const Tensor *src = input(0);
121 const Tensor *weights = input(1);
122
123 ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
124
125 TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
126
127 return output_info;
128 }
129
type() const130 NodeType FusedDepthwiseConvolutionBatchNormalizationNode::type() const
131 {
132 return FusedDepthwiseConvolutionBatchNormalizationNode::node_type;
133 }
134
accept(INodeVisitor & v)135 void FusedDepthwiseConvolutionBatchNormalizationNode::accept(INodeVisitor &v)
136 {
137 v.visit(*this);
138 }
139 } // namespace graph
140 } // namespace arm_compute
141