2 // Copyright (c) 2016-2018 Intel Corporation
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
8 // http://www.apache.org/licenses/LICENSE-2.0
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
17 ///////////////////////////////////////////////////////////////////////////////////////////////////
18 #include "convolution_inst.h"
19 #include "primitive_type_base.h"
20 #include "sliding_window_utils.h"
21 #include "error_handler.h"
22 #include "json_object.h"
26 primitive_type_id convolution_type_id()
28 static primitive_type_base<convolution> instance;
32 layout convolution_inst::calc_output_layout(convolution_node const& node)
34 assert((bool)node.get_primitive()->output_data_type == false
35 && "Output data type forcing is not supported for convolution_node!");
36 auto desc = node.get_primitive();
38 auto input_layout = node.input().get_output_layout();
39 auto weights_layout = node.weights(0).get_output_layout(); //weights are stored after inputs
41 auto input_offset = desc->input_offset;
42 auto stride = desc->stride;
43 auto dilation = desc->dilation;
44 auto split = desc->weights.size();
46 // compute how many outputs in rows and columns will be generate by filter.
47 // outp <= (input_size - (2*input_offset) - kernel_size)/ stride
48 auto filter_size = weights_layout.size;
50 // TODO: Consider moving general parameter verification to arguments constructor.
51 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "Stride spatial X", stride.spatial[0], "value", 0, "Stride spatial X must be positive (>= 1)");
52 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "Stride spatial Y", stride.spatial[1], "value", 0, "Stride spatial Y must be positive (>= 1)");
53 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "Dilatation spatial X", dilation.spatial[0], "value", 0, "Dilatation patial X must be positive (>= 1)");
54 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "Dilatation spatial Y", dilation.spatial[1], "value", 0, "Dilatation spatial Y must be positive (>= 1)");
55 CLDNN_ERROR_GREATER_THAN(node.id(), "Input offset spatial X", 2 * input_offset.spatial[0], "input layout spatial X", input_layout.size.spatial[0], "There is no input data to process");
56 CLDNN_ERROR_GREATER_THAN(node.id(), "Input offset spatial Y", 2 * input_offset.spatial[1], "input layout spatial Y", input_layout.size.spatial[1], "There is no input data to process");
57 CLDNN_ERROR_NOT_EQUAL(node.id(), "Input offset feature", input_offset.feature[0], "", 0, "Input offset in feature is not supported");
58 CLDNN_ERROR_NOT_EQUAL(node.id(), "Input offset batch", input_offset.batch[0], "", 0, "Input offset in batch is not supported");
60 // TODO: FCN and SSD used offset larger than convolution size. does it make sense to support it? do we support it on the ref kernels?
61 // CLDNN_ERROR_GREATER_THAN(node.id(), "Negate input offset spatial X", -input_offset.spatial[0], "input window size spatial X", filter_size.spatial[0], "First convolution is outside of image. please reduce input offset X");
62 // CLDNN_ERROR_GREATER_THAN(node.id(), "Negate input offset spatial Y", -input_offset.spatial[1], "input window size spatial Y", filter_size.spatial[1], "First convolution is outside of image. please reduce input offset Y");
64 if (input_layout.format == format::winograd_2x3_s1_weights || input_layout.format == format::winograd_2x3_s1_fused_weights ||
65 input_layout.format == format::winograd_6x3_s1_fused_weights || input_layout.format == format::image_2d_weights_winograd_6x3_s1_fbxyb || input_layout.format == format::image_2d_weights_winograd_6x3_s1_xfbyb)
66 CLDNN_ERROR_MESSAGE(node.id(), "Input for convolution should not be in windograd weights format - it is reserved for weights only");
68 if (input_layout.format == format::winograd_2x3_s1_data)
70 CLDNN_ERROR_NOT_EQUAL(node.id(), "convolution split", split, "expected value", 1, "Convolution with winograd input only supports split == 1");
71 CLDNN_ERROR_NOT_EQUAL(node.id(), "stride spatial X", stride.spatial[0], "expected value", 1, "Convolution's input in winograd_2x3_s1_data format can only be used with stride 1x1");
72 CLDNN_ERROR_NOT_EQUAL(node.id(), "stride spatial Y", stride.spatial[1], "expected value", 1, "Convolution's input in winograd_2x3_s1_data format can only be used with stride 1x1");
73 CLDNN_ERROR_NOT_EQUAL(node.id(), "Dilatation spatial X", dilation.spatial[0], "expected value", 1, "Winograd 2x3 convolution does not support dilatation");
74 CLDNN_ERROR_NOT_EQUAL(node.id(), "Dilatation spatial Y", dilation.spatial[1], "expected value", 1, "Winograd 2x3 convolution does not support dilatation");
75 if (input_layout.size.feature[0] % 32 != 0)
76 CLDNN_ERROR_MESSAGE(node.id(), "Input for winograd 2x3 convolution should have features count divisable by 32");
77 if (weights_layout.size.batch[0] % 32 != 0)
78 CLDNN_ERROR_MESSAGE(node.id(), "Number of filters (OFM) for winograd 2x3 convolution should be divisable by 32");
80 if (node.get_primitive()->with_activation)
81 CLDNN_ERROR_MESSAGE(node.id(), "Winograd 2x3 convolution should not have activation fused - activation should be performed at transformation from winograd domain stage");
83 CLDNN_ERROR_LESS_THAN(node.id(), "input width", input_layout.size.spatial[0], "filter width", 3, "Convolution input is smaller than weights");
84 CLDNN_ERROR_LESS_THAN(node.id(), "input height", input_layout.size.spatial[1], "filter height", 3, "Convolution input is smaller than weights");
86 constexpr tensor::value_type filter_height = 3; //by definition of format::winograd_2x3_s1_data (our assumption)
87 constexpr tensor::value_type winograd_filter_height = filter_height; //for this format, winograd filter is considered to be a set of 1d filters so its height should remain the same as original filter's
89 return layout{ input_layout.data_type, input_layout.format, tensor{ input_layout.size.batch[0], weights_layout.size.batch[0], input_layout.size.spatial[0], input_layout.size.spatial[1] - winograd_filter_height + 1 }, input_layout.data_padding };
92 // get output feature map from weights. It should be the same as number of biases. Will be verifed in convolution::create()
93 auto number_of_features = weights_layout.size.batch[0] * static_cast<int32_t>(split);
95 if (desc->with_output_size)
97 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "User defined output spatial X", desc->output_size.spatial[0], "value", 0, "must be positive(>= 1)");
98 CLDNN_ERROR_LESS_OR_EQUAL_THAN(node.id(), "User defined output spatial Y", desc->output_size.spatial[1], "value", 0, "must be positive(>= 1)");
100 tensor output_size(input_layout.size.batch[0], number_of_features,
101 desc->output_size.spatial[0], desc->output_size.spatial[1]);
102 return { input_layout.data_type, input_layout.format, output_size };
105 auto output_range = calc_sliding_window_output_range<swor_mode::all>(
106 input_layout.size, filter_size, input_offset, stride, dilation, true, 1);
108 tensor::value_type output_features = desc->output_size.feature[0] != 0 ? desc->output_size.feature[0] : number_of_features;
109 tensor output_size = tensor(input_layout.size.batch[0], output_features,
110 output_range.spatial[0], output_range.spatial[1]);
112 // due to performance reason for using fs_bs_yx_bsv4_fsv32 first convolution have 3 features, so first conv layer will take byxf and return fs_bs_yx_bsv4_fsv32
113 if (input_layout.data_type == data_types::i8 && input_layout.format == format::byx8_f4 && input_layout.size.batch[0] % 4 == 0 && input_layout.size.feature[0] == 3)
115 return layout{ input_layout.data_type, cldnn::format::fs_bs_yx_bsv4_fsv32, output_size };
118 return { input_layout.data_type, input_layout.format, output_size };
121 std::string convolution_inst::to_string(convolution_node const& node)
123 auto desc = node.get_primitive();
124 auto strd = desc->stride;
125 auto split = node.get_split();
126 auto dilation = desc->dilation;
127 auto node_info = node.desc_to_json();
128 auto activation = desc->with_activation ? " true" : "false";
130 std::stringstream primitive_description;
132 json_composite conv_info;
133 conv_info.add("stride", strd.to_string());
134 conv_info.add("input offset", desc->input_offset.to_string());
135 conv_info.add("padding above", desc->padding_above.to_string());
136 conv_info.add("padding below", desc->padding_below.to_string());
137 conv_info.add("split", split);
138 conv_info.add("dilation", dilation.to_string());
139 conv_info.add("with activation", activation);
140 conv_info.add("slope", desc->activation_negative_slope);
141 if (desc->with_output_size)
143 json_composite ud_out_size_info;
144 ud_out_size_info.add("size", desc->output_size.to_string());
145 conv_info.add("with user defined output size", ud_out_size_info);
148 node_info->add("convolution info", conv_info);
149 node_info->dump(primitive_description);
151 return primitive_description.str();
154 convolution_inst::typed_primitive_inst(network_impl& network, convolution_node const& node)
155 : parent(network, node)
157 auto stride = argument.stride;
159 auto input_inst = node.input().get_output_layout();
160 auto output_inst = node.get_output_layout();
161 auto output_size = output_inst.size;
163 CLDNN_ERROR_NOT_EQUAL(node.id(), "Input number of dimensions", input_inst.size.raw.size(), "output number of dimensions", output_inst.size.raw.size(), "Input/output dims mismatch");
164 CLDNN_ERROR_NOT_EQUAL(node.id(), "Stride number of dimensions", stride.raw.size(), "output number of dimensions", output_inst.size.raw.size(), "stride/output dims mismatch");
166 auto split = node.get_split();
167 for (decltype(split) j = 0; j < split; j++)
169 auto filter_inst = node.weights(j).get_output_layout(); //convolution filter
172 auto bias_inst = node.bias(j).get_output_layout();
173 CLDNN_ERROR_NOT_EQUAL(node.id(), "Bias batch[0]", bias_inst.size.batch[0], "expected size of batch", 1, "Biases isn't 1D vector.");
174 CLDNN_ERROR_NOT_EQUAL(node.id(), "Bias feature[0]", bias_inst.size.feature[0], "expected size of feature", 1, "Biases isn't 1D vector.");
175 CLDNN_ERROR_NOT_EQUAL(node.id(), "Bias spatial[1]", bias_inst.size.spatial[1], "expected size of spatial[1]", 1, "Biases isn't 1D vector.");
177 CLDNN_ERROR_NOT_EQUAL(node.id(), "Bias spatial[0]", bias_inst.size.spatial[0], "expected feature map number", output_size.feature[0] / split, "Bias/fm mismatch");
180 auto input_offset = argument.input_offset;
182 CLDNN_ERROR_NOT_EQUAL(node.id(), "Weights number of dimensions", filter_inst.size.raw.size(), "output number of dimensions", output_inst.size.raw.size(), "Weights/output dims mismatch");
183 CLDNN_ERROR_NOT_EQUAL(node.id(), "Convolution padding mode", node.get_output_layout().data_padding.filling_value(), "padding value", 0.0f, "Unknown padding mode.");
184 CLDNN_ERROR_NOT_EQUAL(node.id(), "Input offset number of dimensions", input_offset.raw.size(), "input number of dimensions", input_inst.size.raw.size(), "Input offset/ input size mismatch");
185 CLDNN_ERROR_NOT_EQUAL(node.id(), "Output feature size", output_size.feature.size(), "expected feature size", 1, "Only one-dimensional features are supported");
186 CLDNN_ERROR_NOT_EQUAL(node.id(), "Output batch size", output_size.batch.size(), "expected output size", 1, "Only one-dimensional batch size are supported");
187 CLDNN_ERROR_NOT_EQUAL(node.id(), "Weights spatial size", filter_inst.size.spatial.size(), "expected weights spatial size", 2, "Weights have to have 2 dimensions in spatial domain.");
188 CLDNN_ERROR_LESS_THAN(node.id(), "Weights feature maps number", (input_inst.size.feature[0] - input_offset.feature[0]) / split, "input feature maps number", filter_inst.size.feature[0], "Weights/ifm mismatch");
189 if (filter_inst.format == format::bf_lyx_yx) // local convolution
191 auto local = filter_inst.size.local;
192 CLDNN_ERROR_NOT_EQUAL(node.id(), "Number of local x dimension", local[0], "output x dimension", output_inst.size.spatial[0], "Weights/output dims mismatch");
193 CLDNN_ERROR_NOT_EQUAL(node.id(), "Number of local y dimension", local[1], "output y dimension", output_inst.size.spatial[1], "Weights/output dims mismatch");