inference-engine/tests/unit/engines/mkldnn/graph/layers/internal/graph_reorder_test.cpp

   1 // Copyright (C) 2018-2019 Intel Corporation
   2 // SPDX-License-Identifier: Apache-2.0
   3 //
   4
   5 #include <gtest/gtest.h>
   6 #include <gmock/gmock-spec-builders.h>
   7 #include "mkldnn_plugin/mkldnn_graph.h"
   8
   9 #include "test_graph.hpp"
  10
  11 #include <mock_error_listener.hpp>
  12 #include <mkldnn_plugin/mkldnn_extension_mngr.h>
  13 #include "tests_common.hpp"
  14
  15 using namespace ::testing;
  16 using namespace std;
  17 using namespace mkldnn;
  18
  19 class MKLDNNGraphReorderTests: public TestsCommon {
  20 protected:
  21     virtual void SetUp() {
  22         TestsCommon::SetUp();
  23     }
  24 };
  25
  26 TEST_F(MKLDNNGraphReorderTests, cannotCreatePrimitiveDescriprorsWithoutOtherLayers) {
  27     std::shared_ptr<MKLDNNPlugin::MKLDNNNode> node;
  28     mkldnn::engine eng(mkldnn::engine(mkldnn::engine::kind::cpu, 0));
  29
  30     InferenceEngine::CNNLayerPtr layer(new InferenceEngine::CNNLayer({"TestReorder", "Reorder", InferenceEngine::Precision::FP32}));
  31     node.reset(MKLDNNPlugin::MKLDNNNode::CreateNode(layer, eng, {}));
  32     ASSERT_EQ(MKLDNNPlugin::Type::Reorder, node->getType());
  33
  34     ASSERT_THROW(node->getSupportedDescriptors(), InferenceEngine::details::InferenceEngineException);
  35 }
  36
  37 TEST_F(MKLDNNGraphReorderTests, CreateReorder) {
  38     std::string model = R"V0G0N(
  39 <Net Name="Convolution_Only" version="2" precision="FP32" batch="1">
  40     <layers>
  41         <layer name="in1" type="Input" precision="FP32" id="0">
  42             <output>
  43                 <port id="0">
  44                     <dim>1</dim>
  45                     <dim>9</dim>
  46                     <dim>16</dim>
  47                     <dim>32</dim>
  48                 </port>
  49             </output>
  50         </layer>
  51         <layer name="conv1" id="1" type="Convolution" precision="FP32">
  52             <convolution stride-x="1" stride-y="1"
  53                          pad-x="0"    pad-y="0"
  54                          kernel-x="1" kernel-y="1"
  55                          output="17"   group="1"/>
  56
  57             <weights offset="0" size="612" />
  58             <biases offset="612" size="68" />
  59
  60             <input>
  61                 <port id="1">
  62                     <dim>1</dim>
  63                     <dim>9</dim>
  64                     <dim>16</dim>
  65                     <dim>32</dim>
  66                 </port>
  67             </input>
  68             <output>
  69                 <port id="2">
  70                     <dim>1</dim>
  71                     <dim>17</dim>
  72                     <dim>16</dim>
  73                     <dim>32</dim>
  74                 </port>
  75             </output>
  76         </layer>
  77     </layers>
  78     <edges>
  79         <edge from-layer="0" from-port="0" to-layer="1" to-port="1"/>
  80     </edges>
  81 </Net>
  82 )V0G0N";
  83
  84     InferenceEngine::CNNNetReader net_reader;
  85     ASSERT_NO_THROW(net_reader.ReadNetwork(model.data(), model.length()));
  86
  87     InferenceEngine::TBlob<uint8_t> *weights = new InferenceEngine::TBlob<uint8_t>(InferenceEngine::Precision::U8, InferenceEngine::C,
  88                                                                                    {(1 * 1 * 17 * 9 / 1 + 17)
  89                                                       * sizeof(float)});
  90     weights->allocate();
  91     fill_data((float *) weights->buffer(), weights->size() / sizeof(float));
  92     InferenceEngine::TBlob<uint8_t>::Ptr weights_ptr = InferenceEngine::TBlob<uint8_t>::Ptr(weights);
  93
  94     net_reader.SetWeights(weights_ptr);
  95
  96     MKLDNNGraphTestClass graph;
  97     graph.CreateGraph(net_reader.getNetwork());
  98
  99     auto& nodes = graph.getNodes();
 100     for (int i = 0; i < nodes.size(); i++) {
 101         if (nodes[i]->getType() == MKLDNNPlugin::Reorder) {
 102             ASSERT_EQ(1, nodes[i]->getSupportedPrimitiveDescriptors().size());
 103             ASSERT_EQ(MKLDNNPlugin::impl_desc_type::reorder,
 104                       nodes[i]->getSupportedPrimitiveDescriptors()[0].getImplementationType());
 105             ASSERT_EQ(1, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().inConfs.size());
 106             if (i == 1) {
 107                 ASSERT_EQ(InferenceEngine::Layout::NCHW, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().inConfs[0].desc.getLayout());
 108                 ASSERT_NE(InferenceEngine::Layout::NCHW, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().outConfs[0].desc.getLayout());
 109             } else {
 110                 ASSERT_NE(InferenceEngine::Layout::NCHW, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().inConfs[0].desc.getLayout());
 111                 ASSERT_EQ(InferenceEngine::Layout::NCHW, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().outConfs[0].desc.getLayout());
 112             }
 113             ASSERT_EQ(1, nodes[i]->getSupportedPrimitiveDescriptors()[0].getConfig().outConfs.size());
 114         }
 115     }
 116 }
 117
 118 TEST_F(MKLDNNGraphReorderTests, CreateInPlaceReorder) {
 119     std::string model = R"V0G0N(
 120 <Net Name="InPlaceReorder_Only" version="2" precision="FP32" batch="1">
 121     <layers>
 122         <layer name="in1" type="Input" precision="FP32" id="0">
 123             <output>
 124                 <port id="0">
 125                     <dim>1</dim>
 126                     <dim>9</dim>
 127                     <dim>16</dim>
 128                     <dim>32</dim>
 129                 </port>
 130             </output>
 131         </layer>
 132         <layer name="reshape1" id="1" type="Reshape" precision="FP32">
 133             <input>
 134                 <port id="1">
 135                     <dim>1</dim>
 136                     <dim>9</dim>
 137                     <dim>16</dim>
 138                     <dim>32</dim>
 139                 </port>
 140             </input>
 141             <output>
 142                 <port id="2">
 143                     <dim>32</dim>
 144                     <dim>144</dim>
 145                 </port>
 146             </output>
 147         </layer>
 148         <layer name="reshape2" id="2" type="Reshape" precision="FP32">
 149             <data axis="0" num_axes="-1" dim="1, 4608"/>
 150             <input>
 151                 <port id="1">
 152                     <dim>32</dim>
 153                     <dim>144</dim>
 154                 </port>
 155             </input>
 156             <output>
 157                 <port id="2">
 158                     <dim>1</dim>
 159                     <dim>3</dim>
 160                     <dim>48</dim>
 161                     <dim>32</dim>
 162                 </port>
 163             </output>
 164         </layer>
 165         <layer name="scaleshift" id="3" type="ScaleShift" precision="FP32">
 166             <input>
 167                 <port id="1">
 168                     <dim>1</dim>
 169                     <dim>3</dim>
 170                     <dim>48</dim>
 171                     <dim>32</dim>
 172                 </port>
 173             </input>
 174             <output>
 175                 <port id="2">
 176                     <dim>1</dim>
 177                     <dim>3</dim>
 178                     <dim>48</dim>
 179                     <dim>32</dim>
 180                 </port>
 181             </output>
 182             <blobs>
 183                 <weights offset="0" size="12"/>
 184                 <biases offset="12" size="12"/>
 185             </blobs>
 186         </layer>
 187     </layers>
 188     <edges>
 189         <edge from-layer="0" from-port="0" to-layer="1" to-port="1"/>
 190         <edge from-layer="1" from-port="2" to-layer="2" to-port="1"/>
 191         <edge from-layer="2" from-port="2" to-layer="3" to-port="1"/>
 192     </edges>
 193 </Net>
 194 )V0G0N";
 195
 196     InferenceEngine::CNNNetReader net_reader;
 197     ASSERT_NO_THROW(net_reader.ReadNetwork(model.data(), model.length()));
 198
 199     InferenceEngine::TBlob<uint8_t> *weights = new InferenceEngine::TBlob<uint8_t>(InferenceEngine::Precision::U8, InferenceEngine::C, {24});
 200     weights->allocate();
 201     float *data = weights->buffer().as<float *>();
 202     size_t dataSize = weights->byteSize() / sizeof(float);
 203     for (size_t i = 0; i < dataSize; i++) {
 204         data[i] = 2;
 205     }
 206     InferenceEngine::TBlob<uint8_t>::Ptr weights_ptr = InferenceEngine::TBlob<uint8_t>::Ptr(weights);
 207
 208     net_reader.SetWeights(weights_ptr);
 209     net_reader.getNetwork().addOutput("reshape1");
 210
 211     MKLDNNGraphTestClass graph;
 212     graph.CreateGraph(net_reader.getNetwork());
 213
 214     InferenceEngine::SizeVector dims_src = {1, 9, 16, 32};
 215
 216     InferenceEngine::Blob::Ptr src = InferenceEngine::make_shared_blob<float, const InferenceEngine::SizeVector>(InferenceEngine::Precision::FP32, InferenceEngine::NCHW, dims_src);
 217     src->allocate();
 218     data = src->buffer().as<float *>();
 219     dataSize = src->size();
 220     for (size_t i = 0; i < dataSize; i++) {
 221         data[i] = 1;
 222     }
 223
 224     auto * srcPtr = dynamic_cast<InferenceEngine::TBlob<float>*>(src.get());
 225
 226     if (srcPtr == nullptr)
 227         FAIL() << "Cannot cast blob to TBlob<float>.";
 228
 229     InferenceEngine::BlobMap srcs;
 230     srcs.insert(std::pair<std::string, InferenceEngine::Blob::Ptr>("in1", src));
 231
 232     InferenceEngine::OutputsDataMap out;
 233     out = net_reader.getNetwork().getOutputsInfo();
 234     InferenceEngine::BlobMap outputBlobs;
 235
 236     auto it = out.begin();
 237     std::pair<std::string, InferenceEngine::DataPtr> item = *it;
 238
 239     InferenceEngine::TBlob<float>::Ptr output1;
 240     output1 = InferenceEngine::make_shared_blob<float>(item.second->getTensorDesc());
 241     output1->allocate();
 242     outputBlobs[item.first] = output1;
 243
 244     item = *(++it);
 245
 246     InferenceEngine::TBlob<float>::Ptr output2;
 247     output2 = InferenceEngine::make_shared_blob<float>(item.second->getTensorDesc());
 248     output2->allocate();
 249     outputBlobs[item.first] = output2;
 250
 251     graph.Infer(srcs, outputBlobs);
 252
 253     data = output1->data();
 254     for (size_t i = 0; i < output1->size(); i++) {
 255         ASSERT_EQ(data[i], 1);
 256     }
 257     data = output2->data();
 258     for (size_t i = 0; i < output2->size(); i++) {
 259         ASSERT_EQ(data[i], 4);
 260     }
 261 }