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

   1 // Copyright (C) 2018 Intel Corporation
   2 //
   3 // SPDX-License-Identifier: Apache-2.0
   4 //
   5
   6 #include <gtest/gtest.h>
   7 #include <gmock/gmock-spec-builders.h>
   8 #include "mkldnn_plugin/mkldnn_graph.h"
   9 #include "mock_mkldnn_primitive.hpp"
  10
  11 #include "test_graph.hpp"
  12
  13 #include "single_layer_common.hpp"
  14 #include <mkldnn_plugin/mkldnn_extension_utils.h>
  15 #include <inference_engine/cnn_network_impl.hpp>
  16 #include "tests_common.hpp"
  17
  18
  19 using namespace ::testing;
  20 using namespace std;
  21 using namespace mkldnn;
  22
  23
  24 struct conv_test_params {
  25     struct {
  26         size_t n;
  27         size_t c;
  28         size_t h;
  29         size_t w;
  30     } in;
  31
  32     size_t krn_w;
  33     size_t krn_h;
  34     size_t str_w;
  35     size_t str_h;
  36     size_t pad_w;
  37     size_t pad_h;
  38
  39     size_t out_c;
  40     size_t grp_c;
  41
  42     size_t num_prim_desc;
  43
  44     int selectedType;
  45     std::vector<MKLDNNPlugin::impl_desc_type> preferTypes;
  46
  47     std::vector<std::function<void(MKLDNNPlugin::PrimitiveDescInfo)>> comp;
  48 };
  49
  50 template <typename data_t>
  51 void ref_conv(const InferenceEngine::TBlob<data_t> &src, const data_t *weights, const size_t weightsSize,
  52                 InferenceEngine::TBlob<data_t> &dst, conv_test_params prm) {
  53     size_t KW = prm.krn_w;
  54     size_t KH = prm.krn_h;
  55     size_t GC = prm.grp_c;
  56
  57     size_t IC = src.dims()[1];
  58     size_t IH = src.dims()[2];
  59     size_t IW = src.dims()[3];
  60
  61     size_t OW = (IW + 2 * prm.pad_w - prm.krn_w) / prm.str_w + 1;
  62     size_t OH = (IH + 2 * prm.pad_h - prm.krn_h) / prm.str_h + 1;
  63     size_t OC = prm.out_c;
  64
  65
  66     const data_t *src_data = src.readOnly();
  67     const data_t *weights_data = weights;
  68     const data_t *bias_data = weights_data + KW * KH * OC * IC / GC;
  69     data_t *dst_data = dst.data();
  70
  71     IE_ASSERT(KW * KH * OC * IC / GC + OC == weightsSize);
  72     IE_ASSERT(OW == dst.dims()[0]);
  73     IE_ASSERT(OH == dst.dims()[1]);
  74
  75     for (uint32_t g = 0; g < GC; g++) {
  76         for (uint32_t oc = 0; oc < OC / GC; oc++) {
  77             for (uint32_t oh = 0; oh < OH; oh++) {
  78                 for (uint32_t ow = 0; ow < OW; ow++) {
  79                     size_t oidx = g * OC / GC * OH * OW
  80                                   + oc * OH * OW + oh * OW + ow;
  81                     dst_data[oidx] = bias_data[g * OC / GC + oc];
  82
  83                     for (size_t ic = 0; ic < IC / GC; ic++) {
  84                         for (size_t kh = 0; kh < KH; kh++) {
  85                             for (size_t kw = 0; kw < KW; kw++) {
  86                                 int32_t iw = ow * prm.str_w - prm.pad_w + kw;
  87                                 int32_t ih = oh * prm.str_h - prm.pad_h + kh;
  88                                 if (iw < 0 || iw >= (int32_t)IW || ih < 0
  89                                     || ih >= (int32_t)IH)
  90                                     continue;
  91                                 size_t iidx = g * IC / GC * IH * IW
  92                                               + ic * IH * IW + ih * IW + iw;
  93                                 size_t widx = g * OC / GC * IC / GC * KH * KW
  94                                               + oc * IC / GC * KH * KW
  95                                               + ic * KH * KW + kh * KW + kw;
  96
  97                                 dst_data[ oidx] += src_data[iidx] * weights_data[widx];
  98                             }
  99                         }
 100                     }
 101                 }
 102             }
 103         }
 104     }
 105 }
 106
 107 class MKLDNNGraphConvolutionTests: public TestsCommon,
 108                                    public WithParamInterface<conv_test_params> {
 109     std::string model_t = R"V0G0N(
 110 <Net Name="Convolution_Only" version="2" precision="FP32" batch="1">
 111     <layers>
 112         <layer name="in1" type="Input" precision="FP32" id="0">
 113             <output>
 114                 <port id="0">
 115                     <dim>_IN_</dim>
 116                     <dim>_IC_</dim>
 117                     <dim>_IH_</dim>
 118                     <dim>_IW_</dim>
 119                 </port>
 120             </output>
 121         </layer>
 122         <layer name="conv1" id="1" type="Convolution" precision="FP32">
 123             <convolution stride-x="_SW_" stride-y="_SH_"
 124                          pad-x="_PW_"    pad-y="_PH_"
 125                          kernel-x="_KW_" kernel-y="_KH_"
 126                          output="_OC_"   group="_GC_" PrimitivesPriority="_IMPLS_"/>
 127
 128             <weights offset="0" size="_S1_" />
 129             <biases offset="_S1_" size="_S2_" />
 130
 131             <input>
 132                 <port id="1">
 133                     <dim>_IN_</dim>
 134                     <dim>_IC_</dim>
 135                     <dim>_IH_</dim>
 136                     <dim>_IW_</dim>
 137                 </port>
 138             </input>
 139             <output>
 140                 <port id="2">
 141                     <dim>_IN_</dim>
 142                     <dim>_OC_</dim>
 143                     <dim>_OH_</dim>
 144                     <dim>_OW_</dim>
 145                 </port>
 146             </output>
 147         </layer>
 148     </layers>
 149     <edges>
 150         <edge from-layer="0" from-port="0" to-layer="1" to-port="1"/>
 151     </edges>
 152 </Net>
 153 )V0G0N";
 154
 155 protected:
 156     std::string getModel(conv_test_params p) {
 157         std::string model = model_t;
 158         REPLACE_WITH_NUM(model, "_IW_", p.in.w);
 159         REPLACE_WITH_NUM(model, "_IH_", p.in.h);
 160         REPLACE_WITH_NUM(model, "_IC_", p.in.c);
 161         REPLACE_WITH_NUM(model, "_IN_", p.in.n);
 162
 163         REPLACE_WITH_NUM(model, "_KW_", p.krn_w);
 164         REPLACE_WITH_NUM(model, "_KH_", p.krn_h);
 165         REPLACE_WITH_NUM(model, "_SW_", p.str_w);
 166         REPLACE_WITH_NUM(model, "_SH_", p.str_h);
 167         REPLACE_WITH_NUM(model, "_PW_", p.pad_w);
 168         REPLACE_WITH_NUM(model, "_PH_", p.pad_h);
 169
 170         REPLACE_WITH_NUM(model, "_GC_", p.grp_c);
 171         REPLACE_WITH_NUM(model, "_OC_", p.out_c);
 172         REPLACE_WITH_NUM(model, "_OH_", (p.in.h + 2 * p.pad_h - p.krn_h) / p.str_h + 1);
 173         REPLACE_WITH_NUM(model, "_OW_", (p.in.w + 2 * p.pad_w - p.krn_w) / p.str_w + 1);
 174
 175         size_t w_data_size = (p.krn_w * p.krn_h * p.out_c * p.in.c / p.grp_c) * sizeof(float);
 176         size_t b_data_size = p.out_c * sizeof(float);
 177         REPLACE_WITH_NUM(model, "_S1_", w_data_size);
 178         REPLACE_WITH_NUM(model, "_S2_", b_data_size);
 179         std::string impls;
 180         for (const auto& preferType : p.preferTypes) {
 181             if (!impls.empty())
 182                 impls += ",";
 183             impls += "cpu:" + MKLDNNGraphTestClass::getStrPrimitiveDescriptorType(preferType);
 184         }
 185         REPLACE_WITH_STR(model, "_IMPLS_", impls);
 186         return model;
 187     }
 188
 189     virtual void TearDown() {
 190     }
 191
 192     virtual void SetUp() {
 193         try {
 194             TestsCommon::SetUp();
 195             conv_test_params p = ::testing::WithParamInterface<conv_test_params>::GetParam();
 196             std::string model = getModel(p);
 197
 198             InferenceEngine::CNNNetReader net_reader;
 199             ASSERT_NO_THROW(net_reader.ReadNetwork(model.data(), model.length()));
 200
 201             InferenceEngine::TBlob<uint8_t> *weights = new InferenceEngine::TBlob<uint8_t>(InferenceEngine::Precision::U8, InferenceEngine::C, {(p.krn_w * p.krn_h * p.out_c * p.in.c / p.grp_c + p.out_c)
 202                                                               * sizeof(float)});
 203             weights->allocate();
 204             fill_data((float *) weights->buffer(), weights->size() / sizeof(float));
 205             InferenceEngine::TBlob<uint8_t>::Ptr weights_ptr = InferenceEngine::TBlob<uint8_t>::Ptr(weights);
 206
 207             net_reader.SetWeights(weights_ptr);
 208
 209             MKLDNNGraphTestClass graph;
 210             graph.CreateGraph(net_reader.getNetwork());
 211
 212             auto& nodes = graph.getNodes();
 213             nodes = graph.getNodes();
 214             for (auto &node : nodes) {
 215                 if (node->getType() == MKLDNNPlugin::Convolution) {
 216                     ASSERT_LE(p.num_prim_desc, node->getSupportedPrimitiveDescriptors().size());
 217                     for (size_t j = 0; j < p.num_prim_desc && j < p.comp.size(); j++) {
 218                         p.comp.at(j)(node->getSupportedPrimitiveDescriptors().at(j));
 219                     }
 220                     ASSERT_NE(nullptr, node->getSelectedPrimitiveDescriptor());
 221                     ASSERT_EQ(p.selectedType,
 222                               node->getSelectedPrimitiveDescriptor()->getImplementationType() & p.selectedType);
 223                 }
 224             }
 225
 226             InferenceEngine::SizeVector dims_src = {p.in.n, p.in.c, p.in.h, p.in.w};
 227
 228             InferenceEngine::Blob::Ptr src = InferenceEngine::make_shared_blob<float, const InferenceEngine::SizeVector>(InferenceEngine::Precision::FP32, InferenceEngine::NCHW, dims_src);
 229             src->allocate();
 230             fill_data(src->buffer(), src->size());
 231
 232             auto * srcPtr = dynamic_cast<InferenceEngine::TBlob<float>*>(src.get());
 233
 234             if (srcPtr == nullptr)
 235                 FAIL() << "Cannot cast blob to TBlob<float>.";
 236
 237             InferenceEngine::BlobMap srcs;
 238             srcs.insert(std::pair<std::string, InferenceEngine::Blob::Ptr>("in1", src));
 239
 240             InferenceEngine::OutputsDataMap out;
 241             out = net_reader.getNetwork().getOutputsInfo();
 242             InferenceEngine::BlobMap outputBlobs;
 243
 244             std::pair<std::string, InferenceEngine::DataPtr> item = *out.begin();
 245
 246             InferenceEngine::TBlob<float>::Ptr output;
 247             output = InferenceEngine::make_shared_blob<float>(item.second->getTensorDesc());
 248             output->allocate();
 249             outputBlobs[item.first] = output;
 250
 251             graph.Infer(srcs, outputBlobs);
 252
 253
 254             InferenceEngine::TBlob<float> dst_ref(item.second->getTensorDesc());
 255             dst_ref.allocate();
 256             ref_conv(*srcPtr, (const float *)weights->buffer(), weights->size() / sizeof(float), dst_ref, p);
 257             compare(*output, dst_ref);
 258         } catch (const InferenceEngine::details::InferenceEngineException &e) {
 259             FAIL() << e.what();
 260         }
 261     }
 262 };
 263
 264 TEST_P(MKLDNNGraphConvolutionTests, TestsConvolution) {}
 265
 266 INSTANTIATE_TEST_CASE_P(
 267         TestConvolution, MKLDNNGraphConvolutionTests,
 268         ::testing::Values(
 269                 conv_test_params{{1, 9, 16, 32},
 270                                  1, 1, 1, 1, 0, 0, 17, 1, 7, MKLDNNPlugin::impl_desc_type::jit | MKLDNNPlugin::impl_desc_type::_1x1,
 271                 },
 272                 conv_test_params{{1, 9, 32, 16},
 273                                  2, 4, 1, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 274                 conv_test_params{{1, 9, 32, 16},
 275                                  2, 4, 2, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 276                 conv_test_params{{1, 3, 40, 40},
 277                                  3, 3, 1, 2, 0, 0, 20, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 278                 conv_test_params{{1, 1, 40, 40},
 279                                  3, 3, 1, 2, 0, 0, 20, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 280                 conv_test_params{{1, 1, 32, 16},
 281                                  2, 4, 2, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 282                 conv_test_params{{1, 9, 16, 32},
 283                                  1, 1, 1, 1, 0, 0, 17, 1, 7, MKLDNNPlugin::impl_desc_type::gemm,
 284                                  {MKLDNNPlugin::impl_desc_type::gemm_any,
 285                                   MKLDNNPlugin::impl_desc_type::gemm_blas,
 286                                   MKLDNNPlugin::impl_desc_type::gemm_avx512,
 287                                   MKLDNNPlugin::impl_desc_type::gemm_avx2,
 288                                   MKLDNNPlugin::impl_desc_type::gemm_sse42}
 289                 },
 290                 conv_test_params{{1, 9, 32, 16},
 291                                  2, 4, 1, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::ref_any, {MKLDNNPlugin::impl_desc_type::ref_any} }));
 292
 293 class MKLDNNGraphDynBatchConvolutionTests: public MKLDNNGraphConvolutionTests {
 294 protected:
 295     virtual void SetUp() {
 296         try {
 297             TestsCommon::SetUp();
 298             conv_test_params p = ::testing::WithParamInterface<conv_test_params>::GetParam();
 299             std::string model = getModel(p);
 300             size_t MB = p.in.n;
 301             if (MB < 2)
 302                 MB = 2;
 303
 304             InferenceEngine::CNNNetReader net_reader;
 305             ASSERT_NO_THROW(net_reader.ReadNetwork(model.data(), model.length()));
 306
 307             InferenceEngine::TBlob<uint8_t> *weights = new InferenceEngine::TBlob<uint8_t>(InferenceEngine::Precision::U8, InferenceEngine::C,
 308                     {(p.krn_w * p.krn_h * p.out_c * p.in.c / p.grp_c + p.out_c) * sizeof(float)});
 309             weights->allocate();
 310             fill_data((float *) weights->buffer(), weights->size() / sizeof(float));
 311             InferenceEngine::TBlob<uint8_t>::Ptr weights_ptr = InferenceEngine::TBlob<uint8_t>::Ptr(weights);
 312
 313             net_reader.SetWeights(weights_ptr);
 314             InferenceEngine::CNNNetwork network = net_reader.getNetwork();
 315             auto implNet = dynamic_cast<InferenceEngine::details::CNNNetworkImpl *>(&((InferenceEngine::ICNNNetwork&)network));
 316             ASSERT_NE(nullptr, implNet) << "Failed to cast ICNNNetwork to CNNNetworkImpl";
 317             InferenceEngine::ResponseDesc resp;
 318             InferenceEngine::StatusCode sts  = implNet->setBatchSizeReshape(MB, &resp);
 319             ASSERT_EQ((int)InferenceEngine::StatusCode::OK, sts) << resp.msg;
 320
 321             MKLDNNGraphTestClass graph;
 322             graph.CreateGraph(net_reader.getNetwork());
 323
 324             InferenceEngine::SizeVector dims_src = {MB, p.in.c, p.in.h, p.in.w};
 325
 326             InferenceEngine::Blob::Ptr src = InferenceEngine::make_shared_blob<float, const InferenceEngine::SizeVector>(InferenceEngine::Precision::FP32, InferenceEngine::NCHW, dims_src);
 327             src->allocate();
 328             fill_data(src->buffer(), src->size());
 329
 330             auto * srcPtr = dynamic_cast<InferenceEngine::TBlob<float>*>(src.get());
 331
 332             if (srcPtr == nullptr)
 333                 FAIL() << "Cannot cast blob to TBlob<float>.";
 334
 335             InferenceEngine::BlobMap srcs;
 336             srcs.insert(std::pair<std::string, InferenceEngine::Blob::Ptr>("in1", src));
 337
 338             InferenceEngine::OutputsDataMap out;
 339             out = net_reader.getNetwork().getOutputsInfo();
 340             InferenceEngine::BlobMap outputBlobs;
 341
 342             std::pair<std::string, InferenceEngine::DataPtr> item = *out.begin();
 343
 344             InferenceEngine::TBlob<float>::Ptr output;
 345             output = InferenceEngine::make_shared_blob<float>(item.second->getTensorDesc());
 346             output->allocate();
 347             outputBlobs[item.first] = output;
 348
 349             auto checkConvolution = [](const MKLDNNPlugin::MKLDNNNodePtr& node) {
 350                 return node->getType() == MKLDNNPlugin::Convolution ||
 351                        node->getType() == MKLDNNPlugin::Convolution_Activation ||
 352                        node->getType() == MKLDNNPlugin::Convolution_Sum ||
 353                        node->getType() == MKLDNNPlugin::Convolution_Sum_Activation;
 354             };
 355
 356             graph.checkDynBatch(srcs, outputBlobs, MB, MB, checkConvolution, MKLDNNGraphTestClass::CheckDynBatchType::Child);
 357             graph.checkDynBatch(srcs, outputBlobs, 1, MB, checkConvolution, MKLDNNGraphTestClass::CheckDynBatchType::Child);
 358         } catch (const InferenceEngine::details::InferenceEngineException &e) {
 359             FAIL() << e.what();
 360         }
 361     }
 362 };
 363
 364 TEST_P(MKLDNNGraphDynBatchConvolutionTests, TestsDynBatchConvolution) {}
 365
 366 INSTANTIATE_TEST_CASE_P(
 367         TestDynBatchConvolution, MKLDNNGraphDynBatchConvolutionTests,
 368         ::testing::Values(
 369                 conv_test_params{{1, 8, 16, 32},
 370                                  1, 1, 1, 1, 0, 0, 17, 1, 7, MKLDNNPlugin::impl_desc_type::jit | MKLDNNPlugin::impl_desc_type::_1x1,
 371                 },
 372                 conv_test_params{{1, 9, 32, 16},
 373                                  2, 4, 1, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 374                 conv_test_params{{1, 9, 32, 16},
 375                                  2, 4, 2, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 376                 conv_test_params{{1, 3, 40, 40},
 377                                  3, 3, 1, 2, 0, 0, 20, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 378                 conv_test_params{{1, 1, 40, 40},
 379                                  3, 3, 1, 2, 0, 0, 20, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 380                 conv_test_params{{1, 1, 32, 16},
 381                                  2, 4, 2, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::jit },
 382                 conv_test_params{{1, 9, 16, 32},
 383                                  1, 1, 1, 1, 0, 0, 17, 1, 7, MKLDNNPlugin::impl_desc_type::gemm,
 384                                  {MKLDNNPlugin::impl_desc_type::gemm_any,
 385                                   MKLDNNPlugin::impl_desc_type::gemm_blas,
 386                                   MKLDNNPlugin::impl_desc_type::gemm_avx512,
 387                                   MKLDNNPlugin::impl_desc_type::gemm_avx2,
 388                                   MKLDNNPlugin::impl_desc_type::gemm_sse42}
 389                 },
 390                 conv_test_params{{1, 9, 32, 16},
 391                                  2, 4, 1, 1, 0, 0, 17, 1, 5, MKLDNNPlugin::impl_desc_type::ref_any, {MKLDNNPlugin::impl_desc_type::ref_any} }));