inference-engine/src/mkldnn_plugin/mkldnn_extension_utils.cpp

   1 // Copyright (C) 2018-2019 Intel Corporation
   2 // SPDX-License-Identifier: Apache-2.0
   3 //
   4
   5 #include "mkldnn_extension_utils.h"
   6 #include <limits>
   7 #include <vector>
   8
   9 using namespace mkldnn;
  10 using namespace MKLDNNPlugin;
  11
  12 uint8_t MKLDNNExtensionUtils::sizeOfDataType(mkldnn::memory::data_type dataType) {
  13     switch (dataType) {
  14     case mkldnn::memory::data_type::f32:
  15         return 4;
  16     case mkldnn::memory::data_type::s32:
  17         return 4;
  18     case mkldnn::memory::data_type::s16:
  19         return 2;
  20     case mkldnn::memory::data_type::s8:
  21         return 1;
  22     case mkldnn::memory::data_type::u8:
  23         return 1;
  24     case mkldnn::memory::data_type::bin:
  25         return 1;
  26     case mkldnn::memory::data_type::data_undef:
  27         return 0;
  28
  29     default:
  30         THROW_IE_EXCEPTION << "Unsupported data type.";
  31     }
  32 }
  33
  34 memory::data_type MKLDNNExtensionUtils::IEPrecisionToDataType(InferenceEngine::Precision prec) {
  35     switch (prec) {
  36         case InferenceEngine::Precision::FP32:
  37             return memory::f32;
  38         case InferenceEngine::Precision::I32:
  39             return memory::s32;
  40         case InferenceEngine::Precision::I16:
  41             return memory::s16;
  42         case InferenceEngine::Precision::I8:
  43             return memory::s8;
  44         case InferenceEngine::Precision::U8:
  45             return memory::u8;
  46         case InferenceEngine::Precision::BIN:
  47             return memory::bin;
  48
  49         default: {
  50             THROW_IE_EXCEPTION << "The plugin does not support " << prec.name();
  51         }
  52     }
  53 }
  54
  55 InferenceEngine::Precision MKLDNNExtensionUtils::DataTypeToIEPrecision(memory::data_type dataType) {
  56     switch (dataType) {
  57         case memory::f32:
  58             return InferenceEngine::Precision(InferenceEngine::Precision::FP32);
  59         case memory::s32:
  60             return InferenceEngine::Precision::I32;
  61         case memory::s16:
  62             return InferenceEngine::Precision::I16;
  63         case memory::s8:
  64             return InferenceEngine::Precision::I8;
  65         case memory::u8:
  66             return InferenceEngine::Precision::U8;
  67         case memory::bin:
  68             return InferenceEngine::Precision::BIN;
  69
  70         default: {
  71             THROW_IE_EXCEPTION << "Unsupported data type.";
  72         }
  73     }
  74 }
  75
  76 InferenceEngine::TensorDesc MKLDNNExtensionUtils::getUninitTensorDesc(const InferenceEngine::TensorDesc &desc) {
  77     std::vector<size_t> notInitArr;
  78     std::vector<size_t> zeroArr;
  79     for (size_t i = 0; i < desc.getBlockingDesc().getBlockDims().size(); i++) {
  80         notInitArr.push_back(std::numeric_limits<size_t>::max());
  81         zeroArr.push_back(0);
  82     }
  83     // MKLDNN doesn't support offset_padding_to_data[i] != 0 (assert(src_d_blk.offset_padding_to_data[d] == 0);)
  84     return desc.getLayout() == InferenceEngine::Layout::ANY ? desc :
  85            InferenceEngine::TensorDesc(desc.getPrecision(), desc.getDims(),
  86                                        {desc.getBlockingDesc().getBlockDims(), desc.getBlockingDesc().getOrder(),
  87                                         std::numeric_limits<size_t>::max(), zeroArr, notInitArr});
  88 }
  89
  90 bool MKLDNNExtensionUtils::initTensorsAreEqual(InferenceEngine::TensorDesc desc1, InferenceEngine::TensorDesc desc2) {
  91     if (desc1.getDims() != desc2.getDims() || desc1.getPrecision() != desc2.getPrecision())
  92         return false;
  93     if (desc1.getLayout() == InferenceEngine::Layout::ANY || desc2.getLayout() == InferenceEngine::Layout::ANY)
  94         return true;
  95     bool batch1 = desc1.getDims()[0] == 1;
  96     const auto& in1Block = desc1.getBlockingDesc();
  97     const auto& in2Block = desc2.getBlockingDesc();
  98     size_t uninitNum = std::numeric_limits<size_t>::max();
  99     if (in1Block.getBlockDims().size() != in2Block.getBlockDims().size())
 100         return false;
 101     for (size_t i = 0; i < in1Block.getBlockDims().size(); i++) {
 102         if (in1Block.getBlockDims()[i] != in2Block.getBlockDims()[i] &&
 103                 in1Block.getBlockDims()[i] != uninitNum && in2Block.getBlockDims()[i] != uninitNum)
 104             return false;
 105         if (in1Block.getOffsetPaddingToData()[i] != in2Block.getOffsetPaddingToData()[i] &&
 106                 in1Block.getOffsetPaddingToData()[i] != uninitNum && in2Block.getOffsetPaddingToData()[i] != uninitNum)
 107             return false;
 108         if (i >= batch1 && in1Block.getStrides()[i] != in2Block.getStrides()[i] &&
 109                 in1Block.getStrides()[i] != uninitNum && in2Block.getStrides()[i] != uninitNum)
 110             return false;
 111         if (in1Block.getOrder()[i] != in2Block.getOrder()[i] &&
 112                 in1Block.getOrder()[i] != uninitNum && in2Block.getOrder()[i] != uninitNum)
 113             return false;
 114     }
 115     return !(in1Block.getOffsetPadding() != in2Block.getOffsetPadding() &&
 116         in1Block.getOffsetPadding() != uninitNum && in2Block.getOffsetPadding() != uninitNum);
 117 }