1 // Copyright (C) 2018 Intel Corporation
2 // SPDX-License-Identifier: Apache-2.0
5 #include "mkldnn_tile_node.h"
8 #include <mkldnn_types.h>
9 #include <mkldnn_extension_utils.h>
11 using namespace mkldnn;
12 using namespace MKLDNNPlugin;
13 using namespace InferenceEngine;
15 MKLDNNTileNode::MKLDNNTileNode(const InferenceEngine::CNNLayerPtr& layer, const mkldnn::engine& eng) : MKLDNNNode(layer, eng) {}
17 void MKLDNNTileNode::getSupportedDescriptors() {
18 auto * tileLayer = dynamic_cast<TileLayer*>(getCnnLayer().get());
20 if (tileLayer == nullptr)
21 THROW_IE_EXCEPTION << "Cannot convert tile layer.";
23 if (getParentEdges().size() != 1)
24 THROW_IE_EXCEPTION << "Incorrect number of input edges for layer " << getName();
25 if (!getChildEdges().size())
26 THROW_IE_EXCEPTION << "Incorrect number of output edges for layer " << getName();
28 axis = tileLayer->axis;
29 tiles = tileLayer->tiles;
32 void MKLDNNTileNode::initSupportedPrimitiveDescriptors() {
33 if (!supportedPrimitiveDescriptors.empty())
36 InferenceEngine::Precision precision = getCnnLayer()->insData[0].lock()->getPrecision();
37 if (precision != InferenceEngine::Precision::FP32)
38 precision = InferenceEngine::Precision::FP32;
39 auto inputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
40 precision = getCnnLayer()->outData[0]->getPrecision();
41 if (precision != InferenceEngine::Precision::FP32)
42 precision = InferenceEngine::Precision::FP32;
43 auto outputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
45 auto& inDims = getParentEdgeAt(0)->getDims();
46 memory::format fmt = memory::format::any;
47 if (inDims.ndims() == 2) {
48 fmt = memory::format::nc;
49 } else if (inDims.ndims() == 4) {
50 fmt = memory::format::nchw;
51 } else if (inDims.ndims() == 5) {
52 fmt = memory::format::ncdhw;
54 if (fmt == memory::format::any) {
55 THROW_IE_EXCEPTION << "Tile " << getName() << " supports only 2D, 4D and 5D dimensions!";
58 InferenceEngine::LayerConfig config;
59 config.dynBatchSupport = true;
60 config.inConfs.resize(1);
61 config.outConfs.resize(1);
62 config.inConfs[0].inPlace = -1;
63 config.inConfs[0].constant = false;
64 config.inConfs[0].desc = MKLDNNMemoryDesc(getParentEdgeAt(0)->getDims(), inputDataType, fmt);
65 config.outConfs[0].inPlace = -1;
66 config.outConfs[0].constant = false;
67 config.outConfs[0].desc = MKLDNNMemoryDesc(getChildEdgeAt(0)->getDims(), outputDataType, fmt);
68 supportedPrimitiveDescriptors.push_back({config, impl_desc_type::unknown});
71 void MKLDNNTileNode::createPrimitive() {
72 auto& dstMemPtr = getChildEdgeAt(0)->getMemoryPtr();
73 auto& srcMemPtr = getParentEdgeAt(0)->getMemoryPtr();
74 if (!dstMemPtr || !dstMemPtr->GetPrimitivePtr())
75 THROW_IE_EXCEPTION << "Destination memory didn't allocate.";
76 if (!srcMemPtr || !srcMemPtr->GetPrimitivePtr())
77 THROW_IE_EXCEPTION << "Input memory didn't allocate.";
78 if (getSelectedPrimitiveDescriptor() == nullptr)
79 THROW_IE_EXCEPTION << "Preferable primitive descriptor does not set.";
80 if (getParentEdges().size() != 1)
81 THROW_IE_EXCEPTION << "Incorrect number of input edges for layer " << getName();
84 void MKLDNNTileNode::execute(mkldnn::stream strm) {
85 auto& srcMemory = getParentEdgeAt(0)->getMemory();
87 const float *src_ptr = reinterpret_cast<const float*>(srcMemory.GetData()) +
88 srcMemory.GetDescriptor().data.layout_desc.blocking.offset_padding;
89 float *dst_ptr = reinterpret_cast<float*>(getChildEdgeAt(0)->getMemory().GetData()) +
90 getChildEdgeAt(0)->getMemory().GetDescriptor().data.layout_desc.blocking.offset_padding;
94 memory::dims inDims = srcMemory.GetDims();
95 for (int i=0; i < axis; i++ ) m_outer_dim *= inDims[i];
96 for (int i=axis; i < inDims.size(); i++ ) m_inner_dim *= inDims[i];
98 m_outer_dim /= inDims[0];
99 m_outer_dim *= batchToProcess();
101 m_inner_dim /= inDims[0];
102 m_inner_dim *= batchToProcess();
105 if (m_inner_dim == 1 && m_outer_dim % 8 == 0 && ((inDims.size() == 4 && srcMemory.GetFormat() == memory::nChw8c) ||
106 (inDims.size() == 5 && srcMemory.GetFormat() == memory::nCdhw8c))) {
108 * We may enable tile processing directly to appropriate output format (nChw8c)
112 } else if (m_inner_dim == 1 && m_outer_dim % 16 == 0 &&
113 ((inDims.size() == 4 && srcMemory.GetFormat() == memory::nChw16c) ||
114 (inDims.size() == 5 && srcMemory.GetFormat() == memory::nCdhw16c))) {
116 * We may enable tile processing directly to appropriate output format (nChw16c)
122 for (int i = 0; i < m_outer_dim; ++i) {
123 for (int t = 0; t < tiles; ++t) {
124 memcpy(dst_ptr, src_ptr, m_inner_dim* sizeof(float));
125 dst_ptr += m_inner_dim;
127 src_ptr += m_inner_dim;
131 bool MKLDNNTileNode::created() const {
132 return getType() == Tile;