Review comments

openvinotoolkit · May 25, 2021 · 10f6c7c · 10f6c7c
1 parent bdadc3e
commit 10f6c7c
Show file tree

Hide file tree

Showing 9 changed files with 120 additions and 177 deletions.
diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_graph.cpp
@@ -574,7 +574,7 @@ void MKLDNNGraph::AllocateWithReuse() {
                 && edge->getParent()->isConstant()) {
                 if (edge->getParent()->getType() == Input) {
                     auto constNode = std::static_pointer_cast<MKLDNNInputNode>(edge->getParent());
-                    edge->reuse(constNode->getMemoryPtr());
+                    edge->reuse(std::const_pointer_cast<MKLDNNMemory>(constNode->getMemoryPtr()));
                 } else {
                     edge->externalAllocate(weightsCache);
                 }

diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_graph_optimizer.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_graph_optimizer.cpp
@@ -478,11 +478,11 @@ void MKLDNNGraphOptimizer::FuseConvolutionAndZeroPoints(MKLDNNGraph &graph) {
                 if (zeroPointsConstant == nullptr)
                     IE_THROW() << "Cannot cast to Input node";
 
-                auto zeroPointsBlob = dynamic_cast<const TBlob<uint8_t>*>(zeroPointsConstant->getConstBlob().get());
+                auto zeroPointsBlob = zeroPointsConstant->getMemoryPtr();
                 if (zeroPointsBlob == nullptr)
                     IE_THROW() << "Cannot cast to TBlob internal zero points blob";
 
-                auto zeroPointsData = zeroPointsBlob->cbuffer().as<uint8_t*>();
+                auto zeroPointsData = static_cast<const uint8_t*>(zeroPointsBlob->GetPtr());
                 if (zeroPointsData == nullptr)
                     IE_THROW() << "zeroPointsBlob has not allocated buffer";
 
@@ -515,11 +515,11 @@ void MKLDNNGraphOptimizer::FuseConvolutionAndZeroPoints(MKLDNNGraph &graph) {
         if (!weightsConstant || !weightsConstant->isConstant())
             return;
 
-        auto weightsBlob = dynamic_cast<const TBlob<int8_t>*>(weightsConstant->getConstBlob().get());
+        auto weightsBlob = weightsConstant->getMemoryPtr();
         if (weightsBlob == nullptr)
             IE_THROW() << "Cannot cast to TBlob internal weights blob";
 
-        auto weightsPtr = weightsBlob->cbuffer().as<int8_t*>();
+        auto weightsPtr = static_cast<const int8_t*>(weightsBlob->GetPtr());
         if (weightsPtr == nullptr)
             IE_THROW() << "weightsBlob has not allocated buffer";
 

diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_memory.h b/inference-engine/src/mkldnn_plugin/mkldnn_memory.h
@@ -172,5 +172,6 @@ class MKLDNNMemory {
 };
 
 using MKLDNNMemoryPtr = std::shared_ptr<MKLDNNMemory>;
+using MKLDNNMemoryCPtr = std::shared_ptr<const MKLDNNMemory>;
 
 }  // namespace MKLDNNPlugin
diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_node.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_node.cpp
@@ -1345,10 +1345,14 @@ void MKLDNNNode::fillScalesAndShifts(const MKLDNNNode *parentNode, std::vector<f
     shifts.clear();
     const auto fillValuesFrom = [&](const MKLDNNNodePtr& constInput, std::vector<float>& buffer) {
         auto *constInputNode = dynamic_cast<MKLDNNInputNode *>(constInput.get());
-        auto constBlob = constInputNode->getConstBlob();
-        auto srtPtr = constBlob->cbuffer().as<int8_t *>();
-        buffer.resize(constBlob->size());
-        cpu_convert(srtPtr, &buffer[0], constBlob->getTensorDesc().getPrecision(), Precision::FP32, constBlob->size());
+        auto constBlob = constInputNode->getMemoryPtr();
+        auto const elementsCount = constBlob->GetElementsCount();
+        buffer.resize(elementsCount);
+        cpu_convert(constBlob->GetPtr(),
+                    &buffer[0],
+                    MKLDNNExtensionUtils::DataTypeToIEPrecision(constBlob->GetDataType()),
+                    Precision::FP32,
+                    elementsCount);
     };
 
     const size_t constPort = getParentEdgesAtPort(0)[0]->getParent().get() == parentNode ? 1 : 0;

diff --git a/inference-engine/src/mkldnn_plugin/nodes/mkldnn_deconv_node.cpp b/inference-engine/src/mkldnn_plugin/nodes/mkldnn_deconv_node.cpp
@@ -97,10 +97,12 @@ InferenceEngine::Blob::Ptr MKLDNNDeconvolutionNode::createWeiBlobAsIO(InferenceE
     auto constNode = std::dynamic_pointer_cast<MKLDNNInputNode>(getParentEdgeAt(1)->getParent());
     if (!constNode)
         IE_THROW() << "Cannot cast const input node for node " << getName() << ".";
-    auto blb = constNode->getConstBlob();
+    auto blb = constNode->getMemoryPtr();
     if (!blb)
         IE_THROW() << "Cannot get const weights blob for node " << getName() << ".";
 
+    auto const blbSize = blb->GetSize();
+
     // WA: In int8 case, we are processing weights using internal blob.
     // So we disconnect constant node containing weights from the graph and then don't use it.
     if (getParentEdges().size() == 3) {
@@ -123,7 +125,7 @@ InferenceEngine::Blob::Ptr MKLDNNDeconvolutionNode::createWeiBlobAsIO(InferenceE
         orderForBlockedDesc.push_back(i);
 
     BlockingDesc blkDesc(dimsForBlockedDesc, orderForBlockedDesc);
-    InferenceEngine::TensorDesc tensorDesc(blb->getTensorDesc().getPrecision(), dims, blkDesc);
+    InferenceEngine::TensorDesc tensorDesc(MKLDNNExtensionUtils::DataTypeToIEPrecision(blb->GetDataType()), dims, blkDesc);
 
     Blob::Ptr internalBlob = InferenceEngine::make_shared_blob<int8_t>(tensorDesc);
     internalBlob->allocate();
@@ -132,11 +134,11 @@ InferenceEngine::Blob::Ptr MKLDNNDeconvolutionNode::createWeiBlobAsIO(InferenceE
         IE_THROW(NotAllocated) << "Internal blob was not allocated for node " << getName() << ".";
     size_t intBuffSize = internalBlob->byteSize();
 
-    size_t offset = blb->byteSize();
+    size_t offset = blbSize;
     if (intBuffSize < offset) {
         IE_THROW() << "Cannot create internal buffer. Buffer can be overrun.";
     }
-    cpu_memcpy_s(data, intBuffSize, blb->cbuffer(), blb->byteSize());
+    cpu_memcpy_s(data, intBuffSize, blb->GetPtr(), blbSize);
 
     return internalBlob;
 }