IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[CPU] Fixes for CpuBlockedMemoryDesc constructor and reorder availability checks (#10299)

Browse Source
This commit is contained in:
Maksim Kutakov
2022-02-12 15:29:55 +03:00
committed by GitHub
parent e49370c008
commit 451453c4ce
2 changed files with 17 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/plugins/intel_cpu/src/memory_desc/cpu_blocked_memory_desc.cpp
View File

@@ -8,21 +8,15 @@
using namespace MKLDNNPlugin;
CpuBlockedMemoryDesc::CpuBlockedMemoryDesc(InferenceEngine::Precision prc, const Shape& shape) : MemoryDesc(shape, Blocked), precision(prc) {
auto& dims = shape.getDims();
order.resize(dims.size());
std::iota(order.begin(), order.end(), 0);
blockedDims = dims;
offsetPadding = 0;
offsetPaddingToData.resize(dims.size(), 0);
strides.resize(order.size());
// for empty tensor case we fill all strides with 0 values
strides[strides.size() - 1] = shape.hasZeroDims() ? 0 : 1;
for (size_t i = 2; i <= order.size(); i++) {
strides[strides.size() - i] = strides[strides.size() - (i - 1)] * blockedDims[blockedDims.size() - (i - 1)];
}
static VectorDims makeRange(size_t size) {
VectorDims retVec(size, 0);
std::iota(retVec.begin(), retVec.end(), 0);
return retVec;
}
CpuBlockedMemoryDesc::CpuBlockedMemoryDesc(InferenceEngine::Precision prc, const Shape& shape) :
CpuBlockedMemoryDesc(prc, shape, shape.getDims(), makeRange(shape.getDims().size())) {}
CpuBlockedMemoryDesc::CpuBlockedMemoryDesc(InferenceEngine::Precision prc, const Shape& shape, const VectorDims& blockedDims,
const VectorDims& order, size_t offsetPadding, const VectorDims& offsetPaddingToData,
const VectorDims& strides) : MemoryDesc(shape, Blocked), precision(prc) {

14
src/plugins/intel_cpu/src/mkldnn_graph.cpp
View File

@@ -459,9 +459,13 @@ void MKLDNNGraph::ExecuteConstantNodesOnly() const {
}
}
static bool isReorderAvailable(const MemoryDesc& parentDesc, const MemoryDesc& childDesc, const mkldnn::engine& eng) {
memory::desc dstMemDesc = MemoryDescUtils::convertToDnnlMemoryDesc(childDesc.clone())->getDnnlDesc();
memory::desc srcMemDesc = MemoryDescUtils::convertToDnnlMemoryDesc(parentDesc.clone())->getDnnlDesc();
static bool isReorderAvailable(const MemoryDescPtr& parentDesc, const MemoryDescPtr& childDesc, const mkldnn::engine& eng) {
auto definedParentDesc = parentDesc->isDefined() ? parentDesc : MemoryDescUtils::makeDummyDesc(*parentDesc);
memory::desc srcMemDesc = MemoryDescUtils::convertToDnnlMemoryDesc(definedParentDesc)->getDnnlDesc();
auto definedChildDesc = childDesc->isDefined() ? childDesc : MemoryDescUtils::makeDummyDesc(*childDesc);
memory::desc dstMemDesc = MemoryDescUtils::convertToDnnlMemoryDesc(definedChildDesc)->getDnnlDesc();
mkldnn::primitive_attr attr;
dnnl_primitive_desc_t result = nullptr;
@@ -513,7 +517,9 @@ void MKLDNNGraph::InitEdges() {
MKLDNNEdge::ReorderStatus reorderStatusInternal = MKLDNNEdge::ReorderStatus::Regular;
// Check if there is a reorder that needs the precision conversion
if (edge->getInputDesc().getPrecision() != edge->getOutputDesc().getPrecision() &&
!isReorderAvailable(edge->getInputDesc(), edge->getOutputDesc(), this->getEngine())) {
!isReorderAvailable(edge->getInputPortDesc()->getMemDesc(),
edge->getOutputPortDesc()->getMemDesc(),
this->getEngine())) {
// If we are here, then we need to insert Convert, because there are no reorders that support such type conversion
const auto& inDesc = edge->getInputDesc();
const auto& outDesc = edge->getOutputDesc();