Fix graph copy memory spike

 * Change layer storage of ConstTensors to std::shared_ptr<ConstCpuTensorHandle>
 * Change clone to share ConstTensor rather than copy
 * Remove uses of non-const GetTensor() call
 * Reduce scope of non-optimized network in ExeNet, so memory can be released after use

Signed-off-by: Finn Williams <Finn.Williams@arm.com>

Change-Id: Ibb2c7309d12411d21405bd6024c76bcdf5404545

1 files changed, 12 insertions, 12 deletions

diff --git a/src/armnn/layers/QuantizedLstmLayer.cpp b/src/armnn/layers/QuantizedLstmLayer.cpp
index 578d9eb137..a1ff985abe 100644
--- a/src/armnn/layers/QuantizedLstmLayer.cpp
+++ b/src/armnn/layers/QuantizedLstmLayer.cpp
@@ -49,31 +49,31 @@ QuantizedLstmLayer* QuantizedLstmLayer::Clone(Graph& graph) const
     auto layer = CloneBase<QuantizedLstmLayer>(graph, GetName());
 
     layer->m_QuantizedLstmParameters.m_InputToInputWeights = m_QuantizedLstmParameters.m_InputToInputWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_InputToInputWeights) : nullptr;
+            m_QuantizedLstmParameters.m_InputToInputWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_InputToForgetWeights = m_QuantizedLstmParameters.m_InputToForgetWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_InputToForgetWeights) : nullptr;
+            m_QuantizedLstmParameters.m_InputToForgetWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_InputToCellWeights = m_QuantizedLstmParameters.m_InputToCellWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_InputToCellWeights) : nullptr;
+            m_QuantizedLstmParameters.m_InputToCellWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_InputToOutputWeights = m_QuantizedLstmParameters.m_InputToOutputWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_InputToOutputWeights) : nullptr;
+            m_QuantizedLstmParameters.m_InputToOutputWeights : nullptr;
 
     layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights = m_QuantizedLstmParameters.m_RecurrentToInputWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_RecurrentToInputWeights) : nullptr;
+            m_QuantizedLstmParameters.m_RecurrentToInputWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights = m_QuantizedLstmParameters.m_RecurrentToForgetWeights
-            ? std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_RecurrentToForgetWeights) : nullptr;
+            ? m_QuantizedLstmParameters.m_RecurrentToForgetWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights = m_QuantizedLstmParameters.m_RecurrentToCellWeights ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_RecurrentToCellWeights) : nullptr;
+            m_QuantizedLstmParameters.m_RecurrentToCellWeights : nullptr;
     layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights = m_QuantizedLstmParameters.m_RecurrentToOutputWeights
-            ? std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_RecurrentToOutputWeights) : nullptr;
+            ? m_QuantizedLstmParameters.m_RecurrentToOutputWeights : nullptr;
 
     layer->m_QuantizedLstmParameters.m_InputGateBias = m_QuantizedLstmParameters.m_InputGateBias ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_InputGateBias) : nullptr;
+            m_QuantizedLstmParameters.m_InputGateBias : nullptr;
     layer->m_QuantizedLstmParameters.m_ForgetGateBias = m_QuantizedLstmParameters.m_ForgetGateBias ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_ForgetGateBias) : nullptr;
+            m_QuantizedLstmParameters.m_ForgetGateBias : nullptr;
     layer->m_QuantizedLstmParameters.m_CellBias = m_QuantizedLstmParameters.m_CellBias ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_CellBias) : nullptr;
+            m_QuantizedLstmParameters.m_CellBias : nullptr;
     layer->m_QuantizedLstmParameters.m_OutputGateBias = m_QuantizedLstmParameters.m_OutputGateBias ?
-            std::make_unique<ScopedCpuTensorHandle>(*m_QuantizedLstmParameters.m_OutputGateBias) : nullptr;
+            m_QuantizedLstmParameters.m_OutputGateBias : nullptr;
 
     return std::move(layer);
 }