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, 4 insertions, 3 deletions

diff --git a/src/armnn/optimizations/ConvertFp32NetworkToBf16.hpp b/src/armnn/optimizations/ConvertFp32NetworkToBf16.hpp
index c45ab2cded..a0856a485b 100644
--- a/src/armnn/optimizations/ConvertFp32NetworkToBf16.hpp
+++ b/src/armnn/optimizations/ConvertFp32NetworkToBf16.hpp
@@ -27,9 +27,10 @@ inline LayerT* ConvertWeight(Layer* l)
         {
             std::vector<BFloat16> newValues(info.GetNumElements());
 
-            armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(layer->m_Weight->template GetTensor<float>(),
-                                                                         info.GetNumElements(),
-                                                                         newValues.data());
+            armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(
+                    layer->m_Weight->template GetConstTensor<float>(),
+                    info.GetNumElements(),
+                    newValues.data());
 
             TensorInfo newInfo(info);
             newInfo.SetDataType(DataType::BFloat16);