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

3 files changed, 13 insertions, 12 deletions

diff --git a/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp b/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp
index 26661cfcde..0a5ad9d152 100644
--- a/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp
+++ b/src/armnn/optimizations/AddBroadcastReshapeLayer.hpp
@@ -74,7 +74,7 @@ public:
                 ConstantLayer& constantLayer = static_cast<ConstantLayer&>(parentLayer);
 
                 constantLayer.m_LayerOutput = std::make_unique<ScopedCpuTensorHandle>(
-                                ConstTensor(reshapeInfo,constantLayer.m_LayerOutput.get()->GetTensor<void>()));
+                                ConstTensor(reshapeInfo,constantLayer.m_LayerOutput.get()->GetConstTensor<void>()));
                 constantLayer.GetOutputSlot().SetTensorInfo(reshapeInfo);
 
                 return;
diff --git a/src/armnn/optimizations/ConvertConstants.hpp b/src/armnn/optimizations/ConvertConstants.hpp
index f3ebcdf5d9..df5a5b4f67 100644
--- a/src/armnn/optimizations/ConvertConstants.hpp
+++ b/src/armnn/optimizations/ConvertConstants.hpp
@@ -23,7 +23,7 @@ namespace optimizations
 
 struct BFloat16ToFloat32
 {
-    static void Func(std::unique_ptr<ScopedCpuTensorHandle>& handle)
+    static void Func(std::shared_ptr<ConstCpuTensorHandle>& handle)
     {
         const TensorInfo& info = handle->GetTensorInfo();
 
@@ -31,7 +31,7 @@ struct BFloat16ToFloat32
         {
             std::vector<float> newValues(info.GetNumElements());
 
-            armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(handle->GetTensor<BFloat16>(),
+            armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(handle->GetConstTensor<BFloat16>(),
                                                                          info.GetNumElements(),
                                                                          newValues.data());
 
@@ -44,7 +44,7 @@ struct BFloat16ToFloat32
 
 struct Float16ToFloat32
 {
-    static void Func(std::unique_ptr<ScopedCpuTensorHandle>& handle)
+    static void Func(std::shared_ptr<ConstCpuTensorHandle>& handle)
     {
         const TensorInfo& info = handle->GetTensorInfo();
 
@@ -52,7 +52,7 @@ struct Float16ToFloat32
         {
             std::vector<float> newValues(info.GetNumElements());
 
-            armnnUtils::FloatingPointConverter::ConvertFloat16To32(handle->GetTensor<Half>(),
+            armnnUtils::FloatingPointConverter::ConvertFloat16To32(handle->GetConstTensor<Half>(),
                                                                    info.GetNumElements(),
                                                                    newValues.data());
 
@@ -65,7 +65,7 @@ struct Float16ToFloat32
 
 struct Float32ToBFloat16
 {
-    static void Func(std::unique_ptr<ScopedCpuTensorHandle>& handle)
+    static void Func(std::shared_ptr<ConstCpuTensorHandle>& handle)
     {
         const TensorInfo& info = handle->GetTensorInfo();
 
@@ -73,7 +73,7 @@ struct Float32ToBFloat16
         {
             std::vector<BFloat16> newValues(info.GetNumElements());
 
-            armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(handle->GetTensor<float>(),
+            armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(handle->GetConstTensor<float>(),
                                                                          info.GetNumElements(),
                                                                          newValues.data());
 
@@ -86,7 +86,7 @@ struct Float32ToBFloat16
 
 struct Float32ToFloat16
 {
-    static void Func(std::unique_ptr<ScopedCpuTensorHandle>& handle)
+    static void Func(std::shared_ptr<ConstCpuTensorHandle>& handle)
     {
         const TensorInfo& info = handle->GetTensorInfo();
 
@@ -94,7 +94,7 @@ struct Float32ToFloat16
         {
             std::vector<Half> newValues(info.GetNumElements());
 
-            armnnUtils::FloatingPointConverter::ConvertFloat32To16(handle->GetTensor<float>(),
+            armnnUtils::FloatingPointConverter::ConvertFloat32To16(handle->GetConstTensor<float>(),
                                                                    info.GetNumElements(),
                                                                    newValues.data());
 
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);