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

5 files changed, 31 insertions, 6 deletions

diff --git a/src/armnn/test/GraphTests.cpp b/src/armnn/test/GraphTests.cpp
index 55a01fcc94..602575b352 100644
--- a/src/armnn/test/GraphTests.cpp
+++ b/src/armnn/test/GraphTests.cpp
@@ -590,4 +590,28 @@ BOOST_AUTO_TEST_CASE(DuplicateLayerNames)
     BOOST_TEST(((*std::next(it))->GetType() == armnn::LayerType::Output));
 }
 
+BOOST_AUTO_TEST_CASE(CheckGraphConstTensorSharing)
+{
+    armnn::Graph graph0;
+    const float* sharedWeightPtr;
+
+    {
+        armnn::Graph graph1;
+
+        armnn::FullyConnectedLayer* const fcLayer =
+                graph1.AddLayer<armnn::FullyConnectedLayer>(armnn::FullyConnectedDescriptor(), "fc");
+
+        float weight = 1.0f;
+        armnn::ConstTensor constTensor({{ 1, 1 }, armnn::DataType::Float32}, &weight);
+        fcLayer->m_Weight = std::make_shared<armnn::ScopedCpuTensorHandle>(constTensor);;
+        // point sharedWeightPtr to graph1's const tensor
+        sharedWeightPtr = fcLayer->m_Weight->GetConstTensor<float>();
+
+        graph0 = armnn::Graph(graph1);
+        // graph1 goes out of scope
+    }
+
+    BOOST_TEST(*sharedWeightPtr == 1);
+}
+
 BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/armnn/test/optimizations/ConvertConstantsBFloatTests.cpp b/src/armnn/test/optimizations/ConvertConstantsBFloatTests.cpp
index 5cb89daafd..bb8e674b56 100644
--- a/src/armnn/test/optimizations/ConvertConstantsBFloatTests.cpp
+++ b/src/armnn/test/optimizations/ConvertConstantsBFloatTests.cpp
@@ -57,7 +57,7 @@ BOOST_AUTO_TEST_CASE(ConvertConstantsFloatToBFloatTest)
     BOOST_CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::BFloat16);
 
     // Check whether data matches expected Bf16 data
-    BFloat16* data = fc->m_Weight->GetTensor<BFloat16>();
+    const BFloat16* data = fc->m_Weight->GetConstTensor<BFloat16>();
     BOOST_CHECK(data[0] == BFloat16(0.0f));
     BOOST_CHECK(data[1] == BFloat16(-1.0f));
     BOOST_CHECK(data[2] == BFloat16(3.796875f)); // 0x4073
@@ -113,7 +113,7 @@ BOOST_AUTO_TEST_CASE(ConvertConstantsBFloatToFloatTest)
     BOOST_CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float32);
 
     // Now test the data matches float32 data
-    float* data = fc->m_Weight->GetTensor<float>();
+    const float* data = fc->m_Weight->GetConstTensor<float>();
     BOOST_CHECK(data[0] == 0.0f);
     BOOST_CHECK(data[1] == -1.0f);
     BOOST_CHECK(data[2] == 3.796875f);
diff --git a/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp b/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp
index 84f525f4bf..12df462456 100644
--- a/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp
+++ b/src/armnn/test/optimizations/ConvertConstantsFloatToHalfTests.cpp
@@ -50,7 +50,7 @@ BOOST_AUTO_TEST_CASE(ConvertConstantsFloatToHalfTest)
     BOOST_CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float16);
 
     // Check whether data matches expected fp16 data
-    Half* data = fc->m_Weight->GetTensor<Half>();
+    const Half* data = fc->m_Weight->GetConstTensor<Half>();
     BOOST_CHECK(data[0] == Half(1.0f));
     BOOST_CHECK(data[1] == Half(2.0f));
     BOOST_CHECK(data[2] == Half(3.0f));
diff --git a/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp b/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp
index f2dea53afd..7d7c6b2b0a 100644
--- a/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp
+++ b/src/armnn/test/optimizations/ConvertConstantsHalfToFloatTests.cpp
@@ -50,7 +50,7 @@ BOOST_AUTO_TEST_CASE(ConvertConstantsHalfToFloatTest)
     BOOST_CHECK(fc->m_Weight->GetTensorInfo().GetDataType() == armnn::DataType::Float32);
 
     // Now test the data matches float32 data
-    float* data = fc->m_Weight->GetTensor<float>();
+    const float* data = fc->m_Weight->GetConstTensor<float>();
     BOOST_CHECK(1.0f == data[0]);
     BOOST_CHECK(2.0f == data[1]);
     BOOST_CHECK(3.0f == data[2]);
diff --git a/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp b/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
index b35f983434..a65012eef4 100644
--- a/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
+++ b/src/armnn/test/optimizations/Fp32NetworkToBf16ConverterTests.cpp
@@ -101,7 +101,7 @@ BOOST_AUTO_TEST_CASE(Fp32NetworkToBf16OptimizationConv2DTest)
     BOOST_TEST((outputTensor.GetDataType() == armnn::DataType::Float32));
 
     // Check whether data matches expected Bf16 data
-    armnn::BFloat16* data = conv->m_Weight->GetTensor<armnn::BFloat16>();
+    const armnn::BFloat16* data = conv->m_Weight->GetConstTensor<armnn::BFloat16>();
     BOOST_CHECK(data[0] == armnn::BFloat16(0.0f));
     BOOST_CHECK(data[1] == armnn::BFloat16(-1.0f));
     BOOST_CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
@@ -171,7 +171,7 @@ BOOST_AUTO_TEST_CASE(Fp32NetworkToBf16OptimizationFullyConnectedTest)
     BOOST_TEST((outputTensor.GetDataType() == armnn::DataType::Float32));
 
     // Check whether data matches expected Bf16 data
-    armnn::BFloat16* data = fc->m_Weight->GetTensor<armnn::BFloat16>();
+    const armnn::BFloat16* data = fc->m_Weight->GetConstTensor<armnn::BFloat16>();
     BOOST_CHECK(data[0] == armnn::BFloat16(0.0f));
     BOOST_CHECK(data[1] == armnn::BFloat16(-1.0f));
     BOOST_CHECK(data[2] == armnn::BFloat16(3.796875f)); // 0x4073
@@ -182,4 +182,5 @@ BOOST_AUTO_TEST_CASE(Fp32NetworkToBf16OptimizationFullyConnectedTest)
     BOOST_CHECK(data[7] == armnn::BFloat16(-9.131327E-10f)); // 0xB07B
 }
 
+
 BOOST_AUTO_TEST_SUITE_END()
\ No newline at end of file