IVGCVSW-1199 Disable auto-flattening of Compute Library tensors

This is one of the reasons why the tests in https://review.mlplatform.org/#/c/ml/armnn/+/237/
are failing (but not the only reason).

Change-Id: If485bade2a6dd013cba826cec71d748fc7747249

3 files changed, 15 insertions, 15 deletions

diff --git a/src/backends/aclCommon/ArmComputeTensorUtils.cpp b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
index 9fb78f8a80..ba7cd14f3f 100644
--- a/src/backends/aclCommon/ArmComputeTensorUtils.cpp
+++ b/src/backends/aclCommon/ArmComputeTensorUtils.cpp
@@ -40,7 +40,7 @@ arm_compute::TensorShape BuildArmComputeTensorShape(const armnn::TensorShape& te
     for (unsigned int i = 0; i < tensorShape.GetNumDimensions(); i++)
     {
         // Note that our dimensions are stored in the opposite order to ACL's.
-        shape.set(tensorShape.GetNumDimensions() - i - 1, tensorShape[i]);
+        shape.set(tensorShape.GetNumDimensions() - i - 1, tensorShape[i], false);
 
         // TensorShape::set() flattens leading ones, so that batch size 1 cannot happen.
         // arm_compute tensors expect this.
diff --git a/src/backends/aclCommon/test/CreateWorkloadClNeon.hpp b/src/backends/aclCommon/test/CreateWorkloadClNeon.hpp
index ec2072d243..adabed07f8 100644
--- a/src/backends/aclCommon/test/CreateWorkloadClNeon.hpp
+++ b/src/backends/aclCommon/test/CreateWorkloadClNeon.hpp
@@ -47,7 +47,9 @@ boost::test_tools::predicate_result CompareTensorHandleShape(IComputeTensorHandl
         if (info->dimension(i) != expectedDimension)
         {
             boost::test_tools::predicate_result res(false);
-            res.message() << "Different dimension [" << info->dimension(i) << "!=" << expectedDimension << "]";
+            res.message() << "For dimension " << i <<
+                             " expected size " << expectedDimension <<
+                             " got " << info->dimension(i);
             return res;
         }
 
@@ -104,4 +106,4 @@ void CreateMemCopyWorkloads(IWorkloadFactory& factory)
     BOOST_TEST((outputHandle2->GetTensorInfo() == TensorInfo({2, 3}, DataType::Float32)));
 }
 
-} //namespace
\ No newline at end of file
+} //namespace
diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp
index 47f1515e30..8cef9d78b1 100644
--- a/src/backends/cl/test/ClCreateWorkloadTests.cpp
+++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp
@@ -37,8 +37,8 @@ static void ClCreateActivationWorkloadTest()
     auto inputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
     auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
 
-    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {1}));
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {1}));
+    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {1, 1}));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {1, 1}));
 }
 
 BOOST_AUTO_TEST_CASE(CreateActivationFloatWorkload)
@@ -208,8 +208,8 @@ BOOST_AUTO_TEST_CASE(CreateConvertFp16ToFp32Workload)
     auto inputHandle  = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
     auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
 
-    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {3, 2, 3}));
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {3, 2, 3}));
+    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {1, 3, 2, 3}));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {1, 3, 2, 3}));
     BOOST_TEST((inputHandle->GetTensor().info()->data_type() == arm_compute::DataType::F16));
     BOOST_TEST((outputHandle->GetTensor().info()->data_type() == arm_compute::DataType::F32));
 }
@@ -226,8 +226,8 @@ BOOST_AUTO_TEST_CASE(CreateConvertFp32ToFp16Workload)
     auto inputHandle  = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Inputs[0]);
     auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
 
-    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {3, 2, 3}));
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {3, 2, 3}));
+    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {1, 3, 2, 3}));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {1, 3, 2, 3}));
     BOOST_TEST((inputHandle->GetTensor().info()->data_type() == arm_compute::DataType::F32));
     BOOST_TEST((outputHandle->GetTensor().info()->data_type() == arm_compute::DataType::F16));
 }
@@ -469,7 +469,7 @@ static void ClCreateReshapeWorkloadTest()
     auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
 
     BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {4, 1}));
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {4})); // Leading size 1 dimensions are collapsed by ACL.
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, {1, 4}));
 }
 
 BOOST_AUTO_TEST_CASE(CreateReshapeFloatWorkload)
@@ -537,9 +537,7 @@ static void ClSplitterWorkloadTest()
     BOOST_TEST(CompareIClTensorHandleShape(outputHandle2, {2, 7, 7}));
 
     auto outputHandle0 = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
-    // NOTE: At the moment the CL collapses the tensor to a 2 dim when dimension zero = 1
-    //       we are raising this difference between the NEON and CL libs as an issue with the compute library team.
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle0, {7, 7}));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle0, {1, 7, 7}));
 }
 
 BOOST_AUTO_TEST_CASE(CreateSplitterFloatWorkload)
@@ -784,8 +782,8 @@ static void ClMeanWorkloadTest()
     auto outputHandle = boost::polymorphic_downcast<IClTensorHandle*>(queueDescriptor.m_Outputs[0]);
 
     // The first dimension (batch size) in both input and output is singular thus it has been reduced by ACL.
-    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, { 3, 7, 4 }));
-    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, { 4 }));
+    BOOST_TEST(CompareIClTensorHandleShape(inputHandle, {  1, 3, 7, 4 }));
+    BOOST_TEST(CompareIClTensorHandleShape(outputHandle, { 1, 4 }));
 }
 
 BOOST_AUTO_TEST_CASE(CreateMeanFloat32Workload)