MLCE-82 Add Neon Mean support and unit tests

Factor out new BuildArmComputeReductionCoordinates function
from CL backend into ArmComputeTensorUtils.

Update NEON LayerSupport and WorkloadFactory objects

Change-Id: Icc975ec699199bffafbdb207323df509d35e1e04

8 files changed, 109 insertions, 6 deletions

diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 7efdf159c9..93c1123675 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -24,6 +24,7 @@
 #include "workloads/NeonDepthwiseConvolutionWorkload.hpp"
 #include "workloads/NeonL2NormalizationFloatWorkload.hpp"
 #include "workloads/NeonMaximumWorkload.hpp"
+#include "workloads/NeonMeanWorkload.hpp"
 #include "workloads/NeonMergerWorkload.hpp"
 #include "workloads/NeonMultiplicationFloatWorkload.hpp"
 #include "workloads/NeonNormalizationFloatWorkload.hpp"
@@ -364,11 +365,11 @@ bool NeonLayerSupport::IsMeanSupported(const TensorInfo& input,
                                        const MeanDescriptor& descriptor,
                                        Optional<std::string&> reasonIfUnsupported) const
 {
-    ignore_unused(input);
-    ignore_unused(output);
-    ignore_unused(descriptor);
-    ignore_unused(reasonIfUnsupported);
-    return false;
+    FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMeanWorkloadValidate,
+                                   reasonIfUnsupported,
+                                   input,
+                                   output,
+                                   descriptor);
 }
 
 bool NeonLayerSupport::IsMergerSupported(const std::vector<const TensorInfo*> inputs,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 85e5768571..e635f0cc8d 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -273,7 +273,7 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMaximum(const MaximumQueue
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
 {
-    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
+    return std::make_unique<NeonMeanWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor,
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index fdfd696fbe..d4f414eba1 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -26,6 +26,7 @@ BACKEND_SOURCES := \
         workloads/NeonL2NormalizationFloatWorkload.cpp \
         workloads/NeonLstmFloatWorkload.cpp \
         workloads/NeonMaximumWorkload.cpp \
+        workloads/NeonMeanWorkload.cpp \
         workloads/NeonMergerWorkload.cpp \
         workloads/NeonMultiplicationFloatWorkload.cpp \
         workloads/NeonNormalizationFloatWorkload.cpp \
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
index 37933e0254..5b83b2bfa8 100644
--- a/src/backends/neon/test/NeonLayerTests.cpp
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -399,6 +399,21 @@ ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgNoPeepholeNoProjection,
 ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgWithPeepholeWithProjection,
                      LstmLayerFloat32NoCifgWithPeepholeWithProjectionTest)
 
+// Mean
+ARMNN_AUTO_TEST_CASE(MeanUint8Simple, MeanUint8SimpleTest)
+ARMNN_AUTO_TEST_CASE(MeanUint8SimpleAxis, MeanUint8SimpleAxisTest)
+ARMNN_AUTO_TEST_CASE(MeanUint8KeepDims, MeanUint8KeepDimsTest)
+ARMNN_AUTO_TEST_CASE(MeanUint8MultipleDims, MeanUint8MultipleDimsTest)
+ARMNN_AUTO_TEST_CASE(MeanVtsUint8, MeanVtsUint8Test)
+
+ARMNN_AUTO_TEST_CASE(MeanFloatSimple, MeanFloatSimpleTest)
+ARMNN_AUTO_TEST_CASE(MeanFloatSimpleAxis, MeanFloatSimpleAxisTest)
+ARMNN_AUTO_TEST_CASE(MeanFloatKeepDims, MeanFloatKeepDimsTest)
+ARMNN_AUTO_TEST_CASE(MeanFloatMultipleDims, MeanFloatMultipleDimsTest)
+ARMNN_AUTO_TEST_CASE(MeanVtsFloat1, MeanVtsFloat1Test)
+ARMNN_AUTO_TEST_CASE(MeanVtsFloat2, MeanVtsFloat2Test)
+ARMNN_AUTO_TEST_CASE(MeanVtsFloat3, MeanVtsFloat3Test)
+
 // Max
 ARMNN_AUTO_TEST_CASE(SimpleMaximum, MaximumSimpleTest)
 ARMNN_AUTO_TEST_CASE(MaximumBroadcast1Element, MaximumBroadcast1ElementTest)
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 7b0251ce04..b7dfc3fcfd 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -30,6 +30,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonLstmFloatWorkload.hpp
     NeonMaximumWorkload.cpp
     NeonMaximumWorkload.hpp
+    NeonMeanWorkload.cpp
+    NeonMeanWorkload.hpp
     NeonMergerWorkload.cpp
     NeonMergerWorkload.hpp
     NeonMultiplicationFloatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonMeanWorkload.cpp b/src/backends/neon/workloads/NeonMeanWorkload.cpp
new file mode 100644
index 0000000000..d736e42e0e
--- /dev/null
+++ b/src/backends/neon/workloads/NeonMeanWorkload.cpp
@@ -0,0 +1,53 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonMeanWorkload.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <neon/NeonTensorHandle.hpp>
+
+#include "NeonWorkloadUtils.hpp"
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonMeanWorkloadValidate(const TensorInfo& input,
+                                             const TensorInfo& output,
+                                             const MeanDescriptor& desc)
+{
+    const arm_compute::TensorInfo aclInputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
+                                                                          input.GetNumDimensions(),
+                                                                          desc.m_Axis);
+
+    return arm_compute::NEReduceMean::validate(&aclInputInfo, coords, desc.m_KeepDims, &aclOutputInfo);
+}
+
+NeonMeanWorkload::NeonMeanWorkload(const MeanQueueDescriptor& descriptor, const WorkloadInfo& info)
+    : BaseWorkload<MeanQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonMeanWorkload", 1, 1);
+
+    arm_compute::ITensor& input  = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
+                                                                          info.m_InputTensorInfos[0].GetNumDimensions(),
+                                                                          m_Data.m_Parameters.m_Axis);
+
+    m_Layer.configure(&input, coords, m_Data.m_Parameters.m_KeepDims, &output);
+}
+
+void NeonMeanWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonMeanWorkload_Execute");
+    m_Layer.run();
+}
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonMeanWorkload.hpp b/src/backends/neon/workloads/NeonMeanWorkload.hpp
new file mode 100644
index 0000000000..055b52a011
--- /dev/null
+++ b/src/backends/neon/workloads/NeonMeanWorkload.hpp
@@ -0,0 +1,30 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEReduceMean.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonMeanWorkloadValidate(const TensorInfo& input,
+                                             const TensorInfo& output,
+                                             const MeanDescriptor& desc);
+
+class NeonMeanWorkload : public BaseWorkload<MeanQueueDescriptor>
+{
+public:
+    NeonMeanWorkload(const MeanQueueDescriptor& descriptor, const WorkloadInfo& info);
+
+    void Execute() const override;
+
+private:
+    mutable arm_compute::NEReduceMean m_Layer;
+};
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 1f08d039ae..a5ef0dcb2d 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -17,6 +17,7 @@
 #include "NeonL2NormalizationFloatWorkload.hpp"
 #include "NeonLstmFloatWorkload.hpp"
 #include "NeonMaximumWorkload.hpp"
+#include "NeonMeanWorkload.hpp"
 #include "NeonMergerWorkload.hpp"
 #include "NeonMultiplicationFloatWorkload.hpp"
 #include "NeonNormalizationFloatWorkload.hpp"