MLCE-418 Reduce layer does not support multiple axes

 * Added backend specific optimization to chain new reduces layers
   for each axis to simulate behaviour of a layer with multiple axes.
 * Added function to calculate reduced output shape.
 * Added unit tests.
 * Includes rework to fix IVGCVSW-5987.

Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I154b3698b5e6756b05b2a0b5a3f0896184efce72

2 files changed, 41 insertions, 14 deletions

diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp
index a1299fb458..b496238cf3 100644
--- a/src/backends/neon/NeonBackend.cpp
+++ b/src/backends/neon/NeonBackend.cpp
@@ -29,6 +29,7 @@
 #include "workloads/NeonDivisionWorkload.hpp"
 #include "workloads/NeonFullyConnectedWorkload.hpp"
 #include "workloads/NeonMultiplicationWorkload.hpp"
+#include "workloads/NeonReduceWorkload.hpp"
 #include "workloads/NeonSubtractionWorkload.hpp"
 
 #include <Optimizer.hpp>
@@ -161,6 +162,7 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph
         --it;
         Layer& base = **it;
 
+        // Fuse activation into previous layer if supported by backend
         if ((base.GetType() == LayerType::DepthwiseConvolution2d || base.GetType() == LayerType::Convolution2d
              || base.GetType() == LayerType::BatchNormalization || base.GetType() == LayerType::FullyConnected
              || base.GetType() == LayerType::Addition || base.GetType() == LayerType::Multiplication
@@ -393,6 +395,25 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph
                 }
             }
         }
+
+        // Separate reduce layer with multiple axes into multiple reduce layers with 1 axis.
+        if (base.GetType() == LayerType::Reduce)
+        {
+            ReduceLayer* baseLayer            = PolymorphicDowncast<ReduceLayer*>(&base);
+            ReduceDescriptor reduceDescriptor = baseLayer->GetParameters();
+
+            if (!reduceDescriptor.m_vAxis.empty() && reduceDescriptor.m_vAxis.size() > 1)
+            {
+                // Add new layers to the graph and connect them.
+                std::vector<Layer*> layers = ChainReduceLayers<ReduceLayer>(optimizationViews,
+                                                                            baseLayer,
+                                                                            reduceDescriptor);
+
+                // Replace existing baselayer with new subgraph.
+                ReplaceLayers<ReduceLayer>(optimizationViews, baseLayer, layers);
+                untouched.erase(baseLayer->GetGuid());
+            }
+        }
     }
 
     if (optimizationViews.GetSubstitutions().empty())
diff --git a/src/backends/neon/workloads/NeonReduceWorkload.cpp b/src/backends/neon/workloads/NeonReduceWorkload.cpp
index 0e1b46a3a1..1436cd1192 100644
--- a/src/backends/neon/workloads/NeonReduceWorkload.cpp
+++ b/src/backends/neon/workloads/NeonReduceWorkload.cpp
@@ -20,23 +20,28 @@ arm_compute::Status NeonReduceWorkloadValidate(const TensorInfo& input,
                                                const TensorInfo& output,
                                                const ReduceDescriptor& desc)
 {
-    const arm_compute::TensorInfo aclInputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
-    const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
-    if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1)
+    if ( desc.m_vAxis.size()==1 || desc.m_vAxis.empty())
     {
-        return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
-                                   "NeonReduceWorkload: Reduction is supported only on 1 axis.");
-    }
+        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_vAxis);
+        arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
+                                                                              input.GetNumDimensions(),
+                                                                              desc.m_vAxis);
 
-    return arm_compute::NEReductionOperation::validate(&aclInputInfo,
-                                                       &aclOutputInfo,
-                                                       static_cast<unsigned int>(coords[0]),
-                                                       ConvertReductionOperationToAcl(desc),
-                                                       desc.m_KeepDims);
+        return arm_compute::NEReductionOperation::validate(&aclInputInfo,
+                                                           &aclOutputInfo,
+                                                           static_cast<unsigned int>(coords[0]),
+                                                           ConvertReductionOperationToAcl(desc),
+                                                           desc.m_KeepDims);
+    }
+    else
+    {
+        // Validate layer if there are multiple axes.
+        arm_compute::Status status;
+        IS_MULTI_AXES_REDUCE_SUPPORTED(NeonReduceWorkloadValidate, input, desc, status);
+        return status;
+    }
 }
 
 NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info)
@@ -50,6 +55,7 @@ NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor,
     arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
                                                                           info.m_InputTensorInfos[0].GetNumDimensions(),
                                                                           m_Data.m_Parameters.m_vAxis);
+
     m_Layer.configure(&input,
                       &output,
                       static_cast<unsigned int>(coords[0]),