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.

Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I180b0b111b7bcf3d0c283f1db0b82d5f17757682

2 files changed, 140 insertions, 0 deletions

diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
index a0fca46330..9439ddb61e 100644
--- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
@@ -6,6 +6,9 @@
 #pragma once
 
 #include <armnn/backends/OptimizationViews.hpp>
+#include <armnn/utility/Assert.hpp>
+
+#include <aclCommon/ArmComputeUtils.hpp>
 
 namespace armnn
 {
@@ -147,4 +150,86 @@ LayerType* FuseLayerWithWeightsAndBiases(OptimizationViews& optimizationViews,
     return replacementLayer;
 }
 
+//
+// If reduce layer has multiple axes, add new layer for each axis to simulate the same behaviour
+// as currently only one axis is supported.
+//
+template<typename LayerType>
+void ChainReduceLayers(OptimizationViews& optimizationViews,
+                       LayerType* baseLayer,
+                       ReduceDescriptor& reduceDescriptor)
+{
+    // If layer has single axis don't chain layers.
+    if (!reduceDescriptor.m_vAxis.empty() && reduceDescriptor.m_vAxis.size() > 1)
+    {
+        // Save base layer output shape to compare against the output of the final layer added.
+        const TensorInfo baseLayerInfo = baseLayer->GetOutputSlot(0).GetTensorInfo();
+
+        // Vector of new chained layers, used for substitution.
+        std::vector<Layer*> layers;
+
+        // Vector of axes so each layer is reshaped correctly.
+        std::vector<uint32_t> reduceAxis;
+        unsigned int recalulateAxis = 0;
+
+        for (unsigned int i = 0; i != reduceDescriptor.m_vAxis.size(); ++i)
+        {
+            // Get TensorInfo to populate subsequent layers with.
+            TensorInfo layerInfoToModify = baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
+
+            reduceAxis.emplace_back(reduceDescriptor.m_vAxis[i]);
+
+            // Calculate new shape based on the axes.
+            const TensorShape& reducedShape = ComputeReductionTensorShape(layerInfoToModify,
+                                                                          reduceAxis,
+                                                                          reduceDescriptor.m_KeepDims);
+            layerInfoToModify.SetShape(reducedShape);
+
+            // Create a vector for the single axis to be assigned to the descriptor.
+            // Update axis if keepDims is set reduce layers correctly.
+            std::vector<uint32_t> singleAxis(1, reduceDescriptor.m_vAxis[i] - recalulateAxis);
+
+            // Create a descriptor and assign single axis.
+            ReduceDescriptor newReduceDescriptor = baseLayer->GetParameters();
+            newReduceDescriptor.m_vAxis.assign(singleAxis.begin(), singleAxis.end());
+
+            // Add new layer to graph.
+            std::string layerName = "reduce_layer_" + std::to_string(i);
+            Layer* replacementLayer = optimizationViews.GetGraph().AddLayer<LayerType>(newReduceDescriptor,
+                                                                                       layerName.c_str());
+
+            // Connect previous layer with new layer.
+            // The first and last layer will be connected when the subgraph is replaced.
+            if (!layers.empty())
+            {
+                layers[i - 1]->GetOutputSlot(0).Connect(replacementLayer->GetInputSlot(0));
+            }
+
+            // Set updated tensorInfo for new layer.
+            replacementLayer->GetOutputSlot(0).SetTensorInfo(layerInfoToModify);
+
+            if (!reduceDescriptor.m_KeepDims)
+            {
+                recalulateAxis++;
+            }
+
+            layers.emplace_back(replacementLayer);
+        }
+
+        // Check if the TensorInfo from the last layer equals the inferred output from the original layer.
+        ARMNN_ASSERT(baseLayerInfo == layers.back()->GetOutputSlot().GetTensorInfo());
+
+        std::list<Layer*> replacementLayers(layers.begin(), layers.end());
+
+        // Substitute new chained subgraph for original reduce layer.
+        SubgraphView substitutionSubgraph(baseLayer);
+        SubgraphView replacementSubgraph(CreateInputsFrom({replacementLayers.front()}),
+                                         CreateOutputsFrom({replacementLayers.back()}),
+                                         std::move(replacementLayers));
+
+        optimizationViews.AddSubstitution({substitutionSubgraph, replacementSubgraph});
+
+    }
+}
+
 } // namespace armnn
diff --git a/src/backends/aclCommon/ArmComputeUtils.hpp b/src/backends/aclCommon/ArmComputeUtils.hpp
index d9efab288f..5bc5abcb05 100644
--- a/src/backends/aclCommon/ArmComputeUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeUtils.hpp
@@ -7,6 +7,7 @@
 #include <armnn/Descriptors.hpp>
 #include <armnn/Tensor.hpp>
 #include <armnn/utility/Assert.hpp>
+#include <armnn/utility/NumericCast.hpp>
 #include <backendsCommon/WorkloadData.hpp>
 
 #include <arm_compute/core/Types.h>
@@ -267,4 +268,58 @@ inline arm_compute::ReductionOperation ConvertReductionOperationToAcl(const Redu
     }
 }
 
+/// Function to compute the output tensor shape based on the axes and if keepDims is set.
+inline const TensorShape ComputeReductionTensorShape(const armnn::TensorInfo& input,
+                                                     const std::vector<uint32_t>& vAxis,
+                                                     const bool keepDims)
+{
+    unsigned int rank = input.GetNumDimensions();
+    unsigned int outputRank = 0;
+
+    // Calculate output dimension
+    if (keepDims)
+    {
+        outputRank = rank;
+    }
+    else if (vAxis.empty())
+    {
+        outputRank = 1;
+    }
+    else if (vAxis.size() > input.GetNumDimensions())
+    {
+        throw LayerValidationException("ReduceLayer: Dimensions to reduce can not be bigger than input dimensions");
+    }
+    else
+    {
+        outputRank = input.GetNumDimensions() - armnn::numeric_cast<unsigned int>(vAxis.size());
+        if (outputRank == 0)
+        {
+            outputRank = 1;
+        }
+    }
+
+    std::vector<unsigned int> dimSizes(outputRank, 1);
+    if (!vAxis.empty())
+    {
+        // Skip the dimension that has been reduced unless keepDims is true.
+        unsigned int outputIndex = 0;
+        for (unsigned int i = 0; i < input.GetNumDimensions(); ++i)
+        {
+            if (std::find(vAxis.begin(), vAxis.end(), i) == vAxis.end())
+            {
+                dimSizes[outputIndex] = armnn::numeric_cast<unsigned int>(input.GetShape()[i]);
+                ++outputIndex;
+            }
+            else if (keepDims)
+            {
+                dimSizes[outputIndex] = 1;
+                ++outputIndex;
+            }
+        }
+    }
+
+    const TensorShape inferredShape = TensorShape(outputRank, dimSizes.data());
+    return inferredShape;
+}
+
 } // namespace armnn