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, 178 insertions, 0 deletions

diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
index a0fca46330..521c17cd62 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 +150,85 @@ 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>
+std::vector<Layer*> ChainReduceLayers(OptimizationViews& optimizationViews,
+                                      LayerType* baseLayer,
+                                      ReduceDescriptor& desc)
+{
+    // 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> axes;
+    unsigned int recalulatedAxis = 0;
+
+    for (unsigned int i = 0; i != desc.m_vAxis.size(); ++i)
+    {
+        // Get TensorInfo from base layer and reduce shape using axis.
+        TensorInfo layerInfo = baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
+
+        axes.emplace_back(desc.m_vAxis[i]);
+
+        const TensorInfo& reducedTensorInfo = ComputeReductionTensorShape(layerInfo,
+                                                                          axes,
+                                                                          desc.m_KeepDims);
+
+        // 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, desc.m_vAxis[i] - recalulatedAxis);
+
+        // 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(reducedTensorInfo);
+
+        if (!desc.m_KeepDims)
+        {
+            recalulatedAxis++;
+        }
+
+        layers.emplace_back(replacementLayer);
+    }
+
+    // Check if the TensorInfo from the last layer equals the inferred output from the original layer.
+    ARMNN_ASSERT(baseLayer->GetOutputSlot(0).GetTensorInfo() == layers.back()->GetOutputSlot().GetTensorInfo());
+
+    return layers;
+}
+
+//
+// Substitute baseLayer with new subgraph
+//
+template<typename LayerType>
+void ReplaceLayers(OptimizationViews& optimizationViews,
+                   LayerType* baseLayer,
+                   std::vector<Layer*>& layers)
+{
+    std::list<Layer*> replacementLayers(layers.begin(), layers.end());
+
+    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..624ce5df7a 100644
--- a/src/backends/aclCommon/ArmComputeUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeUtils.hpp
@@ -7,10 +7,19 @@
 #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>
 
+#if defined(ARMCOMPUTENEON_ENABLED)
+#include "neon/workloads/NeonReduceWorkload.hpp"
+#endif
+
+#if defined(ARMCOMPUTECL_ENABLED)
+#include "cl/workloads/ClReduceWorkload.hpp"
+#endif
+
 namespace armnn
 {
 
@@ -267,4 +276,89 @@ 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 TensorInfo ComputeReductionTensorShape(const armnn::TensorInfo& input,
+                                                    const std::vector<uint32_t>& vAxis,
+                                                    const bool keepDims)
+{
+    auto reducedTensorInfo = input;
+    unsigned int rank = reducedTensorInfo.GetNumDimensions();
+    unsigned int outputRank = 0;
+    // Calculate output dimension
+    if (keepDims)
+    {
+        outputRank = rank;
+    }
+    else if (vAxis.empty())
+    {
+        outputRank = 1;
+    }
+    else if (vAxis.size() > reducedTensorInfo.GetNumDimensions())
+    {
+        throw LayerValidationException("ReduceLayer: Dimensions to reduce can not be bigger than input dimensions");
+    }
+    else
+    {
+        outputRank = reducedTensorInfo.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 < reducedTensorInfo.GetNumDimensions(); ++i)
+        {
+            if (std::find(vAxis.begin(), vAxis.end(), i) == vAxis.end())
+            {
+                dimSizes[outputIndex] = armnn::numeric_cast<unsigned int>(reducedTensorInfo.GetShape()[i]);
+                ++outputIndex;
+            }
+            else if (keepDims)
+            {
+                dimSizes[outputIndex] = 1;
+                ++outputIndex;
+            }
+        }
+    }
+    const TensorShape inferredShape = TensorShape(outputRank, dimSizes.data());
+    reducedTensorInfo.SetShape(inferredShape);
+    return reducedTensorInfo;
+}
+
+/// Macro function check if layer with multiple axes is supported on each backend
+#define IS_MULTI_AXES_REDUCE_SUPPORTED(func, input, desc, status)                 \
+    armnn::TensorInfo inputTensorInfo = input;                                    \
+    unsigned int recalulatedAxis = 0;                                             \
+    std::vector<uint32_t> axes;                                                   \
+                                                                                  \
+    for (unsigned int i = 0; i != desc.m_vAxis.size(); ++i)                       \
+    {                                                                             \
+        axes.emplace_back(desc.m_vAxis[i]);                                       \
+                                                                                  \
+        const armnn::TensorInfo& reducedTensorInfo =                              \
+            ComputeReductionTensorShape(input, axes, desc.m_KeepDims);            \
+                                                                                  \
+        std::vector<uint32_t> singleAxis(1, desc.m_vAxis[i] - recalulatedAxis);   \
+                                                                                  \
+        armnn::ReduceDescriptor newReduceDescriptor = desc;                       \
+        newReduceDescriptor.m_vAxis.assign(singleAxis.begin(), singleAxis.end()); \
+                                                                                  \
+        status = func(inputTensorInfo, reducedTensorInfo, newReduceDescriptor);   \
+        if (!status)                                                              \
+        {                                                                         \
+            break;                                                                \
+        }                                                                         \
+                                                                                  \
+        if (!desc.m_KeepDims)                                                     \
+        {                                                                         \
+            recalulatedAxis++;                                                    \
+        }                                                                         \
+                                                                                  \
+        inputTensorInfo = reducedTensorInfo;                                      \
+    }
+
 } // namespace armnn