From d905decd256558bbee165e636ce4242ac3b9c917 Mon Sep 17 00:00:00 2001
From: Matthew Sloyan <matthew.sloyan@arm.com>
Date: Mon, 3 May 2021 12:22:03 +0100
Subject: 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
---
 src/backends/neon/NeonBackend.cpp | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)

(limited to 'src/backends/neon/NeonBackend.cpp')

diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp
index a1299fb458..6d5eab0ddf 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>
@@ -164,7 +165,8 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph
         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
-             || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division)
+             || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division
+             || base.GetType() == LayerType::Reduce)
             && (base.GetAdditionalInformation<ActivationDescriptor>() == nullptr))
         {
             for (auto output = base.BeginOutputSlots(); output != base.EndOutputSlots(); ++output)
@@ -389,6 +391,26 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph
                                 }
                             }
                         }
+
+                        // Separate check for Reduce as we aren't fusing with activation layer
+                        if (base.GetType() == LayerType::Reduce)
+                        {
+                            ReduceLayer* baseLayer = PolymorphicDowncast<ReduceLayer*>(&base);
+
+                            // Get params from base layer
+                            ReduceDescriptor reduceDescriptor = baseLayer->GetParameters();
+
+                            arm_compute::Status status = NeonReduceWorkloadValidate(
+                                    baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(),
+                                    baseLayer->GetOutputSlot(0).GetTensorInfo(),
+                                    reduceDescriptor);
+
+                            if (status)
+                            {
+                                ChainReduceLayers<ReduceLayer>(optimizationViews, baseLayer, reduceDescriptor);
+                                untouched.erase(baseLayer->GetGuid());
+                            }
+                        }
                     }
                 }
             }
-- 
cgit v1.2.1