IVGCVSW-7505 Create FusedLayer and NeonFusedWorkload for AddMulAdd Neon kernel

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ic778d35b001474b44fb1e433a6fe276e4ec9f565

7 files changed, 56 insertions, 1 deletions

diff --git a/include/armnn/BackendHelper.hpp b/include/armnn/BackendHelper.hpp
index 59cbbfced3..986f854636 100644
--- a/include/armnn/BackendHelper.hpp
+++ b/include/armnn/BackendHelper.hpp
@@ -194,6 +194,11 @@ public:
                                    const FullyConnectedDescriptor& descriptor,
                                    Optional<std::string&> reasonIfUnsupported = EmptyOptional());
 
+    bool IsFusedSupported(const std::vector<std::reference_wrapper<TensorInfo>>& inputs,
+                          const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
+                          const FusedDescriptor& descriptor,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional());
+
     bool IsGatherSupported(const TensorInfo& input0,
                            const TensorInfo& input1,
                            const TensorInfo& output,
diff --git a/include/armnn/Descriptors.hpp b/include/armnn/Descriptors.hpp
index f60e8f3bea..30eaefd83b 100644
--- a/include/armnn/Descriptors.hpp
+++ b/include/armnn/Descriptors.hpp
@@ -940,6 +940,27 @@ struct FillDescriptor : BaseDescriptor
     float m_Value;
 };
 
+/// A FusedDescriptor for the FusedLayer.
+struct FusedDescriptor : BaseDescriptor
+{
+    FusedDescriptor(unsigned int numInputSlots = 4u,
+                    unsigned int numOutputSlots = 2u,
+                    FusedKernelType fusedType = FusedKernelType::AddMulAdd)
+        : m_NumInputSlots(numInputSlots), m_NumOutputSlots(numOutputSlots), m_FusedKernelType(fusedType)
+    {}
+
+    bool operator ==(const FusedDescriptor& rhs) const
+    {
+        return m_NumInputSlots  == rhs.m_NumInputSlots &&
+               m_NumOutputSlots == rhs.m_NumOutputSlots &&
+               m_FusedKernelType == rhs.m_FusedKernelType;
+    }
+
+    unsigned int m_NumInputSlots;
+    unsigned int m_NumOutputSlots;
+    FusedKernelType m_FusedKernelType;
+};
+
 /// A GatherDescriptor for the GatherLayer.
 struct GatherDescriptor : BaseDescriptor
 {
diff --git a/include/armnn/DescriptorsFwd.hpp b/include/armnn/DescriptorsFwd.hpp
index be1a3f6782..4b9a3e5060 100644
--- a/include/armnn/DescriptorsFwd.hpp
+++ b/include/armnn/DescriptorsFwd.hpp
@@ -25,6 +25,7 @@ struct ElementwiseUnaryDescriptor;
 struct FakeQuantizationDescriptor;
 struct FillDescriptor;
 struct FullyConnectedDescriptor;
+struct FusedDescriptor;
 struct GatherDescriptor;
 struct InstanceNormalizationDescriptor;
 struct L2NormalizationDescriptor;
diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp
index 1640d7c37d..c2c76e3d97 100644
--- a/include/armnn/INetwork.hpp
+++ b/include/armnn/INetwork.hpp
@@ -477,6 +477,14 @@ public:
     IConnectableLayer* AddFullyConnectedLayer(const FullyConnectedDescriptor& fullyConnectedDescriptor,
                                               const char* name = nullptr);
 
+    /// Adds a Fused layer to the network.
+    /// Method use is for backend users.
+    /// @param fusedDescriptor - FusedDescriptor contains parameters for the Fused layer.
+    /// @param name - Optional name for the layer.
+    /// @return - Interface for configuring the layer.
+    IConnectableLayer* AddFusedLayer(const FusedDescriptor& fusedDescriptor,
+                                     const char* name = nullptr);
+
     /// Adds a permute layer to the network.
     /// @param permuteDescriptor - PermuteDescriptor to configure the permute.
     /// @param name - Optional name for the layer.
diff --git a/include/armnn/Types.hpp b/include/armnn/Types.hpp
index bf4458ee7f..7cb3a859c7 100644
--- a/include/armnn/Types.hpp
+++ b/include/armnn/Types.hpp
@@ -262,6 +262,11 @@ enum class MemBlockStrategyType
     MultiAxisPacking  = 1
 };
 
+enum class FusedKernelType
+{
+    AddMulAdd  = 0
+};
+
 /// Each backend should implement an IBackend.
 class IBackend
 {
@@ -475,6 +480,7 @@ using InferenceTimingPair = std::pair<HighResolutionClock, HighResolutionClock>;
     X(ElementwiseBinary) \
     X(ReverseV2) \
     X(Tile) \
+    X(Fused) \
 
 // New layers should be added at last position to minimize instability.
 
@@ -486,7 +492,7 @@ enum class LayerType
     LIST_OF_LAYER_TYPE
 #undef X
     FirstLayer = Activation,
-    LastLayer = Tile
+    LastLayer = Fused
 };
 
 const char* GetLayerTypeAsCString(LayerType type);
diff --git a/include/armnn/TypesUtils.hpp b/include/armnn/TypesUtils.hpp
index eeb5c9e614..ca098f60fb 100644
--- a/include/armnn/TypesUtils.hpp
+++ b/include/armnn/TypesUtils.hpp
@@ -115,6 +115,15 @@ constexpr char const* GetLogicalBinaryOperationAsCString(LogicalBinaryOperation
     }
 }
 
+constexpr char const* GetFusedTypeAsCString(FusedKernelType type)
+{
+    switch (type)
+    {
+        case FusedKernelType::AddMulAdd:   return "AddMulAdd";
+        default:                           return "Unknown";
+    }
+}
+
 constexpr char const* GetPoolingAlgorithmAsCString(PoolingAlgorithm pooling)
 {
     switch (pooling)
diff --git a/include/armnn/backends/WorkloadData.hpp b/include/armnn/backends/WorkloadData.hpp
index 21a597df8a..86796cbcc0 100644
--- a/include/armnn/backends/WorkloadData.hpp
+++ b/include/armnn/backends/WorkloadData.hpp
@@ -182,6 +182,11 @@ struct FullyConnectedQueueDescriptor : QueueDescriptorWithParameters<FullyConnec
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct FusedQueueDescriptor : QueueDescriptorWithParameters<FusedDescriptor>
+{
+    void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
 // Permute layer workload data.
 struct PermuteQueueDescriptor : QueueDescriptorWithParameters<PermuteDescriptor>
 {