IVGCVSW-1927 Add Neon 8-bit FullyConnected support

Change-Id: Idf4cc7a9a7d3261b9eceb653b999257506cdae76

9 files changed, 37 insertions, 21 deletions

diff --git a/src/backends/NeonLayerSupport.cpp b/src/backends/NeonLayerSupport.cpp
index 30956dfba0..f06db1747d 100644
--- a/src/backends/NeonLayerSupport.cpp
+++ b/src/backends/NeonLayerSupport.cpp
@@ -23,7 +23,7 @@
 #include "NeonWorkloads/NeonL2NormalizationFloatWorkload.hpp"
 #include "NeonWorkloads/NeonMultiplicationFloatWorkload.hpp"
 #include "NeonWorkloads/NeonNormalizationFloatWorkload.hpp"
-#include "NeonWorkloads/NeonFullyConnectedFloatWorkload.hpp"
+#include "NeonWorkloads/NeonFullyConnectedWorkload.hpp"
 #include "NeonWorkloads/NeonPermuteWorkload.hpp"
 #include "NeonWorkloads/NeonPooling2dBaseWorkload.hpp"
 #include "NeonWorkloads/NeonSoftmaxBaseWorkload.hpp"
diff --git a/src/backends/NeonWorkloadFactory.cpp b/src/backends/NeonWorkloadFactory.cpp
index c989121eac..3b994bf049 100644
--- a/src/backends/NeonWorkloadFactory.cpp
+++ b/src/backends/NeonWorkloadFactory.cpp
@@ -116,8 +116,8 @@ std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMerger(const Merger
 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateFullyConnected(
     const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) const
 {
-    return MakeWorkload<NeonFullyConnectedFloatWorkload, NullWorkload>(descriptor, info,
-                                                                         m_MemoryManager.GetIntraLayerManager());
+    return MakeWorkload<NeonFullyConnectedWorkload, NeonFullyConnectedWorkload>(descriptor, info,
+                                                                                m_MemoryManager.GetIntraLayerManager());
 }
 
 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor,
diff --git a/src/backends/NeonWorkloads.hpp b/src/backends/NeonWorkloads.hpp
index 676c23cc4d..e471bbcf35 100644
--- a/src/backends/NeonWorkloads.hpp
+++ b/src/backends/NeonWorkloads.hpp
@@ -21,7 +21,7 @@
 #include "backends/NeonWorkloads/NeonDepthwiseConvolutionFloatWorkload.hpp"
 #include "backends/NeonWorkloads/NeonDepthwiseConvolutionUint8Workload.hpp"
 #include "backends/NeonWorkloads/NeonFloorFloatWorkload.hpp"
-#include "backends/NeonWorkloads/NeonFullyConnectedFloatWorkload.hpp"
+#include "backends/NeonWorkloads/NeonFullyConnectedWorkload.hpp"
 #include "backends/NeonWorkloads/NeonL2NormalizationFloatWorkload.hpp"
 #include "backends/NeonWorkloads/NeonLstmFloatWorkload.hpp"
 #include "backends/NeonWorkloads/NeonMergerFloatWorkload.hpp"
diff --git a/src/backends/NeonWorkloads/CMakeLists.txt b/src/backends/NeonWorkloads/CMakeLists.txt
index 02cb53dff1..ca44bcc1e1 100644
--- a/src/backends/NeonWorkloads/CMakeLists.txt
+++ b/src/backends/NeonWorkloads/CMakeLists.txt
@@ -37,8 +37,8 @@ list(APPEND armnnNeonBackend_sources
     NeonDepthwiseConvolutionUint8Workload.hpp
     NeonFloorFloatWorkload.cpp
     NeonFloorFloatWorkload.hpp
-    NeonFullyConnectedFloatWorkload.cpp
-    NeonFullyConnectedFloatWorkload.hpp
+    NeonFullyConnectedWorkload.cpp
+    NeonFullyConnectedWorkload.hpp
     NeonL2NormalizationFloatWorkload.cpp
     NeonL2NormalizationFloatWorkload.hpp
     NeonLstmFloatWorkload.cpp
diff --git a/src/backends/NeonWorkloads/NeonFullyConnectedFloatWorkload.cpp b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.cpp
index 8fcc0956e2..8cebb4f48f 100644
--- a/src/backends/NeonWorkloads/NeonFullyConnectedFloatWorkload.cpp
+++ b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.cpp
@@ -3,7 +3,7 @@
 // SPDX-License-Identifier: MIT
 //
 
-#include "NeonFullyConnectedFloatWorkload.hpp"
+#include "NeonFullyConnectedWorkload.hpp"
 
 #include <backends/aclCommon/ArmComputeTensorUtils.hpp>
 #include <backends/aclCommon/ArmComputeUtils.hpp>
@@ -42,12 +42,12 @@ arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
                                                         fullyConnectedLayerInfo);
 }
 
-NeonFullyConnectedFloatWorkload::NeonFullyConnectedFloatWorkload(const FullyConnectedQueueDescriptor& descriptor,
+NeonFullyConnectedWorkload::NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor,
     const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
-    : FloatWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
+    : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
     , m_FullyConnectedLayer(memoryManager)
 {
-    m_Data.ValidateInputsOutputs("NeonFullyConnectedFloatWorkload", 1, 1);
+    m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);
 
     arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
@@ -67,11 +67,25 @@ NeonFullyConnectedFloatWorkload::NeonFullyConnectedFloatWorkload(const FullyConn
     m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
 
     // Allocate
-    InitializeArmComputeTensorDataForFloatTypes(*m_WeightsTensor, m_Data.m_Weight);
+    if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QuantisedAsymm8)
+    {
+        InitialiseArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight->GetConstTensor<uint8_t>());
+    }
+    else
+    {
+        InitializeArmComputeTensorDataForFloatTypes(*m_WeightsTensor, m_Data.m_Weight);
+    }
 
     if (m_BiasesTensor)
     {
-        InitializeArmComputeTensorDataForFloatTypes(*m_BiasesTensor, m_Data.m_Bias);
+        if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
+        {
+            InitialiseArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias->GetConstTensor<int32_t>());
+        }
+        else
+        {
+            InitializeArmComputeTensorDataForFloatTypes(*m_BiasesTensor, m_Data.m_Bias);
+        }
     }
 
     // Force Compute Library to perform the necessary copying and reshaping, after which
@@ -80,13 +94,13 @@ NeonFullyConnectedFloatWorkload::NeonFullyConnectedFloatWorkload(const FullyConn
     FreeUnusedTensors();
 }
 
-void NeonFullyConnectedFloatWorkload::Execute() const
+void NeonFullyConnectedWorkload::Execute() const
 {
-    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedFloatWorkload_Execute");
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonFullyConnectedWorkload_Execute");
     m_FullyConnectedLayer.run();
 }
 
-void NeonFullyConnectedFloatWorkload::FreeUnusedTensors()
+void NeonFullyConnectedWorkload::FreeUnusedTensors()
 {
     FreeTensorIfUnused(m_WeightsTensor);
     FreeTensorIfUnused(m_BiasesTensor);
diff --git a/src/backends/NeonWorkloads/NeonFullyConnectedFloatWorkload.hpp b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.hpp
index 27e5717b04..11991f87b5 100644
--- a/src/backends/NeonWorkloads/NeonFullyConnectedFloatWorkload.hpp
+++ b/src/backends/NeonWorkloads/NeonFullyConnectedWorkload.hpp
@@ -20,11 +20,11 @@ arm_compute::Status NeonFullyConnectedWorkloadValidate(const TensorInfo& input,
                                                        const TensorInfo& biases,
                                                        const FullyConnectedDescriptor& descriptor);
 
-class NeonFullyConnectedFloatWorkload : public FloatWorkload<FullyConnectedQueueDescriptor>
+class NeonFullyConnectedWorkload : public BaseWorkload<FullyConnectedQueueDescriptor>
 {
 public:
-    NeonFullyConnectedFloatWorkload(const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info,
-                                    std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
+    NeonFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info,
+                               std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager);
     virtual void Execute() const override;
 
 private:
diff --git a/src/backends/NeonWorkloads/backend.mk b/src/backends/NeonWorkloads/backend.mk
index 06525c9df2..4c7fbe6fe0 100644
--- a/src/backends/NeonWorkloads/backend.mk
+++ b/src/backends/NeonWorkloads/backend.mk
@@ -19,7 +19,7 @@ BACKEND_SOURCES := \
         NeonDepthwiseConvolutionFloatWorkload.cpp \
         NeonDepthwiseConvolutionUint8Workload.cpp \
         NeonFloorFloatWorkload.cpp \
-        NeonFullyConnectedFloatWorkload.cpp \
+        NeonFullyConnectedWorkload.cpp \
         NeonL2NormalizationFloatWorkload.cpp \
         NeonLstmFloatWorkload.cpp \
         NeonMergerFloatWorkload.cpp \
diff --git a/src/backends/test/ArmComputeNeon.cpp b/src/backends/test/ArmComputeNeon.cpp
index f1a2cf65bd..bc3b6b5def 100644
--- a/src/backends/test/ArmComputeNeon.cpp
+++ b/src/backends/test/ArmComputeNeon.cpp
@@ -316,6 +316,8 @@ ARMNN_AUTO_TEST_CASE(SimpleFullyConnectedWithBias, FullyConnectedFloat32Test, tr
 ARMNN_AUTO_TEST_CASE(SimpleFullyConnectedWithTranspose, FullyConnectedFloat32Test, false, true)
 ARMNN_AUTO_TEST_CASE(FullyConnectedLarge, FullyConnectedLargeTest, false)
 ARMNN_AUTO_TEST_CASE(FullyConnectedLargeTransposed, FullyConnectedLargeTest, true)
+ARMNN_AUTO_TEST_CASE(FullyConnectedUint8, FullyConnectedUint8Test, false)
+ARMNN_AUTO_TEST_CASE(FullyConnectedBiasedUint8, FullyConnectedUint8Test, true)
 
 // Add
 ARMNN_AUTO_TEST_CASE(SimpleAdd, AdditionTest)
diff --git a/src/backends/test/CreateWorkloadNeon.cpp b/src/backends/test/CreateWorkloadNeon.cpp
index fbe064e1c4..ce62a02537 100644
--- a/src/backends/test/CreateWorkloadNeon.cpp
+++ b/src/backends/test/CreateWorkloadNeon.cpp
@@ -225,13 +225,13 @@ static void NeonCreateFullyConnectedWorkloadTest()
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 BOOST_AUTO_TEST_CASE(CreateFullyConnectedFloat16Workload)
 {
-    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedFloatWorkload, DataType::Float16>();
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::Float16>();
 }
 #endif
 
 BOOST_AUTO_TEST_CASE(CreateFullyConnectedFloatWorkload)
 {
-    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedFloatWorkload, DataType::Float32>();
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::Float32>();
 }
 
 template <typename NormalizationWorkloadType, typename armnn::DataType DataType>