IVGCVSW-2043 - Merger using ACL for innermost concat axis

     * Add ClMergerWorkload and NeonMergerWorkload to call ACL for innermost concat axis
     * Modify layer support to call ClMergerWorkloadValidate and NeonMergerWorkloadValidate when concat axis is inner most
     * Add m_ConcatAxis to MergerDescriptor
     * Modify MergerQueueDescriptor::Validate to check sub tensor only when using subtensor

!android-nn-driver:166

Change-Id: I56676b43964c8d6d726387b41b3cc34a512c0f0a

24 files changed, 282 insertions, 28 deletions

diff --git a/include/armnn/Descriptors.hpp b/include/armnn/Descriptors.hpp
index 2781786f10..0abc7583b9 100644
--- a/include/armnn/Descriptors.hpp
+++ b/include/armnn/Descriptors.hpp
@@ -63,11 +63,14 @@ struct OriginsDescriptor
     const uint32_t* GetViewOrigin(uint32_t idx) const;
     void ReorderOrigins(unsigned int*  newOrdering, unsigned int numNewOrdering);
     friend void swap(OriginsDescriptor& first, OriginsDescriptor& second);
+    void SetConcatAxis(unsigned int concatAxis);
+    unsigned int GetConcatAxis() const;
 
 private:
-    uint32_t   m_NumViews;
-    uint32_t   m_NumDimensions;
-    uint32_t** m_ViewOrigins;
+    unsigned int m_ConcatAxis;
+    uint32_t     m_NumViews;
+    uint32_t     m_NumDimensions;
+    uint32_t**   m_ViewOrigins;
 };
 
 struct ViewsDescriptor
@@ -138,6 +141,7 @@ OriginsDescriptor CreateMergerDescriptorForConcatenation(TensorShapeIt first, Te
     }
 
     OriginsDescriptor viewsDescriptor(static_cast<uint32_t>(numInputs), numDimensions);
+    viewsDescriptor.SetConcatAxis(concatenationDimension);
 
     uint32_t viewIndex = 0u;
     uint32_t coordAlongConcatDim = 0u;
diff --git a/include/armnn/ILayerSupport.hpp b/include/armnn/ILayerSupport.hpp
index e679f84109..7677971c6b 100644
--- a/include/armnn/ILayerSupport.hpp
+++ b/include/armnn/ILayerSupport.hpp
@@ -132,6 +132,7 @@ public:
                                  Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const;
 
     virtual bool IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                                   const TensorInfo& output,
                                    const OriginsDescriptor& descriptor,
                                    Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const;
 
diff --git a/include/armnn/LayerSupport.hpp b/include/armnn/LayerSupport.hpp
index 7bed5779d3..83d79ec79e 100644
--- a/include/armnn/LayerSupport.hpp
+++ b/include/armnn/LayerSupport.hpp
@@ -147,6 +147,7 @@ bool IsLstmSupported(const BackendId& backend, const TensorInfo& input, const Te
 /// Deprecated in favor of IBackend and ILayerSupport interfaces
 bool IsMergerSupported(const BackendId& backend,
                        const std::vector<const TensorInfo*> inputs,
+                       const TensorInfo& output,
                        const OriginsDescriptor& descriptor,
                        char* reasonIfUnsupported = nullptr,
                        size_t reasonIfUnsupportedMaxLength = 1024);
diff --git a/src/armnn/Descriptors.cpp b/src/armnn/Descriptors.cpp
index a017806782..a200c6462c 100644
--- a/src/armnn/Descriptors.cpp
+++ b/src/armnn/Descriptors.cpp
@@ -72,13 +72,15 @@ PermutationVector::PermutationVector(std::initializer_list<ValueType> dimMapping
 }
 
 OriginsDescriptor::OriginsDescriptor()
-: m_NumViews(0)
+: m_ConcatAxis(1)
+, m_NumViews(0)
 , m_NumDimensions(0)
 , m_ViewOrigins(nullptr)
 {}
 
 OriginsDescriptor::OriginsDescriptor(uint32_t numViews, uint32_t numDimensions /*= 4*/)
-: m_NumViews(numViews)
+: m_ConcatAxis(1)
+, m_NumViews(numViews)
 , m_NumDimensions(numDimensions)
 , m_ViewOrigins(numViews && numDimensions > 0 ? new uint32_t *[numViews]() : nullptr)
 {
@@ -89,7 +91,8 @@ OriginsDescriptor::OriginsDescriptor(uint32_t numViews, uint32_t numDimensions /
 }
 
 OriginsDescriptor::OriginsDescriptor(const OriginsDescriptor& other)
-: m_NumViews(other.m_NumViews)
+: m_ConcatAxis(other.m_ConcatAxis)
+, m_NumViews(other.m_NumViews)
 , m_NumDimensions(other.m_NumDimensions)
 , m_ViewOrigins(other.m_NumViews && other.m_NumDimensions > 0 ? new uint32_t *[other.m_NumViews]() : nullptr)
 {
@@ -121,6 +124,15 @@ OriginsDescriptor& OriginsDescriptor::operator=(OriginsDescriptor rhs)
     return *this;
 }
 
+void OriginsDescriptor::SetConcatAxis(unsigned int concatAxis)
+{
+    m_ConcatAxis = concatAxis;
+}
+unsigned int OriginsDescriptor::GetConcatAxis() const
+{
+    return m_ConcatAxis;
+}
+
 Status OriginsDescriptor::SetViewOriginCoord(uint32_t view, uint32_t coord, uint32_t value)
 {
     if (view >= m_NumViews)
@@ -284,6 +296,7 @@ void swap(OriginsDescriptor& first, OriginsDescriptor& second)
     swap(first.m_NumViews, second.m_NumViews);
     swap(first.m_NumDimensions, second.m_NumDimensions);
     swap(first.m_ViewOrigins, second.m_ViewOrigins);
+    swap(first.m_ConcatAxis, second.m_ConcatAxis);
 }
 
 void swap(ViewsDescriptor& first, ViewsDescriptor& second)
diff --git a/src/armnn/LayerSupport.cpp b/src/armnn/LayerSupport.cpp
index 91aca4ea41..6489fe4f66 100644
--- a/src/armnn/LayerSupport.cpp
+++ b/src/armnn/LayerSupport.cpp
@@ -257,12 +257,13 @@ bool IsLstmSupported(const BackendId& backend, const TensorInfo& input, const Te
 }
 bool IsMergerSupported(const BackendId& backend,
                        std::vector<const TensorInfo*> inputs,
+                       const TensorInfo& output,
                        const OriginsDescriptor& descriptor,
                        char* reasonIfUnsupported,
                        size_t reasonIfUnsupportedMaxLength)
 {
     BOOST_ASSERT(inputs.size() > 0);
-    FORWARD_LAYER_SUPPORT_FUNC(backend, IsMergerSupported, inputs, descriptor);
+    FORWARD_LAYER_SUPPORT_FUNC(backend, IsMergerSupported, inputs, output, descriptor);
 }
 
 bool IsMultiplicationSupported(const BackendId& backend,
diff --git a/src/armnn/layers/MergerLayer.cpp b/src/armnn/layers/MergerLayer.cpp
index e80661a493..545c7e0e0e 100644
--- a/src/armnn/layers/MergerLayer.cpp
+++ b/src/armnn/layers/MergerLayer.cpp
@@ -40,7 +40,10 @@ void MergerLayer::CreateTensorHandles(Graph& graph, const IWorkloadFactory& fact
     //just needs to make sure that the outputs of the prev layer
     //are made subtensors of the output of the merger layer.
     m_OutputHandlers[0].CreateTensorHandles(factory);
-    if (factory.SupportsSubTensors())
+
+    signed long innerAxis = m_Param.GetNumDimensions() - m_Param.GetConcatAxis();
+
+    if (factory.SupportsSubTensors() && innerAxis != 1)
     {
         std::queue<MergerLayer*> m_MergerLayers;
 
diff --git a/src/backends/backendsCommon/ILayerSupport.cpp b/src/backends/backendsCommon/ILayerSupport.cpp
index dc106e344e..55dd447369 100644
--- a/src/backends/backendsCommon/ILayerSupport.cpp
+++ b/src/backends/backendsCommon/ILayerSupport.cpp
@@ -192,6 +192,7 @@ bool ILayerSupport::IsMeanSupported(const TensorInfo& input,
 }
 
 bool ILayerSupport::IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                                      const TensorInfo& output,
                                       const OriginsDescriptor& descriptor,
                                       Optional<std::string&> reasonIfUnsupported) const
 {
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index e1146543ff..18ab4a8709 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -390,6 +390,16 @@ void MergerQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
         throw InvalidArgumentException("MergerQueueDescriptor: At least one TensorInfo output needs to be provided.");
     }
 
+    if(m_Parameters.GetConcatAxis() > workloadInfo.m_InputTensorInfos[0].GetShape().GetNumDimensions())
+    {
+        throw InvalidArgumentException("Invalid Concatenation Axis provided");
+    }
+
+    if (workloadInfo.m_InputTensorInfos[0].GetShape().GetNumDimensions() - m_Parameters.GetConcatAxis() == 1)
+    {
+        return;
+    }
+
     if (workloadInfo.m_InputTensorInfos.size() != m_ViewOrigins.size())
     {
         throw InvalidArgumentException(
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index dc38f1a721..2e2824e521 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -471,7 +471,9 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId,
             auto endPtr = boost::make_transform_iterator(inputs.end(), getTensorInfoPtr);
             std::vector<const TensorInfo*> inputPtrs(beginPtr, endPtr);
 
-            result = layerSupportObject->IsMergerSupported(inputPtrs, cLayer->GetParameters(), reason);
+            const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+
+            result = layerSupportObject->IsMergerSupported(inputPtrs, output, cLayer->GetParameters(), reason);
             break;
         }
         case LayerType::Multiplication:
diff --git a/src/backends/backendsCommon/test/LayerTests.cpp b/src/backends/backendsCommon/test/LayerTests.cpp
index bd8b38da01..f5689e74bf 100755
--- a/src/backends/backendsCommon/test/LayerTests.cpp
+++ b/src/backends/backendsCommon/test/LayerTests.cpp
@@ -2235,6 +2235,7 @@ void Concatenate(
     }
 
     armnn::OriginsDescriptor viewsDescriptor = CreateMergerDescriptorForConcatenation(inputTensorInfos, concatDim);
+    queueDescriptor.m_Parameters = viewsDescriptor;
 
     queueDescriptor.m_ViewOrigins.reserve(viewsDescriptor.GetNumViews());
     for (unsigned int i = 0; i < viewsDescriptor.GetNumViews(); ++i)
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 039f1c24f0..a5030beb47 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -27,6 +27,7 @@
 #include "workloads/ClL2NormalizationFloatWorkload.hpp"
 #include "workloads/ClLstmFloatWorkload.hpp"
 #include "workloads/ClMeanWorkload.hpp"
+#include "workloads/ClMergerWorkload.hpp"
 #include "workloads/ClMultiplicationWorkload.hpp"
 #include "workloads/ClNormalizationFloatWorkload.hpp"
 #include "workloads/ClPadWorkload.hpp"
@@ -366,14 +367,25 @@ bool ClLayerSupport::IsMeanSupported(const TensorInfo& input,
 }
 
 bool ClLayerSupport::IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                                       const TensorInfo& output,
                                        const OriginsDescriptor& descriptor,
                                        Optional<std::string&> reasonIfUnsupported) const
 {
-    ignore_unused(descriptor);
-    return IsSupportedForDataTypeCl(reasonIfUnsupported,
-                                    inputs[0]->GetDataType(),
-                                    &TrueFunc<>,
-                                    &FalseFuncU8<>);
+    if(descriptor.GetNumDimensions() - descriptor.GetConcatAxis() == 1)
+    {
+        FORWARD_WORKLOAD_VALIDATE_FUNC(ClMergerWorkloadValidate,
+                                       reasonIfUnsupported,
+                                       inputs,
+                                       output,
+                                       descriptor);
+    }
+    else
+    {
+        return IsSupportedForDataTypeCl(reasonIfUnsupported,
+                                        inputs[0]->GetDataType(),
+                                        &TrueFunc<>,
+                                        &FalseFuncU8<>);
+    }
 }
 
 bool ClLayerSupport::IsMultiplicationSupported(const TensorInfo& input0,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index 6bdeb5a6f6..b4d278995d 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -117,6 +117,7 @@ public:
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
     bool IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                           const TensorInfo& output,
                            const OriginsDescriptor& descriptor,
                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index fd7ea80f33..551e1a0e4b 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -28,6 +28,7 @@ BACKEND_SOURCES := \
         workloads/ClL2NormalizationFloatWorkload.cpp \
         workloads/ClLstmFloatWorkload.cpp \
         workloads/ClMeanWorkload.cpp \
+        workloads/ClMergerWorkload.cpp \
         workloads/ClMultiplicationWorkload.cpp \
         workloads/ClNormalizationFloatWorkload.cpp \
         workloads/ClPadWorkload.cpp \
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index 901dd748e4..736cf5c4e5 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -32,6 +32,7 @@ list(APPEND armnnClBackendWorkloads_sources
     ClLstmFloatWorkload.hpp
     ClMeanWorkload.cpp
     ClMeanWorkload.hpp
+    ClMergerWorkload.cpp
     ClMergerWorkload.hpp
     ClMultiplicationWorkload.cpp
     ClMultiplicationWorkload.hpp
diff --git a/src/backends/cl/workloads/ClMergerWorkload.cpp b/src/backends/cl/workloads/ClMergerWorkload.cpp
new file mode 100644
index 0000000000..e06d8c51f5
--- /dev/null
+++ b/src/backends/cl/workloads/ClMergerWorkload.cpp
@@ -0,0 +1,85 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#include "ClMergerWorkload.hpp"
+#include "ClWorkloadUtils.hpp"
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <cl/ClTensorHandle.hpp>
+#include <cl/ClLayerSupport.hpp>
+
+#include <boost/polymorphic_pointer_cast.hpp>
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status ClMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
+                                             const TensorInfo& output,
+                                             const MergerDescriptor& descriptor)
+
+{
+    std::vector<arm_compute::TensorInfo> aclInputs;
+    for (const TensorInfo* input : inputs)
+    {
+        arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW);
+        aclInputs.emplace_back(aclInputInfo);
+    }
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+    arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH;
+
+    std::vector<arm_compute::ITensorInfo*> aclInputPtrs;
+    for (arm_compute::ITensorInfo& input : aclInputs)
+    {
+        aclInputPtrs.emplace_back(&input);
+    }
+
+    return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
+
+}
+
+ClMergerWorkload::ClMergerWorkload(const MergerQueueDescriptor& descriptor, const WorkloadInfo& info)
+: BaseWorkload<MergerQueueDescriptor>(descriptor, info)
+{
+    m_Execute = true;
+
+    unsigned int innerAxisOrder = descriptor.m_Parameters.GetNumDimensions() - descriptor.m_Parameters.GetConcatAxis();
+
+    if (innerAxisOrder != 1)
+    {
+        m_Execute = false;
+        return;
+    }
+
+    std::vector<arm_compute::ICLTensor *> aclInputs;
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(armnn::DataLayout::NCHW);
+    for (auto input : m_Data.m_Inputs)
+    {
+        arm_compute::ICLTensor& aclInput  = boost::polymorphic_pointer_downcast<IClTensorHandle>(input)->GetTensor();
+        aclInput.info()->set_data_layout(aclDataLayout);
+        aclInputs.emplace_back(&aclInput);
+    }
+    arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>(
+                                                                         m_Data.m_Outputs[0])->GetTensor();
+    output.info()->set_data_layout(aclDataLayout);
+
+    arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH;
+
+    m_Layer.configure(aclInputs, &output, aclAxis);
+
+    m_Layer.prepare();
+
+}
+
+void ClMergerWorkload::Execute() const
+{
+    if (m_Execute)
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_CL("ClMergerWorkload_Execute");
+        m_Layer.run();
+    }
+
+}
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/cl/workloads/ClMergerWorkload.hpp b/src/backends/cl/workloads/ClMergerWorkload.hpp
index 948fb58bd1..8189a1b24a 100644
--- a/src/backends/cl/workloads/ClMergerWorkload.hpp
+++ b/src/backends/cl/workloads/ClMergerWorkload.hpp
@@ -7,18 +7,25 @@
 
 #include <backendsCommon/Workload.hpp>
 
+#include <arm_compute/runtime/CL/CLFunctions.h>
+
 namespace armnn
 {
 
+arm_compute::Status ClMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
+                                             const TensorInfo& output,
+                                             const MergerDescriptor& descriptor);
+
 class ClMergerWorkload : public BaseWorkload<MergerQueueDescriptor>
 {
 public:
-    using BaseWorkload<MergerQueueDescriptor>::BaseWorkload;
+    ClMergerWorkload(const MergerQueueDescriptor& descriptor, const WorkloadInfo& info);
+
+    void Execute() const override;
 
-    void Execute() const override
-    {
-        // With subtensors, merger is a no-op.
-    }
+private:
+    mutable arm_compute::CLConcatenateLayer m_Layer;
+    bool m_Execute;
 };
 
 } //namespace armnn
diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp
index 28c4b75f2a..fd279e5d55 100644
--- a/src/backends/neon/NeonLayerSupport.cpp
+++ b/src/backends/neon/NeonLayerSupport.cpp
@@ -23,6 +23,7 @@
 #include "workloads/NeonConvolution2dWorkload.hpp"
 #include "workloads/NeonDepthwiseConvolutionWorkload.hpp"
 #include "workloads/NeonL2NormalizationFloatWorkload.hpp"
+#include "workloads/NeonMergerWorkload.hpp"
 #include "workloads/NeonMultiplicationFloatWorkload.hpp"
 #include "workloads/NeonNormalizationFloatWorkload.hpp"
 #include "workloads/NeonFullyConnectedWorkload.hpp"
@@ -334,14 +335,25 @@ bool NeonLayerSupport::IsMeanSupported(const TensorInfo& input,
 }
 
 bool NeonLayerSupport::IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                                         const TensorInfo& output,
                                          const OriginsDescriptor& descriptor,
                                          Optional<std::string&> reasonIfUnsupported) const
 {
-    ignore_unused(descriptor);
-    return IsSupportedForDataTypeNeon(reasonIfUnsupported,
-                                      inputs[0]->GetDataType(),
-                                      &TrueFunc<>,
-                                      &TrueFunc<>);
+    if(descriptor.GetNumDimensions() - descriptor.GetConcatAxis() == 1)
+    {
+        FORWARD_WORKLOAD_VALIDATE_FUNC(NeonMergerWorkloadValidate,
+                                       reasonIfUnsupported,
+                                       inputs,
+                                       output,
+                                       descriptor);
+    }
+    else
+     {
+         return IsSupportedForDataTypeNeon(reasonIfUnsupported,
+                                           inputs[0]->GetDataType(),
+                                           &TrueFunc<>,
+                                           &TrueFunc<>);
+      }
 }
 
 bool NeonLayerSupport::IsMultiplicationSupported(const TensorInfo& input0,
diff --git a/src/backends/neon/NeonLayerSupport.hpp b/src/backends/neon/NeonLayerSupport.hpp
index fb1567c12d..e5cd3cc062 100644
--- a/src/backends/neon/NeonLayerSupport.hpp
+++ b/src/backends/neon/NeonLayerSupport.hpp
@@ -117,6 +117,7 @@ public:
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
     bool IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                           const TensorInfo& output,
                            const OriginsDescriptor& descriptor,
                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
diff --git a/src/backends/neon/backend.mk b/src/backends/neon/backend.mk
index 8f7e72b17c..2f74ecd1ce 100644
--- a/src/backends/neon/backend.mk
+++ b/src/backends/neon/backend.mk
@@ -25,6 +25,7 @@ BACKEND_SOURCES := \
         workloads/NeonFullyConnectedWorkload.cpp \
         workloads/NeonL2NormalizationFloatWorkload.cpp \
         workloads/NeonLstmFloatWorkload.cpp \
+        workloads/NeonMergerWorkload.cpp \
         workloads/NeonMultiplicationFloatWorkload.cpp \
         workloads/NeonNormalizationFloatWorkload.cpp \
         workloads/NeonPermuteWorkload.cpp \
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index f6e8d1c04a..e383b04f25 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -28,6 +28,7 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonL2NormalizationFloatWorkload.hpp
     NeonLstmFloatWorkload.cpp
     NeonLstmFloatWorkload.hpp
+    NeonMergerWorkload.cpp
     NeonMergerWorkload.hpp
     NeonMultiplicationFloatWorkload.cpp
     NeonMultiplicationFloatWorkload.hpp
diff --git a/src/backends/neon/workloads/NeonMergerWorkload.cpp b/src/backends/neon/workloads/NeonMergerWorkload.cpp
new file mode 100644
index 0000000000..f82e24453a
--- /dev/null
+++ b/src/backends/neon/workloads/NeonMergerWorkload.cpp
@@ -0,0 +1,84 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonMergerWorkload.hpp"
+#include <armnn/ArmNN.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <neon/NeonTensorHandle.hpp>
+
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+arm_compute::Status NeonMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
+                                               const TensorInfo& output,
+                                               const MergerDescriptor& descriptor)
+
+{
+    std::vector<arm_compute::TensorInfo> aclInputs;
+    for (const TensorInfo* input : inputs)
+    {
+       arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW);
+       aclInputs.emplace_back(aclInputInfo);
+    }
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+    arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH;
+
+    std::vector<arm_compute::ITensorInfo*> aclInputPtrs;
+    for (arm_compute::ITensorInfo& input : aclInputs)
+    {
+        aclInputPtrs.emplace_back(&input);
+    }
+
+    return arm_compute::NEConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
+
+}
+
+NeonMergerWorkload::NeonMergerWorkload(
+const MergerQueueDescriptor& descriptor, const WorkloadInfo& info)
+        : BaseWorkload<MergerQueueDescriptor>(descriptor, info)
+{
+    m_Execute = true;
+
+    unsigned int innerAxisOrder = descriptor.m_Parameters.GetNumDimensions() - descriptor.m_Parameters.GetConcatAxis();
+
+    if (innerAxisOrder != 1)
+    {
+        m_Execute = false;
+        return;
+    }
+
+    std::vector<arm_compute::ITensor *> aclInputs;
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(armnn::DataLayout::NCHW);
+    for (auto input : m_Data.m_Inputs)
+    {
+        arm_compute::ITensor& aclInput  = boost::polymorphic_pointer_downcast<INeonTensorHandle>(input)->GetTensor();
+        aclInput.info()->set_data_layout(aclDataLayout);
+        aclInputs.emplace_back(&aclInput);
+    }
+    arm_compute::ITensor& output = boost::polymorphic_pointer_downcast<INeonTensorHandle>(
+                                                                       m_Data.m_Outputs[0])->GetTensor();
+    output.info()->set_data_layout(aclDataLayout);
+
+    arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH;
+
+    m_Layer.configure(aclInputs, &output, aclAxis);
+
+    m_Layer.prepare();
+}
+
+void NeonMergerWorkload::Execute() const
+{
+    if (m_Execute)
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonMergerWorkload_Execute");
+        m_Layer.run();
+    }
+}
+
+} //namespace armnn
+
diff --git a/src/backends/neon/workloads/NeonMergerWorkload.hpp b/src/backends/neon/workloads/NeonMergerWorkload.hpp
index 7103d8a469..a4f36d18bc 100644
--- a/src/backends/neon/workloads/NeonMergerWorkload.hpp
+++ b/src/backends/neon/workloads/NeonMergerWorkload.hpp
@@ -6,18 +6,26 @@
 #pragma once
 
 #include <backendsCommon/Workload.hpp>
+#include <neon/workloads/NeonWorkloadUtils.hpp>
 
 namespace armnn
 {
+arm_compute::Status NeonMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
+                                               const TensorInfo& output,
+                                               const MergerDescriptor& descriptor);
+
 class NeonMergerWorkload : public BaseWorkload<MergerQueueDescriptor>
 {
 public:
+    NeonMergerWorkload(const MergerQueueDescriptor& descriptor, const WorkloadInfo& info);
+
     using BaseWorkload<MergerQueueDescriptor>::BaseWorkload;
+    void Execute() const override;
+
+private:
+    mutable arm_compute::NEConcatenateLayer m_Layer;
+    bool m_Execute;
 
-    virtual void Execute() const override
-    {
-        // With subtensors, merger is a no-op.
-    }
 };
 
 } //namespace armnn
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index 167cba54e8..43a2fa2d07 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -322,10 +322,12 @@ bool RefLayerSupport::IsMeanSupported(const TensorInfo& input,
 }
 
 bool RefLayerSupport::IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                                        const TensorInfo& output,
                                         const OriginsDescriptor& descriptor,
                                         Optional<std::string&> reasonIfUnsupported) const
 {
     ignore_unused(descriptor);
+    ignore_unused(output);
     return IsSupportedForDataTypeRef(reasonIfUnsupported,
                                      inputs[0]->GetDataType(),
                                      &TrueFunc<>,
diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp
index 2e86ecee29..a03c89c48c 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -122,6 +122,7 @@ public:
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
     bool IsMergerSupported(const std::vector<const TensorInfo*> inputs,
+                           const TensorInfo& output,
                            const OriginsDescriptor& descriptor,
                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;