6 files changed, 117 insertions, 10 deletions

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