1 files changed, 33 insertions, 18 deletions

diff --git a/src/backends/cl/workloads/ClMergerWorkload.cpp b/src/backends/cl/workloads/ClMergerWorkload.cpp
index e06d8c51f5..610acb91fb 100644
--- a/src/backends/cl/workloads/ClMergerWorkload.cpp
+++ b/src/backends/cl/workloads/ClMergerWorkload.cpp
@@ -9,16 +9,25 @@
 #include <cl/ClTensorHandle.hpp>
 #include <cl/ClLayerSupport.hpp>
 
+#include <arm_compute/core/Types.h>
+
 #include <boost/polymorphic_pointer_cast.hpp>
 
 namespace armnn
 {
 using namespace armcomputetensorutils;
 
+namespace
+{
+size_t CalcAxis(const MergerDescriptor& desc)
+{
+    return (desc.GetNumDimensions() - desc.GetConcatAxis()) - 1;
+}
+} //namespace
+
 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)
@@ -27,59 +36,65 @@ arm_compute::Status ClMergerWorkloadValidate(const std::vector<const TensorInfo*
         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);
     }
 
+    size_t aclAxis = CalcAxis(descriptor);
     return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
-
 }
 
 ClMergerWorkload::ClMergerWorkload(const MergerQueueDescriptor& descriptor, const WorkloadInfo& info)
 : BaseWorkload<MergerQueueDescriptor>(descriptor, info)
 {
-    m_Execute = true;
+    bool allInputsAreSubtensors = true;
 
-    unsigned int innerAxisOrder = descriptor.m_Parameters.GetNumDimensions() - descriptor.m_Parameters.GetConcatAxis();
+    // Check that all inputs are sub-tensors
+    for (auto input : descriptor.m_Inputs)
+    {
+        if (!input->GetParent())
+        {
+            // Non sub-tensor input found so we need to execute the merger function
+            allInputsAreSubtensors = false;
+            break;
+        }
+    }
 
-    if (innerAxisOrder != 1)
+    if (allInputsAreSubtensors)
     {
-        m_Execute = false;
+        // Can skip configuring the merger function since it's not executed
         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);
+    // Create the layer function
+    m_Layer.reset(new arm_compute::CLConcatenateLayer());
 
-    m_Layer.prepare();
+    // Configure input and output tensors
+    size_t aclAxis = CalcAxis(descriptor.m_Parameters);
+    m_Layer->configure(aclInputs, &output, aclAxis);
 
+    // Prepare
+    m_Layer->prepare();
 }
 
 void ClMergerWorkload::Execute() const
 {
-    if (m_Execute)
+    if (m_Layer)
     {
         ARMNN_SCOPED_PROFILING_EVENT_CL("ClMergerWorkload_Execute");
-        m_Layer.run();
+        m_Layer->run();
     }
-
 }
 
 } //namespace armnn
\ No newline at end of file