3 files changed, 126 insertions, 6 deletions
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt
index abbed0e754..a3eedd028e 100644
--- a/src/backends/cl/workloads/CMakeLists.txt
+++ b/src/backends/cl/workloads/CMakeLists.txt
@@ -64,6 +64,7 @@ list(APPEND armnnClBackendWorkloads_sources
     ClSoftmaxUint8Workload.hpp
     ClSpaceToBatchNdWorkload.hpp
     ClSpaceToBatchNdWorkload.cpp
+    ClSplitterWorkload.cpp
     ClSplitterWorkload.hpp
     ClStridedSliceWorkload.cpp
     ClStridedSliceWorkload.hpp
diff --git a/src/backends/cl/workloads/ClSplitterWorkload.cpp b/src/backends/cl/workloads/ClSplitterWorkload.cpp
new file mode 100644
index 0000000000..9bbbcab797
--- /dev/null
+++ b/src/backends/cl/workloads/ClSplitterWorkload.cpp
@@ -0,0 +1,112 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClSplitterWorkload.hpp"
+
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <cl/ClTensorHandle.hpp>
+
+
+namespace armnn
+{
+
+using namespace armcomputetensorutils;
+
+namespace
+{
+    unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int splitAxis)
+    {
+        return (numDimensions - splitAxis) - 1;
+    }
+
+} //namespace
+
+arm_compute::Status ClSplitterWorkloadValidate(const TensorInfo& input,
+                                               const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
+                                               unsigned int splitAxis)
+{
+    const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
+
+    size_t numOutputs = outputs.size();
+
+    std::vector<arm_compute::TensorInfo> aclOutputs;
+    aclOutputs.reserve(numOutputs);
+
+    std::vector<arm_compute::ITensorInfo*> aclOutputPtr;
+    aclOutputPtr.reserve(numOutputs);
+
+    for (size_t i = 0u; i < outputs.size(); ++i)
+    {
+        aclOutputs.emplace_back(BuildArmComputeTensorInfo(outputs[i]));
+        aclOutputPtr.emplace_back(&aclOutputs.back());
+    }
+
+    unsigned int aclAxis = CalcAclAxis(input.GetNumDimensions(), splitAxis);
+    return arm_compute::CLSplit::validate(&aclInputInfo, aclOutputPtr, aclAxis);
+}
+
+ClSplitterWorkload::ClSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info)
+        : BaseWorkload<SplitterQueueDescriptor>(descriptor, info)
+{
+    bool allOutputsAreSubtensors = true;
+
+    // Check that all outputs are sub-tensors
+    for (auto output : m_Data.m_Outputs)
+    {
+        if (output && !output->GetParent())
+        {
+            // Non sub-tensor input found so we need to execute the split function
+            allOutputsAreSubtensors = false;
+            break;
+        }
+    }
+
+    if (allOutputsAreSubtensors)
+    {
+        // Can skip configuring the split function since it's not executed
+        return;
+    }
+
+    arm_compute::ICLTensor& input = boost::polymorphic_pointer_downcast<IClTensorHandle>(
+            m_Data.m_Inputs[0])->GetTensor();
+
+    std::vector<arm_compute::ICLTensor *> aclOutputs;
+    for (auto output : m_Data.m_Outputs)
+    {
+        arm_compute::ICLTensor& aclOutput  = boost::polymorphic_pointer_downcast<IClTensorHandle>(output)->GetTensor();
+        aclOutputs.emplace_back(&aclOutput);
+    }
+
+    // Create the layer function
+    m_Layer.reset(new arm_compute::CLSplit());
+
+    // Configure input and output tensors
+    std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape());
+    if (splitAxis.size() != 1)
+    {
+        throw InvalidArgumentException("Cannot derive split axis from SplitterDescriptor");
+    }
+
+    unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin());
+    m_Layer->configure(&input, aclOutputs, aclAxis);
+
+    // Prepare
+    m_Layer->prepare();
+}
+
+void ClSplitterWorkload::Execute() const
+{
+    if (m_Layer)
+    {
+        ARMNN_SCOPED_PROFILING_EVENT_CL("ClSplitterWorkload_Execute");
+        m_Layer->run();
+    }
+}
+
+} //namespace armnn
diff --git a/src/backends/cl/workloads/ClSplitterWorkload.hpp b/src/backends/cl/workloads/ClSplitterWorkload.hpp
index 950335902e..d024452d78 100644
--- a/src/backends/cl/workloads/ClSplitterWorkload.hpp
+++ b/src/backends/cl/workloads/ClSplitterWorkload.hpp
@@ -7,19 +7,26 @@
 
 #include <backendsCommon/Workload.hpp>
 
+#include <arm_compute/runtime/CL/CLFunctions.h>
+
+#include <functional>
+
 namespace armnn
 {
 
-// Base class template providing an implementation of the Splitter layer common to all data types.
+arm_compute::Status ClSplitterWorkloadValidate(const TensorInfo& input,
+                                               const std::vector<std::reference_wrapper<TensorInfo>>& outputs,
+                                               unsigned int splitAxis);
+
 class ClSplitterWorkload : public BaseWorkload<SplitterQueueDescriptor>
 {
 public:
-    using BaseWorkload<SplitterQueueDescriptor>::BaseWorkload;
+    ClSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info);
+
+    void Execute() const override;
 
-    void Execute() const override
-    {
-        // With subtensors, splitter is a no-op.
-    }
+private:
+    mutable std::unique_ptr<arm_compute::CLSplit> m_Layer;
 };
 
 } //namespace armnn