4 files changed, 111 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 42ac641369..f8d5922b70 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -10,6 +10,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonActivationWorkload.hpp
     NeonAdditionWorkload.cpp
     NeonAdditionWorkload.hpp
+    NeonArgMinMaxWorkload.cpp
+    NeonArgMinMaxWorkload.hpp
     NeonBatchNormalizationWorkload.cpp
     NeonBatchNormalizationWorkload.hpp
     NeonConcatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonArgMinMaxWorkload.cpp b/src/backends/neon/workloads/NeonArgMinMaxWorkload.cpp
new file mode 100644
index 0000000000..e8d537f2ef
--- /dev/null
+++ b/src/backends/neon/workloads/NeonArgMinMaxWorkload.cpp
@@ -0,0 +1,79 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonArgMinMaxWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <TensorUtils.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEArgMinMaxLayer.h>
+
+namespace
+{
+unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int axisIndex)
+{
+    return (numDimensions - axisIndex) - 1;
+}
+
+} //namespace
+
+namespace armnn
+{
+
+arm_compute::Status NeonArgMinMaxWorkloadValidate(const TensorInfo& input,
+                                                  const TensorInfo& output,
+                                                  const ArgMinMaxDescriptor& descriptor)
+{
+    const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    auto numDims = input.GetNumDimensions();
+    auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, descriptor.m_Axis);
+    int aclAxis = boost::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis));
+
+    if (descriptor.m_Function == ArgMinMaxFunction::Max)
+    {
+        return arm_compute::NEArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput,
+                                                       arm_compute::ReductionOperation::ARG_IDX_MAX);
+    }
+    else
+    {
+        return arm_compute::NEArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput,
+                                                       arm_compute::ReductionOperation::ARG_IDX_MIN);
+    }
+}
+
+
+NeonArgMinMaxWorkload::NeonArgMinMaxWorkload(const ArgMinMaxQueueDescriptor& descriptor,
+                                             const WorkloadInfo& info)
+        : BaseWorkload<ArgMinMaxQueueDescriptor>(descriptor, info)
+{
+    arm_compute::ITensor& input = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    auto numDims = info.m_InputTensorInfos[0].GetNumDimensions();
+    auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, m_Data.m_Parameters.m_Axis);
+    int aclAxis = boost::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis));
+
+    if (m_Data.m_Parameters.m_Function == ArgMinMaxFunction::Max)
+    {
+        m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MAX);
+    }
+    else
+    {
+        m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MIN);
+    }
+}
+
+void NeonArgMinMaxWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonArgMinMaxWorkload_Execute");
+    m_ArgMinMaxLayer.run();
+}
+
+} //namespace armnn
+
diff --git a/src/backends/neon/workloads/NeonArgMinMaxWorkload.hpp b/src/backends/neon/workloads/NeonArgMinMaxWorkload.hpp
new file mode 100644
index 0000000000..6301b13718
--- /dev/null
+++ b/src/backends/neon/workloads/NeonArgMinMaxWorkload.hpp
@@ -0,0 +1,29 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/NEON/functions/NEArgMinMaxLayer.h>
+namespace armnn
+{
+
+arm_compute::Status NeonArgMinMaxWorkloadValidate(const TensorInfo& input,
+                                                  const TensorInfo& output,
+                                                  const ArgMinMaxDescriptor& descriptor);
+
+class NeonArgMinMaxWorkload : public BaseWorkload<ArgMinMaxQueueDescriptor>
+{
+public:
+    NeonArgMinMaxWorkload(const ArgMinMaxQueueDescriptor& descriptor, const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+    mutable arm_compute::NEArgMinMaxLayer m_ArgMinMaxLayer;
+};
+
+} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 9d35ed42fe..8044a4f004 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -7,6 +7,7 @@
 #include "NeonAbsWorkload.hpp"
 #include "NeonActivationWorkload.hpp"
 #include "NeonAdditionWorkload.hpp"
+#include "NeonArgMinMaxWorkload.hpp"
 #include "NeonBatchNormalizationWorkload.hpp"
 #include "NeonConstantWorkload.hpp"
 #include "NeonConvertFp16ToFp32Workload.hpp"