IVGCVSW-3846 Add NEON GATHER Workload

Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I1a66fdad63cef16866d9dfcb8a339647f856e1d4

5 files changed, 85 insertions, 2 deletions

diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index 6b3fe67f1f..685d75be12 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -46,6 +46,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonFloorFloatWorkload.hpp
     NeonFullyConnectedWorkload.cpp
     NeonFullyConnectedWorkload.hpp
+    NeonGatherWorkload.cpp
+    NeonGatherWorkload.hpp
     NeonInstanceNormalizationWorkload.cpp
     NeonInstanceNormalizationWorkload.hpp
     NeonL2NormalizationFloatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonConstantWorkload.cpp b/src/backends/neon/workloads/NeonConstantWorkload.cpp
index f7c8a73f78..05fdcf2fdd 100644
--- a/src/backends/neon/workloads/NeonConstantWorkload.cpp
+++ b/src/backends/neon/workloads/NeonConstantWorkload.cpp
@@ -23,7 +23,7 @@ arm_compute::Status NeonConstantWorkloadValidate(const TensorInfo& output)
 {
     const arm_compute::TensorInfo neonOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
 
-    std::array<arm_compute::DataType,8> supportedTypes = {
+    std::array<arm_compute::DataType,9> supportedTypes = {
             arm_compute::DataType::BFLOAT16,
             arm_compute::DataType::F16,
             arm_compute::DataType::F32,
@@ -31,7 +31,8 @@ arm_compute::Status NeonConstantWorkloadValidate(const TensorInfo& output)
             arm_compute::DataType::QASYMM8_SIGNED,
             arm_compute::DataType::QSYMM16,
             arm_compute::DataType::QSYMM8,
-            arm_compute::DataType::QSYMM8_PER_CHANNEL
+            arm_compute::DataType::QSYMM8_PER_CHANNEL,
+            arm_compute::DataType::S32
     };
     auto it = std::find(begin(supportedTypes), end(supportedTypes), neonOutputInfo.data_type());
 
@@ -110,6 +111,11 @@ void NeonConstantWorkload::Execute() const
                 CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int8_t>(), output);
                 break;
             }
+            case arm_compute::DataType::S32:
+            {
+                CopyArmComputeITensorData(data.m_LayerOutput->GetConstTensor<int32_t>(), output);
+                break;
+            }
             default:
             {
                 ARMNN_ASSERT_MSG(false, "Unknown data type");
diff --git a/src/backends/neon/workloads/NeonGatherWorkload.cpp b/src/backends/neon/workloads/NeonGatherWorkload.cpp
new file mode 100644
index 0000000000..2e7c741781
--- /dev/null
+++ b/src/backends/neon/workloads/NeonGatherWorkload.cpp
@@ -0,0 +1,46 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonGatherWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+#include <armnn/utility/PolymorphicDowncast.hpp>
+#include <aclCommon/ArmComputeUtils.hpp>
+
+namespace armnn
+{
+arm_compute::Status NeonGatherWorkloadValidate(const TensorInfo& input,
+                                               const TensorInfo& indices,
+                                               const TensorInfo& output)
+{
+    const arm_compute::TensorInfo aclInput   = BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclIndices = BuildArmComputeTensorInfo(indices);
+    const arm_compute::TensorInfo aclOutput  = BuildArmComputeTensorInfo(output);
+
+    int aclAxis = ComputeAclAxis(0, input);
+
+    return arm_compute::NEGather::validate(&aclInput, &aclIndices, &aclOutput, aclAxis);
+}
+
+NeonGatherWorkload::NeonGatherWorkload(const GatherQueueDescriptor& descriptor,
+                                       const WorkloadInfo& info)
+        : BaseWorkload<GatherQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonGatherWorkload", 1, 1);
+
+    arm_compute::ITensor& input   = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& indices = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
+    arm_compute::ITensor& output  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    int aclAxis = ComputeAclAxis(0, info.m_InputTensorInfos[0]);
+
+    m_Layer.configure(&input, &indices, &output, aclAxis);
+}
+
+void NeonGatherWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonGatherWorkload_Execute");
+    m_Layer.run();
+}
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/neon/workloads/NeonGatherWorkload.hpp b/src/backends/neon/workloads/NeonGatherWorkload.hpp
new file mode 100644
index 0000000000..b1b47a5069
--- /dev/null
+++ b/src/backends/neon/workloads/NeonGatherWorkload.hpp
@@ -0,0 +1,28 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NEGather.h>
+
+namespace armnn
+{
+arm_compute::Status NeonGatherWorkloadValidate(const TensorInfo& input,
+                                               const TensorInfo& indices,
+                                               const TensorInfo& output);
+
+class NeonGatherWorkload : public BaseWorkload<GatherQueueDescriptor>
+{
+public:
+    NeonGatherWorkload(const GatherQueueDescriptor& descriptor, const WorkloadInfo& info);
+    virtual void Execute() const override;
+
+private:
+    mutable arm_compute::NEGather m_Layer;
+};
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index 2da6ea0c01..243f5a46ee 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -26,6 +26,7 @@
 #include "NeonExpWorkload.hpp"
 #include "NeonFloorFloatWorkload.hpp"
 #include "NeonFullyConnectedWorkload.hpp"
+#include "NeonGatherWorkload.hpp"
 #include "NeonInstanceNormalizationWorkload.hpp"
 #include "NeonL2NormalizationFloatWorkload.hpp"
 #include "NeonLstmFloatWorkload.hpp"