IVGCVSW-3830 Add NEON workload for RSQRT
Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: Ida590b7d28dbee40e496794c42b682d740427cca
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index e0831be..9004391 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -62,6 +62,8 @@
NeonReshapeWorkload.hpp
NeonResizeWorkload.cpp
NeonResizeWorkload.hpp
+ NeonRsqrtWorkload.cpp
+ NeonRsqrtWorkload.hpp
NeonSoftmaxBaseWorkload.cpp
NeonSoftmaxBaseWorkload.hpp
NeonSoftmaxFloatWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonRsqrtWorkload.cpp b/src/backends/neon/workloads/NeonRsqrtWorkload.cpp
new file mode 100644
index 0000000..b629283
--- /dev/null
+++ b/src/backends/neon/workloads/NeonRsqrtWorkload.cpp
@@ -0,0 +1,43 @@
+//
+// Copyright © 2019 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonRsqrtWorkload.hpp"
+
+#include "NeonWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeTensorHandle.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <boost/cast.hpp>
+
+namespace armnn
+{
+
+arm_compute::Status NeonRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output)
+{
+ const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+ const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ return arm_compute::NERsqrtLayer::validate(&aclInput, &aclOutput);
+}
+
+NeonRsqrtWorkload::NeonRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info)
+ : BaseWorkload<RsqrtQueueDescriptor>(descriptor, info)
+{
+ m_Data.ValidateInputsOutputs("NeonRsqrtWorkload", 1, 1);
+
+ 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();
+
+ m_RsqrtLayer.configure(&input, &output);
+}
+
+void NeonRsqrtWorkload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonRsqrtWorkload_Execute");
+ m_RsqrtLayer.run();
+}
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonRsqrtWorkload.hpp b/src/backends/neon/workloads/NeonRsqrtWorkload.hpp
new file mode 100644
index 0000000..6f6031e
--- /dev/null
+++ b/src/backends/neon/workloads/NeonRsqrtWorkload.hpp
@@ -0,0 +1,28 @@
+//
+// 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/NEElementwiseUnaryLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output);
+
+class NeonRsqrtWorkload : public BaseWorkload<RsqrtQueueDescriptor>
+{
+public:
+ NeonRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info);
+ virtual void Execute() const override;
+
+private:
+ mutable arm_compute::NERsqrtLayer m_RsqrtLayer;
+};
+
+} // namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index d807e80..4f30f33 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -33,6 +33,7 @@
#include "NeonQuantizeWorkload.hpp"
#include "NeonReshapeWorkload.hpp"
#include "NeonResizeWorkload.hpp"
+#include "NeonRsqrtWorkload.hpp"
#include "NeonSoftmaxFloatWorkload.hpp"
#include "NeonSoftmaxUint8Workload.hpp"
#include "NeonSpaceToDepthWorkload.hpp"