IVGCVSW-2371 Add Rsqrt Ref implementation
*Added Unit Tests
Change-Id: I6cceb8e6dcda35ce08415f8e5ca86019a64d26e3
diff --git a/src/backends/backendsCommon/ILayerSupport.cpp b/src/backends/backendsCommon/ILayerSupport.cpp
index 0b78913..aa1bb50 100644
--- a/src/backends/backendsCommon/ILayerSupport.cpp
+++ b/src/backends/backendsCommon/ILayerSupport.cpp
@@ -336,4 +336,11 @@
return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
}
+bool ILayerSupport::IsRsqrtSupported(const TensorInfo &input,
+ const TensorInfo &output,
+ Optional<std::string &> reasonIfUnsupported) const
+{
+ return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
} // namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index 1dac498..a5db088 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -1040,4 +1040,15 @@
"second input");
}
+void RsqrtQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+ ValidateSingleInput(workloadInfo, "RsqrtQueueDescriptor");
+ ValidateSingleOutput(workloadInfo, "RsqrtQueueDescriptor");
+ ValidateTensorShapesMatch(workloadInfo.m_InputTensorInfos[0],
+ workloadInfo.m_OutputTensorInfos[0],
+ "RsqrtQueueDescriptor",
+ "input",
+ "output");
+}
+
} //namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 9142d87..59e3dfb 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -373,4 +373,9 @@
void Validate(const WorkloadInfo& workloadInfo) const;
};
+struct RsqrtQueueDescriptor : QueueDescriptor
+{
+ void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
} //namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 3b8a7d8..1dc96a5 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -600,6 +600,15 @@
result = layerSupportObject->IsResizeBilinearSupported(OverrideDataType(input, dataType), reason);
break;
}
+ case LayerType::Rsqrt:
+ {
+ const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+ const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+ result = layerSupportObject->IsRsqrtSupported(OverrideDataType(input, dataType),
+ OverrideDataType(output, dataType),
+ reason);
+ break;
+ }
case LayerType::Softmax:
{
auto cLayer = boost::polymorphic_downcast<const SoftmaxLayer*>(&layer);
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index e72987f..aee9f91 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -156,6 +156,9 @@
virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor,
const WorkloadInfo& info) const = 0;
+
+ virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const = 0;
};
} //namespace armnn
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index d4c5fe4..d6528bb 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -378,6 +378,8 @@
DECLARE_LAYER_POLICY_2_PARAM(Reshape)
+DECLARE_LAYER_POLICY_1_PARAM(Rsqrt)
+
DECLARE_LAYER_POLICY_2_PARAM(Softmax)
DECLARE_LAYER_POLICY_2_PARAM(SpaceToBatchNd)
diff --git a/src/backends/backendsCommon/test/LayerTests.cpp b/src/backends/backendsCommon/test/LayerTests.cpp
index 5215007..8e4596b 100755
--- a/src/backends/backendsCommon/test/LayerTests.cpp
+++ b/src/backends/backendsCommon/test/LayerTests.cpp
@@ -6734,6 +6734,170 @@
return result;
}
+LayerTestResult<float, 2> Rsqrt2dTestCommon(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+ const armnn::TensorInfo inputTensorInfo,
+ const armnn::TensorInfo outputTensorInfo,
+ std::vector<float> inputValues,
+ std::vector<float> expectedOutputValues)
+{
+ auto inputTensor = MakeTensor<float, 2>(inputTensorInfo, std::vector<float>(inputValues));
+
+ LayerTestResult<float, 2> result(outputTensorInfo);
+ result.outputExpected = MakeTensor<float, 2>(outputTensorInfo, std::vector<float>(expectedOutputValues));
+
+ std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
+ std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+ armnn::RsqrtQueueDescriptor descriptor;
+
+ armnn::WorkloadInfo info;
+
+ AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get());
+ AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get());
+
+ std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateRsqrt(descriptor, info);
+
+ inputHandle->Allocate();
+ outputHandle->Allocate();
+
+ CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0]);
+
+ workload->Execute();
+
+ CopyDataFromITensorHandle(&result.output[0][0], outputHandle.get());
+
+ return result;
+}
+LayerTestResult<float, 2> Rsqrt2dTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+ const armnn::TensorShape inputShape{ 2, 2 };
+ const armnn::TensorShape outputShape{ 2, 2 };
+
+ const armnn::TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32);
+ const armnn::TensorInfo outputTensorInfo(outputShape, armnn::DataType::Float32);
+
+ std::vector<float> inputValues
+ {
+ 1.f, 4.f,
+ 16.f, 25.f
+ };
+
+ std::vector<float> expectedOutputValues
+ {
+ 1.f, 0.5f,
+ 0.25f, 0.2f
+ };
+
+ return Rsqrt2dTestCommon(workloadFactory, memoryManager,
+ inputTensorInfo, outputTensorInfo,
+ inputValues, expectedOutputValues);
+}
+
+LayerTestResult<float, 3> Rsqrt3dTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+ const armnn::TensorShape inputShape{ 3, 1, 2 };
+ const armnn::TensorShape outputShape{ 3, 1, 2 };
+
+ const armnn::TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32);
+ const armnn::TensorInfo outputTensorInfo(outputShape, armnn::DataType::Float32);
+
+ std::vector<float> inputValues
+ {
+ 1.f, 4.f, 16.f,
+ 25.f, 64.f, 100.f
+ };
+
+ std::vector<float> expectedOutputValues
+ {
+ 1.f, 0.5f, 0.25f,
+ 0.2f, 0.125f, 0.1f
+ };
+
+ auto inputTensor = MakeTensor<float, 3>(inputTensorInfo, std::vector<float>(inputValues));
+
+ LayerTestResult<float, 3> result(outputTensorInfo);
+ result.outputExpected = MakeTensor<float, 3>(outputTensorInfo, std::vector<float >(expectedOutputValues));
+
+ std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
+ std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+ armnn::RsqrtQueueDescriptor descriptor;
+
+ armnn::WorkloadInfo info;
+
+ AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get());
+ AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get());
+
+ std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateRsqrt(descriptor, info);
+
+ inputHandle->Allocate();
+ outputHandle->Allocate();
+
+ CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0][0]);
+
+ workload->Execute();
+
+ CopyDataFromITensorHandle(&result.output[0][0][0], outputHandle.get());
+
+ return result;
+}
+
+LayerTestResult<float, 2> RsqrtZeroTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+ const armnn::TensorShape inputShape{ 1, 2 };
+ const armnn::TensorShape outputShape{ 1, 2 };
+
+ const armnn::TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32);
+ const armnn::TensorInfo outputTensorInfo(outputShape, armnn::DataType::Float32);
+
+ std::vector<float> inputValues
+ {
+ 0.f, -0.f
+ };
+
+ std::vector<float> expectedOutputValues
+ {
+ INFINITY, -INFINITY
+ };
+
+ return Rsqrt2dTestCommon(workloadFactory, memoryManager,
+ inputTensorInfo, outputTensorInfo,
+ inputValues, expectedOutputValues);
+}
+
+LayerTestResult<float, 2> RsqrtNegativeTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+ const armnn::TensorShape inputShape{ 1, 2 };
+ const armnn::TensorShape outputShape{ 1, 2 };
+
+ const armnn::TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32);
+ const armnn::TensorInfo outputTensorInfo(outputShape, armnn::DataType::Float32);
+
+ std::vector<float> inputValues
+ {
+ -25.f, -16.f
+ };
+
+ std::vector<float> expectedOutputValues
+ {
+ -NAN, -NAN
+ };
+
+ return Rsqrt2dTestCommon(workloadFactory, memoryManager,
+ inputTensorInfo, outputTensorInfo,
+ inputValues, expectedOutputValues);
+}
+
LayerTestResult<float, 4> BatchNormTest(
armnn::IWorkloadFactory& workloadFactory,
const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index a871594..98c0806 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -639,6 +639,30 @@
const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
const armnn::DataLayout dataLayout);
+LayerTestResult<float, 2> Rsqrt2dTestCommon(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+ const armnn::TensorInfo inputTensorInfo,
+ const armnn::TensorInfo outputTensorInfo,
+ std::vector<float> inputValues,
+ std::vector<float> expectedOutputValues);
+
+LayerTestResult<float, 2> Rsqrt2dTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 3> Rsqrt3dTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 2> RsqrtZeroTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 2> RsqrtNegativeTest(
+ armnn::IWorkloadFactory& workloadFactory,
+ const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
LayerTestResult<float, 4> BatchNormTest(
armnn::IWorkloadFactory& workloadFactory,
const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index 0653b41..6b03585 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -350,4 +350,10 @@
return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
}
+std::unique_ptr<IWorkload> ClWorkloadFactory::CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const
+{
+ return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
+}
+
} // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index 2844721..85cbd91 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -147,6 +147,9 @@
virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor,
const WorkloadInfo& info) const override;
+ virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const override;
+
private:
template<typename FloatWorkload, typename Uint8Workload, typename QueueDescriptorType, typename... Args>
static std::unique_ptr<IWorkload> MakeWorkload(const QueueDescriptorType& descriptor,
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index e635f0c..ff60efb 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -318,4 +318,10 @@
return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
}
+std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateRsqrt(const RsqrtQueueDescriptor &descriptor,
+ const WorkloadInfo &info) const
+{
+ return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
+}
+
} // namespace armnn
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index 2f7a2db..9a63088 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -148,6 +148,9 @@
virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor,
const WorkloadInfo& info) const override;
+ virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const override;
+
private:
mutable std::shared_ptr<NeonMemoryManager> m_MemoryManager;
};
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index a64339e..56d2e4c 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -495,6 +495,17 @@
&TrueFunc<>);
}
+bool RefLayerSupport::IsRsqrtSupported(const TensorInfo& input,
+ const TensorInfo& output,
+ Optional<std::string&> reasonIfUnsupported) const
+{
+ ignore_unused(output);
+ return IsSupportedForDataTypeRef(reasonIfUnsupported,
+ input.GetDataType(),
+ &TrueFunc<>,
+ &FalseFuncU8<>);
+}
+
bool RefLayerSupport::IsSoftmaxSupported(const TensorInfo& input,
const TensorInfo& output,
const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp
index 3941f4b..188faa8 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -185,6 +185,10 @@
bool IsResizeBilinearSupported(const TensorInfo& input,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+ bool IsRsqrtSupported(const TensorInfo& input,
+ const TensorInfo& output,
+ Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
bool IsSoftmaxSupported(const TensorInfo& input,
const TensorInfo& output,
const SoftmaxDescriptor& descriptor,
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp
index eb8807e..7929363 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -312,4 +312,10 @@
return MakeWorkload<RefDebugFloat32Workload, RefDebugUint8Workload>(descriptor, info);
}
-} // namespace armnn
+std::unique_ptr<IWorkload> RefWorkloadFactory::CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const
+{
+ return MakeWorkload<RefRsqrtFloat32Workload, NullWorkload>(descriptor, info);
+}
+
+} // namespace armnn
\ No newline at end of file
diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp
index f4401cc..f6707f5 100644
--- a/src/backends/reference/RefWorkloadFactory.hpp
+++ b/src/backends/reference/RefWorkloadFactory.hpp
@@ -164,6 +164,9 @@
virtual std::unique_ptr<IWorkload> CreateDebug(const DebugQueueDescriptor& descriptor,
const WorkloadInfo& info) const override;
+
+ virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
+ const WorkloadInfo& info) const override;
private:
template <typename F32Workload, typename U8Workload, typename QueueDescriptorType>
diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk
index b23c752..84f15c9 100644
--- a/src/backends/reference/backend.mk
+++ b/src/backends/reference/backend.mk
@@ -57,6 +57,7 @@
workloads/RefReshapeUint8Workload.cpp \
workloads/RefResizeBilinearFloat32Workload.cpp \
workloads/RefResizeBilinearUint8Workload.cpp \
+ workloads/RefRsqrtFloat32Workload.cpp \
workloads/RefSoftmaxFloat32Workload.cpp \
workloads/RefSoftmaxUint8Workload.cpp \
workloads/RefSpaceToBatchNdWorkload.cpp \
@@ -64,6 +65,7 @@
workloads/RefSplitterFloat32Workload.cpp \
workloads/RefSplitterUint8Workload.cpp \
workloads/ResizeBilinear.cpp \
+ workloads/Rsqrt.cpp \
workloads/SpaceToBatchNd.cpp \
workloads/StridedSlice.cpp \
workloads/StringMapping.cpp \
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index 7223f04..50c47ae 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -380,6 +380,12 @@
ARMNN_AUTO_TEST_CASE(SimpleReshapeFloat32, SimpleReshapeFloat32Test)
ARMNN_AUTO_TEST_CASE(SimpleReshapeUint8, SimpleReshapeUint8Test)
+// Rsqrt
+ARMNN_AUTO_TEST_CASE(Rsqrt2d, Rsqrt2dTest)
+ARMNN_AUTO_TEST_CASE(Rsqrt3d, Rsqrt3dTest)
+ARMNN_AUTO_TEST_CASE(RsqrtZero, RsqrtZeroTest)
+ARMNN_AUTO_TEST_CASE(RsqrtNegative, RsqrtNegativeTest)
+
// Permute
ARMNN_AUTO_TEST_CASE(SimplePermuteFloat32, SimplePermuteFloat32Test)
ARMNN_AUTO_TEST_CASE(SimplePermuteUint8, SimplePermuteUint8Test)
diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index d71e6ea..d15f77d 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -94,6 +94,8 @@
RefResizeBilinearFloat32Workload.hpp
RefResizeBilinearUint8Workload.cpp
RefResizeBilinearUint8Workload.hpp
+ RefRsqrtFloat32Workload.cpp
+ RefRsqrtFloat32Workload.hpp
RefSoftmaxFloat32Workload.cpp
RefSoftmaxFloat32Workload.hpp
RefSoftmaxUint8Workload.cpp
@@ -110,6 +112,8 @@
RefWorkloadUtils.hpp
ResizeBilinear.cpp
ResizeBilinear.hpp
+ Rsqrt.cpp
+ Rsqrt.hpp
Softmax.cpp
Softmax.hpp
SpaceToBatchNd.hpp
diff --git a/src/backends/reference/workloads/RefRsqrtFloat32Workload.cpp b/src/backends/reference/workloads/RefRsqrtFloat32Workload.cpp
new file mode 100644
index 0000000..c08dbf0
--- /dev/null
+++ b/src/backends/reference/workloads/RefRsqrtFloat32Workload.cpp
@@ -0,0 +1,25 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefRsqrtFloat32Workload.hpp"
+
+#include "RefWorkloadUtils.hpp"
+#include "Rsqrt.hpp"
+
+#include <Profiling.hpp>
+
+namespace armnn
+{
+
+void RefRsqrtFloat32Workload::Execute() const
+{
+ ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefRsqrtFloat32Workload_Execute");
+
+ Rsqrt(GetInputTensorDataFloat(0, m_Data),
+ GetOutputTensorDataFloat(0, m_Data),
+ GetTensorInfo(m_Data.m_Inputs[0]));
+}
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefRsqrtFloat32Workload.hpp b/src/backends/reference/workloads/RefRsqrtFloat32Workload.hpp
new file mode 100644
index 0000000..9d1b450
--- /dev/null
+++ b/src/backends/reference/workloads/RefRsqrtFloat32Workload.hpp
@@ -0,0 +1,21 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
+
+namespace armnn
+{
+
+class RefRsqrtFloat32Workload : public Float32Workload<RsqrtQueueDescriptor>
+{
+public:
+ using Float32Workload<RsqrtQueueDescriptor>::Float32Workload;
+ virtual void Execute() const override;
+};
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index ddce68e..8beb03f 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -59,3 +59,4 @@
#include "RefBatchToSpaceNdUint8Workload.hpp"
#include "RefBatchToSpaceNdFloat32Workload.hpp"
#include "RefDebugWorkload.hpp"
+#include "RefRsqrtFloat32Workload.hpp"
diff --git a/src/backends/reference/workloads/Rsqrt.cpp b/src/backends/reference/workloads/Rsqrt.cpp
new file mode 100644
index 0000000..cee38fc
--- /dev/null
+++ b/src/backends/reference/workloads/Rsqrt.cpp
@@ -0,0 +1,23 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "Rsqrt.hpp"
+
+#include <cmath>
+
+namespace armnn
+{
+
+void Rsqrt(const float* in,
+ float* out,
+ const TensorInfo& tensorInfo)
+{
+ for (size_t i = 0; i < tensorInfo.GetNumElements(); i++)
+ {
+ out[i] = 1.f / sqrtf(in[i]);
+ }
+}
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/reference/workloads/Rsqrt.hpp b/src/backends/reference/workloads/Rsqrt.hpp
new file mode 100644
index 0000000..35caced
--- /dev/null
+++ b/src/backends/reference/workloads/Rsqrt.hpp
@@ -0,0 +1,18 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <armnn/Tensor.hpp>
+#include <armnn/Types.hpp>
+
+namespace armnn
+{
+
+/// Performs the reciprocal squareroot function elementwise
+/// on the inputs to give the outputs.
+void Rsqrt(const float* in,
+ float* out,
+ const TensorInfo& tensorInfo);
+
+} //namespace armnn