diff options
Diffstat (limited to 'src/backends')
24 files changed, 368 insertions, 1 deletions
diff --git a/src/backends/backendsCommon/ILayerSupport.cpp b/src/backends/backendsCommon/ILayerSupport.cpp index 0b78913b99..aa1bb5042d 100644 --- a/src/backends/backendsCommon/ILayerSupport.cpp +++ b/src/backends/backendsCommon/ILayerSupport.cpp @@ -336,4 +336,11 @@ bool ILayerSupport::IsGreaterSupported(const TensorInfo& input0, 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 1dac498c11..a5db088be7 100644 --- a/src/backends/backendsCommon/WorkloadData.cpp +++ b/src/backends/backendsCommon/WorkloadData.cpp @@ -1040,4 +1040,15 @@ void GreaterQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const "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 9142d87bfd..59e3dfbf5c 100644 --- a/src/backends/backendsCommon/WorkloadData.hpp +++ b/src/backends/backendsCommon/WorkloadData.hpp @@ -373,4 +373,9 @@ struct DebugQueueDescriptor : QueueDescriptorWithParameters<DebugDescriptor> 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 3b8a7d8f7f..1dc96a5ec3 100644 --- a/src/backends/backendsCommon/WorkloadFactory.cpp +++ b/src/backends/backendsCommon/WorkloadFactory.cpp @@ -600,6 +600,15 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId, 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 e72987fdcb..aee9f91b56 100644 --- a/src/backends/backendsCommon/WorkloadFactory.hpp +++ b/src/backends/backendsCommon/WorkloadFactory.hpp @@ -156,6 +156,9 @@ public: 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 d4c5fe43c6..d6528bb0ae 100644 --- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp +++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp @@ -378,6 +378,8 @@ DECLARE_LAYER_POLICY_2_PARAM(ResizeBilinear) 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 52150075ce..8e4596b703 100755 --- a/src/backends/backendsCommon/test/LayerTests.cpp +++ b/src/backends/backendsCommon/test/LayerTests.cpp @@ -6734,6 +6734,170 @@ LayerTestResult<uint8_t, 4> ResizeBilinearMagUint8Test( 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 a871594900..98c0806ddf 100644 --- a/src/backends/backendsCommon/test/LayerTests.hpp +++ b/src/backends/backendsCommon/test/LayerTests.hpp @@ -639,6 +639,30 @@ LayerTestResult<float, 4> ResizeBilinearMagTest( 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 0653b41643..6b035851b8 100644 --- a/src/backends/cl/ClWorkloadFactory.cpp +++ b/src/backends/cl/ClWorkloadFactory.cpp @@ -350,4 +350,10 @@ std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDebug(const DebugQueueDescri 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 28447216db..85cbd91e11 100644 --- a/src/backends/cl/ClWorkloadFactory.hpp +++ b/src/backends/cl/ClWorkloadFactory.hpp @@ -147,6 +147,9 @@ public: 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 e635f0cc8d..ff60efb86f 100644 --- a/src/backends/neon/NeonWorkloadFactory.cpp +++ b/src/backends/neon/NeonWorkloadFactory.cpp @@ -318,4 +318,10 @@ std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDebug(const DebugQueueDesc 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 2f7a2db315..9a6308871a 100644 --- a/src/backends/neon/NeonWorkloadFactory.hpp +++ b/src/backends/neon/NeonWorkloadFactory.hpp @@ -148,6 +148,9 @@ public: 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 a64339ec69..56d2e4c659 100644 --- a/src/backends/reference/RefLayerSupport.cpp +++ b/src/backends/reference/RefLayerSupport.cpp @@ -495,6 +495,17 @@ bool RefLayerSupport::IsResizeBilinearSupported(const TensorInfo& input, &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 3941f4bc56..188faa84b6 100644 --- a/src/backends/reference/RefLayerSupport.hpp +++ b/src/backends/reference/RefLayerSupport.hpp @@ -185,6 +185,10 @@ public: 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 eb8807eef6..79293635fb 100644 --- a/src/backends/reference/RefWorkloadFactory.cpp +++ b/src/backends/reference/RefWorkloadFactory.cpp @@ -312,4 +312,10 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateDebug(const DebugQueueDescr 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 f4401ccb6e..f6707f5046 100644 --- a/src/backends/reference/RefWorkloadFactory.hpp +++ b/src/backends/reference/RefWorkloadFactory.hpp @@ -164,6 +164,9 @@ public: 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 b23c752b89..84f15c9c80 100644 --- a/src/backends/reference/backend.mk +++ b/src/backends/reference/backend.mk @@ -57,6 +57,7 @@ BACKEND_SOURCES := \ 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 @@ BACKEND_SOURCES := \ 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 7223f04ea9..50c47aecf9 100644 --- a/src/backends/reference/test/RefLayerTests.cpp +++ b/src/backends/reference/test/RefLayerTests.cpp @@ -380,6 +380,12 @@ ARMNN_AUTO_TEST_CASE(SimpleFloor, SimpleFloorTest) 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 d71e6ea032..d15f77d6e4 100644 --- a/src/backends/reference/workloads/CMakeLists.txt +++ b/src/backends/reference/workloads/CMakeLists.txt @@ -94,6 +94,8 @@ list(APPEND armnnRefBackendWorkloads_sources RefResizeBilinearFloat32Workload.hpp RefResizeBilinearUint8Workload.cpp RefResizeBilinearUint8Workload.hpp + RefRsqrtFloat32Workload.cpp + RefRsqrtFloat32Workload.hpp RefSoftmaxFloat32Workload.cpp RefSoftmaxFloat32Workload.hpp RefSoftmaxUint8Workload.cpp @@ -110,6 +112,8 @@ list(APPEND armnnRefBackendWorkloads_sources 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 0000000000..c08dbf0cab --- /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 0000000000..9d1b4505fe --- /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 ddce68e60c..8beb03fe32 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 0000000000..cee38fc1f1 --- /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 0000000000..35cacede66 --- /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 |