diff options
Diffstat (limited to 'src/backends/reference')
-rw-r--r-- | src/backends/reference/RefLayerSupport.cpp | 31 | ||||
-rw-r--r-- | src/backends/reference/RefLayerSupport.hpp | 5 | ||||
-rw-r--r-- | src/backends/reference/RefWorkloadFactory.cpp | 10 | ||||
-rw-r--r-- | src/backends/reference/RefWorkloadFactory.hpp | 3 | ||||
-rw-r--r-- | src/backends/reference/backend.mk | 2 | ||||
-rw-r--r-- | src/backends/reference/test/RefCreateWorkloadTests.cpp | 37 | ||||
-rw-r--r-- | src/backends/reference/test/RefLayerTests.cpp | 7 | ||||
-rw-r--r-- | src/backends/reference/workloads/CMakeLists.txt | 4 | ||||
-rw-r--r-- | src/backends/reference/workloads/RefStackWorkload.cpp | 57 | ||||
-rw-r--r-- | src/backends/reference/workloads/RefStackWorkload.hpp | 22 | ||||
-rw-r--r-- | src/backends/reference/workloads/RefWorkloads.hpp | 1 | ||||
-rw-r--r-- | src/backends/reference/workloads/Stack.cpp | 115 | ||||
-rw-r--r-- | src/backends/reference/workloads/Stack.hpp | 20 |
13 files changed, 314 insertions, 0 deletions
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp index 59c14c4490..b9aa126a8c 100644 --- a/src/backends/reference/RefLayerSupport.cpp +++ b/src/backends/reference/RefLayerSupport.cpp @@ -419,6 +419,7 @@ bool RefLayerSupport::IsConcatSupported(const std::vector<const TensorInfo*> inp "Reference concatenation: output type not supported"); for (const TensorInfo* input : inputs) { + BOOST_ASSERT(input != nullptr); supported &= CheckSupportRule(TypeAnyOf(*input, supportedTypes), reasonIfUnsupported, "Reference concatenation: input type not supported"); @@ -1592,6 +1593,36 @@ bool RefLayerSupport::IsSplitterSupported(const TensorInfo& input, return supported; } +bool RefLayerSupport::IsStackSupported(const std::vector<const TensorInfo*>& inputs, + const TensorInfo& output, + const StackDescriptor& descriptor, + Optional<std::string&> reasonIfUnsupported) const +{ + ignore_unused(descriptor); + + bool supported = true; + std::array<DataType,3> supportedTypes = + { + DataType::Float32, + DataType::QuantisedAsymm8, + DataType::QuantisedSymm16 + }; + + supported &= CheckSupportRule(TypeAnyOf(output, supportedTypes), reasonIfUnsupported, + "Reference stack: output type not supported"); + for (const TensorInfo* input : inputs) + { + BOOST_ASSERT(input != nullptr); + supported &= CheckSupportRule(TypeAnyOf(*input, supportedTypes), reasonIfUnsupported, + "Reference stack: input type not supported"); + + supported &= CheckSupportRule(TypesAreEqual(*input, output), reasonIfUnsupported, + "Reference stack: input and output types mismatched."); + } + + return supported; +} + bool RefLayerSupport::IsStridedSliceSupported(const TensorInfo& input, const TensorInfo& output, const StridedSliceDescriptor& descriptor, diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp index c0bf18824e..f8bbeb78bf 100644 --- a/src/backends/reference/RefLayerSupport.hpp +++ b/src/backends/reference/RefLayerSupport.hpp @@ -241,6 +241,11 @@ public: const ViewsDescriptor& descriptor, Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override; + bool IsStackSupported(const std::vector<const TensorInfo*>& inputs, + const TensorInfo& output, + const StackDescriptor& descriptor, + Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override; + bool IsStridedSliceSupported(const TensorInfo& input, const TensorInfo& output, const StridedSliceDescriptor& descriptor, diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp index 183103c40c..925eb6ad90 100644 --- a/src/backends/reference/RefWorkloadFactory.cpp +++ b/src/backends/reference/RefWorkloadFactory.cpp @@ -506,4 +506,14 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateTransposeConvolution2d( return std::make_unique<RefTransposeConvolution2dWorkload>(descriptor, info); } +std::unique_ptr<IWorkload> RefWorkloadFactory::CreateStack(const StackQueueDescriptor& descriptor, + const WorkloadInfo& info) const +{ + if (IsFloat16(info)) + { + return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info); + } + return std::make_unique<RefStackWorkload>(descriptor, info); +} + } // namespace armnn diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp index 9ef15221ef..b012fbc6f6 100644 --- a/src/backends/reference/RefWorkloadFactory.hpp +++ b/src/backends/reference/RefWorkloadFactory.hpp @@ -203,6 +203,9 @@ public: std::unique_ptr<IWorkload> CreateTransposeConvolution2d(const TransposeConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const override; + std::unique_ptr<IWorkload> CreateStack(const StackQueueDescriptor& 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 411ab7e615..6e1360a1db 100644 --- a/src/backends/reference/backend.mk +++ b/src/backends/reference/backend.mk @@ -62,6 +62,7 @@ BACKEND_SOURCES := \ workloads/RefSoftmaxWorkload.cpp \ workloads/RefSpaceToBatchNdWorkload.cpp \ workloads/RefSpaceToDepthWorkload.cpp \ + workloads/RefStackWorkload.cpp \ workloads/RefStridedSliceWorkload.cpp \ workloads/RefSplitterWorkload.cpp \ workloads/RefTransposeConvolution2dWorkload.cpp \ @@ -69,6 +70,7 @@ BACKEND_SOURCES := \ workloads/Rsqrt.cpp \ workloads/SpaceToBatchNd.cpp \ workloads/SpaceToDepth.cpp \ + workloads/Stack.cpp \ workloads/StridedSlice.cpp \ workloads/StringMapping.cpp \ workloads/Softmax.cpp \ diff --git a/src/backends/reference/test/RefCreateWorkloadTests.cpp b/src/backends/reference/test/RefCreateWorkloadTests.cpp index 2fa6cbfd5d..f7999d0ffe 100644 --- a/src/backends/reference/test/RefCreateWorkloadTests.cpp +++ b/src/backends/reference/test/RefCreateWorkloadTests.cpp @@ -990,4 +990,41 @@ BOOST_AUTO_TEST_CASE(CreateSpaceToDepthWorkloadQSymm16) RefCreateSpaceToDepthWorkloadTest<RefSpaceToDepthWorkload, armnn::DataType::QuantisedSymm16>(); } +static void RefCreateStackWorkloadTest(const armnn::TensorShape& inputShape, + const armnn::TensorShape& outputShape, + unsigned int axis, + unsigned int numInputs, + armnn::DataType dataType) +{ + armnn::Graph graph; + RefWorkloadFactory factory; + auto workload = CreateStackWorkloadTest<RefStackWorkload>(factory, + graph, + inputShape, + outputShape, + axis, + numInputs, + dataType); + + // Check output is as expected + auto queueDescriptor = workload->GetData(); + auto outputHandle = boost::polymorphic_downcast<RefTensorHandle*>(queueDescriptor.m_Outputs[0]); + BOOST_TEST((outputHandle->GetTensorInfo() == TensorInfo(outputShape, dataType))); +} + +BOOST_AUTO_TEST_CASE(CreateStackFloat32Workload) +{ + RefCreateStackWorkloadTest({ 3, 4, 5 }, { 3, 4, 2, 5 }, 2, 2, armnn::DataType::Float32); +} + +BOOST_AUTO_TEST_CASE(CreateStackUint8Workload) +{ + RefCreateStackWorkloadTest({ 3, 4, 5 }, { 3, 4, 2, 5 }, 2, 2, armnn::DataType::QuantisedAsymm8); +} + +BOOST_AUTO_TEST_CASE(CreateStackUint16Workload) +{ + RefCreateStackWorkloadTest({ 3, 4, 5 }, { 3, 4, 2, 5 }, 2, 2, armnn::DataType::QuantisedSymm16); +} + BOOST_AUTO_TEST_SUITE_END() diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp index 509dbf7960..4f46d27883 100644 --- a/src/backends/reference/test/RefLayerTests.cpp +++ b/src/backends/reference/test/RefLayerTests.cpp @@ -1175,4 +1175,11 @@ ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dUint8Nchw, UnbiasedStr ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dInt16Nhwc, UnbiasedStridedTransposeConvolution2dInt16NhwcTest) ARMNN_AUTO_TEST_CASE(UnbiasedStridedTransposeConvolution2dInt16Nchw, UnbiasedStridedTransposeConvolution2dInt16NchwTest) +// Stack +ARMNN_AUTO_TEST_CASE(Stack0Axis, Stack0AxisTest<armnn::DataType::Float32>) +ARMNN_AUTO_TEST_CASE(Stack4dOutput1Axis, Stack4dOutput1AxisTest<armnn::DataType::Float32>) +ARMNN_AUTO_TEST_CASE(Stack4dOutput2Axis, Stack4dOutput2AxisTest<armnn::DataType::Float32>) +ARMNN_AUTO_TEST_CASE(Stack4dOutput3Axis, Stack4dOutput3AxisTest<armnn::DataType::Float32>) +ARMNN_AUTO_TEST_CASE(Stack3dOutput1Axis3Input, Stack3dOutput1Axis3InputTest<armnn::DataType::Float32>) + BOOST_AUTO_TEST_SUITE_END() diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt index 696605d662..c9db057be5 100644 --- a/src/backends/reference/workloads/CMakeLists.txt +++ b/src/backends/reference/workloads/CMakeLists.txt @@ -109,6 +109,8 @@ list(APPEND armnnRefBackendWorkloads_sources RefSpaceToDepthWorkload.hpp RefSplitterWorkload.cpp RefSplitterWorkload.hpp + RefStackWorkload.cpp + RefStackWorkload.hpp RefStridedSliceWorkload.cpp RefStridedSliceWorkload.hpp RefTransposeConvolution2dWorkload.cpp @@ -127,6 +129,8 @@ list(APPEND armnnRefBackendWorkloads_sources SpaceToDepth.cpp Splitter.hpp Splitter.cpp + Stack.cpp + Stack.hpp StridedSlice.hpp StridedSlice.cpp StringMapping.cpp diff --git a/src/backends/reference/workloads/RefStackWorkload.cpp b/src/backends/reference/workloads/RefStackWorkload.cpp new file mode 100644 index 0000000000..be36f40633 --- /dev/null +++ b/src/backends/reference/workloads/RefStackWorkload.cpp @@ -0,0 +1,57 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "RefStackWorkload.hpp" + +#include "RefWorkloadUtils.hpp" +#include "Stack.hpp" + +#include <Profiling.hpp> + +namespace armnn +{ + +RefStackWorkload::RefStackWorkload(const StackQueueDescriptor& descriptor, + const WorkloadInfo& info) + : BaseWorkload(descriptor, info) +{} + +void RefStackWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefStackWorkload_Execute"); + + // Can perform a simple concatenation when axis == 0 + if (!m_Data.m_Parameters.m_Axis) + { + float* output = GetOutputTensorData<float>(0, m_Data); + BOOST_ASSERT(output != nullptr); + + unsigned int numInputs = m_Data.m_Parameters.m_NumInputs; + unsigned int inputLength = GetTensorInfo(m_Data.m_Inputs[0]).GetNumElements(); + + for (unsigned int inputIdx=0; inputIdx<numInputs; ++inputIdx) + { + const float* input = GetInputTensorData<float>(inputIdx, m_Data); + for (unsigned int elmt=0; elmt<inputLength; ++elmt) + { + output[(inputIdx * inputLength) + elmt] = input[elmt]; + } + } + return; + } + + std::vector<std::unique_ptr<Decoder<float>>> inputDecoders; + for (unsigned int i=0; i<m_Data.m_Inputs.size(); ++i) + { + inputDecoders.push_back(MakeDecoder<float>(GetTensorInfo(m_Data.m_Inputs[i]), + m_Data.m_Inputs[i]->Map())); + } + std::unique_ptr<Encoder<float>> outputEncoder = MakeEncoder<float>(GetTensorInfo(m_Data.m_Outputs[0]), + m_Data.m_Outputs[0]->Map()); + + Stack(m_Data, inputDecoders, *outputEncoder); +} + +} // namespace armnn diff --git a/src/backends/reference/workloads/RefStackWorkload.hpp b/src/backends/reference/workloads/RefStackWorkload.hpp new file mode 100644 index 0000000000..ceb27d9f60 --- /dev/null +++ b/src/backends/reference/workloads/RefStackWorkload.hpp @@ -0,0 +1,22 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include <backendsCommon/Workload.hpp> +#include <backendsCommon/WorkloadData.hpp> + +namespace armnn +{ + +class RefStackWorkload : public BaseWorkload<StackQueueDescriptor> +{ +public: + explicit RefStackWorkload(const StackQueueDescriptor& descriptor, + const WorkloadInfo& info); + virtual void Execute() const override; +}; + +} // namespace armnn diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp index 4bdf05daa8..e86dccd5bf 100644 --- a/src/backends/reference/workloads/RefWorkloads.hpp +++ b/src/backends/reference/workloads/RefWorkloads.hpp @@ -46,6 +46,7 @@ #include "RefSplitterWorkload.hpp" #include "RefSoftmaxWorkload.hpp" #include "RefSpaceToBatchNdWorkload.hpp" +#include "RefStackWorkload.hpp" #include "RefStridedSliceWorkload.hpp" #include "RefSpaceToDepthWorkload.hpp" #include "RefTransposeConvolution2dWorkload.hpp" diff --git a/src/backends/reference/workloads/Stack.cpp b/src/backends/reference/workloads/Stack.cpp new file mode 100644 index 0000000000..386c8992eb --- /dev/null +++ b/src/backends/reference/workloads/Stack.cpp @@ -0,0 +1,115 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "Stack.hpp" +#include "RefWorkloadUtils.hpp" + +namespace armnn +{ + +void Stack(const StackQueueDescriptor& data, + std::vector<std::unique_ptr<Decoder<float>>>& inputs, + Encoder<float>& output) +{ + const TensorInfo& outputInfo = GetTensorInfo(data.m_Outputs[0]); + const TensorInfo& inputInfo = GetTensorInfo(data.m_Inputs[0]); + + unsigned int outputNumDims = outputInfo.GetNumDimensions(); + unsigned int inputNumDims = inputInfo.GetNumDimensions(); + + const armnn::TensorShape& outputDims = outputInfo.GetShape(); + const armnn::TensorShape& inputDims = inputInfo.GetShape(); + + unsigned int axis = data.m_Parameters.m_Axis; + + // Initialise output data + unsigned int numOutputElements = 1; + for (unsigned int i=0; i<outputNumDims; ++i) + { + numOutputElements *= outputDims[i]; + } + + const unsigned int iNumTensors = static_cast<unsigned int>(data.m_Inputs.size()); + const unsigned int iBatchSize = inputDims[0]; + const unsigned int iChannels = (inputNumDims > 1) ? inputDims[1] : 1; + const unsigned int iHeight = (inputNumDims > 2) ? inputDims[2] : 1; + const unsigned int iWidth = (inputNumDims > 3) ? inputDims[3] : 1; + + const unsigned int oBatchSize = outputDims[1]; + const unsigned int oChannels = (outputNumDims > 2) ? outputDims[2] : 1; + const unsigned int oHeight = (outputNumDims > 3) ? outputDims[3] : 1; + const unsigned int oWidth = (outputNumDims > 4) ? outputDims[4] : 1; + + // Array to store the input coordinates + // iCoordinates[0] = i, iCoordinates[1] = bi, iCoordinates[2] = ci + // iCoordinates[3] = hi, iCoordinates[4] = wi, iCoordinates[5] = 0 + // iCoordinates[5] will be always zero and used for not incrementing + // the output when the input has less than 4 dimensions + std::array<unsigned int, 6> iCoordinates{ 0 }; + + // Array of pointers used to map the output coordinates to the input ones, in accordance with the axis + // This array is initialized with &iCoordinates[5] since this will be always zero + std::array<unsigned int *, 5> oCoordinates = { &iCoordinates[5], + &iCoordinates[5], + &iCoordinates[5], + &iCoordinates[5], + &iCoordinates[5] }; + + // Set the axis coordinate + oCoordinates[axis] = &iCoordinates[0]; + + // Map the output coordinates, accounting for the axis + unsigned int dim_shift = 0; + for(unsigned int dim = 0; dim < inputNumDims; ++dim) + { + if(dim == axis) + { + dim_shift++; + } + oCoordinates[dim + dim_shift] = &iCoordinates[dim + 1]; + } + + // Alias for the input coordinates + unsigned int &i = iCoordinates[0]; + unsigned int &bi = iCoordinates[1]; + unsigned int &ci = iCoordinates[2]; + unsigned int &hi = iCoordinates[3]; + unsigned int &wi = iCoordinates[4]; + + // Alias for the output coordinates + unsigned int &o = *(oCoordinates[0]); + unsigned int &bo = *(oCoordinates[1]); + unsigned int &co = *(oCoordinates[2]); + unsigned int &ho = *(oCoordinates[3]); + unsigned int &wo = *(oCoordinates[4]); + + // Stack tensors + for(; i < iNumTensors; ++(i)) + { + for(bi = 0; bi < iBatchSize; ++(bi)) + { + for(ci = 0; ci < iChannels; ++(ci)) + { + for(hi = 0; hi < iHeight; ++(hi)) + { + for(wi = 0; wi < iWidth; ++(wi)) + { + output[o * oWidth * oHeight * oChannels * oBatchSize + + bo * oWidth * oHeight * oChannels + + co * oWidth * oHeight + + ho * oWidth + + wo]; + + output.Set(inputs[i]->Get()); + + ++(*(inputs[i])); + } + } + } + } + } +} + +} // namespace armnn diff --git a/src/backends/reference/workloads/Stack.hpp b/src/backends/reference/workloads/Stack.hpp new file mode 100644 index 0000000000..cd86d41552 --- /dev/null +++ b/src/backends/reference/workloads/Stack.hpp @@ -0,0 +1,20 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "Encoders.hpp" +#include "Decoders.hpp" + +#include <backendsCommon/WorkloadData.hpp> + +namespace armnn +{ + +void Stack (const StackQueueDescriptor& data, + std::vector<std::unique_ptr<Decoder<float>>>& inputs, + Encoder<float>& output); + +} // namespace armnn |