diff options
author | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-07-13 19:46:11 +0100 |
---|---|---|
committer | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-08-06 09:25:26 +0000 |
commit | 81beae3a870004795275e9266bc43d845b9f78db (patch) | |
tree | 70af86f3c36c8e330c72770e6f1419ca7b2a4bb8 /src/backends/reference | |
parent | 95e9efc28ce70a8cda93e722f5ce90ebc96bdd95 (diff) | |
download | armnn-81beae3a870004795275e9266bc43d845b9f78db.tar.gz |
IVGCVSW-6119 ConstTensorsAsInput: FullyConnected
* Constant weights and biases are now stored as Constant layers.
* Updated Serializer, Deserializer and unit tests to reflect this.
* Updated TfLiteDelegate, TfLiteParser and OnnxParser.
* Updated Schema with IsConstant and ConstantTensorsAsInputs.
* Updated Ref backend to handle constant weights and
bias as inputs rather than reading from member variables.
* Added dynamic or constant input EndToEnd tests.
!android-nn-driver:5959
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: Ibf3cf437df1100e4b322b0d303c575c6339f9696
Diffstat (limited to 'src/backends/reference')
4 files changed, 42 insertions, 33 deletions
diff --git a/src/backends/reference/RefBackend.hpp b/src/backends/reference/RefBackend.hpp index 441f4ebdf4..2855957e31 100644 --- a/src/backends/reference/RefBackend.hpp +++ b/src/backends/reference/RefBackend.hpp @@ -12,7 +12,8 @@ namespace armnn const BackendCapabilities cpuRefCapabilities("CpuRef", { {"NonConstWeights", true}, - {"AsyncExecution", true} + {"AsyncExecution", true}, + {"ConstantTensorsAsInputs", true} }); const std::set<armnn::BackendCapability> oldCpuRefCapabilities { diff --git a/src/backends/reference/test/RefCreateWorkloadTests.cpp b/src/backends/reference/test/RefCreateWorkloadTests.cpp index 4293ef54f3..fae8d0cdd4 100644 --- a/src/backends/reference/test/RefCreateWorkloadTests.cpp +++ b/src/backends/reference/test/RefCreateWorkloadTests.cpp @@ -486,6 +486,24 @@ TEST_CASE("RefCreateFullyConnectedWithBlobWorkloadTest") TensorInfo({ 3, 7 }, armnn::DataType::Float32, outputQScale)); } +TEST_CASE("CreateFullyConnectedWorkloadWeightsBiasesAsInputsFloat32") +{ + Graph graph; + RefWorkloadFactory factory = GetFactory(); + + auto workload = + CreateFullyConnectedWorkloadWeightsBiasesAsInputsTest<RefFullyConnectedWorkload, + armnn::DataType::Float32>(factory, graph); + + // Checks that outputs and inputs are as we expect them (see definition of CreateFullyConnectedWorkloadTest). + float inputsQScale = 0.0f; + float outputQScale = 0.0f; + CheckInputsOutput(std::move(workload), + TensorInfo({ 3, 1, 4, 5 }, armnn::DataType::Float32, inputsQScale), + TensorInfo({ 7, 20 }, armnn::DataType::Float32, inputsQScale), + TensorInfo({ 3, 7 }, armnn::DataType::Float32, outputQScale)); +} + template <typename FullyConnectedWorkloadType, armnn::DataType DataType> static void RefCreateFullyConnectedWorkloadTest() { diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp index 69a2048078..424df977c8 100644 --- a/src/backends/reference/test/RefEndToEndTests.cpp +++ b/src/backends/reference/test/RefEndToEndTests.cpp @@ -600,11 +600,21 @@ TEST_CASE("RefFillEndToEndTestInt32") FillEndToEnd<armnn::DataType::Signed32>(defaultBackends); } -TEST_CASE("RefFullyConnectedEndToEndTestInt32") +TEST_CASE("RefFullyConnectedEndToEndTestFloat32") { FullyConnectedWithDynamicWeightsEndToEnd<armnn::DataType::Float32>(defaultBackends); } +TEST_CASE("RefFullyConnectedEndToEndTestNonConstantWeightsConstantBiasesFloat32") +{ + FullyConnectedWithDynamicOrConstantInputsEndToEnd<armnn::DataType::Float32>(defaultBackends, true, true); +} + +TEST_CASE("RefFullyConnectedEndToEndTestConstantWeightsNonConstantBiasesFloat32") +{ + FullyConnectedWithDynamicOrConstantInputsEndToEnd<armnn::DataType::Float32>(defaultBackends, true, false); +} + TEST_CASE("RefGatherFloatTest") { GatherEndToEnd<armnn::DataType::Float32>(defaultBackends); diff --git a/src/backends/reference/workloads/RefFullyConnectedWorkload.cpp b/src/backends/reference/workloads/RefFullyConnectedWorkload.cpp index 99e3eab075..5a7951ec48 100644 --- a/src/backends/reference/workloads/RefFullyConnectedWorkload.cpp +++ b/src/backends/reference/workloads/RefFullyConnectedWorkload.cpp @@ -16,20 +16,6 @@ RefFullyConnectedWorkload::RefFullyConnectedWorkload( const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info) { - if (descriptor.m_Parameters.m_ConstantWeights) - { - m_Weight = std::make_unique<ScopedTensorHandle>(*(descriptor.m_Weight)); - const TensorInfo& rWeightInfo = m_Weight->GetTensorInfo(); - m_WeightShape = rWeightInfo.GetShape(); - m_WeightDecoder = MakeDecoder<float>(rWeightInfo, m_Weight->Map(true)); - - if (descriptor.m_Parameters.m_BiasEnabled) - { - m_Bias = std::make_unique<ScopedTensorHandle>(*(descriptor.m_Bias)); - const TensorInfo& biasInfo = m_Bias->GetTensorInfo(); - m_BiasDecoder = MakeDecoder<float>(biasInfo, m_Bias->Map(true)); - } - } } void RefFullyConnectedWorkload::PostAllocationConfigure() @@ -44,18 +30,15 @@ void RefFullyConnectedWorkload::PostAllocationConfigure(std::vector<ITensorHandl ARMNN_ASSERT(inputInfo.GetNumDimensions() > 1); m_InputShape = inputInfo.GetShape(); - if (!m_Data.m_Parameters.m_ConstantWeights) + const TensorInfo& rWeightInfo = GetTensorInfo(inputs[1]); + ARMNN_ASSERT(inputInfo.GetNumDimensions() > 1); + m_WeightShape = rWeightInfo.GetShape(); + m_WeightDecoder = MakeDecoder<float>(rWeightInfo); + + if (m_Data.m_Parameters.m_BiasEnabled) { - const TensorInfo& rWeightInfo = GetTensorInfo(inputs[1]); - ARMNN_ASSERT(inputInfo.GetNumDimensions() > 1); - m_WeightShape = rWeightInfo.GetShape(); - m_WeightDecoder = MakeDecoder<float>(rWeightInfo); - - if (m_Data.m_Parameters.m_BiasEnabled) - { - const TensorInfo& biasInfo = GetTensorInfo(inputs[2]); - m_BiasDecoder = MakeDecoder<float>(biasInfo); - } + const TensorInfo& biasInfo = GetTensorInfo(inputs[2]); + m_BiasDecoder = MakeDecoder<float>(biasInfo); } const TensorInfo& outputInfo = GetTensorInfo(outputs[0]); @@ -87,13 +70,10 @@ void RefFullyConnectedWorkload::Execute(std::vector<ITensorHandle*> inputs, std: std::unique_ptr<Decoder<float>> inputDecoder = MakeDecoder<float>(GetTensorInfo(inputs[0]), inputs[0]->Map()); std::unique_ptr<Encoder<float>> OutputEncoder = MakeEncoder<float>(GetTensorInfo(outputs[0]), outputs[0]->Map()); - if (!m_Data.m_Parameters.m_ConstantWeights) + m_WeightDecoder->Reset(inputs[1]->Map()); + if (m_Data.m_Parameters.m_BiasEnabled) { - m_WeightDecoder->Reset(inputs[1]->Map()); - if (m_Data.m_Parameters.m_BiasEnabled) - { - m_BiasDecoder->Reset(inputs[2]->Map()); - } + m_BiasDecoder->Reset(inputs[2]->Map()); } FullyConnected(m_InputShape, |