From f89964ec2f5e66a0c87ca4cc1535f616a7c38afa Mon Sep 17 00:00:00 2001 From: James Ward Date: Mon, 9 Nov 2020 11:57:47 +0000 Subject: IVGCVSW-5385 TfLiteDelegate: Implement the Transpose operator Signed-off-by: James Ward Change-Id: Iea3d7ecccb82d85ec2d2c5cfdcdaf692236a60aa --- delegate/CMakeLists.txt | 2 + delegate/src/Transpose.hpp | 81 +++++++++++++- delegate/src/test/TransposeTest.cpp | 46 ++++++++ delegate/src/test/TransposeTestHelper.hpp | 174 ++++++++++++++++++++++++++++++ 4 files changed, 300 insertions(+), 3 deletions(-) create mode 100644 delegate/src/test/TransposeTest.cpp create mode 100644 delegate/src/test/TransposeTestHelper.hpp diff --git a/delegate/CMakeLists.txt b/delegate/CMakeLists.txt index 0a3015aff1..de6566ac33 100644 --- a/delegate/CMakeLists.txt +++ b/delegate/CMakeLists.txt @@ -108,6 +108,8 @@ if(BUILD_UNIT_TESTS) src/test/QuantizationTestHelper.hpp src/test/ResizeTest.cpp src/test/ResizeTestHelper.hpp + src/test/TransposeTest.cpp + src/test/TransposeTestHelper.hpp src/test/TestUtils.hpp) add_executable(DelegateUnitTests ${armnnDelegate_unittest_sources}) diff --git a/delegate/src/Transpose.hpp b/delegate/src/Transpose.hpp index 2d5823da84..c44c0d2773 100644 --- a/delegate/src/Transpose.hpp +++ b/delegate/src/Transpose.hpp @@ -9,6 +9,7 @@ #include #include #include +#include namespace armnnDelegate { @@ -17,9 +18,83 @@ TfLiteStatus VisitTransposeOperator(DelegateData& delegateData, TfLiteContext* tfLiteContext, TfLiteNode* tfLiteNode, int nodeIndex, - int32_t operatorCode) + int32_t tfliteTransposeOperatorCode) { - return kTfLiteError; -} + TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex)); + TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex)); + + const TfLiteTensor *tfLiteTensors = tfLiteContext->tensors; + const TfLiteTensor& tfLiteInputTensor0 = tfLiteTensors[tfLiteNode->inputs->data[0]]; + if (IsDynamicTensor(tfLiteInputTensor0)) + { + TF_LITE_MAYBE_KERNEL_LOG(tfLiteContext, + "TfLiteArmnnDelegate: Dynamic input tensors are not supported in " + "operator #%d node #%d: ", + tfliteTransposeOperatorCode, nodeIndex); + return kTfLiteError; + } + + const TfLiteTensor& tfLiteInputTensor1 = tfLiteTensors[tfLiteNode->inputs->data[1]]; + if (IsDynamicTensor(tfLiteInputTensor1)) + { + TF_LITE_MAYBE_KERNEL_LOG(tfLiteContext, + "TfLiteArmnnDelegate: Dynamic input tensors are not supported in " + "operator #%d node #%d: ", + tfliteTransposeOperatorCode, nodeIndex); + return kTfLiteError; + } + + const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]]; + if (IsDynamicTensor(tfLiteOutputTensor)) + { + TF_LITE_MAYBE_KERNEL_LOG(tfLiteContext, + "TfLiteArmnnDelegate: Dynamic output tensors are not supported in " + "operator #%d node #%d: ", + tfliteTransposeOperatorCode, nodeIndex); + return kTfLiteError; + } + + const armnn::TensorInfo& inputTensorInfo0 = GetTensorInfoForTfLiteTensor(tfLiteInputTensor0); + const armnn::TensorInfo& inputTensorInfo1 = GetTensorInfoForTfLiteTensor(tfLiteInputTensor1); //permutation tensor + const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor); + + auto* permTensorDataPtr = tflite::GetTensorData(&tfLiteInputTensor1); + auto numEl = tfLiteInputTensor1.dims->data[0]; + + ARMNN_ASSERT( numEl <= armnn::MaxNumOfTensorDimensions); + ARMNN_ASSERT( tfLiteInputTensor1.dims->size == 1); // ensure only single dimension to the permutation tensor + armnn::TransposeDescriptor descriptor(armnn::PermutationVector( + reinterpret_cast (permTensorDataPtr), + (armnn::PermutationVector::SizeType)(numEl))); + + bool isSupported = false; + + auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported) + { + FORWARD_LAYER_SUPPORT_FUNC(__func__, + tfLiteContext, + IsTransposeSupported, + delegateData.m_Backends, + isSupported, + inputTensorInfo0, + outputTensorInfo, + descriptor); + }; + + if (!delegateData.m_Network) + { + validateFunc(outputTensorInfo, isSupported); + return isSupported ? kTfLiteOk : kTfLiteError; + } + + armnn::IConnectableLayer* transposeLayer = delegateData.m_Network->AddTransposeLayer(descriptor); + ARMNN_ASSERT(transposeLayer != nullptr); + ARMNN_ASSERT(transposeLayer->GetNumInputSlots() == 1); // permutation vector given to descriptor object + + armnn::IOutputSlot& outputSlot = transposeLayer->GetOutputSlot(0); + outputSlot.SetTensorInfo(outputTensorInfo); + + return Connect(transposeLayer, tfLiteNode, delegateData); +} } // namespace armnnDelegate diff --git a/delegate/src/test/TransposeTest.cpp b/delegate/src/test/TransposeTest.cpp new file mode 100644 index 0000000000..67751e325a --- /dev/null +++ b/delegate/src/test/TransposeTest.cpp @@ -0,0 +1,46 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "TransposeTestHelper.hpp" + +#include + +#include +#include + +namespace armnnDelegate +{ + +TEST_SUITE ("Transpose_GpuAccTests") +{ + +TEST_CASE ("Transpose_Float32_GpuAcc_Test") +{ + std::vector backends = {armnn::Compute::GpuAcc}; + TransposeFP32Test(backends); +} + +} + +TEST_SUITE ("Transpose_CpuAccTests") +{ + +TEST_CASE ("Transpose_Float32_CpuAcc_Test") +{ + std::vector backends = {armnn::Compute::CpuAcc}; + TransposeFP32Test(backends); +} + +} + +TEST_SUITE ("Transpose_CpuRefTests") +{ +TEST_CASE ("Transpose_Float32_CpuRef_Test") +{ + std::vector backends = { armnn::Compute::CpuRef }; + TransposeFP32Test(backends); +} +} +} // namespace armnnDelegate diff --git a/delegate/src/test/TransposeTestHelper.hpp b/delegate/src/test/TransposeTestHelper.hpp new file mode 100644 index 0000000000..d63a854fbf --- /dev/null +++ b/delegate/src/test/TransposeTestHelper.hpp @@ -0,0 +1,174 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +#include +#include +#include +#include +#include +#include + +#include + +namespace +{ +std::vector CreateTransposeTfLiteModel(tflite::TensorType tensorType, + const std::vector & input0TensorShape, + const std::vector & inputPermVecShape, + const std::vector & outputTensorShape, + const std::vector& inputPermVec) +{ + using namespace tflite; + flatbuffers::FlatBufferBuilder flatBufferBuilder; + std::array, 2> buffers; + buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})); + buffers[1] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast(inputPermVec.data()), + sizeof(int32_t) * inputPermVec.size())); + std::array, 3> tensors; + tensors[0] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(input0TensorShape.data(), + input0TensorShape.size()), + tensorType, 0); + tensors[1] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(inputPermVecShape.data(), + inputPermVecShape.size()), + tflite::TensorType_INT32, 1, + flatBufferBuilder.CreateString("permutation_vector")); + tensors[2] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(outputTensorShape.data(), + outputTensorShape.size()), + tensorType); + const std::vector operatorInputs{ {0, 1} }; + const std::vector operatorOutputs{{2}}; + flatbuffers::Offset transposeOperator = + CreateOperator(flatBufferBuilder, + 0, + flatBufferBuilder.CreateVector(operatorInputs.data(), operatorInputs.size()), + flatBufferBuilder.CreateVector(operatorOutputs.data(), operatorOutputs.size()), + BuiltinOptions_TransposeOptions, + CreateTransposeOptions(flatBufferBuilder).Union()); + const std::vector subgraphInputs{ {0, 1} }; + const std::vector subgraphOutputs{{2}}; + flatbuffers::Offset subgraph = + CreateSubGraph(flatBufferBuilder, + flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), + flatBufferBuilder.CreateVector(subgraphInputs.data(), subgraphInputs.size()), + flatBufferBuilder.CreateVector(subgraphOutputs.data(), subgraphOutputs.size()), + flatBufferBuilder.CreateVector(&transposeOperator, 1)); + flatbuffers::Offset modelDescription = + flatBufferBuilder.CreateString("ArmnnDelegate: Transpose Operator Model"); + flatbuffers::Offset operatorCode = CreateOperatorCode(flatBufferBuilder, + tflite::BuiltinOperator_TRANSPOSE); + flatbuffers::Offset flatbufferModel = + CreateModel(flatBufferBuilder, + TFLITE_SCHEMA_VERSION, + flatBufferBuilder.CreateVector(&operatorCode, 1), + flatBufferBuilder.CreateVector(&subgraph, 1), + modelDescription, + flatBufferBuilder.CreateVector(buffers.data(), buffers.size())); + flatBufferBuilder.Finish(flatbufferModel); + return std::vector(flatBufferBuilder.GetBufferPointer(), + flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); +} + +void TransposeFP32Test(std::vector& backends) +{ + using namespace tflite; + + // set test input data + std::vector input0Shape {4, 2, 3}; + std::vector inputPermVecShape {3}; + std::vector outputShape {2, 3, 4}; + + std::vector input0Values = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, + 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23}; + std::vector inputPermVec = {2, 0, 1}; + std::vector expectedOutputValues = {0, 3, 6, 9, 12, 15, 18, 21, 1, 4, 7, 10, + 13, 16, 19, 22, 2, 5, 8, 11, 14, 17, 20, 23}; + + // create model + std::vector modelBuffer = CreateTransposeTfLiteModel(::tflite::TensorType_FLOAT32, + input0Shape, + inputPermVecShape, + outputShape, + inputPermVec); + + const Model* tfLiteModel = GetModel(modelBuffer.data()); + // Create TfLite Interpreters + std::unique_ptr armnnDelegateInterpreter; + CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) + (&armnnDelegateInterpreter) == kTfLiteOk); + CHECK(armnnDelegateInterpreter != nullptr); + CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk); + + std::unique_ptr tfLiteInterpreter; + CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) + (&tfLiteInterpreter) == kTfLiteOk); + CHECK(tfLiteInterpreter != nullptr); + CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk); + + // Create the ArmNN Delegate + armnnDelegate::DelegateOptions delegateOptions(backends); + std::unique_ptr + theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), + armnnDelegate::TfLiteArmnnDelegateDelete); + CHECK(theArmnnDelegate != nullptr); + // Modify armnnDelegateInterpreter to use armnnDelegate + CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk); + + // Set input data for tflite + auto tfLiteInterpreterInput0Id = tfLiteInterpreter->inputs()[0]; + auto tfLiteInterpreterInput0Data = tfLiteInterpreter->typed_tensor(tfLiteInterpreterInput0Id); + for (unsigned int i = 0; i < input0Values.size(); ++i) + { + tfLiteInterpreterInput0Data[i] = input0Values[i]; + } + + auto tfLiteInterpreterInput1Id = tfLiteInterpreter->inputs()[1]; + auto tfLiteInterpreterInput1Data = tfLiteInterpreter->typed_tensor(tfLiteInterpreterInput1Id); + for (unsigned int i = 0; i < inputPermVec.size(); ++i) + { + tfLiteInterpreterInput1Data[i] = inputPermVec[i]; + } + + //Set input data for armnn delegate + auto armnnDelegateInput0Id = armnnDelegateInterpreter->inputs()[0]; + auto armnnDelegateInput0Data = armnnDelegateInterpreter->typed_tensor(armnnDelegateInput0Id); + for (unsigned int i = 0; i < input0Values.size(); ++i) + { + armnnDelegateInput0Data[i] = input0Values[i]; + } + + auto armnnDelegateInput1Id = armnnDelegateInterpreter->inputs()[1]; + auto armnnDelegateInput1Data = armnnDelegateInterpreter->typed_tensor(armnnDelegateInput1Id); + for (unsigned int i = 0; i < inputPermVec.size(); ++i) + { + armnnDelegateInput1Data[i] = inputPermVec[i]; + } + + // Run EnqueWorkload + CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk); + CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk); + + // Compare output data + auto tfLiteInterpreterOutputId = tfLiteInterpreter->outputs()[0]; + auto tfLiteInterpreterOutputData = tfLiteInterpreter->typed_tensor(tfLiteInterpreterOutputId); + auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0]; + auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor(armnnDelegateOutputId); + for (size_t i = 0; i < expectedOutputValues.size(); ++i) + { + CHECK(expectedOutputValues[i] == armnnDelegateOutputData[i]); + CHECK(tfLiteInterpreterOutputData[i] == expectedOutputValues[i]); + CHECK(tfLiteInterpreterOutputData[i] == armnnDelegateOutputData[i]); + } + + armnnDelegateInterpreter.reset(nullptr); +} +} -- cgit v1.2.1