diff options
author | Teresa Charlin <teresa.charlinreyes@arm.com> | 2023-03-14 12:10:28 +0000 |
---|---|---|
committer | Teresa Charlin <teresa.charlinreyes@arm.com> | 2023-03-28 11:41:55 +0100 |
commit | ad1b3d7518429e2d16a2695d9b0bbf81b6565ac9 (patch) | |
tree | a5b8e1ad68a2437f007338f0b6195ca5ed2bddc3 /delegate/test/QuantizationTestHelper.hpp | |
parent | 9cb3466b677a1048b8abb24661e92c4c83fdda04 (diff) | |
download | armnn-ad1b3d7518429e2d16a2695d9b0bbf81b6565ac9.tar.gz |
IVGCVSW-7555 Restructure Delegate
* New folders created:
* common is for common code where TfLite API is not used
* classic is for existing delegate implementations
* opaque is for new opaque delegate implementation,
* tests is for shared between existing Delegate and Opaque Delegate which have test utils to work which delegate to use.
* Existing delegate is built to libarmnnDelegate.so and opaque delegate is built as libarmnnOpaqueDelegate.so
* Opaque structure is introduced but no API is added yet.
* CmakeList.txt and delegate/CMakeList.txt have been modified and 2 new CmakeList.txt added
* Rename BUILD_ARMNN_TFLITE_DELEGATE as BUILD_CLASSIC_DELEGATE
* Rename BUILD_ARMNN_TFLITE_OPAQUE_DELEGATE as BUILD_OPAQUE_DELEGATE
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ib682b9ad0ac8d8acdc4ec6d9099bb0008a9fe8ed
Diffstat (limited to 'delegate/test/QuantizationTestHelper.hpp')
-rw-r--r-- | delegate/test/QuantizationTestHelper.hpp | 200 |
1 files changed, 200 insertions, 0 deletions
diff --git a/delegate/test/QuantizationTestHelper.hpp b/delegate/test/QuantizationTestHelper.hpp new file mode 100644 index 0000000000..af898f332d --- /dev/null +++ b/delegate/test/QuantizationTestHelper.hpp @@ -0,0 +1,200 @@ +// +// Copyright © 2020, 2023 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include <armnn_delegate.hpp> + +#include <flatbuffers/flatbuffers.h> +#include <tensorflow/lite/interpreter.h> +#include <tensorflow/lite/kernels/register.h> +#include <tensorflow/lite/model.h> +#include <schema_generated.h> +#include <tensorflow/lite/version.h> + +#include <doctest/doctest.h> + +namespace +{ + +std::vector<char> CreateQuantizationTfLiteModel(tflite::BuiltinOperator quantizationOperatorCode, + tflite::TensorType inputTensorType, + tflite::TensorType outputTensorType, + const std::vector <int32_t>& inputTensorShape, + const std::vector <int32_t>& outputTensorShape, + float quantScale = 1.0f, + int quantOffset = 0) +{ + using namespace tflite; + flatbuffers::FlatBufferBuilder flatBufferBuilder; + + std::vector<flatbuffers::Offset<tflite::Buffer>> buffers; + buffers.push_back(CreateBuffer(flatBufferBuilder)); + buffers.push_back(CreateBuffer(flatBufferBuilder)); + buffers.push_back(CreateBuffer(flatBufferBuilder)); + + + auto quantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector<float>({ quantScale }), + flatBufferBuilder.CreateVector<int64_t>({ quantOffset }), + QuantizationDetails_CustomQuantization); + + std::array<flatbuffers::Offset<Tensor>, 2> tensors; + tensors[0] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(), + inputTensorShape.size()), + inputTensorType, + 1, + flatBufferBuilder.CreateString("input"), + quantizationParameters); + tensors[1] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(), + outputTensorShape.size()), + outputTensorType, + 2, + flatBufferBuilder.CreateString("output"), + quantizationParameters); + + // create operator + tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_NONE; + flatbuffers::Offset<void> operatorBuiltinOptions = 0; + switch (quantizationOperatorCode) + { + case BuiltinOperator_QUANTIZE: + { + operatorBuiltinOptionsType = BuiltinOptions_QuantizeOptions; + operatorBuiltinOptions = CreateQuantizeOptions(flatBufferBuilder).Union(); + break; + } + case BuiltinOperator_DEQUANTIZE: + { + operatorBuiltinOptionsType = BuiltinOptions_DequantizeOptions; + operatorBuiltinOptions = CreateDequantizeOptions(flatBufferBuilder).Union(); + break; + } + default: + break; + } + + const std::vector<int32_t> operatorInputs{0}; + const std::vector<int32_t> operatorOutputs{1}; + flatbuffers::Offset <Operator> quantizationOperator = + CreateOperator(flatBufferBuilder, + 0, + flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()), + flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()), + operatorBuiltinOptionsType, + operatorBuiltinOptions); + + const std::vector<int> subgraphInputs{0}; + const std::vector<int> subgraphOutputs{1}; + flatbuffers::Offset <SubGraph> subgraph = + CreateSubGraph(flatBufferBuilder, + flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), + flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()), + flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()), + flatBufferBuilder.CreateVector(&quantizationOperator, 1)); + + flatbuffers::Offset <flatbuffers::String> modelDescription = + flatBufferBuilder.CreateString("ArmnnDelegate: Quantization Operator Model"); + flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, quantizationOperatorCode); + + flatbuffers::Offset <Model> 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<char>(flatBufferBuilder.GetBufferPointer(), + flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); +} + +template <typename InputT, typename OutputT> +void QuantizationTest(tflite::BuiltinOperator quantizeOperatorCode, + tflite::TensorType inputTensorType, + tflite::TensorType outputTensorType, + std::vector<armnn::BackendId>& backends, + std::vector<int32_t>& inputShape, + std::vector<int32_t>& outputShape, + std::vector<InputT>& inputValues, + std::vector<OutputT>& expectedOutputValues, + float quantScale = 1.0f, + int quantOffset = 0) +{ + using namespace tflite; + std::vector<char> modelBuffer = CreateQuantizationTfLiteModel(quantizeOperatorCode, + inputTensorType, + outputTensorType, + inputShape, + outputShape, + quantScale, + quantOffset); + + const Model* tfLiteModel = GetModel(modelBuffer.data()); + + // Create TfLite Interpreters + std::unique_ptr<Interpreter> armnnDelegateInterpreter; + CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) + (&armnnDelegateInterpreter) == kTfLiteOk); + CHECK(armnnDelegateInterpreter != nullptr); + CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk); + + std::unique_ptr<Interpreter> 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<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)> + theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), + armnnDelegate::TfLiteArmnnDelegateDelete); + CHECK(theArmnnDelegate != nullptr); + + // Modify armnnDelegateInterpreter to use armnnDelegate + CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk); + + // Set input data + auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0]; + auto tfLiteDelageInputData = tfLiteInterpreter->typed_tensor<InputT>(tfLiteDelegateInputId); + for (unsigned int i = 0; i < inputValues.size(); ++i) + { + tfLiteDelageInputData[i] = inputValues[i]; + } + + auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0]; + auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<InputT>(armnnDelegateInputId); + for (unsigned int i = 0; i < inputValues.size(); ++i) + { + armnnDelegateInputData[i] = inputValues[i]; + } + + // Run EnqueWorkload + CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk); + CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk); + + // Compare output data + auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0]; + auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor<OutputT>(tfLiteDelegateOutputId); + auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0]; + auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<OutputT>(armnnDelegateOutputId); + + for (size_t i = 0; i < expectedOutputValues.size(); i++) + { + CHECK(expectedOutputValues[i] == armnnDelegateOutputData[i]); + CHECK(tfLiteDelageOutputData[i] == expectedOutputValues[i]); + CHECK(tfLiteDelageOutputData[i] == armnnDelegateOutputData[i]); + } +} + +} // anonymous namespace
\ No newline at end of file |