diff options
author | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-10-12 10:26:30 +0100 |
---|---|---|
committer | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-10-22 13:49:51 +0000 |
commit | 81ec994a3ebc8ad02c4a622846cf64b70e1182bd (patch) | |
tree | 83e961072eb7f75986f953e708fbd322ec90f511 /delegate/src/test/ConvolutionTestHelper.hpp | |
parent | 4692e11e5af29b97748a1585a092df6800a0a831 (diff) | |
download | armnn-81ec994a3ebc8ad02c4a622846cf64b70e1182bd.tar.gz |
IVGCVSW-6166 Add Support for Conv3d to TFLite Delegate
* Conv3d is only correctly supported for external delegates
from TF v2.6, as there was a breaking bug in v2.5.
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: Ib7941307f4c7b0d3dbb7deaa5a90aceb63c1162f
Diffstat (limited to 'delegate/src/test/ConvolutionTestHelper.hpp')
-rw-r--r-- | delegate/src/test/ConvolutionTestHelper.hpp | 258 |
1 files changed, 258 insertions, 0 deletions
diff --git a/delegate/src/test/ConvolutionTestHelper.hpp b/delegate/src/test/ConvolutionTestHelper.hpp index 1b33c1d74d..ce1f951d21 100644 --- a/delegate/src/test/ConvolutionTestHelper.hpp +++ b/delegate/src/test/ConvolutionTestHelper.hpp @@ -5,6 +5,8 @@ #pragma once +#include "TestUtils.hpp" + #include <armnn_delegate.hpp> #include <flatbuffers/flatbuffers.h> @@ -221,6 +223,7 @@ void ConvolutionTest(tflite::BuiltinOperator convolutionOperatorCode, using namespace tflite; std::vector<char> modelBuffer; + modelBuffer = CreateConv2dTfLiteModel(convolutionOperatorCode, tensorType, strideX, @@ -301,6 +304,261 @@ void ConvolutionTest(tflite::BuiltinOperator convolutionOperatorCode, } } +// Conv3d is only correctly supported for external delegates from TF Lite v2.6, as there was a breaking bug in v2.5. +#if defined(ARMNN_POST_TFLITE_2_5) +template <typename T, typename B = float> +std::vector<char> CreateConv3dTfLiteModel(tflite::BuiltinOperator convolutionOperatorCode, + tflite::TensorType tensorType, + std::vector<uint32_t> strides, + std::vector<uint32_t> dilation, + tflite::Padding padding, + tflite::ActivationFunctionType fused_activation_function, + const std::vector<int32_t>& inputTensorShape, + const std::vector<int32_t>& filterTensorShape, + const std::vector<int32_t>& biasTensorShape, + const std::vector<int32_t>& outputTensorShape, + const std::vector<T>& filterData, + const std::vector<B>& biasData, + const std::vector<float> biasScales = {1.0f}, + const std::vector<int64_t> biasOffsets = {0}, + const std::vector<float> filterScales = {1.0f}, + const std::vector<int64_t> filterOffsets = {0}, + float outputQuantScale = 2.0f, + int outputQuantOffset = 0, + float quantScale = 1.0f, + int quantOffset = 0, + int32_t depth_multiplier = 1, + int32_t filterQuantizationDim = 0) +{ + using namespace tflite; + flatbuffers::FlatBufferBuilder flatBufferBuilder; + + std::array<flatbuffers::Offset<tflite::Buffer>, 3> buffers; + buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})); + buffers[1] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(filterData.data()), + sizeof(T) * filterData.size())); + + buffers[2] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(biasData.data()), + sizeof(B) * biasData.size())); + + auto quantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector<float>({ quantScale }), + flatBufferBuilder.CreateVector<int64_t>({ quantOffset })); + auto outputQuantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector<float>({ outputQuantScale }), + flatBufferBuilder.CreateVector<int64_t>({ outputQuantOffset })); + + auto filterQuantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector<float>(filterScales), + flatBufferBuilder.CreateVector<int64_t>(filterOffsets), + tflite::QuantizationDetails_NONE, + 0, + filterQuantizationDim); + + auto biasQuantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector<float>(biasScales), + flatBufferBuilder.CreateVector<int64_t>(biasOffsets)); + + std::array<flatbuffers::Offset<Tensor>, 4> tensors; + tensors[0] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(), + inputTensorShape.size()), + tensorType, + 0, + flatBufferBuilder.CreateString("input"), + quantizationParameters); + tensors[1] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(filterTensorShape.data(), + filterTensorShape.size()), + tensorType, + 1, + flatBufferBuilder.CreateString("filter"), + filterQuantizationParameters); + + auto biasTensorType = ::tflite::TensorType_FLOAT32; + if (tensorType == ::tflite::TensorType_INT8 || tensorType == ::tflite::TensorType_UINT8) + { + biasTensorType = ::tflite::TensorType_INT32; + } + tensors[2] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(biasTensorShape.data(), biasTensorShape.size()), + biasTensorType, + 2, + flatBufferBuilder.CreateString("bias"), + biasQuantizationParameters); + tensors[3] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(), + outputTensorShape.size()), + tensorType, + 0, + flatBufferBuilder.CreateString("output"), + outputQuantizationParameters); + + tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_Conv3DOptions; + flatbuffers::Offset<void> operatorBuiltinOptions = CreateConv3DOptions(flatBufferBuilder, + padding, + strides[2], // Depth + strides[0], // Width + strides[1], // Height + fused_activation_function, + dilation[2], + dilation[0], + dilation[1]).Union(); + + // Create operator + const std::vector<int> operatorInputs{0, 1, 2}; + const std::vector<int> operatorOutputs{3}; + flatbuffers::Offset <Operator> convolutionOperator = + 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, 1, 2}; + const std::vector<int> subgraphOutputs{3}; + 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(&convolutionOperator, 1)); + + flatbuffers::Offset <flatbuffers::String> modelDescription = + flatBufferBuilder.CreateString("ArmnnDelegate: Convolution 3d Operator Model"); + + // If using an operator with a code greater than 127 then the enum value should be passed as the fifth + // parameter rather than the second like in other tests. + flatbuffers::Offset <OperatorCode> operatorCode = + CreateOperatorCode(flatBufferBuilder, 0, 0, 1, tflite::BuiltinOperator_CONV_3D); + + 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 T, typename B = float> +void Convolution3dTest(tflite::BuiltinOperator convolutionOperatorCode, + tflite::TensorType tensorType, + std::vector<uint32_t> strides, + std::vector<uint32_t> dilation, + tflite::Padding padding, + tflite::ActivationFunctionType fused_activation_function, + std::vector<armnn::BackendId>& backends, + std::vector<int32_t>& inputShape, + std::vector<int32_t>& filterShape, + std::vector<int32_t>& outputShape, + std::vector<T>& inputValues, + std::vector<T>& filterValues, + std::vector<T>& expectedOutputValues, + const std::vector<int32_t>& biasShape = {}, + const std::vector<B>& biasValues = {}, + const std::vector<float> biasScales = {1.0f}, + const std::vector<int64_t> biasOffsets = {0}, + const std::vector<float> filterScales = {1.0f}, + const std::vector<int64_t> filterOffsets = {0}, + float outputQuantScale = 2.0f, + int outputQuantOffset = 0, + float quantScale = 1.0f, + int quantOffset = 0, + int32_t depth_multiplier = 1, + int32_t filterQuantizationDim = 3) +{ + using namespace tflite; + + std::vector<char> modelBuffer; + modelBuffer = CreateConv3dTfLiteModel(convolutionOperatorCode, + tensorType, + strides, + dilation, + padding, + fused_activation_function, + inputShape, + filterShape, + biasShape, + outputShape, + filterValues, + biasValues, + biasScales, + biasOffsets, + filterScales, + filterOffsets, + outputQuantScale, + outputQuantOffset, + quantScale, + quantOffset, + depth_multiplier, + filterQuantizationDim); + + 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 + armnnDelegate::FillInput<T>(tfLiteInterpreter, 0, inputValues); + armnnDelegate::FillInput<T>(armnnDelegateInterpreter, 0, inputValues); + + // Run EnqueueWorkload + CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk); + CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk); + + // Compare output data + auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0]; + auto tfLiteDelagateOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateOutputId); + auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0]; + auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId); + + armnnDelegate::CompareData(expectedOutputValues.data(), armnnDelegateOutputData, expectedOutputValues.size(), 1); + armnnDelegate::CompareData(expectedOutputValues.data(), tfLiteDelagateOutputData, expectedOutputValues.size(), 1); + armnnDelegate::CompareData(tfLiteDelagateOutputData, armnnDelegateOutputData, expectedOutputValues.size(), 1); +} +#endif + template <typename T> std::vector<char> CreateTransposeConvTfLiteModel(tflite::TensorType tensorType, uint32_t strideX, |