diff options
Diffstat (limited to 'libs/tflite/src')
-rw-r--r-- | libs/tflite/src/Diff.cpp | 598 | ||||
-rw-r--r-- | libs/tflite/src/FeatureView.cpp | 70 | ||||
-rw-r--r-- | libs/tflite/src/Quantization.cpp | 22 | ||||
-rw-r--r-- | libs/tflite/src/TensorShapeUtils.cpp | 48 | ||||
-rw-r--r-- | libs/tflite/src/TensorView.test.cpp | 53 | ||||
-rw-r--r-- | libs/tflite/src/ext/kernels/Abs.cpp | 103 | ||||
-rw-r--r-- | libs/tflite/src/ext/kernels/SquaredDifference.cpp | 112 | ||||
-rw-r--r-- | libs/tflite/src/ext/kernels/TensorFlowMax.cpp | 405 | ||||
-rw-r--r-- | libs/tflite/src/ext/kernels/TensorFlowSum.cpp | 400 | ||||
-rw-r--r-- | libs/tflite/src/ext/kernels/register.cpp | 221 | ||||
-rw-r--r-- | libs/tflite/src/ext/nnapi_delegate.cpp | 1209 | ||||
-rw-r--r-- | libs/tflite/src/ext/nnapi_delegate_ex_AddOpsAndParams_lambda.inc | 106 | ||||
-rw-r--r-- | libs/tflite/src/interp/FlatBufferBuilder.cpp | 40 | ||||
-rw-r--r-- | libs/tflite/src/interp/FunctionBuilder.cpp | 34 |
14 files changed, 3421 insertions, 0 deletions
diff --git a/libs/tflite/src/Diff.cpp b/libs/tflite/src/Diff.cpp new file mode 100644 index 000000000..45ef06110 --- /dev/null +++ b/libs/tflite/src/Diff.cpp @@ -0,0 +1,598 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/Diff.h" +#include "tflite/ext/nnapi_delegate.h" + +#include "misc/fp32.h" + +#include "misc/tensor/IndexIterator.h" +#include "misc/tensor/IndexFormatter.h" +#include "misc/tensor/Zipper.h" +#include "misc/tensor/Comparator.h" + +#include "misc/environment.h" + +#include <iostream> +#include <cassert> + +class DiffSummary : public nnfw::misc::tensor::Comparator::Observer +{ +public: + DiffSummary() + : max_abs_diff_index(0), max_abs_diff_expected{0.0f}, max_abs_diff_obtained{0.0f}, + max_abs_diff_value{0.0f}, max_rel_diff_index(0), max_rel_diff_expected{0.0f}, + max_rel_diff_obtained{0.0f}, max_rel_diff_value{0.0f} + { + // DO NOTHING + } + +public: + void notify(const nnfw::misc::tensor::Index &index, float expected, float obtained) override; + +public: + nnfw::misc::tensor::Index max_abs_diff_index; + float max_abs_diff_expected; + float max_abs_diff_obtained; + float max_abs_diff_value; + + nnfw::misc::tensor::Index max_rel_diff_index; + float max_rel_diff_expected; + float max_rel_diff_obtained; + float max_rel_diff_value; +}; + +void DiffSummary::notify(const nnfw::misc::tensor::Index &index, float expected, float obtained) +{ + const auto abs_diff_value = std::fabs(expected - obtained); + + if (max_abs_diff_value < abs_diff_value) + { + max_abs_diff_index = index; + max_abs_diff_value = abs_diff_value; + max_abs_diff_expected = expected; + max_abs_diff_obtained = obtained; + } + + const auto rel_diff_value = nnfw::misc::fp32::relative_diff(expected, obtained); + + if (max_rel_diff_value < rel_diff_value) + { + max_rel_diff_index = index; + max_rel_diff_value = rel_diff_value; + max_rel_diff_expected = expected; + max_rel_diff_obtained = obtained; + } +} + +template <typename T> +bool TfLiteInterpMatchApp::compareSingleTensorView(const nnfw::tflite::TensorView<T> &expected, + const nnfw::tflite::TensorView<T> &obtained, + int id) const +{ + std::vector<nnfw::misc::tensor::Diff<T>> diffs; + assert(expected.shape() == obtained.shape()); + + using nnfw::misc::tensor::zip; + using nnfw::misc::tensor::Index; + + zip(expected.shape(), expected, obtained) + << [&](const Index &index, T expected_value, T obtained_value) { + if (expected_value != obtained_value) + { + diffs.emplace_back(index, expected_value, obtained_value); + } + }; + + // TODO Unify summary generation code + if (diffs.size() == 0) + { + std::cout << " Tensor #" << id << ": MATCHED" << std::endl; + } + else + { + std::cout << " Tensor #" << id << ": UNMATCHED" << std::endl; + std::cout << " " << diffs.size() << " diffs are detected" << std::endl; + } + + if (diffs.size() > 0 && _verbose != 0) + { + std::cout << " ---- Details ---" << std::endl; + for (const auto &diff : diffs) + { + std::cout << " Diff at [" << nnfw::misc::tensor::IndexFormatter(diff.index) << "]" + << std::endl; + std::cout << " expected: " << diff.expected << std::endl; + std::cout << " obtained: " << diff.obtained << std::endl; + } + } + + return diffs.size() == 0; +} + +template <> +bool TfLiteInterpMatchApp::compareSingleTensorView<float>( + const nnfw::tflite::TensorView<float> &expected, + const nnfw::tflite::TensorView<float> &obtained, int id) const +{ + DiffSummary summary; + + assert(expected.shape() == obtained.shape()); + auto diffs = _comparator.compare(expected.shape(), expected, obtained, &summary); + + // TODO Unify summary generation code + if (diffs.size() == 0) + { + std::cout << " Tensor #" << id << ": MATCHED" << std::endl; + } + else + { + std::cout << " Tensor #" << id << ": UNMATCHED" << std::endl; + std::cout << " " << diffs.size() << " diffs are detected" << std::endl; + } + + // Print out max_diff + if (summary.max_abs_diff_value > 0) + { + std::cout << " Max absolute diff at [" + << nnfw::misc::tensor::IndexFormatter(summary.max_abs_diff_index) << "]" << std::endl; + std::cout << " expected: " << summary.max_abs_diff_expected << std::endl; + std::cout << " obtained: " << summary.max_abs_diff_obtained << std::endl; + std::cout << " absolute diff: " << summary.max_abs_diff_value << std::endl; + } + + if (summary.max_rel_diff_value > 0) + { + const auto tolerance_level = summary.max_rel_diff_value / FLT_EPSILON; + + std::cout << " Max relative diff at [" + << nnfw::misc::tensor::IndexFormatter(summary.max_rel_diff_index) << "]" << std::endl; + std::cout << " expected: " << summary.max_rel_diff_expected << std::endl; + std::cout << " obtained: " << summary.max_rel_diff_obtained << std::endl; + std::cout << " relative diff: " << summary.max_rel_diff_value << std::endl; + std::cout << " (tolerance level = " << tolerance_level << ")" << std::endl; + } + + if (diffs.size() > 0) + { + if (_verbose != 0) + { + std::cout << " ---- Details ---" << std::endl; + for (const auto &diff : diffs) + { + const auto absolute_diff = std::fabs(diff.expected - diff.obtained); + const auto relative_diff = nnfw::misc::fp32::relative_diff(diff.expected, diff.obtained); + const auto tolerance_level = relative_diff / FLT_EPSILON; + + std::cout << " Diff at [" << nnfw::misc::tensor::IndexFormatter(diff.index) << "]" + << std::endl; + std::cout << " expected: " << diff.expected << std::endl; + std::cout << " obtained: " << diff.obtained << std::endl; + std::cout << " absolute diff: " << absolute_diff << std::endl; + std::cout << " relative diff: " << relative_diff << std::endl; + std::cout << " (tolerance level = " << tolerance_level << ")" << std::endl; + } + } + + return false; + } + return true; +} + +#include <map> + +bool TfLiteInterpMatchApp::run(::tflite::Interpreter &interp, ::tflite::Interpreter &nnapi) const +{ + assert(interp.outputs() == nnapi.outputs()); + + bool all_matched = true; + + using Comparator = std::function<bool(int id, ::tflite::Interpreter &, ::tflite::Interpreter &)>; + + std::map<TfLiteType, Comparator> comparators; + + comparators[kTfLiteUInt8] = [this](int id, ::tflite::Interpreter &interp, + ::tflite::Interpreter &nnapi) { + const auto expected = nnfw::tflite::TensorView<uint8_t>::make(interp, id); + const auto obtained = nnfw::tflite::TensorView<uint8_t>::make(nnapi, id); + + return compareSingleTensorView(expected, obtained, id); + }; + + comparators[kTfLiteInt32] = [this](int id, ::tflite::Interpreter &interp, + ::tflite::Interpreter &nnapi) { + const auto expected = nnfw::tflite::TensorView<int32_t>::make(interp, id); + const auto obtained = nnfw::tflite::TensorView<int32_t>::make(nnapi, id); + + return compareSingleTensorView(expected, obtained, id); + }; + + comparators[kTfLiteFloat32] = [this](int id, ::tflite::Interpreter &interp, + ::tflite::Interpreter &nnapi) { + const auto expected = nnfw::tflite::TensorView<float>::make(interp, id); + const auto obtained = nnfw::tflite::TensorView<float>::make(nnapi, id); + + return compareSingleTensorView(expected, obtained, id); + }; + + comparators[kTfLiteBool] = [this](int id, ::tflite::Interpreter &interp, + ::tflite::Interpreter &nnapi) { + const auto expected = nnfw::tflite::TensorView<bool>::make(interp, id); + const auto obtained = nnfw::tflite::TensorView<bool>::make(nnapi, id); + + return compareSingleTensorView(expected, obtained, id); + }; + + for (const auto &id : interp.outputs()) + { + assert(interp.tensor(id)->type == nnapi.tensor(id)->type); + + auto it = comparators.find(interp.tensor(id)->type); + + if (it == comparators.end()) + { + throw std::runtime_error{"Not supported output type"}; + } + + const auto &comparator = it->second; + + if (!comparator(id, interp, nnapi)) + { + all_matched = false; + } + } + + return all_matched; +} + +#include "misc/tensor/Object.h" + +using namespace std::placeholders; + +template <> uint8_t RandomGenerator::generate<uint8_t>(void) +{ + // The value of type_range is 255. + float type_range = static_cast<float>(std::numeric_limits<uint8_t>::max()) - + static_cast<float>(std::numeric_limits<uint8_t>::min()); + // Most _dist values range from -5.0 to 5.0. + float min_range = -5.0f; + float max_range = 5.0f; + return static_cast<uint8_t>((_dist(_rand) - min_range) * type_range / (max_range - min_range)); +} + +#include "tflite/TensorLogger.h" +// +// Random Test Runner +// +int RandomTestRunner::run(const nnfw::tflite::Builder &builder) +{ + auto tfl_interp = builder.build(); + auto nnapi = builder.build(); + + tfl_interp->UseNNAPI(false); + + // Allocate Tensors + tfl_interp->AllocateTensors(); + nnapi->AllocateTensors(); + + assert(tfl_interp->inputs() == nnapi->inputs()); + + using ::tflite::Interpreter; + using Initializer = std::function<void(int id, Interpreter *, Interpreter *)>; + + std::map<TfLiteType, Initializer> initializers; + std::map<TfLiteType, Initializer> reseters; + + // Generate singed 32-bit integer (s32) input + initializers[kTfLiteInt32] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteInt32); + assert(nnapi->tensor(id)->type == kTfLiteInt32); + + auto tfl_interp_view = nnfw::tflite::TensorView<int32_t>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<int32_t>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + int32_t value = 0; + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + // TODO Generate random values + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + ++value; + }; + }; + + // Generate singed 32-bit integer (s32) input + reseters[kTfLiteInt32] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteInt32); + assert(nnapi->tensor(id)->type == kTfLiteInt32); + + auto tfl_interp_view = nnfw::tflite::TensorView<int32_t>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<int32_t>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + int32_t value = 0; + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + // TODO Generate random values + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + initializers[kTfLiteUInt8] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteUInt8); + assert(nnapi->tensor(id)->type == kTfLiteUInt8); + + auto tfl_interp_view = nnfw::tflite::TensorView<uint8_t>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<uint8_t>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<uint8_t (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<uint8_t>); + const nnfw::misc::tensor::Object<uint8_t> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + assert(tfl_interp_view.shape() == data.shape()); + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + const auto value = data.at(ind); + + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + reseters[kTfLiteUInt8] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteUInt8); + assert(nnapi->tensor(id)->type == kTfLiteUInt8); + + auto tfl_interp_view = nnfw::tflite::TensorView<uint8_t>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<uint8_t>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<uint8_t (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<uint8_t>); + const nnfw::misc::tensor::Object<uint8_t> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + assert(tfl_interp_view.shape() == data.shape()); + + uint8_t value = 0; + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + initializers[kTfLiteFloat32] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteFloat32); + assert(nnapi->tensor(id)->type == kTfLiteFloat32); + + auto tfl_interp_view = nnfw::tflite::TensorView<float>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<float>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<float (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<float>); + const nnfw::misc::tensor::Object<float> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + + assert(tfl_interp_view.shape() == data.shape()); + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + const auto value = data.at(ind); + + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + reseters[kTfLiteFloat32] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteFloat32); + assert(nnapi->tensor(id)->type == kTfLiteFloat32); + + auto tfl_interp_view = nnfw::tflite::TensorView<float>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<float>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<float (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<float>); + const nnfw::misc::tensor::Object<float> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + + assert(tfl_interp_view.shape() == data.shape()); + + float value = 0; + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + initializers[kTfLiteBool] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteBool); + assert(nnapi->tensor(id)->type == kTfLiteBool); + + auto tfl_interp_view = nnfw::tflite::TensorView<bool>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<bool>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<bool (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<bool>); + const nnfw::misc::tensor::Object<bool> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + + assert(tfl_interp_view.shape() == data.shape()); + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + const auto value = data.at(ind); + + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + reseters[kTfLiteBool] = [&](int id, Interpreter *tfl_interp, Interpreter *nnapi) { + assert(tfl_interp->tensor(id)->type == kTfLiteBool); + assert(nnapi->tensor(id)->type == kTfLiteBool); + + auto tfl_interp_view = nnfw::tflite::TensorView<bool>::make(*tfl_interp, id); + auto nnapi_view = nnfw::tflite::TensorView<bool>::make(*nnapi, id); + + assert(tfl_interp_view.shape() == nnapi_view.shape()); + + auto fp = static_cast<bool (RandomGenerator::*)(const ::nnfw::misc::tensor::Shape &, + const ::nnfw::misc::tensor::Index &)>( + &RandomGenerator::generate<bool>); + const nnfw::misc::tensor::Object<bool> data(tfl_interp_view.shape(), + std::bind(fp, _randgen, _1, _2)); + + assert(tfl_interp_view.shape() == data.shape()); + + bool value = false; + + nnfw::misc::tensor::iterate(tfl_interp_view.shape()) + << [&](const nnfw::misc::tensor::Index &ind) { + tfl_interp_view.at(ind) = value; + nnapi_view.at(ind) = value; + }; + }; + + // Fill IFM with random numbers + for (const auto id : tfl_interp->inputs()) + { + assert(tfl_interp->tensor(id)->type == nnapi->tensor(id)->type); + + auto it = initializers.find(tfl_interp->tensor(id)->type); + + if (it == initializers.end()) + { + throw std::runtime_error{"Not supported input type"}; + } + + it->second(id, tfl_interp.get(), nnapi.get()); + } + + // Fill OFM with 0 + for (const auto id : tfl_interp->outputs()) + { + assert(tfl_interp->tensor(id)->type == nnapi->tensor(id)->type); + + auto it = reseters.find(tfl_interp->tensor(id)->type); + + if (it == reseters.end()) + { + throw std::runtime_error{"Not supported input type"}; + } + + it->second(id, tfl_interp.get(), nnapi.get()); + } + + std::cout << "[NNAPI TEST] Run T/F Lite Interpreter without NNAPI" << std::endl; + tfl_interp->Invoke(); + + std::cout << "[NNAPI TEST] Run T/F Lite Interpreter with NNAPI" << std::endl; + + char *env = getenv("UPSTREAM_DELEGATE"); + + if (env && !std::string(env).compare("1")) + { + nnapi->UseNNAPI(true); + nnapi->Invoke(); + } + else + { + nnfw::tflite::NNAPIDelegate d; + + if (d.BuildGraph(nnapi.get())) + { + throw std::runtime_error{"Failed to BuildGraph"}; + } + + if (d.Invoke(nnapi.get())) + { + throw std::runtime_error{"Failed to BuildGraph"}; + } + } + + // Compare OFM + std::cout << "[NNAPI TEST] Compare the result" << std::endl; + + const auto tolerance = _param.tolerance; + + auto equals = [tolerance](float lhs, float rhs) { + // NOTE Hybrid approach + // TODO Allow users to set tolerance for absolute_epsilon_equal + if (nnfw::misc::fp32::absolute_epsilon_equal(lhs, rhs)) + { + return true; + } + + return nnfw::misc::fp32::epsilon_equal(lhs, rhs, tolerance); + }; + + nnfw::misc::tensor::Comparator comparator(equals); + TfLiteInterpMatchApp app(comparator); + + app.verbose() = _param.verbose; + + bool res = app.run(*tfl_interp, *nnapi); + + if (!res) + { + return 255; + } + + std::cout << "[NNAPI TEST] PASSED" << std::endl; + + if (_param.tensor_logging) + nnfw::tflite::TensorLogger::instance().save(_param.log_path, *tfl_interp); + + return 0; +} + +RandomTestRunner RandomTestRunner::make(int seed) +{ + RandomTestParam param; + + param.verbose = 0; + param.tolerance = 1; + + nnfw::misc::env::IntAccessor("VERBOSE").access(param.verbose); + nnfw::misc::env::IntAccessor("TOLERANCE").access(param.tolerance); + + return RandomTestRunner{seed, param}; +} diff --git a/libs/tflite/src/FeatureView.cpp b/libs/tflite/src/FeatureView.cpp new file mode 100644 index 000000000..fdf5a4b00 --- /dev/null +++ b/libs/tflite/src/FeatureView.cpp @@ -0,0 +1,70 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/FeatureView.h" +#include "tflite/TensorUtils.h" + +#include <cassert> + +namespace nnfw +{ +namespace tflite +{ + +nnfw::misc::feature::Shape getFeatureShape(const TfLiteTensor *tensor) +{ + nnfw::misc::feature::Shape shape{tensor->dims->data[3], tensor->dims->data[1], + tensor->dims->data[2]}; + + return shape; +} + +FeatureView<float>::FeatureView(::tflite::Interpreter &interp, const InputIndex &index) +{ + const auto tensor_index = interp.inputs().at(index.asInt()); + auto tensor_ptr = interp.tensor(tensor_index); + + assert(isFloatTensor(tensor_ptr)); + assert(isFeatureTensor(tensor_ptr)); + + _shape = getFeatureShape(tensor_ptr); + _base = interp.typed_tensor<float>(tensor_index); +} + +FeatureView<float>::FeatureView(::tflite::Interpreter &interp, const OutputIndex &index) +{ + const auto tensor_index = interp.outputs().at(index.asInt()); + auto tensor_ptr = interp.tensor(tensor_index); + + assert(isFloatTensor(tensor_ptr)); + assert(isFeatureTensor(tensor_ptr)); + + _shape = getFeatureShape(tensor_ptr); + _base = interp.typed_tensor<float>(tensor_index); +} + +float FeatureView<float>::at(uint32_t ch, uint32_t row, uint32_t col) const +{ + return *(_base + getElementOffset(ch, row, col)); +} + +float &FeatureView<float>::at(uint32_t ch, uint32_t row, uint32_t col) +{ + return *(_base + getElementOffset(ch, row, col)); +} + +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/Quantization.cpp b/libs/tflite/src/Quantization.cpp new file mode 100644 index 000000000..9c162c342 --- /dev/null +++ b/libs/tflite/src/Quantization.cpp @@ -0,0 +1,22 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/Quantization.h" + +TfLiteQuantizationParams make_default_quantization(void) +{ + return TfLiteQuantizationParams{0.0f, 0}; +} diff --git a/libs/tflite/src/TensorShapeUtils.cpp b/libs/tflite/src/TensorShapeUtils.cpp new file mode 100644 index 000000000..b5d906719 --- /dev/null +++ b/libs/tflite/src/TensorShapeUtils.cpp @@ -0,0 +1,48 @@ +#include "tflite/TensorShapeUtils.h" + +namespace nnfw +{ +namespace tflite +{ + +nnfw::misc::tensor::Shape broadcast(const nnfw::misc::tensor::Shape &lhs_shape, + const nnfw::misc::tensor::Shape &rhs_shape) +{ + const uint32_t lhs_rank = lhs_shape.rank(); + const uint32_t rhs_rank = rhs_shape.rank(); + const uint32_t out_rank = std::max(lhs_rank, rhs_rank); + + // TODO Simplify implementation + std::vector<int32_t> lhs_normalized_dims; + std::vector<int32_t> rhs_normalized_dims; + + for (uint32_t n = 0; n < out_rank - lhs_rank; ++n) + { + lhs_normalized_dims.emplace_back(1); + } + for (uint32_t axis = 0; axis < lhs_rank; ++axis) + { + lhs_normalized_dims.emplace_back(lhs_shape.dim(axis)); + } + + for (uint32_t n = 0; n < out_rank - rhs_rank; ++n) + { + rhs_normalized_dims.emplace_back(1); + } + for (uint32_t axis = 0; axis < rhs_rank; ++axis) + { + rhs_normalized_dims.emplace_back(rhs_shape.dim(axis)); + } + + nnfw::misc::tensor::Shape out_shape(out_rank); + + for (uint32_t axis = 0; axis < out_rank; ++axis) + { + out_shape.dim(axis) = std::max(lhs_normalized_dims.at(axis), rhs_normalized_dims.at(axis)); + } + + return out_shape; +} + +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/TensorView.test.cpp b/libs/tflite/src/TensorView.test.cpp new file mode 100644 index 000000000..c710b3c33 --- /dev/null +++ b/libs/tflite/src/TensorView.test.cpp @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/TensorView.h" + +#include <cassert> + +void int_test(void) +{ + int value[6] = {1, 2, 3, 4, 5, 6}; + + const nnfw::misc::tensor::Shape shape{2, 3}; + const nnfw::tflite::TensorView<int> view{shape, value}; + + assert(view.at(nnfw::misc::tensor::Index{0, 0}) == 1); + assert(view.at(nnfw::misc::tensor::Index{0, 1}) == 2); + assert(view.at(nnfw::misc::tensor::Index{0, 2}) == 3); + assert(view.at(nnfw::misc::tensor::Index{1, 0}) == 4); + assert(view.at(nnfw::misc::tensor::Index{1, 1}) == 5); + assert(view.at(nnfw::misc::tensor::Index{1, 2}) == 6); +} + +int main(int argc, char **argv) +{ + float value[6] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}; + + const nnfw::misc::tensor::Shape shape{2, 3}; + const nnfw::tflite::TensorView<float> view{shape, value}; + + assert(view.at(nnfw::misc::tensor::Index{0, 0}) == 1.0f); + assert(view.at(nnfw::misc::tensor::Index{0, 1}) == 2.0f); + assert(view.at(nnfw::misc::tensor::Index{0, 2}) == 3.0f); + assert(view.at(nnfw::misc::tensor::Index{1, 0}) == 4.0f); + assert(view.at(nnfw::misc::tensor::Index{1, 1}) == 5.0f); + assert(view.at(nnfw::misc::tensor::Index{1, 2}) == 6.0f); + + int_test(); + + return 0; +} diff --git a/libs/tflite/src/ext/kernels/Abs.cpp b/libs/tflite/src/ext/kernels/Abs.cpp new file mode 100644 index 000000000..7e9c2338d --- /dev/null +++ b/libs/tflite/src/ext/kernels/Abs.cpp @@ -0,0 +1,103 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/ext/kernels/Abs.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" + +#include <iostream> +#include <cmath> + +namespace nnfw +{ +namespace tflite +{ +namespace custom +{ +namespace Abs +{ + +void *InitAbs(TfLiteContext *context, const char *buffer, size_t length) { return nullptr; } + +void FreeAbs(TfLiteContext *context, void *buffer) {} + +TfLiteStatus PrepareAbs(TfLiteContext *context, TfLiteNode *node) +{ + TF_LITE_ENSURE_EQ(context, ::tflite::NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, ::tflite::NumOutputs(node), 1); + + const TfLiteTensor *input = ::tflite::GetInput(context, node, 0); + TfLiteTensor *output = ::tflite::GetOutput(context, node, 0); + + TF_LITE_ENSURE_EQ(context, input->type, output->type); + + return context->ResizeTensor(context, output, TfLiteIntArrayCopy(input->dims)); +} + +TfLiteStatus EvalAbs(TfLiteContext *context, TfLiteNode *node) +{ + const TfLiteTensor *input = ::tflite::GetInput(context, node, 0); + TfLiteTensor *output = ::tflite::GetOutput(context, node, 0); + size_t elements = ::tflite::NumElements(input); + switch (input->type) + { + case kTfLiteFloat32: + { + auto *in = input->data.f; + auto *in_end = in + elements; + auto *out = output->data.f; + for (; in < in_end; in++, out++) + *out = std::abs(*in); + return kTfLiteOk; + } + case kTfLiteInt32: + { + auto *in = input->data.i32; + auto *in_end = in + elements; + auto *out = output->data.i32; + for (; in < in_end; in++, out++) + *out = std::abs(*in); + return kTfLiteOk; + } + case kTfLiteInt64: + { + auto *in = input->data.i64; + auto *in_end = in + elements; + auto *out = output->data.i64; + for (; in < in_end; in++, out++) + *out = std::abs(*in); + return kTfLiteOk; + } + case kTfLiteUInt8: + { + auto *in = input->data.uint8; + auto *in_end = in + elements; + auto *out = output->data.uint8; + for (; in < in_end; in++, out++) + *out = std::abs(*in); + return kTfLiteOk; + } + default: + { + context->ReportError(context, "Input type %d is not supported", input->type); + return kTfLiteError; + } + } +} + +} // namespace Abs +} // namespace custom +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/ext/kernels/SquaredDifference.cpp b/libs/tflite/src/ext/kernels/SquaredDifference.cpp new file mode 100644 index 000000000..8ac2b1de0 --- /dev/null +++ b/libs/tflite/src/ext/kernels/SquaredDifference.cpp @@ -0,0 +1,112 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/ext/kernels/SquaredDifference.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" + +#include <iostream> + +namespace nnfw +{ +namespace tflite +{ +namespace custom +{ +namespace SquaredDifference +{ + +void *InitSquaredDifference(TfLiteContext *context, const char *buffer, size_t length) +{ + return nullptr; +} + +void FreeSquaredDifference(TfLiteContext *context, void *buffer) {} + +TfLiteStatus PrepareSquaredDifference(TfLiteContext *context, TfLiteNode *node) +{ + TF_LITE_ENSURE_EQ(context, ::tflite::NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, ::tflite::NumOutputs(node), 1); + + const TfLiteTensor *input1 = ::tflite::GetInput(context, node, 0); + const TfLiteTensor *input2 = ::tflite::GetInput(context, node, 1); + TfLiteTensor *output = ::tflite::GetOutput(context, node, 0); + + TF_LITE_ENSURE_EQ(context, input1->type, input2->type); + TF_LITE_ENSURE_EQ(context, input1->type, output->type); + + return context->ResizeTensor(context, output, TfLiteIntArrayCopy(input1->dims)); +} + +TfLiteStatus EvalSquaredDifference(TfLiteContext *context, TfLiteNode *node) +{ + + const TfLiteTensor *input1 = ::tflite::GetInput(context, node, 0); + const TfLiteTensor *input2 = ::tflite::GetInput(context, node, 1); + + TfLiteTensor *output = ::tflite::GetOutput(context, node, 0); + + size_t elements = ::tflite::NumElements(input1); + + switch (input1->type) + { + case kTfLiteFloat32: + { + const float *in1 = input1->data.f; + const float *in2 = input2->data.f; + const float *in_end1 = in1 + elements; + float *out = output->data.f; + + for (; in1 < in_end1; in1++, in2++, out++) + *out = ((*in1 - *in2) * (*in1 - *in2)); + + return kTfLiteOk; + } + case kTfLiteInt32: + { + const int *in1 = input1->data.i32; + const int *in2 = input2->data.i32; + const int *in_end1 = in1 + elements; + int *out = output->data.i32; + + for (; in1 < in_end1; in1++, in2++, out++) + *out = ((*in1 - *in2) * (*in1 - *in2)); + + return kTfLiteOk; + } + case kTfLiteInt64: + { + const int64_t *in1 = input1->data.i64; + const int64_t *in2 = input1->data.i64; + const int64_t *in_end1 = in1 + elements; + int64_t *out = output->data.i64; + + for (; in1 < in_end1; in1++, in2++, out++) + *out = ((*in1 - *in2) * (*in1 - *in2)); + + return kTfLiteOk; + } + default: + { + context->ReportError(context, "InputType is %d Unsupported", input1->type); + return kTfLiteError; + } + } +} + +} // namespace SquaredDifference +} // namespace custom +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/ext/kernels/TensorFlowMax.cpp b/libs/tflite/src/ext/kernels/TensorFlowMax.cpp new file mode 100644 index 000000000..d72ad242c --- /dev/null +++ b/libs/tflite/src/ext/kernels/TensorFlowMax.cpp @@ -0,0 +1,405 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/ext/kernels/TensorFlowMax.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" + +#include <iostream> + +namespace nnfw +{ +namespace tflite +{ +namespace custom +{ +namespace TensorFlowMax +{ + +struct TensorFlowMaxOp +{ + TensorFlowMaxOp(TfLiteContext *context, TfLiteNode *node) + { + input = ::tflite::GetInput(context, node, 0); + axis = ::tflite::GetInput(context, node, 1); + output = ::tflite::GetOutput(context, node, 0); + } + const TfLiteTensor *input; + const TfLiteTensor *axis; + TfLiteTensor *output; +}; + +void *InitTensorFlowMax(TfLiteContext *context, const char *buffer, size_t length) +{ + // Creates two temp tensors to store index and axis for internal + // implementation only. + auto *scratch_tensor_index = new int; + context->AddTensors(context, 2, scratch_tensor_index); + return scratch_tensor_index; +} + +void FreeTensorFlowMax(TfLiteContext *context, void *buffer) +{ + delete static_cast<TensorFlowMaxOp *>(buffer); +} + +// Resizes the temp tensor that stores resolved axis. +TfLiteStatus ResizeTempAxis(TfLiteContext *context, TensorFlowMaxOp *op_context, + TfLiteTensor *resolved_axis) +{ + TfLiteIntArray *axis_size = TfLiteIntArrayCreate(1); + axis_size->data[0] = static_cast<int>(::tflite::NumElements(op_context->axis)); + return context->ResizeTensor(context, resolved_axis, axis_size); +} + +// Resizes output array based on the input size and resolved axis. +TfLiteStatus ResizeOutputTensor(TfLiteContext *context, TensorFlowMaxOp *op_context) +{ + size_t num_axis = ::tflite::NumElements(op_context->axis); + TfLiteIntArray *input_dims = op_context->input->dims; + int input_num_dims = ::tflite::NumDimensions(op_context->input); + const int *axis = op_context->axis->data.i32; + + { + // Calculates size of reducing axis. + int num_reduce_axis = num_axis; + for (int i = 0; i < num_axis; ++i) + { + int current = axis[i]; + if (current < 0) + { + current += input_num_dims; + } + TF_LITE_ENSURE(context, current >= 0 && current < input_num_dims); + for (int j = 0; j < i; ++j) + { + int previous = axis[j]; + if (previous < 0) + { + previous += input_num_dims; + } + if (current == previous) + { + --num_reduce_axis; + break; + } + } + } + // Determines output dimensions. + int output_num_dims = ::tflite::NumDimensions(op_context->output); + TF_LITE_ENSURE(context, (input_num_dims == output_num_dims) || + (input_num_dims - num_reduce_axis == output_num_dims)); + + if (input_num_dims == output_num_dims) + { + TfLiteIntArray *output_dims = TfLiteIntArrayCopy(input_dims); + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + int current = axis[axis_idx]; + output_dims->data[current] = 1; + } + return context->ResizeTensor(context, op_context->output, output_dims); + } + else + { + TfLiteIntArray *output_dims = TfLiteIntArrayCreate(output_num_dims); + int num_skip_axis = 0; + for (int idx = 0; idx < input_num_dims; ++idx) + { + bool is_axis = false; + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (axis[axis_idx] == idx || axis[axis_idx] + input_num_dims == idx) + { + ++num_skip_axis; + is_axis = true; + break; + } + } + if (!is_axis) + { + output_dims->data[idx - num_skip_axis] = input_dims->data[idx]; + } + } + return context->ResizeTensor(context, op_context->output, output_dims); + } + } +} + +// Initializes temp tensors to store index and resolved axis. +TfLiteStatus InitializeTemporaries(TfLiteContext *context, TfLiteNode *node, + TensorFlowMaxOp *op_context) +{ + // Creates a temp index to iterate through input data. + int *scratch_tensor_index = reinterpret_cast<int *>(node->user_data); + TfLiteIntArrayFree(node->temporaries); + node->temporaries = TfLiteIntArrayCreate(2); + node->temporaries->data[0] = *scratch_tensor_index; + TfLiteTensor *scratch_tensor = &context->tensors[node->temporaries->data[0]]; + scratch_tensor->type = kTfLiteInt32; + scratch_tensor->allocation_type = kTfLiteArenaRw; + TfLiteIntArray *index_size = TfLiteIntArrayCreate(1); + index_size->data[0] = ::tflite::NumDimensions(op_context->input); + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_tensor, index_size)); + + // Creates a temp tensor to store resolved axis given input data. + node->temporaries->data[1] = *scratch_tensor_index + 1; + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + resolved_axis->type = kTfLiteInt32; + return kTfLiteOk; +} + +TfLiteStatus PrepareTensorFlowMax(TfLiteContext *context, TfLiteNode *node) +{ + TF_LITE_ENSURE_EQ(context, ::tflite::NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, ::tflite::NumOutputs(node), 1); + + TensorFlowMaxOp op_context(context, node); + TF_LITE_ENSURE_OK(context, InitializeTemporaries(context, node, &op_context)); + + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + // Leaves work to Eval if axis is not constant; else resizes output. + if (!::tflite::IsConstantTensor(op_context.axis)) + { + ::tflite::SetTensorToDynamic(op_context.output); + ::tflite::SetTensorToDynamic(resolved_axis); + return kTfLiteOk; + } + resolved_axis->allocation_type = kTfLiteArenaRw; + TF_LITE_ENSURE_OK(context, ResizeTempAxis(context, &op_context, resolved_axis)); + return ResizeOutputTensor(context, &op_context); +} + +// Gets offset of index if expanded on axis. When expanded, the flattened offset +// will not change, if the output index changes on the given axis. For example, +// if you have a 2D tensor and you are expanding to 3D on axis 0, +// then index (0, 1, 2) and index (1, 1, 2) will map from the same flattened +// offset. +inline size_t ExpandedInputOffset(const int num_dims, const int *dims, const int *index, + const int num_axis, const int *axis) +{ + size_t offset = 0; + int out_idx = 0; + for (int in_idx = 0; in_idx < num_dims; ++in_idx) + { + // if we need to expand this axis + bool is_axis = false; + if (axis != nullptr) + { + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (in_idx == axis[axis_idx]) + { + is_axis = true; + break; + } + } + } + if (!is_axis) + { + offset = offset * static_cast<size_t>(dims[in_idx]) + static_cast<size_t>(index[out_idx]); + out_idx++; + } + else + { + offset = offset * static_cast<size_t>(dims[in_idx]); + } + } + return offset; +} + +// Gets offset of index if reducing on axis. When reducing, the flattened offset +// will not change, if the input index changes on the given axis. For example, +// if you have a 3D tensor and you are reducing to 2D by eliminating axis 0, +// then index (0, 1, 2) and index (1, 1, 2) will map to the same flattened +// offset. +// TODO(kanlig): uses Dims to represent dimensions. +inline size_t ReducedOutputOffset(const int num_dims, const int *dims, const int *index, + const int num_axis, const int *axis) +{ + size_t offset = 0; + for (int idx = 0; idx < num_dims; ++idx) + { + // if we need to skip this axis + bool is_axis = false; + if (axis != nullptr) + { + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (idx == axis[axis_idx]) + { + is_axis = true; + break; + } + } + } + if (!is_axis) + { + offset = offset * static_cast<size_t>(dims[idx]) + static_cast<size_t>(index[idx]); + } + } + return offset; +} + +// Gets next index to iterate through a multidimensional array. +inline bool NextIndex(TfLiteContext *context, const int num_dims, const int *dims, int *current) +{ + int carry = 1; + for (int idx = num_dims - 1; idx >= 0; --idx) + { + int current_val = current[idx] + carry; + TF_LITE_ENSURE(context, (dims[idx] >= current_val)); + if (dims[idx] == current_val) + { + current[idx] = 0; + } + else + { + current[idx] = current_val; + carry = 0; + break; + } + } + return (carry == 0); +} + +template <typename T> +inline TfLiteStatus +CustomMax(TfLiteContext *context, T *input_data, const int *input_dims, const int input_num_dims, + T *output_data, const int *output_dims, const int output_num_dims, const int *axis, + const int num_axis_dimensions, bool keep_dims, int *temp_index, int *resolved_axis) +{ + // resolves axis. + int num_resolved_axis = 0; + for (int idx = 0; idx < num_axis_dimensions; ++idx) + { + int current = axis[idx]; + TF_LITE_ENSURE(context, (current < input_num_dims && current + input_num_dims >= 0)); + if (current < 0) + { + current += input_num_dims; + } + bool is_dup = false; + for (int j = 0; j < num_resolved_axis; ++j) + { + if (resolved_axis[j] == current) + { + is_dup = true; + break; + } + } + if (!is_dup) + { + resolved_axis[num_resolved_axis++] = current; + } + } + + TF_LITE_ENSURE(context, (input_num_dims > 0)); + TF_LITE_ENSURE(context, (input_dims != nullptr)); + TF_LITE_ENSURE(context, (temp_index != nullptr)); + + // resets output data. + for (int idx = 0; idx < output_num_dims; ++idx) + { + temp_index[idx] = 0; + } + for (bool has_next = true; has_next; + has_next = NextIndex(context, output_num_dims, output_dims, temp_index)) + { + size_t output_offset = + ReducedOutputOffset(output_num_dims, output_dims, temp_index, 0, nullptr); + size_t input_offset = ExpandedInputOffset(input_num_dims, input_dims, temp_index, + num_resolved_axis, resolved_axis); + output_data[output_offset] = input_data[input_offset]; + } + + // resets temp index. + for (int idx = 0; idx < input_num_dims; ++idx) + { + temp_index[idx] = 0; + } + + // iterates through input_data. + for (bool has_next = true; has_next; + has_next = NextIndex(context, input_num_dims, input_dims, temp_index)) + { + size_t input_offset = ReducedOutputOffset(input_num_dims, input_dims, temp_index, 0, nullptr); + size_t output_offset = ReducedOutputOffset(input_num_dims, input_dims, temp_index, + num_resolved_axis, resolved_axis); + if (output_data[output_offset] < input_data[input_offset]) + { + output_data[output_offset] = input_data[input_offset]; + } + } + + return kTfLiteOk; +} + +TfLiteStatus EvalTensorFlowMax(TfLiteContext *context, TfLiteNode *node) +{ + + TensorFlowMaxOp op_context(context, node); + int num_axis = static_cast<int>(::tflite::NumElements(op_context.axis)); + TfLiteTensor *temp_index = &context->tensors[node->temporaries->data[0]]; + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + // Resize the output tensor if the output tensor is dynamic. + if (::tflite::IsDynamicTensor(op_context.output)) + { + TF_LITE_ENSURE_OK(context, ResizeTempAxis(context, &op_context, resolved_axis)); + TF_LITE_ENSURE_OK(context, ResizeOutputTensor(context, &op_context)); + } + + TfLiteStatus returnStatus = kTfLiteOk; + switch (op_context.input->type) + { + case kTfLiteFloat32: + returnStatus = CustomMax<float>( + context, op_context.input->data.f, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.f, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, num_axis, false, + temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteInt32: + returnStatus = CustomMax<int>(context, op_context.input->data.i32, + op_context.input->dims->data, op_context.input->dims->size, + op_context.output->data.i32, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, + num_axis, false, temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteUInt8: + returnStatus = CustomMax<uint8_t>( + context, op_context.input->data.uint8, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.uint8, + op_context.output->dims->data, op_context.output->dims->size, op_context.axis->data.i32, + num_axis, false, temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteInt64: + returnStatus = CustomMax<int64_t>( + context, op_context.input->data.i64, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.i64, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, num_axis, false, + temp_index->data.i32, resolved_axis->data.i32); + break; + default: + returnStatus = kTfLiteError; + } + + return returnStatus; +} + +} // namespace TensorFlowMax +} // namespace custom +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/ext/kernels/TensorFlowSum.cpp b/libs/tflite/src/ext/kernels/TensorFlowSum.cpp new file mode 100644 index 000000000..cbf97970c --- /dev/null +++ b/libs/tflite/src/ext/kernels/TensorFlowSum.cpp @@ -0,0 +1,400 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/ext/kernels/TensorFlowSum.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" + +#include <iostream> + +namespace nnfw +{ +namespace tflite +{ +namespace custom +{ +namespace TensorFlowSum +{ + +struct TensorFlowSumOp +{ + TensorFlowSumOp(TfLiteContext *context, TfLiteNode *node) + { + input = ::tflite::GetInput(context, node, 0); + axis = ::tflite::GetInput(context, node, 1); + output = ::tflite::GetOutput(context, node, 0); + } + const TfLiteTensor *input; + const TfLiteTensor *axis; + TfLiteTensor *output; +}; + +void *InitTensorFlowSum(TfLiteContext *context, const char *buffer, size_t length) +{ + // Creates two temp tensors to store index and axis for internal + // implementation only. + auto *scratch_tensor_index = new int; + context->AddTensors(context, 2, scratch_tensor_index); + return scratch_tensor_index; +} + +void FreeTensorFlowSum(TfLiteContext *context, void *buffer) +{ + delete static_cast<TensorFlowSumOp *>(buffer); +} + +// Resizes the temp tensor that stores resolved axis. +TfLiteStatus ResizeTempAxis(TfLiteContext *context, TensorFlowSumOp *op_context, + TfLiteTensor *resolved_axis) +{ + TfLiteIntArray *axis_size = TfLiteIntArrayCreate(1); + axis_size->data[0] = static_cast<int>(::tflite::NumElements(op_context->axis)); + return context->ResizeTensor(context, resolved_axis, axis_size); +} + +// Resizes output array based on the input size and resolved axis. +TfLiteStatus ResizeOutputTensor(TfLiteContext *context, TensorFlowSumOp *op_context) +{ + size_t num_axis = ::tflite::NumElements(op_context->axis); + TfLiteIntArray *input_dims = op_context->input->dims; + int input_num_dims = ::tflite::NumDimensions(op_context->input); + const int *axis = op_context->axis->data.i32; + + { + // Calculates size of reducing axis. + int num_reduce_axis = num_axis; + for (int i = 0; i < num_axis; ++i) + { + int current = axis[i]; + if (current < 0) + { + current += input_num_dims; + } + TF_LITE_ENSURE(context, current >= 0 && current < input_num_dims); + for (int j = 0; j < i; ++j) + { + int previous = axis[j]; + if (previous < 0) + { + previous += input_num_dims; + } + if (current == previous) + { + --num_reduce_axis; + break; + } + } + } + // Determines output dimensions. + int output_num_dims = ::tflite::NumDimensions(op_context->output); + TF_LITE_ENSURE(context, (input_num_dims == output_num_dims) || + (input_num_dims - num_reduce_axis == output_num_dims)); + + if (input_num_dims == output_num_dims) + { + TfLiteIntArray *output_dims = TfLiteIntArrayCopy(input_dims); + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + int current = axis[axis_idx]; + output_dims->data[current] = 1; + } + return context->ResizeTensor(context, op_context->output, output_dims); + } + else + { + TfLiteIntArray *output_dims = TfLiteIntArrayCreate(output_num_dims); + int num_skip_axis = 0; + for (int idx = 0; idx < input_num_dims; ++idx) + { + bool is_axis = false; + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (axis[axis_idx] == idx || axis[axis_idx] + input_num_dims == idx) + { + ++num_skip_axis; + is_axis = true; + break; + } + } + if (!is_axis) + { + output_dims->data[idx - num_skip_axis] = input_dims->data[idx]; + } + } + return context->ResizeTensor(context, op_context->output, output_dims); + } + } +} + +// Initializes temp tensors to store index and resolved axis. +TfLiteStatus InitializeTemporaries(TfLiteContext *context, TfLiteNode *node, + TensorFlowSumOp *op_context) +{ + // Creates a temp index to iterate through input data. + int *scratch_tensor_index = reinterpret_cast<int *>(node->user_data); + TfLiteIntArrayFree(node->temporaries); + node->temporaries = TfLiteIntArrayCreate(2); + node->temporaries->data[0] = *scratch_tensor_index; + TfLiteTensor *scratch_tensor = &context->tensors[node->temporaries->data[0]]; + scratch_tensor->type = kTfLiteInt32; + scratch_tensor->allocation_type = kTfLiteArenaRw; + TfLiteIntArray *index_size = TfLiteIntArrayCreate(1); + index_size->data[0] = ::tflite::NumDimensions(op_context->input); + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_tensor, index_size)); + + // Creates a temp tensor to store resolved axis given input data. + node->temporaries->data[1] = *scratch_tensor_index + 1; + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + resolved_axis->type = kTfLiteInt32; + return kTfLiteOk; +} + +TfLiteStatus PrepareTensorFlowSum(TfLiteContext *context, TfLiteNode *node) +{ + TF_LITE_ENSURE_EQ(context, ::tflite::NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, ::tflite::NumOutputs(node), 1); + + TensorFlowSumOp op_context(context, node); + TF_LITE_ENSURE_OK(context, InitializeTemporaries(context, node, &op_context)); + + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + // Leaves work to Eval if axis is not constant; else resizes output. + if (!::tflite::IsConstantTensor(op_context.axis)) + { + ::tflite::SetTensorToDynamic(op_context.output); + ::tflite::SetTensorToDynamic(resolved_axis); + return kTfLiteOk; + } + resolved_axis->allocation_type = kTfLiteArenaRw; + TF_LITE_ENSURE_OK(context, ResizeTempAxis(context, &op_context, resolved_axis)); + return ResizeOutputTensor(context, &op_context); +} + +// Gets offset of index if expanded on axis. When expanded, the flattened offset +// will not change, if the output index changes on the given axis. For example, +// if you have a 2D tensor and you are expanding to 3D on axis 0, +// then index (0, 1, 2) and index (1, 1, 2) will map from the same flattened +// offset. +inline size_t ExpandedInputOffset(const int num_dims, const int *dims, const int *index, + const int num_axis, const int *axis) +{ + size_t offset = 0; + int out_idx = 0; + for (int in_idx = 0; in_idx < num_dims; ++in_idx) + { + // if we need to expand this axis + bool is_axis = false; + if (axis != nullptr) + { + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (in_idx == axis[axis_idx]) + { + is_axis = true; + break; + } + } + } + if (!is_axis) + { + offset = offset * static_cast<size_t>(dims[in_idx]) + static_cast<size_t>(index[out_idx]); + out_idx++; + } + else + { + offset = offset * static_cast<size_t>(dims[in_idx]); + } + } + return offset; +} + +// Gets offset of index if reducing on axis. When reducing, the flattened offset +// will not change, if the input index changes on the given axis. For example, +// if you have a 3D tensor and you are reducing to 2D by eliminating axis 0, +// then index (0, 1, 2) and index (1, 1, 2) will map to the same flattened +// offset. +// TODO(kanlig): uses Dims to represent dimensions. +inline size_t ReducedOutputOffset(const int num_dims, const int *dims, const int *index, + const int num_axis, const int *axis) +{ + size_t offset = 0; + for (int idx = 0; idx < num_dims; ++idx) + { + // if we need to skip this axis + bool is_axis = false; + if (axis != nullptr) + { + for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) + { + if (idx == axis[axis_idx]) + { + is_axis = true; + break; + } + } + } + if (!is_axis) + { + offset = offset * static_cast<size_t>(dims[idx]) + static_cast<size_t>(index[idx]); + } + } + return offset; +} + +// Gets next index to iterate through a multidimensional array. +inline bool NextIndex(TfLiteContext *context, const int num_dims, const int *dims, int *current) +{ + int carry = 1; + for (int idx = num_dims - 1; idx >= 0; --idx) + { + int current_val = current[idx] + carry; + TF_LITE_ENSURE(context, (dims[idx] >= current_val)); + if (dims[idx] == current_val) + { + current[idx] = 0; + } + else + { + current[idx] = current_val; + carry = 0; + break; + } + } + return (carry == 0); +} + +template <typename T> +inline TfLiteStatus +CustomSum(TfLiteContext *context, T *input_data, const int *input_dims, const int input_num_dims, + T *output_data, const int *output_dims, const int output_num_dims, const int *axis, + const int num_axis_dimensions, bool keep_dims, int *temp_index, int *resolved_axis) +{ + // resolves axis. + int num_resolved_axis = 0; + for (int idx = 0; idx < num_axis_dimensions; ++idx) + { + int current = axis[idx]; + TF_LITE_ENSURE(context, (current < input_num_dims && current + input_num_dims >= 0)); + if (current < 0) + { + current += input_num_dims; + } + bool is_dup = false; + for (int j = 0; j < num_resolved_axis; ++j) + { + if (resolved_axis[j] == current) + { + is_dup = true; + break; + } + } + if (!is_dup) + { + resolved_axis[num_resolved_axis++] = current; + } + } + + TF_LITE_ENSURE(context, (input_num_dims > 0)); + TF_LITE_ENSURE(context, (input_dims != nullptr)); + TF_LITE_ENSURE(context, (temp_index != nullptr)); + + // resets output data. + for (int idx = 0; idx < output_num_dims; ++idx) + { + temp_index[idx] = 0; + } + for (bool has_next = true; has_next; + has_next = NextIndex(context, output_num_dims, output_dims, temp_index)) + { + size_t output_offset = + ReducedOutputOffset(output_num_dims, output_dims, temp_index, 0, nullptr); + output_data[output_offset] = 0; + } + + // resets temp index. + for (int idx = 0; idx < input_num_dims; ++idx) + { + temp_index[idx] = 0; + } + + // iterates through input_data. + for (bool has_next = true; has_next; + has_next = NextIndex(context, input_num_dims, input_dims, temp_index)) + { + size_t input_offset = ReducedOutputOffset(input_num_dims, input_dims, temp_index, 0, nullptr); + size_t output_offset = ReducedOutputOffset(input_num_dims, input_dims, temp_index, + num_resolved_axis, resolved_axis); + output_data[output_offset] += input_data[input_offset]; + } + + return kTfLiteOk; +} + +TfLiteStatus EvalTensorFlowSum(TfLiteContext *context, TfLiteNode *node) +{ + + TensorFlowSumOp op_context(context, node); + int num_axis = static_cast<int>(::tflite::NumElements(op_context.axis)); + TfLiteTensor *temp_index = &context->tensors[node->temporaries->data[0]]; + TfLiteTensor *resolved_axis = &context->tensors[node->temporaries->data[1]]; + // Resize the output tensor if the output tensor is dynamic. + if (::tflite::IsDynamicTensor(op_context.output)) + { + TF_LITE_ENSURE_OK(context, ResizeTempAxis(context, &op_context, resolved_axis)); + TF_LITE_ENSURE_OK(context, ResizeOutputTensor(context, &op_context)); + } + + TfLiteStatus returnStatus = kTfLiteOk; + switch (op_context.input->type) + { + case kTfLiteFloat32: + returnStatus = CustomSum<float>( + context, op_context.input->data.f, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.f, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, num_axis, false, + temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteInt32: + returnStatus = CustomSum<int>(context, op_context.input->data.i32, + op_context.input->dims->data, op_context.input->dims->size, + op_context.output->data.i32, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, + num_axis, false, temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteUInt8: + returnStatus = CustomSum<uint8_t>( + context, op_context.input->data.uint8, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.uint8, + op_context.output->dims->data, op_context.output->dims->size, op_context.axis->data.i32, + num_axis, false, temp_index->data.i32, resolved_axis->data.i32); + break; + case kTfLiteInt64: + returnStatus = CustomSum<int64_t>( + context, op_context.input->data.i64, op_context.input->dims->data, + op_context.input->dims->size, op_context.output->data.i64, op_context.output->dims->data, + op_context.output->dims->size, op_context.axis->data.i32, num_axis, false, + temp_index->data.i32, resolved_axis->data.i32); + break; + default: + returnStatus = kTfLiteError; + } + + return returnStatus; +} + +} // namespace TensorFlowSum +} // namespace custom +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/ext/kernels/register.cpp b/libs/tflite/src/ext/kernels/register.cpp new file mode 100644 index 000000000..b822bd616 --- /dev/null +++ b/libs/tflite/src/ext/kernels/register.cpp @@ -0,0 +1,221 @@ +/* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// NOTE To minimize diff with upstream tensorflow, disable clang-format +// clang-format off + +// NOTE This code is derived from the following file (in TensorFlow) +// 'externals/tensorflow/tensorflow/contrib/lite/kernels/register.cc' +#include "tflite/ext/kernels/register.h" +#include "tflite/ext/kernels/CustomOps.h" + +namespace tflite { +namespace ops { +namespace builtin { + +TfLiteRegistration *Register_RELU(); +TfLiteRegistration *Register_RELU_N1_TO_1(); +TfLiteRegistration *Register_RELU6(); +TfLiteRegistration *Register_TANH(); +TfLiteRegistration *Register_LOGISTIC(); +TfLiteRegistration *Register_AVERAGE_POOL_2D(); +TfLiteRegistration *Register_MAX_POOL_2D(); +TfLiteRegistration *Register_L2_POOL_2D(); +TfLiteRegistration *Register_CONV_2D(); +TfLiteRegistration *Register_DEPTHWISE_CONV_2D(); +TfLiteRegistration *Register_SVDF(); +TfLiteRegistration *Register_RNN(); +TfLiteRegistration *Register_BIDIRECTIONAL_SEQUENCE_RNN(); +TfLiteRegistration *Register_UNIDIRECTIONAL_SEQUENCE_RNN(); +TfLiteRegistration *Register_EMBEDDING_LOOKUP(); +TfLiteRegistration *Register_EMBEDDING_LOOKUP_SPARSE(); +TfLiteRegistration *Register_FULLY_CONNECTED(); +TfLiteRegistration *Register_LSH_PROJECTION(); +TfLiteRegistration *Register_HASHTABLE_LOOKUP(); +TfLiteRegistration *Register_SOFTMAX(); +TfLiteRegistration *Register_CONCATENATION(); +TfLiteRegistration *Register_ADD(); +TfLiteRegistration *Register_SPACE_TO_BATCH_ND(); +TfLiteRegistration *Register_DIV(); +TfLiteRegistration *Register_SUB(); +TfLiteRegistration *Register_BATCH_TO_SPACE_ND(); +TfLiteRegistration *Register_MUL(); +TfLiteRegistration *Register_L2_NORMALIZATION(); +TfLiteRegistration *Register_LOCAL_RESPONSE_NORMALIZATION(); +TfLiteRegistration *Register_LSTM(); +TfLiteRegistration *Register_BIDIRECTIONAL_SEQUENCE_LSTM(); +TfLiteRegistration *Register_UNIDIRECTIONAL_SEQUENCE_LSTM(); +TfLiteRegistration *Register_PAD(); +TfLiteRegistration *Register_PADV2(); +TfLiteRegistration *Register_RESHAPE(); +TfLiteRegistration *Register_RESIZE_BILINEAR(); +TfLiteRegistration *Register_SKIP_GRAM(); +TfLiteRegistration *Register_SPACE_TO_DEPTH(); +TfLiteRegistration *Register_GATHER(); +TfLiteRegistration *Register_TRANSPOSE(); +TfLiteRegistration *Register_MEAN(); +TfLiteRegistration *Register_SPLIT(); +TfLiteRegistration *Register_SQUEEZE(); +TfLiteRegistration *Register_STRIDED_SLICE(); +TfLiteRegistration *Register_EXP(); +TfLiteRegistration *Register_TOPK_V2(); +TfLiteRegistration *Register_LOG_SOFTMAX(); +TfLiteRegistration *Register_CAST(); +TfLiteRegistration *Register_DEQUANTIZE(); +TfLiteRegistration *Register_PRELU(); +TfLiteRegistration *Register_MAXIMUM(); +TfLiteRegistration *Register_MINIMUM(); +TfLiteRegistration *Register_ARG_MAX(); +TfLiteRegistration *Register_GREATER(); +TfLiteRegistration *Register_GREATER_EQUAL(); +TfLiteRegistration *Register_LESS(); +TfLiteRegistration *Register_LESS_EQUAL(); +TfLiteRegistration *Register_FLOOR(); +TfLiteRegistration *Register_NEG(); +TfLiteRegistration *Register_SELECT(); +TfLiteRegistration *Register_SLICE(); +TfLiteRegistration *Register_SIN(); +TfLiteRegistration *Register_TRANSPOSE_CONV(); +TfLiteRegistration *Register_SPARSE_TO_DENSE(); +#ifndef OBS_BUILD +TfLiteRegistration *Register_SUM(); +TfLiteRegistration *Register_REDUCE_MAX(); +TfLiteRegistration *Register_REDUCE_MIN(); +TfLiteRegistration *Register_EQUAL(); +TfLiteRegistration *Register_NOT_EQUAL(); +TfLiteRegistration *Register_SQRT(); +TfLiteRegistration *Register_RSQRT(); +TfLiteRegistration *Register_SHAPE(); +TfLiteRegistration *Register_POW(); +TfLiteRegistration *Register_FAKE_QUANT(); +TfLiteRegistration *Register_PACK(); +TfLiteRegistration *Register_ONE_HOT(); +TfLiteRegistration *Register_LOGICAL_OR(); +TfLiteRegistration *Register_LOGICAL_AND(); +TfLiteRegistration *Register_LOGICAL_NOT(); +TfLiteRegistration *Register_UNPACK(); +TfLiteRegistration *Register_FLOOR_DIV(); +TfLiteRegistration *Register_SQUARE(); +TfLiteRegistration *Register_ZEROS_LIKE(); +#endif // OBS_BUILD + +} // namespace builtin +} // namespace ops +} // namespace tflite + +namespace nnfw { +namespace tflite { + +BuiltinOpResolver::BuiltinOpResolver() +{ + // Using namespace directive to minimize diff with upstream tensorflow + using namespace ::tflite::ops::builtin; + using namespace ::tflite; + + AddBuiltin(BuiltinOperator_RELU, Register_RELU()); + AddBuiltin(BuiltinOperator_RELU_N1_TO_1, Register_RELU_N1_TO_1()); + AddBuiltin(BuiltinOperator_RELU6, Register_RELU6()); + AddBuiltin(BuiltinOperator_TANH, Register_TANH()); + AddBuiltin(BuiltinOperator_LOGISTIC, Register_LOGISTIC()); + AddBuiltin(BuiltinOperator_AVERAGE_POOL_2D, Register_AVERAGE_POOL_2D()); + AddBuiltin(BuiltinOperator_MAX_POOL_2D, Register_MAX_POOL_2D()); + AddBuiltin(BuiltinOperator_L2_POOL_2D, Register_L2_POOL_2D()); + AddBuiltin(BuiltinOperator_CONV_2D, Register_CONV_2D()); + AddBuiltin(BuiltinOperator_DEPTHWISE_CONV_2D, Register_DEPTHWISE_CONV_2D()); + AddBuiltin(BuiltinOperator_SVDF, Register_SVDF()); + AddBuiltin(BuiltinOperator_RNN, Register_RNN()); + AddBuiltin(BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN, Register_BIDIRECTIONAL_SEQUENCE_RNN()); + AddBuiltin(BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN, Register_UNIDIRECTIONAL_SEQUENCE_RNN()); + AddBuiltin(BuiltinOperator_EMBEDDING_LOOKUP, Register_EMBEDDING_LOOKUP()); + AddBuiltin(BuiltinOperator_EMBEDDING_LOOKUP_SPARSE, Register_EMBEDDING_LOOKUP_SPARSE()); + AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED()); + AddBuiltin(BuiltinOperator_LSH_PROJECTION, Register_LSH_PROJECTION()); + AddBuiltin(BuiltinOperator_HASHTABLE_LOOKUP, Register_HASHTABLE_LOOKUP()); + AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX()); + AddBuiltin(BuiltinOperator_CONCATENATION, Register_CONCATENATION()); + AddBuiltin(BuiltinOperator_ADD, Register_ADD()); + AddBuiltin(BuiltinOperator_SPACE_TO_BATCH_ND, Register_SPACE_TO_BATCH_ND()); + AddBuiltin(BuiltinOperator_BATCH_TO_SPACE_ND, Register_BATCH_TO_SPACE_ND()); + AddBuiltin(BuiltinOperator_MUL, Register_MUL()); + AddBuiltin(BuiltinOperator_L2_NORMALIZATION, Register_L2_NORMALIZATION()); + AddBuiltin(BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION, Register_LOCAL_RESPONSE_NORMALIZATION()); + AddBuiltin(BuiltinOperator_LSTM, Register_LSTM()); + AddBuiltin(BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM, Register_BIDIRECTIONAL_SEQUENCE_LSTM()); + AddBuiltin(BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM, Register_UNIDIRECTIONAL_SEQUENCE_LSTM()); + AddBuiltin(BuiltinOperator_PAD, Register_PAD()); + AddBuiltin(BuiltinOperator_PADV2, Register_PADV2()); + AddBuiltin(BuiltinOperator_RESHAPE, Register_RESHAPE()); + AddBuiltin(BuiltinOperator_RESIZE_BILINEAR, Register_RESIZE_BILINEAR()); + AddBuiltin(BuiltinOperator_SKIP_GRAM, Register_SKIP_GRAM()); + AddBuiltin(BuiltinOperator_SPACE_TO_DEPTH, Register_SPACE_TO_DEPTH()); + AddBuiltin(BuiltinOperator_GATHER, Register_GATHER()); + AddBuiltin(BuiltinOperator_TRANSPOSE, Register_TRANSPOSE()); + AddBuiltin(BuiltinOperator_MEAN, Register_MEAN()); + AddBuiltin(BuiltinOperator_DIV, Register_DIV()); + AddBuiltin(BuiltinOperator_SUB, Register_SUB()); + AddBuiltin(BuiltinOperator_SPLIT, Register_SPLIT()); + AddBuiltin(BuiltinOperator_SQUEEZE, Register_SQUEEZE()); + AddBuiltin(BuiltinOperator_STRIDED_SLICE, Register_STRIDED_SLICE()); + AddBuiltin(BuiltinOperator_EXP, Register_EXP()); + AddBuiltin(BuiltinOperator_TOPK_V2, Register_TOPK_V2()); + AddBuiltin(BuiltinOperator_LOG_SOFTMAX, Register_LOG_SOFTMAX()); + AddBuiltin(BuiltinOperator_CAST, Register_CAST()); + AddBuiltin(BuiltinOperator_DEQUANTIZE, Register_DEQUANTIZE()); + AddBuiltin(BuiltinOperator_PRELU, Register_PRELU()); + AddBuiltin(BuiltinOperator_MAXIMUM, Register_MAXIMUM()); + AddBuiltin(BuiltinOperator_MINIMUM, Register_MINIMUM()); + AddBuiltin(BuiltinOperator_ARG_MAX, Register_ARG_MAX()); + AddBuiltin(BuiltinOperator_GREATER, Register_GREATER()); + AddBuiltin(BuiltinOperator_GREATER_EQUAL, Register_GREATER_EQUAL()); + AddBuiltin(BuiltinOperator_LESS, Register_LESS()); + AddBuiltin(BuiltinOperator_LESS_EQUAL, Register_LESS_EQUAL()); + AddBuiltin(BuiltinOperator_FLOOR, Register_FLOOR()); + AddBuiltin(BuiltinOperator_NEG, Register_NEG()); + AddBuiltin(BuiltinOperator_SELECT, Register_SELECT()); + AddBuiltin(BuiltinOperator_SLICE, Register_SLICE()); + AddBuiltin(BuiltinOperator_SIN, Register_SIN()); +#ifndef OBS_BUILD + AddBuiltin(BuiltinOperator_SUM, Register_SUM()); + AddBuiltin(BuiltinOperator_REDUCE_MAX, Register_REDUCE_MAX()); + AddBuiltin(BuiltinOperator_REDUCE_MIN, Register_REDUCE_MIN()); + AddBuiltin(BuiltinOperator_TRANSPOSE_CONV, Register_TRANSPOSE_CONV()); + AddBuiltin(BuiltinOperator_SPARSE_TO_DENSE, Register_SPARSE_TO_DENSE()); + AddBuiltin(BuiltinOperator_EQUAL, Register_EQUAL()); + AddBuiltin(BuiltinOperator_NOT_EQUAL, Register_NOT_EQUAL()); + AddBuiltin(BuiltinOperator_SQRT, Register_SQRT()); + AddBuiltin(BuiltinOperator_RSQRT, Register_RSQRT()); + AddBuiltin(BuiltinOperator_SHAPE, Register_SHAPE()); + AddBuiltin(BuiltinOperator_POW, Register_POW()); + AddBuiltin(BuiltinOperator_FAKE_QUANT, Register_FAKE_QUANT(), 1, 2); + AddBuiltin(BuiltinOperator_PACK, Register_PACK()); + AddBuiltin(BuiltinOperator_ONE_HOT, Register_ONE_HOT()); + AddBuiltin(BuiltinOperator_LOGICAL_OR, Register_LOGICAL_OR()); + AddBuiltin(BuiltinOperator_LOGICAL_AND, Register_LOGICAL_AND()); + AddBuiltin(BuiltinOperator_LOGICAL_NOT, Register_LOGICAL_NOT()); + AddBuiltin(BuiltinOperator_UNPACK, Register_UNPACK()); + AddBuiltin(BuiltinOperator_FLOOR_DIV, Register_FLOOR_DIV()); + AddBuiltin(BuiltinOperator_SQUARE, Register_SQUARE()); + AddBuiltin(BuiltinOperator_ZEROS_LIKE, Register_ZEROS_LIKE()); +#endif // OBS_BUILD + + AddCustom("TensorFlowMax", nnfw::tflite::custom::Register_TensorFlowMax()); + AddCustom("SquaredDifference", nnfw::tflite::custom::Register_SquaredDifference()); + AddCustom("TensorFlowSum", nnfw::tflite::custom::Register_TensorFlowSum()); + AddCustom("Abs", nnfw::tflite::custom::Register_Abs()); +} + +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/ext/nnapi_delegate.cpp b/libs/tflite/src/ext/nnapi_delegate.cpp new file mode 100644 index 000000000..25858a7b4 --- /dev/null +++ b/libs/tflite/src/ext/nnapi_delegate.cpp @@ -0,0 +1,1209 @@ +/* Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// NOTE To minimize diff with upstream tensorflow, disable clang-format +// clang-format off + +// NOTE This code is derived from the following file (in TensorFlow v1.12) +// 'externals/tensorflow/tensorflow/contrib/lite/nnapi_delegate.cc' +#include "tflite/ext/nnapi_delegate.h" +#include <fcntl.h> +#include <sys/mman.h> +#include <sys/stat.h> +#include <sys/types.h> +#ifdef OBS_BUILD +#include "tensorflow/contrib/lite/builtin_op_data.h" +#include "tensorflow/contrib/lite/error_reporter.h" +#else +#include "tensorflow/contrib/lite/c/builtin_op_data.h" +#include "tensorflow/contrib/lite/core/api/error_reporter.h" +#endif +#include "tensorflow/contrib/lite/model.h" +#include "NeuralNetworksShim.h" +#include "NeuralNetworksExShim.h" + +#ifdef __ANDROID__ +#include <android/log.h> +#include <sys/system_properties.h> +#endif + +namespace nnfw { +namespace tflite { + +void logError(const char* format, ...) { + // stderr is convenient for native tests, but is not captured for apps + va_list args_for_stderr; + va_start(args_for_stderr, format); + vfprintf(stderr, format, args_for_stderr); + va_end(args_for_stderr); + fprintf(stderr, "\n"); + fflush(stderr); +#ifdef __ANDROID__ + // produce logcat output for general consumption + va_list args_for_log; + va_start(args_for_log, format); + __android_log_vprint(ANDROID_LOG_ERROR, "tflite", format, args_for_log); + va_end(args_for_log); +#endif +} + +#define FATAL(...) \ + logError(__VA_ARGS__); \ + exit(1); + +// TODO(aselle): Change the error model to use status codes. +#define CHECK_TFLITE_SUCCESS(x) \ + if (x != kTfLiteOk) { \ + FATAL("Aborting since tflite returned failure nnapi_delegate.cc:%d.", \ + __LINE__); \ + } + +#define CHECK_NN(x) \ + if (x != ANEURALNETWORKS_NO_ERROR) { \ + FATAL("Aborting since NNAPI returned failure nnapi_delegate.cc:%d", \ + __LINE__); \ + } + +#define RETURN_ERROR_IF_TFLITE_FAILED(x) \ + if (x != kTfLiteOk) { \ + logError( \ + "Returning error since TFLite returned failure nnapi_delegate.cc:%d.", \ + __LINE__); \ + return kTfLiteError; \ + } + +#define RETURN_ERROR_IF_NN_FAILED(x) \ + if (x != ANEURALNETWORKS_NO_ERROR) { \ + logError( \ + "Returning error since NNAPI returned failure nnapi_delegate.cc:%d.", \ + __LINE__); \ + return kTfLiteError; \ + } + +// Tracking of NNAPI operand ids +static const int64_t kOperandIdNotSet = -1; +static const int64_t kOperandNotNeeded = -2; + +namespace { + +int32_t GetAndroidSdkVersion() { +#ifdef __ANDROID__ + const char* sdkProp = "ro.build.version.sdk"; + char sdkVersion[PROP_VALUE_MAX]; + int length = __system_property_get(sdkProp, sdkVersion); + if (length != 0) { + for (int i = 0; i < length; ++i) { + int digit = sdkVersion[i] - '0'; + if (digit < 0 || digit > 9) { + // Non-numeric SDK version, assume it's higher then expected; + return 0xFFFF; + } + } + return atoi(sdkVersion); + } + FATAL("No %s prop", sdkProp); +#endif // __ANDROID__ + return 0; +} + +int32_t GetAndroidSdkVersionCached() { + static int32_t androidSdkVersion = GetAndroidSdkVersion(); + return androidSdkVersion; +} + +static const uint32_t dimension_for_scalar[1] = {1}; + +} // namespace + +NNAPIAllocation::NNAPIAllocation(const char* filename, + ::tflite::ErrorReporter* error_reporter) + : MMAPAllocation(filename, error_reporter) { + if (mmapped_buffer_ != MAP_FAILED) + CHECK_NN(ANeuralNetworksMemory_createFromFd(buffer_size_bytes_, PROT_READ, + mmap_fd_, 0, &handle_)); +} + +NNAPIAllocation::~NNAPIAllocation() { + if (handle_) { + ANeuralNetworksMemory_free(handle_); + } +} + +NNAPIDelegate::~NNAPIDelegate() { + if (nn_compiled_model_) { + ANeuralNetworksCompilation_free(nn_compiled_model_); + nn_compiled_model_ = nullptr; + } + if (nn_model_) { + ANeuralNetworksModel_free(nn_model_); + nn_model_ = nullptr; + // TODO(aselle): Is this thread-safe and callable multiple times? + } + // ANeuralNetworksShutdown(); +} + +// Adds the tensors of the interpreter to the NN API model. +TfLiteStatus addTensorOperands(::tflite::Interpreter* interpreter, + ANeuralNetworksModel* nn_model, + uint32_t* no_of_operands_added, + std::vector<int64_t>* nnapi_ids) { + uint32_t next_id = 0; + for (size_t i = 0; i < interpreter->tensors_size(); i++) { + // Skip temporaries and RNN back-edges. + if ((*nnapi_ids)[i] == kOperandNotNeeded) continue; + + (*nnapi_ids)[i] = int64_t(next_id); + + int32_t nn_type = 0; + // NNAPI requires 32-bit float scale to be zero, tflite doesn't care + float scale = 0.0f; + int32_t zeroPoint = 0; + TfLiteTensor* tensor = interpreter->tensor(i); + switch (tensor->type) { + case kTfLiteNoType: + // Tensors added during initialization of Ops don't have a type yet and + // should not be registered with the NNAPI. + continue; + case kTfLiteFloat32: + nn_type = ANEURALNETWORKS_TENSOR_FLOAT32; + break; + case kTfLiteUInt8: + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM; + scale = tensor->params.scale; + // FIXME The next line is a workaround because currently zero scale is + // passed down from TF + // Lite. Note that the latest NeuralNetworks.h (see + // https://android.googlesource.com/platform/frameworks/ml/+/master/nn/runtime/include/NeuralNetworks.h) + // requires scale to be greater than zero. Remove this workaround + // when the scale + // value is correctly passed. + scale = (scale == 0.0f) ? 1.0f : scale; + zeroPoint = tensor->params.zero_point; + break; + case kTfLiteInt32: + nn_type = ANEURALNETWORKS_TENSOR_INT32; + scale = tensor->params.scale; + zeroPoint = tensor->params.zero_point; + break; + case kTfLiteBool: + // Workaround to pass bool type under NNAPI + // Use bool type using ANEURALNETWORKS_TENSOR_QUANT8_ASYMM with scale = 1.0f and zero_point = 0 + nn_type = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM; + scale = 1.0f; + zeroPoint = 0; + break; + default: + logError("Unsupported tensor type %d", tensor->type); + return kTfLiteError; + } + if (tensor->dims->size == 0) { + // WORKAROUND Some model have dimension zero + switch (tensor->type) { + case kTfLiteFloat32: + nn_type = ANEURALNETWORKS_TENSOR_FLOAT32; + break; + case kTfLiteInt32: + nn_type = ANEURALNETWORKS_TENSOR_INT32; + break; + default: + logError("NNAPI doesn't support tensors with rank 0 (index %d name %s)", + i, tensor->name); + return kTfLiteError; + } + } + if (tensor->dims->size > 4) { + logError("NNAPI doesn't support tensors with rank > 4 (index %d name %s)", + i, tensor->name); + return kTfLiteError; + } + // TODO(aselle): Note, many of these are intermediate results. Do I need + // to ever specify these sizes. I am currently below doing setValue + // on all of them, but I shouldn't in the future. + // Answer(jeanluc): If all the operators can set the dimension correctly, + // you won't need to. + ANeuralNetworksOperandType operand_type{ + nn_type, static_cast<uint32_t>(tensor->dims->size), + reinterpret_cast<uint32_t*>(tensor->dims->data), scale, zeroPoint}; + if (tensor->dims->size == 0) { + // WORKAROUND Some model have dimension zero + // Consider scalar as vector size 1 + operand_type.dimensions = dimension_for_scalar; + operand_type.dimensionCount = 1; + } + RETURN_ERROR_IF_NN_FAILED( + ANeuralNetworksModel_addOperand(nn_model, &operand_type)); + // TODO(aselle): Based on Michael's suggestion, limiting this to read + // only memory + if (tensor->allocation_type == kTfLiteMmapRo) { + if (const NNAPIAllocation* alloc = dynamic_cast<const NNAPIAllocation*>( + static_cast<const ::tflite::Allocation*>(tensor->allocation))) { + RETURN_ERROR_IF_NN_FAILED( + ANeuralNetworksModel_setOperandValueFromMemory( + nn_model, next_id, alloc->memory(), + alloc->offset(tensor->data.raw), tensor->bytes)); + } else { + RETURN_ERROR_IF_NN_FAILED(ANeuralNetworksModel_setOperandValue( + nn_model, next_id, tensor->data.raw, tensor->bytes)); + } + } else if (tensor->bytes == 0) { + // These size 0 tensors are optional tensors reserved. + RETURN_ERROR_IF_NN_FAILED( + ANeuralNetworksModel_setOperandValue(nn_model, next_id, nullptr, 0)); + } + + ++next_id; + } + *no_of_operands_added = next_id; + return kTfLiteOk; +} + +void MapAndAddTensorIds(const int* from_ids_buf, size_t from_ids_count, + std::vector<uint32_t>* into, + const std::vector<int64_t>& map) { + for (size_t i = 0; i < from_ids_count; i++) { + int from_id = from_ids_buf[i]; + if (from_id == kOptionalTensor) { + into->push_back(from_id); + } else { + into->push_back(map[from_id]); + } + } +} + +// Adds the operations and their parameters to the NN API model. +// 'next-id' is the operand ID of the next operand of the model. +TfLiteStatus AddOpsAndParams( + ::tflite::Interpreter* interpreter, ANeuralNetworksModel* nn_model, + uint32_t next_id, std::vector<int>* model_state_inputs, + std::vector<int>* model_state_outputs, + const std::vector<int64_t>& tensor_id_to_nnapi_id) { + for (size_t i = 0; i < interpreter->nodes_size(); i++) { + const auto* node_and_registration = interpreter->node_and_registration(i); + const TfLiteNode& node = node_and_registration->first; + const TfLiteRegistration& registration = node_and_registration->second; + ::tflite::BuiltinOperator builtin = + static_cast<::tflite::BuiltinOperator>(registration.builtin_code); + + // Add the parameters. + std::vector<uint32_t> augmented_inputs, augmented_outputs; + MapAndAddTensorIds(node.inputs->data, node.inputs->size, &augmented_inputs, + tensor_id_to_nnapi_id); + MapAndAddTensorIds(node.outputs->data, node.outputs->size, + &augmented_outputs, tensor_id_to_nnapi_id); + + auto add_scalar_int32 = [&nn_model, &augmented_inputs, + &next_id](int value) { + ANeuralNetworksOperandType operand_type{.type = ANEURALNETWORKS_INT32}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)) + CHECK_NN(ANeuralNetworksModel_setOperandValue(nn_model, next_id, &value, + sizeof(int32_t))) + augmented_inputs.push_back(next_id++); + }; + + auto add_scalar_float32 = [&nn_model, &augmented_inputs, + &next_id](float value) { + ANeuralNetworksOperandType operand_type{.type = ANEURALNETWORKS_FLOAT32}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)) + CHECK_NN(ANeuralNetworksModel_setOperandValue(nn_model, next_id, &value, + sizeof(float))) + augmented_inputs.push_back(next_id++); + }; + + auto add_vector_int32 = [&](const int* values, uint32_t num_values) { + ANeuralNetworksOperandType operand_type{ + .type = ANEURALNETWORKS_TENSOR_INT32, + .dimensionCount = 1, + .dimensions = &num_values}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)) + CHECK_NN(ANeuralNetworksModel_setOperandValue( + nn_model, next_id, values, sizeof(int32_t) * num_values)); + augmented_inputs.push_back(next_id++); + }; + + // Handle state tensors of RNN, LSTM, SVDF. + // For each state_out tensor, a corresponding state_in operand needs to be + // created for NNAPI. + auto duplicate_state_tensor_float32 = + [interpreter, &nn_model, &next_id, &augmented_inputs, + &model_state_inputs, &model_state_outputs](int tensor_id) { + const TfLiteTensor* tensor = interpreter->tensor(tensor_id); + ANeuralNetworksOperandType operand_type{ + ANEURALNETWORKS_TENSOR_FLOAT32, + static_cast<uint32_t>(tensor->dims->size), + reinterpret_cast<uint32_t*>(tensor->dims->data), + tensor->params.scale, tensor->params.zero_point}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)); + augmented_inputs.push_back(next_id); + model_state_inputs->push_back(next_id); + model_state_outputs->push_back(tensor_id); + next_id++; + }; + auto check_and_add_activation = [&add_scalar_int32](int activation) { + if (activation > kTfLiteActRelu6) { + logError("NNAPI only supports RELU, RELU1 and RELU6 activations"); + return kTfLiteError; + } + add_scalar_int32(activation); + return kTfLiteOk; + }; + + auto add_add_params = [&add_scalar_int32](void* data) { + auto* builtin = reinterpret_cast<TfLiteAddParams*>(data); + if (builtin->activation > kTfLiteActRelu6) { + logError("NNAPI only supports RELU, RELU1 and RELU6 activations"); + return kTfLiteError; + } + add_scalar_int32(builtin->activation); + return kTfLiteOk; + }; + + auto add_pooling_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { + auto builtin = reinterpret_cast<TfLitePoolParams*>(data); + add_scalar_int32(builtin->padding); + add_scalar_int32(builtin->stride_width); + add_scalar_int32(builtin->stride_height); + add_scalar_int32(builtin->filter_width); + add_scalar_int32(builtin->filter_height); + return check_and_add_activation(builtin->activation); + }; + + auto add_convolution_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { + auto builtin = reinterpret_cast<TfLiteConvParams*>(data); + add_scalar_int32(builtin->padding); + add_scalar_int32(builtin->stride_width); + add_scalar_int32(builtin->stride_height); + return check_and_add_activation(builtin->activation); + }; + + auto add_depthwise_conv_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { + auto builtin = reinterpret_cast<TfLiteDepthwiseConvParams*>(data); + add_scalar_int32(builtin->padding); + add_scalar_int32(builtin->stride_width); + add_scalar_int32(builtin->stride_height); + add_scalar_int32(builtin->depth_multiplier); + return check_and_add_activation(builtin->activation); + }; + + auto add_fully_connected_params = [&check_and_add_activation](void* data) { + auto builtin = reinterpret_cast<TfLiteFullyConnectedParams*>(data); + return check_and_add_activation(builtin->activation); + }; + + auto add_concatenation_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteConcatenationParams*>(data); + add_scalar_int32(builtin->axis); + if (builtin->activation != kTfLiteActNone) { + logError("Concatenation does not support fused activation in NNAPI"); + return kTfLiteError; + } + return kTfLiteOk; + }; + + auto add_softmax_params = [&add_scalar_float32](void* data) { + auto builtin = reinterpret_cast<TfLiteSoftmaxParams*>(data); + add_scalar_float32(builtin->beta); + }; + + auto add_space_to_depth_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteSpaceToDepthParams*>(data); + add_scalar_int32(builtin->block_size); + }; + + auto add_lstm_params = [&add_scalar_int32, + &add_scalar_float32](void* data) { + auto builtin = reinterpret_cast<TfLiteLSTMParams*>(data); + add_scalar_int32(builtin->activation); + add_scalar_float32(builtin->cell_clip); + add_scalar_float32(builtin->proj_clip); + }; + + // LSTM in NNAPI requires scratch tensor as an output operand. + auto add_lstm_scratch_tensor_float32 = [interpreter, &node, &nn_model, + &next_id, &augmented_outputs]() { + if (node.temporaries->size == 0) return; + int scratch_buffer_index = node.temporaries->data[0]; + const TfLiteTensor* tensor = interpreter->tensor(scratch_buffer_index); + ANeuralNetworksOperandType operand_type{ + ANEURALNETWORKS_TENSOR_FLOAT32, + static_cast<uint32_t>(tensor->dims->size), + reinterpret_cast<uint32_t*>(tensor->dims->data), tensor->params.scale, + tensor->params.zero_point}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)); + augmented_outputs.insert(augmented_outputs.begin(), next_id++); + }; + + auto add_mean_params = [&add_scalar_int32](void* data) { +#ifdef OBS_BUILD + auto builtin = reinterpret_cast<TfLiteMeanParams*>(data); +#else + auto builtin = reinterpret_cast<TfLiteReducerParams*>(data); +#endif + add_scalar_int32(builtin->keep_dims); + }; + + auto add_svdf_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteSVDFParams*>(data); + add_scalar_int32(builtin->rank); + add_scalar_int32(builtin->activation); + }; + + auto add_rnn_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteRNNParams*>(data); + add_scalar_int32(builtin->activation); + }; + + auto add_squeeze_params = [&](void* data) { + const auto* builtin = reinterpret_cast<TfLiteSqueezeParams*>(data); + // Note that we add the squeeze dimensions even if the dimensions were + // unspecified (empty), as NNAPI requires the operand. + add_vector_int32(builtin->squeeze_dims, + static_cast<uint32_t>(builtin->num_squeeze_dims)); + }; + + // Handle optional input tensors. + auto add_optional_tensors = [&nn_model, &augmented_inputs, + &next_id](int nn_type) { + for (size_t idx = 0; idx < augmented_inputs.size(); idx++) { + if (augmented_inputs[idx] == kOptionalTensor) { + const std::vector<uint32_t> dim = {0, 0}; + ANeuralNetworksOperandType operand_type{nn_type, 2, dim.data(), 0, 0}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)) + CHECK_NN(ANeuralNetworksModel_setOperandValue(nn_model, next_id, + nullptr, 0)) + augmented_inputs[idx] = next_id++; + } + } + }; + + int nnapi_version = 10; +#include "nnapi_delegate_ex_AddOpsAndParams_lambda.inc" + + ANeuralNetworksOperationType nn_op_type; + + // Using namespace directive to minimize diff with upstream tensorflow + namespace tflite = ::tflite; + + switch (builtin) { + case tflite::BuiltinOperator_ADD: + nn_op_type = ANEURALNETWORKS_ADD; + RETURN_ERROR_IF_TFLITE_FAILED(add_add_params(node.builtin_data)); + break; + case tflite::BuiltinOperator_MUL: + nn_op_type = ANEURALNETWORKS_MUL; + RETURN_ERROR_IF_TFLITE_FAILED(add_add_params(node.builtin_data)); + break; + case tflite::BuiltinOperator_AVERAGE_POOL_2D: + RETURN_ERROR_IF_TFLITE_FAILED(add_pooling_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_AVERAGE_POOL_2D; + break; + case tflite::BuiltinOperator_MAX_POOL_2D: + RETURN_ERROR_IF_TFLITE_FAILED(add_pooling_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_MAX_POOL_2D; + break; + case tflite::BuiltinOperator_L2_POOL_2D: + RETURN_ERROR_IF_TFLITE_FAILED(add_pooling_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_L2_POOL_2D; + break; + case tflite::BuiltinOperator_CONV_2D: { + auto builtin = reinterpret_cast<TfLiteConvParams*>(node.builtin_data); + if (builtin->dilation_width_factor != 1 || + builtin->dilation_height_factor != 1 || node.inputs->size != 3) { + logError("NNAPI does not support dilated Conv2D."); + return kTfLiteError; + } + } + RETURN_ERROR_IF_TFLITE_FAILED( + add_convolution_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_CONV_2D; + break; + case tflite::BuiltinOperator_RELU: + nn_op_type = ANEURALNETWORKS_RELU; + break; + case tflite::BuiltinOperator_RELU_N1_TO_1: + nn_op_type = ANEURALNETWORKS_RELU1; + break; + case tflite::BuiltinOperator_RELU6: + nn_op_type = ANEURALNETWORKS_RELU6; + break; + case tflite::BuiltinOperator_TANH: + nn_op_type = ANEURALNETWORKS_TANH; + break; + case tflite::BuiltinOperator_FLOOR: + nn_op_type = ANEURALNETWORKS_FLOOR; + break; + case tflite::BuiltinOperator_LOGISTIC: + nn_op_type = ANEURALNETWORKS_LOGISTIC; + break; + case tflite::BuiltinOperator_DEPTHWISE_CONV_2D: + RETURN_ERROR_IF_TFLITE_FAILED( + add_depthwise_conv_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_DEPTHWISE_CONV_2D; + break; + case tflite::BuiltinOperator_CONCATENATION: + RETURN_ERROR_IF_TFLITE_FAILED( + add_concatenation_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_CONCATENATION; + break; + case tflite::BuiltinOperator_SOFTMAX: + add_softmax_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_SOFTMAX; + break; + case tflite::BuiltinOperator_FULLY_CONNECTED: + RETURN_ERROR_IF_TFLITE_FAILED( + add_fully_connected_params(node.builtin_data)); + nn_op_type = ANEURALNETWORKS_FULLY_CONNECTED; + break; + case tflite::BuiltinOperator_RESHAPE: + if (node.inputs->size != 2) { + logError("NNAPI only supports 2-input RESHAPE"); + return kTfLiteError; + } + nn_op_type = ANEURALNETWORKS_RESHAPE; + // add_reshape_params(node.builtin_data); + break; + case tflite::BuiltinOperator_RESIZE_BILINEAR: + add_resize_bilinear_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_RESIZE_BILINEAR; + break; + case tflite::BuiltinOperator_SPACE_TO_DEPTH: + add_space_to_depth_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_SPACE_TO_DEPTH; + break; + case tflite::BuiltinOperator_LSTM: { + if (node.inputs->size + /* no of params */ 3 != 21) { + logError("NNAPI only supports 21-input LSTMs"); + return kTfLiteError; + } + duplicate_state_tensor_float32( + node.outputs->data[/*kOutputStateTensor*/ 0]); + duplicate_state_tensor_float32( + node.outputs->data[/*kCellStateTensor*/ 1]); + add_lstm_params(node.builtin_data); + add_lstm_scratch_tensor_float32(); + add_optional_tensors(ANEURALNETWORKS_TENSOR_FLOAT32); + nn_op_type = ANEURALNETWORKS_LSTM; + break; + } + case tflite::BuiltinOperator_DEQUANTIZE: + nn_op_type = ANEURALNETWORKS_DEQUANTIZE; + break; + case tflite::BuiltinOperator_SVDF: { + duplicate_state_tensor_float32(node.outputs->data[/*kStateTensor*/ 0]); + add_svdf_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_SVDF; + break; + } + case tflite::BuiltinOperator_RNN: { + duplicate_state_tensor_float32( + node.outputs->data[/*kHiddenStateTensor*/ 0]); + add_rnn_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_RNN; + break; + } + case tflite::BuiltinOperator_EMBEDDING_LOOKUP: + nn_op_type = ANEURALNETWORKS_EMBEDDING_LOOKUP; + break; + case tflite::BuiltinOperator_PAD: + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_PAD; + break; + case tflite::BuiltinOperator_MEAN: + nnapi_version = 11; // require NNAPI 1.1 + add_mean_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_MEAN; + break; + case tflite::BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION: + nn_op_type = ANEURALNETWORKS_LOCAL_RESPONSE_NORMALIZATION; + add_lrn_params(node.builtin_data); + break; + case tflite::BuiltinOperator_DIV: + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_DIV; + RETURN_ERROR_IF_TFLITE_FAILED(check_and_add_activation( + reinterpret_cast<TfLiteDivParams*>(node.builtin_data)->activation)); + break; + case tflite::BuiltinOperator_SUB: + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_SUB; + RETURN_ERROR_IF_TFLITE_FAILED(check_and_add_activation( + reinterpret_cast<TfLiteSubParams*>(node.builtin_data)->activation)); + break; + case tflite::BuiltinOperator_SQUEEZE: + nnapi_version = 11; // requires NNAPI 1.1 + add_squeeze_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_SQUEEZE; + break; + case tflite::BuiltinOperator_TRANSPOSE: + // The permutation input tensor value dictates the output dimensions. + // TODO(b/110888333): Support dynamically-sized tensors in delegates. + if ((node.inputs->size > 1) && + (interpreter->tensor(node.inputs->data[1])->allocation_type != + kTfLiteMmapRo)) { + logError("NNAPI does not yet support dynamic tensors."); + return kTfLiteError; + } + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_TRANSPOSE; + break; + case tflite::BuiltinOperator_L2_NORMALIZATION: + nn_op_type = ANEURALNETWORKS_L2_NORMALIZATION; + if (reinterpret_cast<TfLiteL2NormParams*>(node.builtin_data) + ->activation != kTfLiteActNone) { + logError( + "NNAPI does not support L2Normalization with fused activations"); + return kTfLiteError; + } + if ((node.inputs->size > 0) && + (interpreter->tensor(node.inputs->data[0])->dims->size != 4)) { + logError("NNAPI only supports input rank 4 for L2Normalization"); + return kTfLiteError; + } + break; + case tflite::BuiltinOperator_HASHTABLE_LOOKUP: + if (interpreter->tensor(node.outputs->data[0])->type != + kTfLiteFloat32) { + logError("NNAPI only support HASHTABLE_LOOKUP with float32 output", + builtin); + return kTfLiteError; + } + nn_op_type = ANEURALNETWORKS_HASHTABLE_LOOKUP; + break; + case tflite::BuiltinOperator_STRIDED_SLICE: + add_strided_slice_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_STRIDED_SLICE; + break; + case tflite::BuiltinOperator_SPACE_TO_BATCH_ND: + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_SPACE_TO_BATCH_ND; + break; + case tflite::BuiltinOperator_BATCH_TO_SPACE_ND: + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_BATCH_TO_SPACE_ND; + check_batch_to_space_params(); + break; + case tflite::BuiltinOperator_CAST: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_CAST_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_TOPK_V2: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_TOPK_V2_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_GATHER: + add_gather_ex_params(node.builtin_data); + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_GATHER_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_SPLIT: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_SPLIT_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_NEG: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_NEG_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_EXP: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_EXP_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_TRANSPOSE_CONV: + add_transpose_conv_params(node.builtin_data); + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_TRANSPOSE_CONV_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_PRELU: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_PRELU_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_ARG_MAX: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_ARGMAX_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; +#ifndef OBS_BUILD + case tflite::BuiltinOperator_PACK: + add_pack_ex_params(node.builtin_data); + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_PACK_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_UNPACK: + add_unpack_ex_params(node.builtin_data); + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_UNPACK_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_SQRT: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_SQRT_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_RSQRT: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_RSQRT_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_EQUAL: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_EQUAL_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_NOT_EQUAL: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_NOT_EQUAL_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_SUM: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_REDUCE_SUM_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_REDUCE_MAX: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_TENSORFLOW_MAX_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_REDUCE_MIN: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_REDUCE_MIN_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_LOGICAL_AND: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_LOGICAL_AND_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + case tflite::BuiltinOperator_LOGICAL_OR: + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_LOGICAL_OR_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; +#endif + case tflite::BuiltinOperator_CONCAT_EMBEDDINGS: + case tflite::BuiltinOperator_LSH_PROJECTION: + case tflite::BuiltinOperator_BIDIRECTIONAL_SEQUENCE_RNN: + case tflite::BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_RNN: + case tflite::BuiltinOperator_EMBEDDING_LOOKUP_SPARSE: + case tflite::BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: + case tflite::BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM: + //case tflite::BuiltinOperator_LOCAL_RESPONSE_NORMALIZATION: + case tflite::BuiltinOperator_PADV2: + //case tflite::BuiltinOperator_RESIZE_BILINEAR: + case tflite::BuiltinOperator_CALL: + case tflite::BuiltinOperator_SKIP_GRAM: + //case tflite::BuiltinOperator_RELU_N1_TO_1: + //case tflite::BuiltinOperator_GATHER: + //case tflite::BuiltinOperator_SPACE_TO_BATCH_ND: + //case tflite::BuiltinOperator_BATCH_TO_SPACE_ND: + //case tflite::BuiltinOperator_TOPK_V2: + //case tflite::BuiltinOperator_SPLIT: + //case tflite::BuiltinOperator_STRIDED_SLICE: + //case tflite::BuiltinOperator_EXP: + case tflite::BuiltinOperator_LOG_SOFTMAX: + //case tflite::BuiltinOperator_DEQUANTIZE: + case tflite::BuiltinOperator_DELEGATE: + //case tflite::BuiltinOperator_CAST: + //case tflite::BuiltinOperator_PRELU: + case tflite::BuiltinOperator_MAXIMUM: + case tflite::BuiltinOperator_MINIMUM: +#ifndef OBS_BUILD + case tflite::BuiltinOperator_ARG_MIN: +#endif + case tflite::BuiltinOperator_GREATER: + case tflite::BuiltinOperator_GREATER_EQUAL: + case tflite::BuiltinOperator_LESS: + case tflite::BuiltinOperator_LESS_EQUAL: + //case tflite::BuiltinOperator_NEG: + case tflite::BuiltinOperator_SELECT: + case tflite::BuiltinOperator_SLICE: + case tflite::BuiltinOperator_SIN: + //case tflite::BuiltinOperator_LOG: + //case tflite::BuiltinOperator_TRANSPOSE_CONV: +#ifndef OBS_BUILD + case tflite::BuiltinOperator_TILE: + case tflite::BuiltinOperator_EXPAND_DIMS: + case tflite::BuiltinOperator_SPARSE_TO_DENSE: + //case tflite::BuiltinOperator_EQUAL: + //case tflite::BuiltinOperator_NOT_EQUAL: + //case tflite::BuiltinOperator_SUM: + //case tflite::BuiltinOperator_REDUCE_MAX: + //case tflite::BuiltinOperator_REDUCE_MIN: + case tflite::BuiltinOperator_REDUCE_PROD: + //case tflite::BuiltinOperator_SQRT: + //case tflite::BuiltinOperator_RSQRT: + case tflite::BuiltinOperator_SHAPE: + case tflite::BuiltinOperator_POW: + case tflite::BuiltinOperator_FAKE_QUANT: + //case tflite::BuiltinOperator_PACK: + //case tflite::BuiltinOperator_LOGICAL_OR: + case tflite::BuiltinOperator_ONE_HOT: + //case tflite::BuiltinOperator_LOGICAL_AND: + case tflite::BuiltinOperator_LOGICAL_NOT: + //case tflite::BuiltinOperator_UNPACK: + case tflite::BuiltinOperator_FLOOR_DIV: + case tflite::BuiltinOperator_REDUCE_ANY: + case tflite::BuiltinOperator_SQUARE: + case tflite::BuiltinOperator_ZEROS_LIKE: + case tflite::BuiltinOperator_FILL: +#endif + logError("Op code %d is currently not delegated to NNAPI", builtin); + return kTfLiteError; + break; + case tflite::BuiltinOperator_CUSTOM: { + std::string custom_name(registration.custom_name); + if (custom_name.compare("TensorFlowMax") == 0) { + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_TENSORFLOW_MAX_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + } + else if (custom_name.compare("SquaredDifference") == 0) { + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_SQUARED_DIFFERENCE_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + } + else if (custom_name.compare("TensorFlowSum") == 0) { + CHECK_NN(ANeuralNetworksModel_addOperationEx( + nn_model, ANEURALNETWORKS_REDUCE_SUM_EX, + static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(node.outputs->size), + reinterpret_cast<uint32_t*>(node.outputs->data))); + continue; + } + logError("Custom operations are not supported when using NNAPI."); + return kTfLiteError; + break; + } +#ifdef OBS_BUILD + default: + logError("Op code %d is currently not delegated to NNAPI", builtin); + return kTfLiteError; + break; +#endif + } + + //if (nnapi_version == 11 && GetAndroidSdkVersionCached() < 28) { + // FATAL("Op %d needs NNAPI1.1", builtin); + //} + + // Add the operation. + RETURN_ERROR_IF_NN_FAILED(ANeuralNetworksModel_addOperation( + nn_model, nn_op_type, static_cast<uint32_t>(augmented_inputs.size()), + augmented_inputs.data(), + static_cast<uint32_t>(augmented_outputs.size()), + reinterpret_cast<uint32_t*>(augmented_outputs.data()))); + } + return kTfLiteOk; +} + +TfLiteStatus NNAPIDelegate::BuildGraph(::tflite::Interpreter* interpreter) { + if (nn_model_ && nn_compiled_model_) return model_status_; + + // TODO(aselle): This is not correct. need to handle resize invalidation. + if (!nn_model_) { + CHECK_NN(ANeuralNetworksModel_create(&nn_model_)); + + // Find which tensors should be added to NNAPI. TFLite has temporaries + // and RNN back-edges which are are not valid for NNAPI. We look through all + // inputs and outputs and mark the mapping in tensor_id_to_nnapi_id with + // kOperandIdNotSet. addTensorOperands will replace those with the + // corresponding NNAPI operand ids and skip kOperandNotNeeded entries. + std::vector<int64_t> tensor_id_to_nnapi_id(interpreter->tensors_size(), + kOperandNotNeeded); + auto set_ids_to_not_set = [&tensor_id_to_nnapi_id](const int* buf, + size_t count) { + for (int j = 0; j < count; j++) { + auto tensor_id = buf[j]; + if (tensor_id != kOptionalTensor) { + tensor_id_to_nnapi_id[tensor_id] = kOperandIdNotSet; + } + } + }; + for (size_t i = 0; i < interpreter->nodes_size(); i++) { + const auto* node_and_registration = interpreter->node_and_registration(i); + const TfLiteNode& node = node_and_registration->first; + set_ids_to_not_set(node.inputs->data, node.inputs->size); + set_ids_to_not_set(node.outputs->data, node.outputs->size); + } + set_ids_to_not_set(interpreter->inputs().data(), + interpreter->inputs().size()); + set_ids_to_not_set(interpreter->outputs().data(), + interpreter->outputs().size()); + + uint32_t next_id = 0; + RETURN_ERROR_IF_TFLITE_FAILED(addTensorOperands( + interpreter, nn_model_, &next_id, &tensor_id_to_nnapi_id)); + RETURN_ERROR_IF_TFLITE_FAILED( + AddOpsAndParams(interpreter, nn_model_, next_id, &model_states_inputs_, + &model_states_outputs_, tensor_id_to_nnapi_id)); + + std::vector<uint32_t> augmented_inputs; + MapAndAddTensorIds(interpreter->inputs().data(), + interpreter->inputs().size(), &augmented_inputs, + tensor_id_to_nnapi_id); + augmented_inputs.insert(augmented_inputs.end(), + model_states_inputs_.begin(), + model_states_inputs_.end()); + std::vector<uint32_t> augmented_outputs; + MapAndAddTensorIds(interpreter->outputs().data(), + interpreter->outputs().size(), &augmented_outputs, + tensor_id_to_nnapi_id); + MapAndAddTensorIds(model_states_outputs_.data(), + model_states_outputs_.size(), &augmented_outputs, + tensor_id_to_nnapi_id); + + CHECK_NN(ANeuralNetworksModel_identifyInputsAndOutputs( + nn_model_, static_cast<uint32_t>(augmented_inputs.size()), + reinterpret_cast<const uint32_t*>(augmented_inputs.data()), + static_cast<uint32_t>(augmented_outputs.size()), + reinterpret_cast<const uint32_t*>(augmented_outputs.data()))); + + // TODO Support ANeuralNetworksModel_relaxComputationFloat32toFloat16 + //if (GetAndroidSdkVersionCached() >= 28) { + // CHECK_NN(ANeuralNetworksModel_relaxComputationFloat32toFloat16( + // nn_model_, interpreter->GetAllowFp16PrecisionForFp32())); + //} + CHECK_NN(ANeuralNetworksModel_finish(nn_model_)); + } + if (!nn_compiled_model_) { + CHECK_NN(ANeuralNetworksCompilation_create(nn_model_, &nn_compiled_model_)); + CHECK_NN(ANeuralNetworksCompilation_finish(nn_compiled_model_)); + } + return kTfLiteOk; +} + +#include <unordered_map> + +TfLiteStatus NNAPIDelegate::Invoke(::tflite::Interpreter* interpreter) { + if (!nn_model_) { + model_status_ = BuildGraph(interpreter); + if (model_status_ != kTfLiteOk) { + logError("Failed to build graph for NNAPI"); + } + } + if (model_status_ != kTfLiteOk) { + return model_status_; + } + + ANeuralNetworksExecution* execution = nullptr; + CHECK_NN(ANeuralNetworksExecution_create(nn_compiled_model_, &execution)); + + // Allocate temporary buffer to save casted boolean tensor + std::unordered_map<size_t, uint8_t*> input_boolean_tensors; + std::unordered_map<size_t, uint8_t*> output_boolean_tensors; + for (size_t i = 0; i < interpreter->inputs().size(); i++) + { + int input = interpreter->inputs()[i]; + TfLiteTensor* tensor = interpreter->tensor(input); + if (tensor->type == kTfLiteBool) + { + size_t elements = tensor->bytes / sizeof(bool); + uint8_t* temp_tensor = new uint8_t[tensor->bytes / sizeof(bool)]; + input_boolean_tensors[i] = temp_tensor; + for (size_t idx = 0; idx < elements; idx++) + { + temp_tensor[idx] = (tensor->data.b[idx] ? 0x00 : 0xff); + } + } + } + for (size_t i = 0; i < interpreter->outputs().size(); i++) + { + int output = interpreter->outputs()[i]; + TfLiteTensor* tensor = interpreter->tensor(output); + if (tensor->type == kTfLiteBool) + { + uint8_t* temp_tensor = new uint8_t[tensor->bytes / sizeof(bool)]; + output_boolean_tensors[i] = temp_tensor; + } + } + + // Currently perform deep copy of input buffer + for (size_t i = 0; i < interpreter->inputs().size(); i++) { + int input = interpreter->inputs()[i]; + // TODO(aselle): Is this what we want or do we want input instead? + // TODO(aselle): This should be called setInputValue maybe to be cons. + TfLiteTensor* tensor = interpreter->tensor(input); + if (tensor->type == kTfLiteBool) + { + CHECK_NN(ANeuralNetworksExecution_setInput( + execution, i, nullptr, input_boolean_tensors[i], tensor->bytes * sizeof(uint8_t) / sizeof(bool))); + } + else + { + CHECK_NN(ANeuralNetworksExecution_setInput( + execution, i, nullptr, tensor->data.raw, tensor->bytes)); + } + } + + // Tell nn api where to place final data. + for (size_t i = 0; i < interpreter->outputs().size(); i++) { + int output = interpreter->outputs()[i]; + TfLiteTensor* tensor = interpreter->tensor(output); + + if (tensor->type == kTfLiteBool) + { + CHECK_NN(ANeuralNetworksExecution_setOutput( + execution, i, nullptr, output_boolean_tensors[i], tensor->bytes * sizeof(uint8_t) / sizeof(bool))); + } + else + { + CHECK_NN(ANeuralNetworksExecution_setOutput( + execution, i, nullptr, tensor->data.raw, tensor->bytes)); + } + } + + // The state_out of previous invocation need to be mapped to state_in of + // current invocation. + for (size_t i = 0; i < model_states_outputs_.size(); i++) { + int state_tensor_idx = model_states_outputs_[i]; + TfLiteTensor* tensor = interpreter->tensor(state_tensor_idx); + // Here we are using a deep copy for state_in tensors so that we are not + // reading and writing into the same buffer during a invocation. + // TODO(miaowang): using double shared buffer to minimize the copies. + CHECK_NN(ANeuralNetworksExecution_setInput( + execution, i + interpreter->inputs().size(), nullptr, tensor->data.raw, + tensor->bytes)); + // Tell NNAPI where to output the state_out. + CHECK_NN(ANeuralNetworksExecution_setOutput( + execution, i + interpreter->outputs().size(), nullptr, tensor->data.raw, + tensor->bytes)); + } + + // Currently use blocking compute. + ANeuralNetworksEvent* event = nullptr; + CHECK_NN(ANeuralNetworksExecution_startCompute(execution, &event)); + CHECK_NN(ANeuralNetworksEvent_wait(event)); + ANeuralNetworksEvent_free(event); + ANeuralNetworksExecution_free(execution); + + // Tell nn api where to place final data. + for (size_t i = 0; i < interpreter->inputs().size(); i++) { + int input = interpreter->inputs()[i]; + TfLiteTensor* tensor = interpreter->tensor(input); + + if (tensor->type == kTfLiteBool) + { + uint8_t* temp_tensor = input_boolean_tensors[i]; + input_boolean_tensors[i] = nullptr; + delete temp_tensor; + } + } + for (size_t i = 0; i < interpreter->outputs().size(); i++) { + int output = interpreter->outputs()[i]; + TfLiteTensor* tensor = interpreter->tensor(output); + + if (tensor->type == kTfLiteBool) + { + uint8_t* temp_tensor = output_boolean_tensors[i]; + size_t elements = tensor->bytes / sizeof(bool); + for (size_t idx = 0; idx < elements; idx++) + { + tensor->data.b[idx] = ((temp_tensor[idx] == 0x00) ? false : true); + } + output_boolean_tensors[i] = nullptr; + delete temp_tensor; + } + } + +#if 0 + printf("From the NN API:\n"); + TfLiteTensor* tensor = interpreter->tensor(interpreter->outputs()[0]); + if (float* data = + interpreter->typed_tensor<float>(interpreter->outputs()[0])) { + size_t num = tensor->bytes / sizeof(float); + for (float* p = data; p < data + num; p++) { + printf(" %f", *p); + } + printf("\n"); + } +#endif + + return kTfLiteOk; +} + +bool NNAPIDelegate::IsSupported() { return nnfw::NNAPIExists(); } + +} // namespace tflite +} // namespace nnfw + +// clang-format on diff --git a/libs/tflite/src/ext/nnapi_delegate_ex_AddOpsAndParams_lambda.inc b/libs/tflite/src/ext/nnapi_delegate_ex_AddOpsAndParams_lambda.inc new file mode 100644 index 000000000..a91e4de60 --- /dev/null +++ b/libs/tflite/src/ext/nnapi_delegate_ex_AddOpsAndParams_lambda.inc @@ -0,0 +1,106 @@ +// This file is included from AddOpsAndParams defined in nnapi_delegate.cc +// and contains lambda for extened implementation to original Tensorflow Lite. + auto add_resize_bilinear_params = [&add_scalar_int32, &interpreter, &augmented_inputs](void* data) { + auto builtin = reinterpret_cast<TfLiteResizeBilinearParams*>(data); + if (builtin->align_corners) { + FATAL("Resize bilinear does not support align corners in NNAPI"); + } + + TfLiteTensor* tensor = interpreter->tensor(augmented_inputs.back()); + assert(tensor->type == kTfLiteInt32); + assert(tensor->bytes == sizeof(int)*2); + augmented_inputs.pop_back(); + + int height = ((int*)(tensor->data.raw))[1]; + int width = ((int*)(tensor->data.raw))[0]; + add_scalar_int32(height); + add_scalar_int32(width); + }; + + auto check_l2normalization_params = [interpreter, &node](void* data) { + auto builtin = reinterpret_cast<TfLiteL2NormParams*>(data); + if (builtin->activation != kTfLiteActNone) { + FATAL("NNAPI does not support L2Normalization with fused activations"); + } + if ((node.inputs->size > 0) && + (interpreter->tensor(node.inputs->data[0])->dims->size != 4)) { + FATAL("NNAPI only supports input rank 4 for L2Normalization"); + } + }; + + auto add_transpose_conv_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteTransposeConvParams*>(data); + add_scalar_int32(builtin->padding); + add_scalar_int32(builtin->stride_width); + add_scalar_int32(builtin->stride_height); + }; + + auto add_lrn_params = [&add_scalar_int32, + &add_scalar_float32](void* data) { + auto builtin = reinterpret_cast<TfLiteLocalResponseNormParams*>(data); + add_scalar_int32(builtin->radius); + add_scalar_float32(builtin->bias); + add_scalar_float32(builtin->alpha); + add_scalar_float32(builtin->beta); + }; + + auto add_strided_slice_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteStridedSliceParams*>(data); + add_scalar_int32(builtin->begin_mask); + add_scalar_int32(builtin->end_mask); + // ellipsis_mask and new_axis_mask are not supported on nn runtime + // cf) tflite interpreter supports both operations + if (builtin->ellipsis_mask) { + FATAL("STRIDE_SLICE does not support ellipsis_mask in NNAPI"); + } + if (builtin->new_axis_mask) { + FATAL("STRIDE_SLICE does not support new_axis_mask in NNAPI"); + } + add_scalar_int32(builtin->shrink_axis_mask); + }; + + auto add_gather_ex_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteGatherParams*>(data); + add_scalar_int32(builtin->axis); + if (builtin->axis != 0) { + FATAL("GATHER does not support axis>0 in NNAPI"); + } + }; + +#ifndef OBS_BUILD + auto add_pack_ex_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLitePackParams*>(data); + add_scalar_int32(builtin->values_count); + add_scalar_int32(builtin->axis); + }; + + auto add_unpack_ex_params = [&add_scalar_int32](void* data) { + auto builtin = reinterpret_cast<TfLiteUnpackParams*>(data); + add_scalar_int32(builtin->num); + add_scalar_int32(builtin->axis); + }; +#endif + + auto check_batch_to_space_params = [interpreter, &node, &augmented_inputs]() { + + //If there are 3 inputs, check if crops is having default values {0, 0, 0, 0} + //Else unsupported by NNAPI + + if(augmented_inputs.size() == 3) + { + const uint32_t crops_buffer_index = node.inputs->data[2]; + const TfLiteTensor* crops = interpreter->tensor(crops_buffer_index); + const int *crops_value = crops->data.i32; + + //Check if crops is having default values {0, 0, 0, 0} + if(crops_value[0] != 0 || crops_value[1] != 0 || crops_value[2] != 0 || crops_value[3] != 0) + { + FATAL("BATCH_TO_SPACE_ND does not support Explicit crops in NNAPI"); + } + else + { + //Restrict crops input and pass only other two inputs + augmented_inputs.pop_back(); + } + } + }; diff --git a/libs/tflite/src/interp/FlatBufferBuilder.cpp b/libs/tflite/src/interp/FlatBufferBuilder.cpp new file mode 100644 index 000000000..4b9cde719 --- /dev/null +++ b/libs/tflite/src/interp/FlatBufferBuilder.cpp @@ -0,0 +1,40 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/interp/FlatBufferBuilder.h" + +#include "tflite/ext/kernels/register.h" + +namespace nnfw +{ +namespace tflite +{ + +std::unique_ptr<::tflite::Interpreter> FlatBufferBuilder::build(void) const +{ + std::unique_ptr<::tflite::Interpreter> interpreter; + + nnfw::tflite::BuiltinOpResolver resolver; + + ::tflite::InterpreterBuilder builder(_model, resolver); + + builder(&interpreter); + + return std::move(interpreter); +} + +} // namespace tflite +} // namespace nnfw diff --git a/libs/tflite/src/interp/FunctionBuilder.cpp b/libs/tflite/src/interp/FunctionBuilder.cpp new file mode 100644 index 000000000..eab940c18 --- /dev/null +++ b/libs/tflite/src/interp/FunctionBuilder.cpp @@ -0,0 +1,34 @@ +/* + * Copyright (c) 2018 Samsung Electronics Co., Ltd. All Rights Reserved + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#include "tflite/interp/FunctionBuilder.h" + +namespace nnfw +{ +namespace tflite +{ + +std::unique_ptr<::tflite::Interpreter> FunctionBuilder::build(void) const +{ + auto res = std::unique_ptr<::tflite::Interpreter>{new ::tflite::Interpreter}; + + _fn(*res); + + return std::move(res); +} + +} // namespace tflite +} // namespace nnfw |