From 534556391ae153b1e8fdd11b233e2dc6ecd1e530 Mon Sep 17 00:00:00 2001 From: Michael Butler Date: Mon, 25 Oct 2021 11:20:33 -0700 Subject: Create NN AIDL adapter This change adds the following adapters: * nn::IDevice -> BnDevice * nn::IPreparedModel -> BnPreparedModel * nn::IBurst -> BnBurst * nn::IBuffer -> BnBuffer Bug: N/A Test: mma Test: locally created a binderized service with this adapter code, which passed VtsHalNeuralnetworksTargetTest Change-Id: I966f65a1e4d75284c050b77f3f40c515e4970130 --- .../utils/adapter/aidl/src/PreparedModel.cpp | 225 +++++++++++++++++++++ 1 file changed, 225 insertions(+) create mode 100644 neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp (limited to 'neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp') diff --git a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp new file mode 100644 index 0000000000..71ed1a857b --- /dev/null +++ b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp @@ -0,0 +1,225 @@ +/* + * Copyright (C) 2021 The Android Open Source Project + * + * 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 "PreparedModel.h" + +#include "Burst.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +namespace aidl::android::hardware::neuralnetworks::adapter { +namespace { + +class FencedExecutionCallback : public BnFencedExecutionCallback { + public: + FencedExecutionCallback(nn::ExecuteFencedInfoCallback callback) + : kCallback(std::move(callback)) {} + + ndk::ScopedAStatus getExecutionInfo(Timing* timingLaunched, Timing* timingFenced, + ErrorStatus* errorStatus) override { + const auto result = kCallback(); + if (result.ok()) { + const auto& [nnTimingLaunched, nnTimingFenced] = result.value(); + *timingLaunched = utils::convert(nnTimingLaunched).value(); + *timingFenced = utils::convert(nnTimingFenced).value(); + *errorStatus = ErrorStatus::NONE; + } else { + constexpr auto kNoTiming = Timing{.timeOnDeviceNs = -1, .timeInDriverNs = -1}; + const auto& [message, code] = result.error(); + LOG(ERROR) << "getExecutionInfo failed with " << code << ": " << message; + const auto aidlStatus = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + *timingLaunched = kNoTiming; + *timingFenced = kNoTiming; + *errorStatus = aidlStatus; + } + return ndk::ScopedAStatus::ok(); + } + + private: + const nn::ExecuteFencedInfoCallback kCallback; +}; + +template +auto convertInput(const Type& object) -> decltype(nn::convert(std::declval())) { + auto result = nn::convert(object); + if (!result.has_value()) { + result.error().code = nn::ErrorStatus::INVALID_ARGUMENT; + } + return result; +} + +nn::GeneralResult> convertSyncFences( + const std::vector& waitFor) { + auto handles = NN_TRY(convertInput(waitFor)); + + constexpr auto valid = [](const nn::SharedHandle& handle) { + return handle != nullptr && handle->ok(); + }; + if (!std::all_of(handles.begin(), handles.end(), valid)) { + return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid sync fence"; + } + + std::vector syncFences; + syncFences.reserve(waitFor.size()); + for (auto& handle : handles) { + syncFences.push_back(nn::SyncFence::create(std::move(handle)).value()); + } + return syncFences; +} + +nn::Duration makeDuration(int64_t durationNs) { + return nn::Duration(std::chrono::nanoseconds(durationNs)); +} + +nn::GeneralResult makeOptionalDuration(int64_t durationNs) { + if (durationNs < -1) { + return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid duration " << durationNs; + } + return durationNs < 0 ? nn::OptionalDuration{} : makeDuration(durationNs); +} + +nn::GeneralResult makeOptionalTimePoint(int64_t durationNs) { + if (durationNs < -1) { + return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid time point " << durationNs; + } + return durationNs < 0 ? nn::OptionalTimePoint{} : nn::TimePoint(makeDuration(durationNs)); +} + +nn::ExecutionResult executeSynchronously(const nn::IPreparedModel& preparedModel, + const Request& request, + bool measureTiming, int64_t deadlineNs, + int64_t loopTimeoutDurationNs) { + const auto nnRequest = NN_TRY(convertInput(request)); + const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO; + const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs)); + const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs)); + + const auto result = + preparedModel.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration); + + if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) { + const auto& [message, code, outputShapes] = result.error(); + LOG(ERROR) << "executeSynchronously failed with " << code << ": " << message; + return ExecutionResult{.outputSufficientSize = false, + .outputShapes = utils::convert(outputShapes).value(), + .timing = {.timeInDriverNs = -1, .timeOnDeviceNs = -1}}; + } + + const auto& [outputShapes, timing] = NN_TRY(result); + return ExecutionResult{.outputSufficientSize = true, + .outputShapes = utils::convert(outputShapes).value(), + .timing = utils::convert(timing).value()}; +} + +nn::GeneralResult executeFenced( + const nn::IPreparedModel& preparedModel, const Request& request, + const std::vector& waitFor, bool measureTiming, + int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs) { + const auto nnRequest = NN_TRY(convertInput(request)); + const auto nnWaitFor = NN_TRY(convertSyncFences(waitFor)); + const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO; + const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs)); + const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs)); + const auto nnDuration = NN_TRY(makeOptionalDuration(durationNs)); + + auto [syncFence, executeFencedInfoCallback] = NN_TRY(preparedModel.executeFenced( + nnRequest, nnWaitFor, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, nnDuration)); + + ndk::ScopedFileDescriptor fileDescriptor; + if (syncFence.hasFd()) { + auto uniqueFd = NN_TRY(nn::dupFd(syncFence.getFd())); + fileDescriptor = ndk::ScopedFileDescriptor(uniqueFd.release()); + } + + return FencedExecutionResult{.callback = ndk::SharedRefBase::make( + std::move(executeFencedInfoCallback)), + .syncFence = std::move(fileDescriptor)}; +} + +} // namespace + +PreparedModel::PreparedModel(nn::SharedPreparedModel preparedModel) + : kPreparedModel(std::move(preparedModel)) { + CHECK(kPreparedModel != nullptr); +} + +ndk::ScopedAStatus PreparedModel::executeSynchronously(const Request& request, bool measureTiming, + int64_t deadlineNs, + int64_t loopTimeoutDurationNs, + ExecutionResult* executionResult) { + auto result = adapter::executeSynchronously(*kPreparedModel, request, measureTiming, deadlineNs, + loopTimeoutDurationNs); + if (!result.has_value()) { + const auto& [message, code, _] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus PreparedModel::executeFenced( + const Request& request, const std::vector& waitFor, + bool measureTiming, int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs, + FencedExecutionResult* executionResult) { + auto result = adapter::executeFenced(*kPreparedModel, request, waitFor, measureTiming, + deadlineNs, loopTimeoutDurationNs, durationNs); + if (!result.has_value()) { + const auto& [message, code] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus PreparedModel::configureExecutionBurst(std::shared_ptr* burst) { + auto result = kPreparedModel->configureExecutionBurst(); + if (!result.has_value()) { + const auto& [message, code] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *burst = ndk::SharedRefBase::make(std::move(result).value()); + return ndk::ScopedAStatus::ok(); +} + +nn::SharedPreparedModel PreparedModel::getUnderlyingPreparedModel() const { + return kPreparedModel; +} + +} // namespace aidl::android::hardware::neuralnetworks::adapter -- cgit v1.2.3 From 7f5c7d293c2dad462dc9c0f1f1a160fb2c2c9a9b Mon Sep 17 00:00:00 2001 From: Xusong Wang Date: Tue, 5 Oct 2021 14:10:41 -0700 Subject: Reusable execution at HAL level -- HAL. This CL modifies the canonical/AIDL adapter to use IExecution object if available. Bug: 202405342 Bug: 202431255 Test: NNT_static Test: CtsNNAPITestCases Test: VtsHalNeuralnetworksTargetTest Change-Id: I6aac3c57f97ac87a5ba3f78cfd843fcc403decff --- .../utils/adapter/aidl/src/PreparedModel.cpp | 99 ++++++++++++++++++++++ 1 file changed, 99 insertions(+) (limited to 'neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp') diff --git a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp index 71ed1a857b..5cab62c625 100644 --- a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp +++ b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp @@ -17,6 +17,7 @@ #include "PreparedModel.h" #include "Burst.h" +#include "Execution.h" #include #include @@ -26,6 +27,7 @@ #include #include #include +#include #include #include #include @@ -167,6 +169,56 @@ nn::GeneralResult executeFenced( .syncFence = std::move(fileDescriptor)}; } +nn::GeneralResult createReusableExecution( + const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming, + int64_t loopTimeoutDurationNs) { + const auto nnRequest = NN_TRY(convertInput(request)); + const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO; + const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs)); + return preparedModel.createReusableExecution(nnRequest, nnMeasureTiming, nnLoopTimeoutDuration); +} + +nn::ExecutionResult executeSynchronously(const nn::IExecution& execution, + int64_t deadlineNs) { + const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs)); + + const auto result = execution.compute(nnDeadline); + + if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) { + const auto& [message, code, outputShapes] = result.error(); + LOG(ERROR) << "executeSynchronously failed with " << code << ": " << message; + return ExecutionResult{.outputSufficientSize = false, + .outputShapes = utils::convert(outputShapes).value(), + .timing = {.timeInDriverNs = -1, .timeOnDeviceNs = -1}}; + } + + const auto& [outputShapes, timing] = NN_TRY(result); + return ExecutionResult{.outputSufficientSize = true, + .outputShapes = utils::convert(outputShapes).value(), + .timing = utils::convert(timing).value()}; +} + +nn::GeneralResult executeFenced( + const nn::IExecution& execution, const std::vector& waitFor, + int64_t deadlineNs, int64_t durationNs) { + const auto nnWaitFor = NN_TRY(convertSyncFences(waitFor)); + const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs)); + const auto nnDuration = NN_TRY(makeOptionalDuration(durationNs)); + + auto [syncFence, executeFencedInfoCallback] = + NN_TRY(execution.computeFenced(nnWaitFor, nnDeadline, nnDuration)); + + ndk::ScopedFileDescriptor fileDescriptor; + if (syncFence.hasFd()) { + auto uniqueFd = NN_TRY(nn::dupFd(syncFence.getFd())); + fileDescriptor = ndk::ScopedFileDescriptor(uniqueFd.release()); + } + + return FencedExecutionResult{.callback = ndk::SharedRefBase::make( + std::move(executeFencedInfoCallback)), + .syncFence = std::move(fileDescriptor)}; +} + } // namespace PreparedModel::PreparedModel(nn::SharedPreparedModel preparedModel) @@ -222,4 +274,51 @@ nn::SharedPreparedModel PreparedModel::getUnderlyingPreparedModel() const { return kPreparedModel; } +ndk::ScopedAStatus PreparedModel::createReusableExecution(const Request& request, + bool measureTiming, + int64_t loopTimeoutDurationNs, + std::shared_ptr* execution) { + auto result = adapter::createReusableExecution(*kPreparedModel, request, measureTiming, + loopTimeoutDurationNs); + if (!result.has_value()) { + const auto& [message, code] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *execution = ndk::SharedRefBase::make(std::move(result).value()); + return ndk::ScopedAStatus::ok(); +} + +Execution::Execution(nn::SharedExecution execution) : kExecution(std::move(execution)) { + CHECK(kExecution != nullptr); +} + +ndk::ScopedAStatus Execution::executeSynchronously(int64_t deadlineNs, + ExecutionResult* executionResult) { + auto result = adapter::executeSynchronously(*kExecution, deadlineNs); + if (!result.has_value()) { + const auto& [message, code, _] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus Execution::executeFenced(const std::vector& waitFor, + int64_t deadlineNs, int64_t durationNs, + FencedExecutionResult* executionResult) { + auto result = adapter::executeFenced(*kExecution, waitFor, deadlineNs, durationNs); + if (!result.has_value()) { + const auto& [message, code] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + } // namespace aidl::android::hardware::neuralnetworks::adapter -- cgit v1.2.3