diff options
Diffstat (limited to 'neuralnetworks/utils/adapter/aidl')
6 files changed, 140 insertions, 33 deletions
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h index f2687c4a69..8d42e2fc7f 100644 --- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h +++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h @@ -46,6 +46,11 @@ class Burst : public BnBurst { bool measureTiming, int64_t deadlineNs, int64_t loopTimeoutDurationNs, ExecutionResult* executionResult) override; + ndk::ScopedAStatus executeSynchronouslyWithConfig( + const Request& request, const std::vector<int64_t>& memoryIdentifierTokens, + const ExecutionConfig& config, int64_t deadlineNs, + ExecutionResult* executionResult) override; + ndk::ScopedAStatus releaseMemoryResource(int64_t memoryIdentifierToken) override; class ThreadSafeMemoryCache { diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h index aa29d63b6b..c94f27085f 100644 --- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h +++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h @@ -31,6 +31,7 @@ #include <aidl/android/hardware/neuralnetworks/IPreparedModelParcel.h> #include <aidl/android/hardware/neuralnetworks/Model.h> #include <aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.h> +#include <aidl/android/hardware/neuralnetworks/PrepareModelConfig.h> #include <aidl/android/hardware/neuralnetworks/Priority.h> #include <android/binder_auto_utils.h> #include <nnapi/IDevice.h> @@ -72,6 +73,9 @@ class Device : public BnDevice { const std::vector<ndk::ScopedFileDescriptor>& dataCache, const std::vector<uint8_t>& token, const std::shared_ptr<IPreparedModelCallback>& callback) override; + ndk::ScopedAStatus prepareModelWithConfig( + const Model& model, const PrepareModelConfig& config, + const std::shared_ptr<IPreparedModelCallback>& callback) override; protected: const ::android::nn::SharedDevice kDevice; diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h index f92b0bc783..d1359d6ac1 100644 --- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h +++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h @@ -51,9 +51,17 @@ class PreparedModel : public BnPreparedModel { int64_t loopTimeoutDurationNs, int64_t durationNs, FencedExecutionResult* executionResult) override; ndk::ScopedAStatus configureExecutionBurst(std::shared_ptr<IBurst>* burst) override; - ndk::ScopedAStatus createReusableExecution(const Request& request, bool measureTiming, - int64_t loopTimeoutDurationNs, + ndk::ScopedAStatus createReusableExecution(const Request& request, + const ExecutionConfig& config, std::shared_ptr<IExecution>* execution) override; + ndk::ScopedAStatus executeSynchronouslyWithConfig(const Request& request, + const ExecutionConfig& config, + int64_t deadlineNs, + ExecutionResult* executionResult) override; + ndk::ScopedAStatus executeFencedWithConfig( + const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor, + const ExecutionConfig& config, int64_t deadlineNs, int64_t durationNs, + FencedExecutionResult* executionResult) override; ::android::nn::SharedPreparedModel getUnderlyingPreparedModel() const; diff --git a/neuralnetworks/utils/adapter/aidl/src/Burst.cpp b/neuralnetworks/utils/adapter/aidl/src/Burst.cpp index 4fabb20635..a4a80faf2a 100644 --- a/neuralnetworks/utils/adapter/aidl/src/Burst.cpp +++ b/neuralnetworks/utils/adapter/aidl/src/Burst.cpp @@ -93,7 +93,8 @@ std::vector<nn::IBurst::OptionalCacheHold> ensureAllMemoriesAreCached( nn::ExecutionResult<ExecutionResult> executeSynchronously( const nn::IBurst& burst, const Burst::ThreadSafeMemoryCache& cache, const Request& request, const std::vector<int64_t>& memoryIdentifierTokens, bool measureTiming, int64_t deadlineNs, - int64_t loopTimeoutDurationNs) { + int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints, + const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) { if (request.pools.size() != memoryIdentifierTokens.size()) { return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "request.pools.size() != memoryIdentifierTokens.size()"; @@ -107,11 +108,13 @@ nn::ExecutionResult<ExecutionResult> executeSynchronously( const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO; const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs)); const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs)); + auto nnHints = NN_TRY(convertInput(hints)); + auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix)); const auto hold = ensureAllMemoriesAreCached(&nnRequest, memoryIdentifierTokens, burst, cache); - const auto result = - burst.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration); + const auto result = burst.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, + nnHints, nnExtensionNameToPrefix); if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) { const auto& [message, code, outputShapes] = result.error(); @@ -155,7 +158,24 @@ ndk::ScopedAStatus Burst::executeSynchronously(const Request& request, ExecutionResult* executionResult) { auto result = adapter::executeSynchronously(*kBurst, kMemoryCache, request, memoryIdentifierTokens, - measureTiming, deadlineNs, loopTimeoutDurationNs); + measureTiming, deadlineNs, loopTimeoutDurationNs, {}, {}); + if (!result.has_value()) { + auto [message, code, _] = std::move(result).error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast<int32_t>(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus Burst::executeSynchronouslyWithConfig( + const Request& request, const std::vector<int64_t>& memoryIdentifierTokens, + const ExecutionConfig& config, int64_t deadlineNs, ExecutionResult* executionResult) { + auto result = adapter::executeSynchronously( + *kBurst, kMemoryCache, request, memoryIdentifierTokens, config.measureTiming, + deadlineNs, config.loopTimeoutDurationNs, config.executionHints, + config.extensionNameToPrefix); if (!result.has_value()) { auto [message, code, _] = std::move(result).error(); const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); diff --git a/neuralnetworks/utils/adapter/aidl/src/Device.cpp b/neuralnetworks/utils/adapter/aidl/src/Device.cpp index 763be7f3fa..84aaddbe9d 100644 --- a/neuralnetworks/utils/adapter/aidl/src/Device.cpp +++ b/neuralnetworks/utils/adapter/aidl/src/Device.cpp @@ -148,13 +148,14 @@ void notify(IPreparedModelCallback* callback, PrepareModelResult result) { } } -nn::GeneralResult<void> prepareModel(const nn::SharedDevice& device, const Executor& executor, - const Model& model, ExecutionPreference preference, - Priority priority, int64_t deadlineNs, - const std::vector<ndk::ScopedFileDescriptor>& modelCache, - const std::vector<ndk::ScopedFileDescriptor>& dataCache, - const std::vector<uint8_t>& token, - const std::shared_ptr<IPreparedModelCallback>& callback) { +nn::GeneralResult<void> prepareModel( + const nn::SharedDevice& device, const Executor& executor, const Model& model, + ExecutionPreference preference, Priority priority, int64_t deadlineNs, + const std::vector<ndk::ScopedFileDescriptor>& modelCache, + const std::vector<ndk::ScopedFileDescriptor>& dataCache, const std::vector<uint8_t>& token, + const std::vector<TokenValuePair>& hints, + const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix, + const std::shared_ptr<IPreparedModelCallback>& callback) { if (callback.get() == nullptr) { return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid callback"; } @@ -166,12 +167,16 @@ nn::GeneralResult<void> prepareModel(const nn::SharedDevice& device, const Execu auto nnModelCache = NN_TRY(convertInput(modelCache)); auto nnDataCache = NN_TRY(convertInput(dataCache)); const auto nnToken = NN_TRY(convertCacheToken(token)); + auto nnHints = NN_TRY(convertInput(hints)); + auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix)); Task task = [device, nnModel = std::move(nnModel), nnPreference, nnPriority, nnDeadline, nnModelCache = std::move(nnModelCache), nnDataCache = std::move(nnDataCache), - nnToken, callback] { - auto result = device->prepareModel(nnModel, nnPreference, nnPriority, nnDeadline, - nnModelCache, nnDataCache, nnToken); + nnToken, nnHints = std::move(nnHints), + nnExtensionNameToPrefix = std::move(nnExtensionNameToPrefix), callback] { + auto result = + device->prepareModel(nnModel, nnPreference, nnPriority, nnDeadline, nnModelCache, + nnDataCache, nnToken, nnHints, nnExtensionNameToPrefix); notify(callback.get(), std::move(result)); }; executor(std::move(task), nnDeadline); @@ -273,8 +278,9 @@ ndk::ScopedAStatus Device::prepareModel(const Model& model, ExecutionPreference const std::vector<ndk::ScopedFileDescriptor>& dataCache, const std::vector<uint8_t>& token, const std::shared_ptr<IPreparedModelCallback>& callback) { - const auto result = adapter::prepareModel(kDevice, kExecutor, model, preference, priority, - deadlineNs, modelCache, dataCache, token, callback); + const auto result = + adapter::prepareModel(kDevice, kExecutor, model, preference, priority, deadlineNs, + modelCache, dataCache, token, {}, {}, callback); if (!result.has_value()) { const auto& [message, code] = result.error(); const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); @@ -301,4 +307,21 @@ ndk::ScopedAStatus Device::prepareModelFromCache( return ndk::ScopedAStatus::ok(); } +ndk::ScopedAStatus Device::prepareModelWithConfig( + const Model& model, const PrepareModelConfig& config, + const std::shared_ptr<IPreparedModelCallback>& callback) { + const auto result = adapter::prepareModel( + kDevice, kExecutor, model, config.preference, config.priority, config.deadlineNs, + config.modelCache, config.dataCache, config.cacheToken, config.compilationHints, + config.extensionNameToPrefix, callback); + if (!result.has_value()) { + const auto& [message, code] = result.error(); + const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); + callback->notify(aidlCode, nullptr); + return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage( + static_cast<int32_t>(aidlCode), message.c_str()); + } + return ndk::ScopedAStatus::ok(); +} + } // namespace aidl::android::hardware::neuralnetworks::adapter diff --git a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp index 5cab62c625..790558fde4 100644 --- a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp +++ b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp @@ -118,17 +118,20 @@ nn::GeneralResult<nn::OptionalTimePoint> makeOptionalTimePoint(int64_t durationN return durationNs < 0 ? nn::OptionalTimePoint{} : nn::TimePoint(makeDuration(durationNs)); } -nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IPreparedModel& preparedModel, - const Request& request, - bool measureTiming, int64_t deadlineNs, - int64_t loopTimeoutDurationNs) { +nn::ExecutionResult<ExecutionResult> executeSynchronously( + const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming, + int64_t deadlineNs, int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints, + const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) { 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)); + auto nnHints = NN_TRY(convertInput(hints)); + auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix)); const auto result = - preparedModel.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration); + preparedModel.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, + nnHints, nnExtensionNameToPrefix); if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) { const auto& [message, code, outputShapes] = result.error(); @@ -147,16 +150,21 @@ nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IPreparedMod nn::GeneralResult<FencedExecutionResult> executeFenced( const nn::IPreparedModel& preparedModel, const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor, bool measureTiming, - int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs) { + int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs, + const std::vector<TokenValuePair>& hints, + const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) { 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 nnHints = NN_TRY(convertInput(hints)); + auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix)); auto [syncFence, executeFencedInfoCallback] = NN_TRY(preparedModel.executeFenced( - nnRequest, nnWaitFor, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, nnDuration)); + nnRequest, nnWaitFor, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, nnDuration, + nnHints, nnExtensionNameToPrefix)); ndk::ScopedFileDescriptor fileDescriptor; if (syncFence.hasFd()) { @@ -171,11 +179,16 @@ nn::GeneralResult<FencedExecutionResult> executeFenced( nn::GeneralResult<nn::SharedExecution> createReusableExecution( const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming, - int64_t loopTimeoutDurationNs) { + int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints, + const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) { 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); + auto nnHints = NN_TRY(convertInput(hints)); + auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix)); + + return preparedModel.createReusableExecution(nnRequest, nnMeasureTiming, nnLoopTimeoutDuration, + nnHints, nnExtensionNameToPrefix); } nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IExecution& execution, @@ -231,7 +244,7 @@ ndk::ScopedAStatus PreparedModel::executeSynchronously(const Request& request, b int64_t loopTimeoutDurationNs, ExecutionResult* executionResult) { auto result = adapter::executeSynchronously(*kPreparedModel, request, measureTiming, deadlineNs, - loopTimeoutDurationNs); + loopTimeoutDurationNs, {}, {}); if (!result.has_value()) { const auto& [message, code, _] = result.error(); const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); @@ -247,7 +260,41 @@ ndk::ScopedAStatus PreparedModel::executeFenced( bool measureTiming, int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs, FencedExecutionResult* executionResult) { auto result = adapter::executeFenced(*kPreparedModel, request, waitFor, measureTiming, - deadlineNs, loopTimeoutDurationNs, durationNs); + 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<int32_t>(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus PreparedModel::executeSynchronouslyWithConfig(const Request& request, + const ExecutionConfig& config, + int64_t deadlineNs, + ExecutionResult* executionResult) { + auto result = adapter::executeSynchronously( + *kPreparedModel, request, config.measureTiming, deadlineNs, + config.loopTimeoutDurationNs, config.executionHints, config.extensionNameToPrefix); + 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<int32_t>(aidlCode), message.c_str()); + } + *executionResult = std::move(result).value(); + return ndk::ScopedAStatus::ok(); +} + +ndk::ScopedAStatus PreparedModel::executeFencedWithConfig( + const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor, + const ExecutionConfig& config, int64_t deadlineNs, int64_t durationNs, + FencedExecutionResult* executionResult) { + auto result = adapter::executeFenced(*kPreparedModel, request, waitFor, config.measureTiming, + deadlineNs, config.loopTimeoutDurationNs, durationNs, + config.executionHints, config.extensionNameToPrefix); if (!result.has_value()) { const auto& [message, code] = result.error(); const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); @@ -275,11 +322,11 @@ nn::SharedPreparedModel PreparedModel::getUnderlyingPreparedModel() const { } ndk::ScopedAStatus PreparedModel::createReusableExecution(const Request& request, - bool measureTiming, - int64_t loopTimeoutDurationNs, + const ExecutionConfig& config, std::shared_ptr<IExecution>* execution) { - auto result = adapter::createReusableExecution(*kPreparedModel, request, measureTiming, - loopTimeoutDurationNs); + auto result = adapter::createReusableExecution( + *kPreparedModel, request, config.measureTiming, config.loopTimeoutDurationNs, + config.executionHints, config.extensionNameToPrefix); if (!result.has_value()) { const auto& [message, code] = result.error(); const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE); |