xref: /frameworks/ml/nn/runtime/VersionedInterfaces.h

/*
 * Copyright (C) 2018 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.
 */

#ifndef ANDROID_FRAMEWORKS_ML_NN_RUNTIME_VERSIONED_INTERFACES_H
#define ANDROID_FRAMEWORKS_ML_NN_RUNTIME_VERSIONED_INTERFACES_H

#include <android-base/macros.h>

#include <cstddef>
#include <functional>
#include <memory>
#include <optional>
#include <shared_mutex>
#include <string>
#include <tuple>
#include <utility>
#include <vector>

#include "Callbacks.h"
#include "HalInterfaces.h"
#include "Utils.h"

namespace android {
namespace nn {

// forward declarations
class ExecutionBurstController;
class IDeviceDeathHandler;
class IPreparedModelDeathHandler;
class MetaModel;
class VersionedIPreparedModel;

/**
 * Each class (VersionedIDevice, VersionedIPreparedModel) wraps a HIDL interface
 * of any version to abstract away version differences. It allows the remainder
 * of the runtime to always use the most up-to-date version of all HIDL types.
 * As such, any reference to a HIDL type in the rest of the runtime
 * will--by default--be the latest HIDL version.
 *
 * Each class will attempt to call the latest version of each interface method
 * if possible. If the latest method is unavailable, the versioned class
 * will attempt to upcast the type (e.g., V1_1::Model to V1_0::Model), and
 * invoke the latest interface method possible. If the versioned class
 * fails to find a matching applicable function, it will return an error.
 */

/** This class wraps an IDevice object of any version. */
class VersionedIDevice {
    DISALLOW_IMPLICIT_CONSTRUCTORS(VersionedIDevice);

    // forward declaration of nested class
    class Core;

   public:
    /**
     * Create a VersionedIDevice object.
     *
     * Prefer using this function over the constructor, as it adds more
     * protections.
     *
     * @param serviceName The name of the service that provides "device".
     * @param makeDevice A device factory function that returns a device object
     *                   that is at least version 1.0 of the IDevice interface.
     * @return A valid VersionedIDevice object, otherwise nullptr.
     */
    static std::shared_ptr<VersionedIDevice> create(std::string serviceName,
                                                    const hal::DeviceFactory& makeDevice);

    /**
     * Constructor for the VersionedIDevice object.
     *
     * VersionedIDevice will default to using the latest version of all IDevice
     * interface methods automatically.
     *
     * @param capabilities Performance capabilities of the driver.
     * @param supportedExtensions Extensions supported by the driver.
     * @param type The device type of the driver.
     * @param versionString The version string of the driver.
     * @param numberOfCacheFilesNeeded Number of model cache and data cache
     *     files needed by the driver.
     * @param serviceName The name of the service that provides core.getDevice<V1_0::IDevice>().
     * @param makeDevice A device factory function that returns a device object
     *                   that is at least version 1.0 of the IDevice interface.
     * @param core An object that encapsulates a V1_0::IDevice, any appropriate downcasts to
     *             newer interfaces, and a hidl_death_recipient that will proactively handle
     *             the case when the service containing the IDevice object crashes.
     */
    VersionedIDevice(hal::Capabilities capabilities,
                     std::vector<hal::Extension> supportedExtensions, int32_t type,
                     std::string versionString,
                     std::pair<uint32_t, uint32_t> numberOfCacheFilesNeeded,
                     std::string serviceName, const hal::DeviceFactory& makeDevice, Core core);

    /**
     * Gets the capabilities of a driver.
     *
     * @return capabilities Capabilities of the driver.
     */
    const hal::Capabilities& getCapabilities() const;

    /**
     * Gets information about extensions supported by the driver implementation.
     *
     * Extensions of category ExtensionCategory::BASE must not appear
     * in the list.
     *
     * All extension operations and operands must be fully supported for the
     * extension to appear in the list of supported extensions.
     *
     * @return extensions A list of supported extensions.
     */
    const std::vector<hal::Extension>& getSupportedExtensions() const;

    /**
     * Gets the supported operations in a MetaModel.
     *
     * getSupportedOperations indicates which operations of
     * MetaModel::getModel() are fully supported by the vendor driver. If an
     * operation may not be supported for any reason, getSupportedOperations
     * must return false for that operation.
     *
     * @param metaModel A MetaModel whose operations--and their corresponding
     *                  operands--are to be verified by the driver.  When
     *                  metaModel.getModel() is not compliant with the HAL
     *                  version of the vendor driver, the MetaModel's slicing
     *                  functionality (MetaModel::getSlice*()) is employed
     *                  to query the vendor driver about which of the subset of
     *                  compliant operations are supported.  See the MetaModel
     *                  class in MetaModel.h for more details.
     * @return status Error status of the call, must be:
     *                - NONE if successful
     *                - DEVICE_UNAVAILABLE if driver is offline or busy
     *                - GENERAL_FAILURE if there is an unspecified error
     *                - INVALID_ARGUMENT if provided model is invalid
     * @return supportedOperations A list of supported operations, where true
     *                             indicates the operation is supported and
     *                             false indicates the operation is not
     *                             supported. The index of "supported"
     *                             corresponds with the index of the operation
     *                             it is describing.
     */
    std::pair<hal::ErrorStatus, hal::hidl_vec<bool>> getSupportedOperations(
            const MetaModel& metaModel) const;

    /**
     * Creates a prepared model for execution.
     *
     * prepareModel is used to make any necessary transformations or alternative
     * representations to a model for execution, possibly including
     * transformations on the constant data, optimization on the model's graph,
     * or compilation into the device's native binary format. The model itself
     * is not changed.
     *
     * Optionally, caching information may be provided for the driver to either:
     * - load the prepared model from cache, bypassing full model preparation
     * - save the prepared model to cache for faster model compilation time when
     *     the same model preparation is requested in the future
     *
     * The prepareModel function must verify the inputs to the prepareModel
     * function are correct. If there is an error, prepareModel must immediately
     * return the appropriate result code and nullptr for the
     * VersionedIPreparedModel. If the inputs to the prepareModel function are
     * valid and there is no error, prepareModel must prepare the model.
     *
     * If the model was prepared successfully, prepareModel must return
     * ANEURALNETWORKS_NO_ERROR and the produced VersionedIPreparedModel object.
     * If an error occurred preparing the model, prepareModel must return the
     * appropriate result code and nullptr for the VersionedIPreparedModel.
     *
     * The only information that may be unknown to the model at this stage is
     * the shape of the tensors, which may only be known at execution time. As
     * such, some driver services may return partially prepared models, where
     * the prepared model may only be finished when it is paired with a set of
     * inputs to the model. Note that the same prepared model object may be
     * used with different shapes of inputs on different (possibly concurrent)
     * executions.
     *
     * Multiple threads may call prepareModel on the same model concurrently.
     *
     * @param makeModel Factory function to create the model to be prepared for
     *     execution.
     * @param preference Indicates the intended execution behavior of a prepared
     *     model.
     * @param priority Priority of the prepared model relative to other prepared
     *     models owned by an application.
     * @param deadline Optional time point. If provided, prepareModel is
     *     expected to complete by this time point. If it is not able to be
     *     completed by the deadline, the execution may be aborted.
     * @param cacheDir String specifying the cache directory.
     * @param maybeToken An optional caching token of length
     *     Constant::BYTE_SIZE_OF_CACHE_TOKEN identifying the prepared model.
     *     The same token will be provided when retrieving the prepared model
     *     from the cache files with prepareModelFromCache. Tokens should be
     *     chosen to have a low rate of collision for a particular application.
     *     The driver cannot detect a collision; a collision will result in a
     *     failed execution or in a successful execution that produces incorrect
     *     output values. If both modelCache and dataCache are empty indicating
     *     that caching information is not provided, this token must be ignored.
     * @return A pair of:
     *     - Result code of preparing the model; must be:
     *         - ANEURALNETWORKS_NO_ERROR if preparation succeeded
     *         - ANEURALNETWORKS_UNAVAILABLE_DEVICE if driver is offline or busy
     *         - ANEURALNETWORKS_OP_FAILED if there is an unspecified error
     *         - ANEURALNETWORKS_BAD_DATA if one of the input arguments related
     *             to preparing the model is invalid
     *     - preparedModel A VersionedIPreparedModel object representing a model
     *         that has been prepared for execution, else nullptr.
     */
    std::pair<int, std::shared_ptr<VersionedIPreparedModel>> prepareModel(
            const hal::ModelFactory& makeModel, hal::ExecutionPreference preference, hal::Priority,
            const std::optional<Deadline>& deadline, const std::string& cacheDir,
            const std::optional<hal::CacheToken>& maybeToken) const;

    /**
     * Returns the feature level of a driver.
     *
     * @return featureLevel The API level of the most advanced feature this driver implements.
     *                      For example, if the driver implements the features introduced in
     *                      Android P, the value would be 28.
     *                      Return -1 if the driver is offline or busy, or the query resulted in
     *                      an unspecified error.
     */
    int64_t getFeatureLevel() const;

    /**
     * Returns the device type of a driver.
     *
     * @return deviceType The type of a given device, which can help application
     *     developers to distribute Machine Learning workloads and other
     *     workloads such as graphical rendering. E.g., for an app which renders
     *     AR scenes based on real time object detection results, the developer
     *     could choose an ACCELERATOR type device for ML workloads, and reserve
     *     GPU for graphical rendering.
     */
    int32_t getType() const;

    /**
     * Get the version string of the driver implementation.
     *
     * The version string must be a unique token among the set of version strings of
     * drivers of a specific device. The token identifies the device driver's
     * implementation. The token must not be confused with the feature level which is solely
     * defined by the interface version. This API is opaque to the Android framework, but the
     * Android framework may use the information for debugging or to pass on to NNAPI applications.
     *
     * Application developers sometimes have specific requirements to ensure good user experiences,
     * and they need more information to make intelligent decisions when the Android framework
     * cannot. For example, combined with the device name and other information, the token can help
     * NNAPI applications filter devices based on their needs:
     *     - An application demands a certain level of performance, but a specific version of
     *       the driver cannot meet that requirement because of a performance regression.
     *       The application can disallow the driver based on the version provided.
     *     - An application has a minimum precision requirement, but certain versions of
     *       the driver cannot meet that requirement because of bugs or certain optimizations.
     *       The application can filter out versions of these drivers.
     *
     * @return version The version string of the device implementation.
     */
    const std::string& getVersionString() const;

    /**
     * Gets the caching requirements of the driver implementation.
     *
     * There are two types of cache file descriptors provided to the driver: model cache
     * and data cache.
     *
     * The data cache is for caching constant data, possibly including preprocessed
     * and transformed tensor buffers. Any modification to the data cache should
     * have no worse effect than generating bad output values at execution time.
     *
     * The model cache is for caching security-sensitive data such as compiled
     * executable machine code in the device's native binary format. A modification
     * to the model cache may affect the driver's execution behavior, and a malicious
     * client could make use of this to execute beyond the granted permission. Thus,
     * the driver must always check whether the model cache is corrupted before
     * preparing the model from cache.
     *
     * getNumberOfCacheFilesNeeded returns how many of each type of cache files the driver
     * implementation needs to cache a single prepared model. Returning 0 for both types
     * indicates compilation caching is not supported by this driver. The driver may
     * still choose not to cache certain compiled models even if it reports that caching
     * is supported.
     *
     * If the device reports that caching is not supported, the user may avoid calling
     * IDevice::prepareModelFromCache or providing cache file descriptors to
     * IDevice::prepareModel_1_2.
     *
     * @return numModelCache An unsigned integer indicating how many files for model cache
     *                       the driver needs to cache a single prepared model. It must
     *                       be less than or equal to Constant::MAX_NUMBER_OF_CACHE_FILES.
     * @return numDataCache An unsigned integer indicating how many files for data cache
     *                      the driver needs to cache a single prepared model. It must
     *                      be less than or equal to Constant::MAX_NUMBER_OF_CACHE_FILES.
     */
    std::pair<uint32_t, uint32_t> getNumberOfCacheFilesNeeded() const;

    /**
     * Returns the name of the service.
     *
     * @return Name of the service.
     */
    const std::string& getName() const;

    /**
     * Allocates a driver-managed buffer with the properties specified by the descriptor as well as
     * the input and output roles of prepared models.
     *
     * The allocate function must verify the inputs to the allocate function are correct. If there
     * is an error, or if a certain role or property is not supported by the driver, the allocate
     * function must return with an appropriate ErrorStatus, a nullptr as the IBuffer, and 0 as the
     * buffer token. If the allocation is successful, this method must return with ErrorStatus::NONE
     * and the produced IBuffer with a positive token identifying the allocated buffer. A successful
     * allocation must accommodate all of the specified roles and buffer properties.
     *
     * The buffer is allocated as an uninitialized state. An uninitialized buffer may only be used
     * in ways that are specified by outputRoles. A buffer is initialized after it is used as an
     * output in a successful execution, or after a successful invocation of IBuffer::copyFrom on
     * the buffer. An initialized buffer may be used according to all roles specified in inputRoles
     * and outputRoles. A buffer will return to the uninitialized state if it is used as an output
     * in a failed execution, or after a failed invocation of IBuffer::copyFrom on the buffer.
     *
     * The driver may deduce the dimensions of the buffer according to the buffer descriptor as
     * well as the input and output roles. The dimensions or rank of the buffer may be unknown at
     * this stage. As such, some driver services may only create a placeholder and defer the actual
     * allocation until execution time. Note that the same buffer may be used for different shapes
     * of outputs on different executions. When the buffer is used as an input, the input shape
     * must be the same as the output shape from the last execution using this buffer as an output.
     *
     * The driver must apply proper validatation upon every usage of the buffer, and fail the
     * execution immediately if the usage is illegal.
     *
     * @param desc A buffer descriptor specifying the properties of the buffer to allocate.
     * @param preparedModels A vector of IPreparedModel objects. Must only contain IPreparedModel
     *     objects from the same IDevice as this method invoked on.
     * @param inputRoles A vector of roles with each specifying an input to a prepared model.
     * @param outputRoles A vector of roles with each specifying an output to a prepared model.
     *     Each role specified in inputRoles and outputRoles must be unique. The corresponding
     *     model operands of the roles must have the same OperandType, scale, zero point, and
     *     ExtraParams. The dimensions of the operands and the dimensions specified in the buffer
     *     descriptor must be compatible with each other. Two dimensions are incompatible if there
     *     is at least one axis that is fully specified in both but has different values.
     * @return A tuple consisting of:
     *     - Error status of the buffer allocation. Must be:
     *         - NONE if successful
     *         - DEVICE_UNAVAILABLE if driver is offline or busy
     *         - GENERAL_FAILURE if a certain buffer property or a certain role is not supported,
     *           or if there is an unspecified error
     *         - INVALID_ARGUMENT if one of the input arguments is invalid
     *     - The allocated IBuffer object. If the buffer was unable to be allocated
     *       due to an error, nullptr must be returned.
     *     - A positive token identifying the allocated buffer. The same token will be
     *       provided when referencing the buffer as one of the memory pools in the request of an
     *       execution. If the buffer was unable to be allocated due to an error, the token must be
     *       0.
     */
    std::tuple<hal::ErrorStatus, sp<hal::IBuffer>, uint32_t> allocate(
            const hal::BufferDesc& desc,
            const std::vector<std::shared_ptr<VersionedIPreparedModel>>& preparedModels,
            const hal::hidl_vec<hal::BufferRole>& inputRoles,
            const hal::hidl_vec<hal::BufferRole>& outputRoles) const;

    /**
     * Blocks until the device is not in a bad state.
     *
     * @return Error code after waiting. ANEURALNETWORKS_NO_ERROR if device is
     *     not in a bad state.
     */
    int wait() const;

   private:
    // Cached initialization results.
    const hal::Capabilities kCapabilities;
    const std::vector<hal::Extension> kSupportedExtensions;
    const int32_t kType;
    const std::string kVersionString;
    const std::pair<uint32_t, uint32_t> kNumberOfCacheFilesNeeded;

    // internal methods to prepare a model
    std::pair<int, std::shared_ptr<VersionedIPreparedModel>> prepareModelInternal(
            const hal::Model& model, hal::ExecutionPreference preference, hal::Priority priority,
            const std::optional<Deadline>& deadline, const std::string& cacheDir,
            const std::optional<hal::CacheToken>& maybeToken) const;
    std::pair<int, std::shared_ptr<VersionedIPreparedModel>> prepareModelFromCacheInternal(
            const std::optional<Deadline>& deadline, const std::string& cacheDir,
            const hal::CacheToken& token) const;

    /**
     * This is a utility class for VersionedIDevice that encapsulates a
     * V1_0::IDevice, any appropriate downcasts to newer interfaces, and a
     * hidl_death_recipient that will proactively handle the case when the
     * service containing the IDevice object crashes.
     *
     * This is a convenience class to help VersionedIDevice recover from an
     * IDevice object crash: It bundles together all the data that needs to
     * change when recovering from a crash, and simplifies the process of
     * instantiating that data (at VersionedIDevice creation time) and
     * re-instantiating that data (at crash recovery time).
     */
    class Core {
       public:
        /**
         * Constructor for the Core object.
         *
         * Core is constructed with a V1_0::IDevice object, which represents a
         * device that is at least v1.0 of the interface. The constructor
         * downcasts to the latest version of the IDevice interface, allowing
         * VersionedIDevice to default to using the latest version of all
         * IDevice interface methods automatically.
         *
         * @param device A device object that is at least version 1.0 of the IDevice
         *               interface.
         * @param deathHandler A hidl_death_recipient that will proactively handle
         *                     the case when the service containing the IDevice
         *                     object crashes.
         */
        Core(sp<hal::V1_0::IDevice> device, sp<IDeviceDeathHandler> deathHandler);

        /**
         * Destructor for the Core object.
         *
         * This destructor unlinksToDeath this object's hidl_death_recipient as it
         * no longer needs to handle the case where the IDevice's service crashes.
         */
        ~Core();

        // Support move but not copy
        Core(Core&&) noexcept;
        Core& operator=(Core&&) noexcept;
        Core(const Core&) = delete;
        Core& operator=(const Core&) = delete;

        /**
         * Create a Core object.
         *
         * Prefer using this function over the constructor, as it adds more
         * protections.
         *
         * This call linksToDeath a hidl_death_recipient that can
         * proactively handle the case when the service containing the IDevice
         * object crashes.
         *
         * @param device A device object that is at least version 1.0 of the IDevice
         *               interface.
         * @return A valid Core object, otherwise nullopt.
         */
        static std::optional<Core> create(sp<hal::V1_0::IDevice> device);

        /**
         * Returns sp<*::IDevice> that is a downcast of the sp<V1_0::IDevice>
         * passed to the constructor.  This will be nullptr if that IDevice is
         * not actually of the specified downcast type.
         */
        template <typename T_IDevice>
        sp<T_IDevice> getDevice() const;
        template <>
        sp<hal::V1_0::IDevice> getDevice() const {
            return mDeviceV1_0;
        }
        template <>
        sp<hal::V1_1::IDevice> getDevice() const {
            return mDeviceV1_1;
        }
        template <>
        sp<hal::V1_2::IDevice> getDevice() const {
            return mDeviceV1_2;
        }
        template <>
        sp<hal::V1_3::IDevice> getDevice() const {
            return mDeviceV1_3;
        }

        /**
         * Returns sp<*::IDevice> (as per getDevice()) and the
         * hidl_death_recipient that will proactively handle the case when the
         * service containing the IDevice object crashes.
         */
        template <typename T_IDevice>
        std::pair<sp<T_IDevice>, sp<IDeviceDeathHandler>> getDeviceAndDeathHandler() const;

       private:
        /**
         * All versions of IDevice are necessary because the driver could be v1.0,
         * v1.1, or a later version. All these pointers logically represent the same
         * object.
         *
         * The general strategy is: HIDL returns a V1_0 device object, which
         * (if not nullptr) could be v1.0, v1.1, or a greater version. The V1_0
         * object is then "dynamically cast" to a V1_1 object. If successful,
         * mDeviceV1_1 will point to the same object as mDeviceV1_0; otherwise,
         * mDeviceV1_1 will be nullptr.
         *
         * In general:
         * * If the device is truly v1.0, mDeviceV1_0 will point to a valid object
         *   and mDeviceV1_1 will be nullptr.
         * * If the device is truly v1.1 or later, both mDeviceV1_0 and mDeviceV1_1
         *   will point to the same valid object.
         *
         * Idiomatic usage: if mDeviceV1_1 is non-null, do V1_1 dispatch; otherwise,
         * do V1_0 dispatch.
         */
        sp<hal::V1_0::IDevice> mDeviceV1_0;
        sp<hal::V1_1::IDevice> mDeviceV1_1;
        sp<hal::V1_2::IDevice> mDeviceV1_2;
        sp<hal::V1_3::IDevice> mDeviceV1_3;

        /**
         * HIDL callback to be invoked if the service for mDeviceV1_0 crashes.
         *
         * nullptr if this Core instance is a move victim and hence has no
         * callback to be unlinked.
         */
        sp<IDeviceDeathHandler> mDeathHandler;
    };

    // This method retrieves the appropriate mCore.mDevice* field, under a read lock.
    template <typename T_IDevice>
    sp<T_IDevice> getDevice() const EXCLUDES(mMutex) {
        std::shared_lock lock(mMutex);
        return mCore.getDevice<T_IDevice>();
    }

    // This method retrieves the appropriate mCore.mDevice* fields, under a read lock.
    template <typename T_IDevice>
    auto getDeviceAndDeathHandler() const EXCLUDES(mMutex) {
        std::shared_lock lock(mMutex);
        return mCore.getDeviceAndDeathHandler<T_IDevice>();
    }

    // This method calls the function fn in a manner that supports recovering
    // from a driver crash: If the driver implementation is dead because the
    // driver crashed either before the call to fn or during the call to fn, we
    // will attempt to obtain a new instance of the same driver and call fn
    // again.
    //
    // If a callback is provided, this method protects it against driver death
    // and waits for it (callback->wait()).
    template <typename T_Return, typename T_IDevice, typename T_Callback = std::nullptr_t>
    hal::Return<T_Return> recoverable(
            const char* context,
            const std::function<hal::Return<T_Return>(const sp<T_IDevice>&)>& fn,
            const T_Callback& callback = nullptr) const EXCLUDES(mMutex);

    // The name of the service that implements the driver.
    const std::string kServiceName;

    // Factory function object to generate an IDevice object.
    const hal::DeviceFactory kMakeDevice;

    // Guards access to mCore.
    mutable std::shared_mutex mMutex;

    // Data that can be rewritten during driver recovery.  Guarded againt
    // synchronous access by a mutex: Any number of concurrent read accesses is
    // permitted, but a write access excludes all other accesses.
    mutable Core mCore GUARDED_BY(mMutex);
};

/** This class wraps an IPreparedModel object of any version. */
class VersionedIPreparedModel {
    DISALLOW_IMPLICIT_CONSTRUCTORS(VersionedIPreparedModel);

   public:
    /**
     * Constructor for the VersionedIPreparedModel object.
     *
     * This constructor should not be used directly. Instead,
     * VersionedIPreparedModel should be created via
     * VersionedIDevice::prepareModel*.
     *
     * VersionedIPreparedModel is constructed with the V1_0::IPreparedModel object, which
     * represents a device that is at least v1.0 of the interface. The constructor downcasts
     * to the latest version of the IPreparedModel interface, and will default to using the
     * latest version of all IPreparedModel interface methods automatically.
     *
     * @param preparedModel A prepared model object that is least version 1.0 of the
     *                      IPreparedModel interface.
     * @param deathHandler A hidl_death_recipient that will proactively handle
     *                     the case when the service containing the IDevice
     *                     object crashes.
     */
    VersionedIPreparedModel(sp<hal::V1_0::IPreparedModel> preparedModel,
                            sp<IPreparedModelDeathHandler> deathHandler);

    /**
     * Destructor for the VersionedIPreparedModel object.
     *
     * This destructor unlinksToDeath this object's hidl_death_recipient as it
     * no longer needs to handle the case where the IPreparedModel's service
     * crashes.
     */
    ~VersionedIPreparedModel();

    /**
     * Performs a synchronous execution on a prepared model.
     *
     * The execution is performed synchronously with respect to the caller.
     * VersionedIPreparedModel::execute must verify the inputs to the function
     * are correct. If there is an error, VersionedIPreparedModel::execute must
     * immediately return with the appropriate result code. If the inputs to the
     * function are valid and there is no error,
     * VersionedIPreparedModel::execute must perform the execution, and must not
     * return until the execution is complete.
     *
     * If the prepared model was prepared from a model wherein all tensor
     * operands have fully specified dimensions, and the inputs to the function
     * are valid, and at execution time every operation's input operands have
     * legal values, then the execution should complete successfully
     * (ANEURALNETWORKS_NO_ERROR): There must be no failure unless the device
     * itself is in a bad state.
     *
     * execute may be called with an optional deadline. If the execution is not
     * able to be completed before the provided deadline, the execution may be
     * aborted, and either {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or
     * {@link ErrorStatus::MISSED_DEADLINE_PERSISTENT} must be returned. The
     * error due to an abort must be sent the same way as other errors,
     * described above.
     *
     * Any number of calls to the VersionedIPreparedModel::execute function, in
     * any combination, may be made concurrently, even on the same
     * VersionedIPreparedModel object.
     *
     * @param request The input and output information on which the prepared
     *     model is to be executed.
     * @param measure Specifies whether or not to measure duration of the
     *     execution.
     * @param deadline Optional time point. If provided, prepareModel is
     *     expected to complete by this time point. If it is not able to be
     *     completed by the deadline, the execution may be aborted.
     * @param loopTimeoutDuration The maximum amount of time that should be spent
     *     executing a {@link OperationType::WHILE} operation. If a loop
     *     condition model does not output false within this duration, the
     *     execution must be aborted. If no loop timeout duration is provided,
     *     the maximum amount of time is {@link LoopTimeoutDurationNs::DEFAULT}.
     *     When provided, the duration must not exceed {@link
     *     LoopTimeoutDurationNs::MAXIMUM}.
     * @param preferSynchronous 'true' to perform synchronous HAL execution when
     *     possible, 'false' to force asynchronous HAL execution.
     * @return A tuple consisting of:
     *     - Result code of the execution, must be:
     *         - ANEURALNETWORKS_NO_ERROR if execution is performed successfully
     *         - ANEURALNETWORKS_UNAVAILABLE_DEVICE if driver is offline or busy
     *         - ANEURALNETWORKS_OP_FAILED if there is an unspecified error
     *         - ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE if at least one output
     *             operand buffer is not large enough to store the corresponding
     *             output
     *         - ANEURALNETWORKS_BAD_DATA if one of the input arguments is
     *             invalid
     *     - A list of shape information of model output operands.
     *         The index into "outputShapes" corresponds to the index of the
     *         output operand in the Request outputs vector. outputShapes must
     *         be empty unless the result code is either
     *         ANEURALNETWORKS_NO_ERROR or
     *         ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE. outputShapes may be
     *         empty if the result code is ANEURALNETWORKS_NO_ERROR and all
     *         model output operands are fully-specified at execution time.
     *         outputShapes must have the same number of elements as the number
     *         of model output operands if the result code is
     *         ANEURALNETWORKS_OUTPUT_INSUFFICIENT_SIZE, or if the result code
     *         is ANEURALNETWORKS_NO_ERROR and the model has at least one output
     *         operand that is not fully-specified.
     *     - Duration of execution. Unless measure is YES and result code is
     *         ANEURALNETWORKS_NO_ERROR, all times must be reported as
     *         UINT64_MAX. A driver may choose to report any time as UINT64_MAX,
     *         indicating that measurement is not available.
     */
    std::tuple<int, std::vector<hal::OutputShape>, hal::Timing> execute(
            const hal::Request& request, hal::MeasureTiming measure,
            const std::optional<Deadline>& deadline,
            const hal::OptionalTimeoutDuration& loopTimeoutDuration, bool preferSynchronous) const;

    /**
     * Creates a burst controller on a prepared model.
     *
     * @param preferPowerOverLatency 'true' if the Burst object should run in a
     *                               more power efficient mode, 'false' if more
     *                               power can be used to possibly reduce
     *                               burst compute latency.
     * @return ExecutionBurstController Execution burst controller object.
     *                                  nullptr is returned if the burst cannot
     *                                  be configured for any reason.
     */
    std::shared_ptr<ExecutionBurstController> configureExecutionBurst(
            bool preferPowerOverLatency) const;

    /**
     * Launch a fenced asynchronous execution on a prepared model.
     *
     * The execution is performed asynchronously with respect to the caller.
     * executeFenced must fully validate the request. If there is an error during validation,
     * executeFenced must immediately return with the corresponding ErrorStatus. If the inputs
     * to the function are valid and there is no error and there is no error launching,
     * executeFenced must dispatch an asynchronous task to perform the execution in the
     * background, and immediately return with ErrorStatus::NONE, a sync fence that will be
     * signaled once the execution is completed, and a callback that can be used by the client
     * to query the duration and runtime error status. If the task has finished
     * before the call returns, empty handle may be returned for the syncFence. If the
     * asynchronous task fails to launch, executeFenced must immediately return with
     * ErrorStatus::GENERAL_FAILURE, an empty handle for the syncFence, and nullptr
     * for callback. The execution must wait for all the sync fences (if any) in waitFor to be
     * signaled before starting the actual execution.
     *
     * If any of sync fences in waitFor changes to error status after the executeFenced
     * call succeeds, the driver must immediately set the returned syncFence to error status.
     *
     * When the asynchronous task has finished its execution, it must
     * immediately signal the syncFence returned from executeFenced call. After
     * the syncFence is signaled, the task must not modify the content of
     * any data object referenced by 'request' (described by the
     * {@link @1.0::DataLocation} of a {@link @1.0::RequestArgument}).
     *
     * executeFenced may be called with an optional deadline and an optional
     * timeoutDurationAfterFence. If the execution is not able to be completed
     * before the provided deadline or within the timeoutDurationAfterFence,
     * whichever comes earlier, the execution may be aborted, and either {@link
     * ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link
     * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due
     * to an abort must be sent the same way as other errors, described above.
     *
     * Any number of calls to the executeFenced, execute* and executeSynchronously*
     * functions, in any combination, may be made concurrently, even on the same
     * IPreparedModel object.
     *
     * @param request The input and output information on which the prepared
     *                model is to be executed.
     * @param waitFor A vector of sync fence file descriptors. The execution must
     *                wait for all sync fence to be signaled before starting the
     *                task.
     * @param measure Specifies whether or not to measure duration of the execution.
     * @param deadline The time by which execution is expected to complete. If
     *                 the execution cannot be finished by the deadline, the
     *                 execution may be aborted.
     * @param loopTimeoutDuration The maximum amount of time that should be spent
     *     executing a {@link OperationType::WHILE} operation. If a loop
     *     condition model does not output false within this duration, the
     *     execution must be aborted. If no loop timeout duration is provided,
     *     the maximum amount of time is {@link LoopTimeoutDurationNs::DEFAULT}.
     *     When provided, the duration must not exceed {@link
     *     LoopTimeoutDurationNs::MAXIMUM}.
     * @param timeoutDurationAfterFence The timeout duration within which the
     *                                  execution is expected to complete after
     *                                  all sync fences in waitFor are signaled.
     * @return A tuple consisting of:
     *         - Error code of the dispatch call.
     *         - A sync_fence that will be triggered when the task is completed.
     *           The sync_fence will be set to error if critical error occurs when doing
     *           actual evaluation.
     *         - A callback can be used to query information like duration
     *           and detailed runtime error status when the task is completed.
     *         - Optional timing information. Only useful if the call is simulated using
     *           sync execution. Either IFencedExecutionCallback will be
     *           returned or optional timing information is returned
     */
    std::tuple<int, hal::hidl_handle, sp<hal::IFencedExecutionCallback>, hal::Timing> executeFenced(
            const hal::Request& request, const hal::hidl_vec<hal::hidl_handle>& waitFor,
            hal::MeasureTiming measure, const std::optional<Deadline>& deadline,
            const hal::OptionalTimeoutDuration& loopTimeoutDuration,
            const hal::OptionalTimeoutDuration& timeoutDurationAfterFence);

   private:
    friend class VersionedIDevice;

    std::tuple<int, std::vector<hal::OutputShape>, hal::Timing> executeAsynchronously(
            const hal::Request& request, hal::MeasureTiming timing,
            const std::optional<Deadline>& deadline,
            const hal::OptionalTimeoutDuration& loopTimeoutDuration) const;
    std::tuple<int, std::vector<hal::OutputShape>, hal::Timing> executeSynchronously(
            const hal::Request& request, hal::MeasureTiming measure,
            const std::optional<Deadline>& deadline,
            const hal::OptionalTimeoutDuration& loopTimeoutDuration) const;

    /**
     * Returns sp<V1_3::IPreparedModel> that is a downcast of the sp<V1_0::IPreparedModel>
     * passed to the constructor.  This will be nullptr if that IPreparedModel is
     * not actually of the specified downcast type.
     */
    sp<hal::V1_3::IPreparedModel> getV1_3() const { return mPreparedModelV1_3; }

    /**
     * All versions of IPreparedModel are necessary because the preparedModel could be v1.0,
     * v1.2, or a later version. All these pointers logically represent the same object.
     *
     * The general strategy is: HIDL returns a V1_0 prepared model object, which
     * (if not nullptr) could be v1.0, v1.2, or a greater version. The V1_0
     * object is then "dynamically cast" to objects of later versions. If successful,
     * mPreparedModel* will point to the same object as mPreparedModelV1_0; otherwise,
     * mPreparedModel* will be nullptr.
     *
     * In general:
     * * If the prepared model is truly v1.0, mPreparedModelV1_0 will point to a valid object,
     *   both mPreparedModelV1_2 and mPreparedModelV1_3 will be nullptr.
     * * If the prepared model is truly v1.2, both mPreparedModelV1_0 and mPreparedModelV1_2
     *   will point to the same valid object, but mPreparedModelV1_3 will be nullptr.
     * * If the prepared model is truly v1.3 or later, all of mPreparedModelV1_0,
     *   mPreparedModelV1_2, and mPreparedModelV1_3 will point to the same valid object.
     *
     * Idiomatic usage: if mPreparedModelV1_3 is non-null, do V1_3 dispatch;
     *       otherwise, if mPreparedModelV1_2 is non-null, do V1_2 dispatch;
     *       otherwise, do V1_0 dispatch.
     */
    sp<hal::V1_0::IPreparedModel> mPreparedModelV1_0;
    sp<hal::V1_2::IPreparedModel> mPreparedModelV1_2;
    sp<hal::V1_3::IPreparedModel> mPreparedModelV1_3;

    /**
     * HIDL callback to be invoked if the service for mPreparedModelV1_0 crashes.
     */
    const sp<IPreparedModelDeathHandler> mDeathHandler;
};

}  // namespace nn
}  // namespace android

#endif  // ANDROID_FRAMEWORKS_ML_NN_RUNTIME_VERSIONED_INTERFACES_H