tree: c4b3fe3b6443e336c662bb23bed38aca45e89218 [path history] [tgz]
  1. OWNERS
  2. README.md
  3. adapter/
  4. common/
  5. service/
neuralnetworks/utils/README.md

NNAPI Conversions

convert fails if either the source type or the destination type is invalid, and it yields a valid object if the conversion succeeds. For example, let's say that an enumeration in the current version has fewer possible values than the “same” canonical enumeration, such as OperationType. The new value of HARD_SWISH (introduced in Android R / NN HAL 1.3) does not map to any valid existing value in OperationType, but an older value of ADD (introduced in Android OC-MR1 / NN HAL 1.0) is valid. This can be seen in the following model conversions:

// Unsuccessful conversion
const nn::Model canonicalModel = createModelWhichHasV1_3Operations();
const nn::Result<V1_0::Model> maybeVersionedModel = V1_0::utils::convert(canonicalModel);
EXPECT_FALSE(maybeVersionedModel.has_value());
// Successful conversion
const nn::Model canonicalModel = createModelWhichHasOnlyV1_0Operations();
const nn::Result<V1_0::Model> maybeVersionedModel = V1_0::utils::convert(canonicalModel);
ASSERT_TRUE(maybeVersionedModel.has_value());
const V1_0::Model& versionedModel = maybeVersionedModel.value();
EXPECT_TRUE(V1_0::utils::valid(versionedModel));

V1_X::utils::convert does not guarantee that all information is preserved. For example, In the case of nn::ErrorStatus, the new value of MISSED_DEADLINE_TRANSIENT can be represented by the existing value of V1_0::GENERAL_FAILURE:

// Lossy Canonical -> HAL -> Canonical conversion
const nn::ErrorStatus canonicalBefore = nn::ErrorStatus::MISSED_DEADLINE_TRANSIENT;
const V1_0::ErrorStatus versioned = V1_0::utils::convert(canonicalBefore).value();
const nn::ErrorStatus canonicalAfter = nn::convert(versioned).value();
EXPECT_NE(canonicalBefore, canonicalAfter);

However, nn::convert is guaranteed to preserve all information:

// Lossless HAL -> Canonical -> HAL conversion
const V1_0::ErrorStatus versionedBefore = V1_0::ErrorStatus::GENERAL_FAILURE;
const nn::ErrorStatus canonical = nn::convert(versionedBefore).value();
const V1_0::ErrorStatus versionedAfter = V1_0::utils::convert(canonical).value();
EXPECT_EQ(versionedBefore, versionedAfter);

The convert functions operate only on types that are used in a HIDL method call directly. The unvalidatedConvert functions operate on types that are either used in a HIDL method call directly (i.e., not as a nested class) or used in a subsequent version of the NN HAL. Prefer using convert over unvalidatedConvert.

Interface Lifetimes across Processes

HIDL

Some notes about HIDL interface objects and lifetimes across processes:

All HIDL interface objects inherit from IBase, which itself inherits from ::android::RefBase. As such, all HIDL interface objects are reference counted and must be owned through ::android::sp (or referenced through ::android::wp). Allocating RefBase objects on the stack will log errors and may result in crashes, and deleting a RefBase object through another means (e.g., “delete”, “free”, or RAII-cleanup through std::unique_ptr or some equivalent) will result in double-free and/or use-after-free undefined behavior.

HIDL/Binder manages the reference count of HIDL interface objects automatically across processes. If a process that references (but did not create) the HIDL interface object dies, HIDL/Binder ensures any reference count it held is properly released. (Caveat: it might be possible that HIDL/Binder behave strangely with ::android::wp references.)

If the process which created the HIDL interface object dies, any call on this object from another process will result in a HIDL transport error with the code DEAD_OBJECT.

AIDL

We use NDK backend for AIDL interfaces. Handling of lifetimes is generally the same with the following differences:

  • Interfaces inherit from ndk::ICInterface, which inherits from ndk::SharedRefBase. The latter is an analog of ::android::RefBase using std::shared_ptr for reference counting.
  • AIDL calls return ndk::ScopedAStatus which wraps fields of types binder_status_t and binder_exception_t. In case the call is made on a dead object, the call will return ndk::ScopedAStatus with exception EX_TRANSACTION_FAILED and binder status STATUS_DEAD_OBJECT.

Protecting Asynchronous Calls

Across HIDL

Some notes about asynchronous calls across HIDL:

For synchronous calls across HIDL, if an error occurs after the function was called but before it returns, HIDL will return a transport error. For example, if the message cannot be delivered to the server process or if the server process dies before returning a result, HIDL will return from the function with the appropriate transport error in the Return<> object, which can be queried with Return<>::isOk(), Return<>::isDeadObject(), Return<>::description(), etc.

However, HIDL offers no such error management in the case of asynchronous calls. By default, if the client launches an asynchronous task and the server fails to return a result through the callback, the client will be left waiting indefinitely for a result it will never receive.

In the NNAPI, IDevice::prepareModel* and IPreparedModel::execute* (but not IPreparedModel::executeSynchronously*) are asynchronous calls across HIDL. Specifically, these asynchronous functions are called with a HIDL interface callback object (IPrepareModelCallback for IDevice::prepareModel* and IExecutionCallback for IPreparedModel::execute*) and are expected to quickly return, and the results are returned at a later time through these callback objects.

To protect against the case when the server dies after the asynchronous task was called successfully but before the results could be returned, HIDL provides an object called a “hidl_death_recipient,” which can be used to detect when an interface object (and more generally, the server process) has died. nnapi/hal/ProtectCallback.h's DeathHandler uses hidl_death_recipients to detect when the driver process has died, and DeathHandler will unblock any thread waiting on the results of an IProtectedCallback callback object that may otherwise not be signaled. In order for this to work, the IProtectedCallback object must have been registered via DeathHandler::protectCallback().

Across AIDL

We use NDK backend for AIDL interfaces. Handling of asynchronous calls is generally the same with the following differences:

  • AIDL calls return ndk::ScopedAStatus which wraps fields of types binder_status_t and binder_exception_t. In case the call is made on a dead object, the call will return ndk::ScopedAStatus with exception EX_TRANSACTION_FAILED and binder status STATUS_DEAD_OBJECT.
  • AIDL interface doesn't contain asynchronous IPreparedModel::execute.
  • Service death is handled using AIBinder_DeathRecipient object which is linked to an interface object using AIBinder_linkToDeath. nnapi/hal/aidl/ProtectCallback.h provides DeathHandler object that is a direct analog of HIDL DeathHandler, only using libbinder_ndk objects for implementation.