Support sync fence in NNAPI
- Add IPreparedModel::dispatchRequest to NNAPI 1.3 HAL
- Add IDispatchExecutionCallback to allow clients query information
related to the actual evaluation.
Bug: 142778241
Test: mm
Change-Id: I87cbb7f2aee87342b0418fce04eb4050e2bc1920
diff --git a/neuralnetworks/1.3/IPreparedModel.hal b/neuralnetworks/1.3/IPreparedModel.hal
index bce6ee2..f84bcf4 100644
--- a/neuralnetworks/1.3/IPreparedModel.hal
+++ b/neuralnetworks/1.3/IPreparedModel.hal
@@ -24,6 +24,7 @@
import OptionalTimePoint;
import Request;
import IExecutionCallback;
+import IFencedExecutionCallback;
/**
* IPreparedModel describes a model that has been prepared for execution and
@@ -91,7 +92,8 @@
* execution cannot be finished by the deadline, the
* execution must be aborted.
* @param callback A callback object used to return the error status of
- * the execution. The callback object's notify function must
+ * the execution, shape information of model output operands, and
+ * duration of execution. The callback object's notify function must
* be called exactly once, even if the execution was
* unsuccessful.
* @return status Error status of the call, must be:
@@ -187,4 +189,57 @@
OptionalTimePoint deadline)
generates (ErrorStatus status, vec<OutputShape> outputShapes,
Timing timing);
+
+ /**
+ * 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, and only accept one that is
+ * guaranteed to be completed, unless a hardware failure or kernel panic happens on the device.
+ * If there is an error during validation, executeFenced must immediately return with
+ * the corresponding ErrorStatus. If the request is valid 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 sync fence. If the
+ * asynchronous task fails to launch, executeFenced must immediately return with
+ * ErrorStatus::GENERAL_FAILURE, and empty handle for the sync fence and nullptr
+ * for callback. The execution must wait for all the sync fences (if any) in wait_for to be
+ * signaled before starting the actual execution.
+ *
+ * If any of sync fences in wait_for changes to error status after the executeFenced
+ * call succeeds, the driver must immediately set the returned sync fence to error status.
+ *
+ * When the asynchronous task has finished its execution, it must
+ * immediately signal the sync_fence created when dispatching. After
+ * the sync_fence 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}).
+ *
+ * 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.
+ * Execution must not start until all sync fences have been signaled.
+ * @param measure Specifies whether or not to measure duration of the execution.
+ * The duration runs from the time the driver sees the call
+ * to the executeFenced function to the time sync_fence is triggered.
+ * @return status Error status of the call, must be:
+ * - NONE if task is successfully launched
+ * - DEVICE_UNAVAILABLE if driver is offline or busy
+ * - GENERAL_FAILURE if there is an unspecified error
+ * - INVALID_ARGUMENT if one of the input arguments is invalid, including
+ * fences in error states.
+ * @return syncFence The sync fence that will be triggered when the task is completed.
+ * The sync fence will be set to error if a critical error,
+ * e.g. hardware failure or kernel panic, occurs when doing execution.
+ * @return callback The IFencedExecutionCallback can be used to query information like duration
+ * and error status when the execution is completed.
+ */
+ executeFenced(Request request, vec<handle> waitFor, MeasureTiming measure)
+ generates (ErrorStatus status, handle syncFence, IFencedExecutionCallback callback);
};