Update neuralnetworks HAL to allow collecting execution duration.
Test: VtsHalNeuralnetworksV1_0TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.0::IDevice/sample-all
Test: VtsHalNeuralnetworksV1_1TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.1::IDevice/sample-all
Test: VtsHalNeuralnetworksV1_2TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Bug: 115390094
Change-Id: If67a5ffe39cfdd78498e01f26251734fdc8e66c7
diff --git a/neuralnetworks/1.2/IExecutionCallback.hal b/neuralnetworks/1.2/IExecutionCallback.hal
index 47de1b6..7f6c9ee 100644
--- a/neuralnetworks/1.2/IExecutionCallback.hal
+++ b/neuralnetworks/1.2/IExecutionCallback.hal
@@ -18,7 +18,6 @@
import @1.0::ErrorStatus;
import @1.0::IExecutionCallback;
-import OutputShape;
/**
* IExecutionCallback must be used to return the error status result from an
@@ -50,6 +49,11 @@
* of the output operand in the Request outputs vector.
* outputShapes must be empty unless the status is either
* NONE or OUTPUT_INSUFFICIENT_SIZE.
+ * @return Timing Duration of execution. Unless MeasureTiming::YES was passed when
+ * launching the execution and status is NONE, all times must
+ * be reported as UINT64_MAX. A driver may choose to report
+ * any time as UINT64_MAX, indicating that particular measurement is
+ * not available.
*/
- oneway notify_1_2(ErrorStatus status, vec<OutputShape> outputShapes);
+ oneway notify_1_2(ErrorStatus status, vec<OutputShape> outputShapes, Timing timing);
};
diff --git a/neuralnetworks/1.2/IPreparedModel.hal b/neuralnetworks/1.2/IPreparedModel.hal
index 2d4e572..5d2d80f 100644
--- a/neuralnetworks/1.2/IPreparedModel.hal
+++ b/neuralnetworks/1.2/IPreparedModel.hal
@@ -59,6 +59,10 @@
*
* @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.
+ * The duration runs from the time the driver sees the call
+ * to the execute_1_2 function to the time the driver invokes
+ * the callback.
* @param callback A callback object used to return the error status of
* the execution. The callback object's notify function must
* be called exactly once, even if the execution was
@@ -72,7 +76,7 @@
* - INVALID_ARGUMENT if one of the input arguments is
* invalid
*/
- execute_1_2(Request request, IExecutionCallback callback)
+ execute_1_2(Request request, MeasureTiming measure, IExecutionCallback callback)
generates (ErrorStatus status);
/**
@@ -98,6 +102,10 @@
*
* @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.
+ * The duration runs from the time the driver sees the call
+ * to the executeSynchronously function to the time the driver
+ * returns from the function.
* @return status Error status of the execution, must be:
* - NONE if execution is performed successfully
* - DEVICE_UNAVAILABLE if driver is offline or busy
@@ -112,9 +120,13 @@
* of the output operand in the Request outputs vector.
* outputShapes must be empty unless the status is either
* NONE or OUTPUT_INSUFFICIENT_SIZE.
+ * @return Timing Duration of execution. Unless measure is YES and status is
+ * NONE, 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.
*/
- executeSynchronously(Request request)
- generates (ErrorStatus status, vec<OutputShape> outputShapes);
+ executeSynchronously(Request request, MeasureTiming measure)
+ generates (ErrorStatus status, vec<OutputShape> outputShapes, Timing timing);
/**
* Configure a Burst object used to execute multiple inferences on a
diff --git a/neuralnetworks/1.2/types.hal b/neuralnetworks/1.2/types.hal
index ce993d7..8bc28b4 100644
--- a/neuralnetworks/1.2/types.hal
+++ b/neuralnetworks/1.2/types.hal
@@ -447,8 +447,34 @@
};
/**
- * FmqRequestDatum is a single element of a serialized representation of a
- * {@link @1.0::Request} object which is sent across FastMessageQueue.
+ * Specifies whether or not to measure timing information during execution.
+ */
+enum MeasureTiming : int32_t {
+ NO = 0,
+ YES = 1,
+};
+
+/**
+
+ * Timing information measured during execution. Each time is a duration from
+ * the beginning of some task to the end of that task, including time when that
+ * task is not active (for example, preempted by some other task, or
+ * waiting for some resource to become available).
+ *
+ * Times are measured in microseconds.
+ * When a time is not available, it must be reported as UINT64_MAX.
+ */
+struct Timing {
+ /** Execution time on device (not driver, which runs on host processor). */
+ uint64_t timeOnDevice;
+ /** Execution time in driver (including time on device). */
+ uint64_t timeInDriver;
+};
+
+/**
+ * FmqRequestDatum is a single element of a serialized representation of an
+ * execution request (a {@link @1.0::Request} object and a {@link MeasureTiming}
+ * value) which is sent across FastMessageQueue.
*
* The serialized representation for a particular execution is referred to later
* in these descriptions as a 'packet'.
@@ -456,7 +482,7 @@
* FastMessageQueue can only pass HIDL-defined types that do not involve nested
* buffers, handles, or interfaces.
*
- * The {@link @1.0::Request} is serialized as follows:
+ * The request is serialized as follows:
* 1) 'packetInformation'
* 2) For each input operand:
* 2.1) 'inputOperandInformation'
@@ -468,6 +494,7 @@
* 3.2.1) 'outputOperandDimensionValue'
* 4) For each pool:
* 4.1) 'poolIdentifier'
+ * 5) 'measureTiming'
*/
safe_union FmqRequestDatum {
/**
@@ -561,12 +588,21 @@
* identifier.
*/
int32_t poolIdentifier;
+
+ /**
+ * Specifies whether or not to measure duration of the execution. The
+ * duration runs from the time the driver dequeues the request from a
+ * FastMessageQueue to the time the driver enqueues results to a
+ * FastMessageQueue.
+ */
+ MeasureTiming measureTiming;
};
/**
* FmqResultDatum is a single element of a serialized representation of the
- * values returned from an execution ({@link @1.0::ErrorStatus} and
- * vec<{@link OutputShape}>) which is returned via FastMessageQueue.
+ * values returned from an execution ({@link @1.0::ErrorStatus},
+ * vec<{@link OutputShape}>, and {@link Timing}) which is returned via
+ * FastMessageQueue.
*
* The serialized representation for a particular execution is referred to later
* in these descriptions as a 'packet'.
@@ -581,6 +617,7 @@
* 2.1) 'operandInformation'
* 2.2) For each dimension element of the operand:
* 2.2.1) 'operandDimensionValue'
+ * 3) 'executionTiming'
*/
safe_union FmqResultDatum {
/**
@@ -636,4 +673,12 @@
* Element of the dimensions vector.
*/
uint32_t operandDimensionValue;
+
+ /**
+ * Duration of execution. Unless measurement was requested and execution
+ * succeeds, 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.
+ */
+ Timing executionTiming;
};
diff --git a/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp b/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
index 1eaea4b..00a7c3e 100644
--- a/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
+++ b/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
@@ -42,6 +42,10 @@
///////////////////////// UTILITY FUNCTIONS /////////////////////////
+static bool badTiming(Timing timing) {
+ return timing.timeOnDevice == UINT64_MAX && timing.timeInDriver == UINT64_MAX;
+}
+
static void createPreparedModel(const sp<IDevice>& device, const Model& model,
sp<IPreparedModel>* preparedModel) {
ASSERT_NE(nullptr, preparedModel);
@@ -98,31 +102,46 @@
Request request, const std::function<void(Request*)>& mutation) {
mutation(&request);
+ // We'd like to test both with timing requested and without timing
+ // requested. Rather than running each test both ways, we'll decide whether
+ // to request timing by hashing the message. We do not use std::hash because
+ // it is not guaranteed stable across executions.
+ char hash = 0;
+ for (auto c : message) {
+ hash ^= c;
+ };
+ MeasureTiming measure = (hash & 1) ? MeasureTiming::YES : MeasureTiming::NO;
+
{
SCOPED_TRACE(message + " [execute_1_2]");
sp<ExecutionCallback> executionCallback = new ExecutionCallback();
ASSERT_NE(nullptr, executionCallback.get());
Return<ErrorStatus> executeLaunchStatus =
- preparedModel->execute_1_2(request, executionCallback);
+ preparedModel->execute_1_2(request, measure, executionCallback);
ASSERT_TRUE(executeLaunchStatus.isOk());
ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, static_cast<ErrorStatus>(executeLaunchStatus));
executionCallback->wait();
ErrorStatus executionReturnStatus = executionCallback->getStatus();
const auto& outputShapes = executionCallback->getOutputShapes();
+ Timing timing = executionCallback->getTiming();
ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, executionReturnStatus);
ASSERT_EQ(outputShapes.size(), 0);
+ ASSERT_TRUE(badTiming(timing));
}
{
SCOPED_TRACE(message + " [executeSynchronously]");
Return<void> executeStatus = preparedModel->executeSynchronously(
- request, [](ErrorStatus error, const hidl_vec<OutputShape>& outputShapes) {
- ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, error);
- EXPECT_EQ(outputShapes.size(), 0);
- });
+ request, measure,
+ [](ErrorStatus error, const hidl_vec<OutputShape>& outputShapes,
+ const Timing& timing) {
+ ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, error);
+ EXPECT_EQ(outputShapes.size(), 0);
+ EXPECT_TRUE(badTiming(timing));
+ });
ASSERT_TRUE(executeStatus.isOk());
}
}