Add 1.2 NN HAL interface for dynamic output shape.
Let notify_1_2() notify output shapes.
Document unspecified dimensions and rank.
Bug: 73506513
Bug: 77234888
Test: NeuralNetworksTest_static
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Change-Id: I01108913212d9f4aa47daf2f293ea19259925865
diff --git a/neuralnetworks/1.0/vts/functional/Callbacks.cpp b/neuralnetworks/1.0/vts/functional/Callbacks.cpp
index a1c5a1a..03afcd0 100644
--- a/neuralnetworks/1.0/vts/functional/Callbacks.cpp
+++ b/neuralnetworks/1.0/vts/functional/Callbacks.cpp
@@ -139,8 +139,10 @@
return Void();
}
-Return<void> ExecutionCallback::notify_1_2(ErrorStatus errorStatus) {
+Return<void> ExecutionCallback::notify_1_2(ErrorStatus errorStatus,
+ const hidl_vec<OutputShape>& outputShapes) {
mErrorStatus = errorStatus;
+ mOutputShapes = outputShapes;
CallbackBase::notify();
return Void();
}
@@ -150,6 +152,11 @@
return mErrorStatus;
}
+const std::vector<OutputShape>& ExecutionCallback::getOutputShapes() {
+ wait();
+ return mOutputShapes;
+}
+
} // namespace implementation
} // namespace V1_2
} // namespace neuralnetworks
diff --git a/neuralnetworks/1.0/vts/functional/Callbacks.h b/neuralnetworks/1.0/vts/functional/Callbacks.h
index e89980d..46f29a6 100644
--- a/neuralnetworks/1.0/vts/functional/Callbacks.h
+++ b/neuralnetworks/1.0/vts/functional/Callbacks.h
@@ -275,8 +275,9 @@
* Either IExecutionCallback::notify or IExecutionCallback::notify_1_2 must
* be called exactly once on a given ExecutionCallback object.
*
- * @param status Error status returned from asynchronously preparing the
- * model; will be:
+ * @param status Error status returned from launching the asynchronous task
+ * (if the launch fails) or from the asynchronous task itself
+ * (if the launch succeeds). Must be:
* - NONE if the asynchronous execution was successful
* - DEVICE_UNAVAILABLE if driver is offline or busy
* - GENERAL_FAILURE if there is an unspecified error
@@ -285,27 +286,73 @@
* - INVALID_ARGUMENT if the input request is invalid
*/
Return<void> notify(ErrorStatus status) override;
- Return<void> notify_1_2(ErrorStatus status) override;
+
+ /**
+ * Similar to IExecutionCallback::notify, but for V1_2::IPreparedModel to
+ * also notify output shapes along with error status.
+ *
+ * @param status Error status returned from launching the asynchronous task
+ * (if the launch fails) or from the asynchronous task itself
+ * (if the launch succeeds). Must be:
+ * - NONE if the asynchronous execution was successful
+ * - DEVICE_UNAVAILABLE if driver is offline or busy
+ * - GENERAL_FAILURE if the asynchronous task resulted in an
+ * unspecified error
+ * - OUTPUT_INSUFFICIENT_SIZE if at least one output
+ * operand buffer is not large enough to store the
+ * corresponding output
+ * - INVALID_ARGUMENT if one of the input arguments to
+ * prepareModel is invalid
+ * @param outputShapes 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 status is either
+ * NONE or OUTPUT_INSUFFICIENT_SIZE.
+ */
+ Return<void> notify_1_2(ErrorStatus status, const hidl_vec<OutputShape>& outputShapes) override;
/**
* Retrieves the error status returned from the asynchronous task launched
- * by IPreparedModel::execute. If IPreparedModel::execute has not finished
+ * by either IPreparedModel::execute or IPreparedModel::execute_1_2. If
+ * IPreparedModel::execute or IPreparedModel::execute_1_2 has not finished
* asynchronously executing, this call will block until the asynchronous task
* notifies the object.
*
- * @return status Error status returned from asynchronously preparing the
- * model; will be:
+ * @return status Error status returned from launching the asynchronous task
+ * (if the launch fails) or from the asynchronous task itself
+ * (if the launch succeeds). Must be:
* - NONE if the asynchronous execution was successful
* - DEVICE_UNAVAILABLE if driver is offline or busy
- * - GENERAL_FAILURE if there is an unspecified error
- * - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is
- * not large enough to store the resultant values
- * - INVALID_ARGUMENT if the input request is invalid
+ * - GENERAL_FAILURE if the asynchronous task resulted in an
+ * unspecified error
+ * - OUTPUT_INSUFFICIENT_SIZE if at least one output
+ * operand buffer is not large enough to store the
+ * corresponding output
+ * - INVALID_ARGUMENT if one of the input arguments to
+ * prepareModel is invalid
*/
ErrorStatus getStatus();
- private:
+ /**
+ * Retrieves the output shapes returned from the asynchronous task launched
+ * by IPreparedModel::execute_1_2. If IPreparedModel::execute_1_2 has not finished
+ * asynchronously executing, this call will block until the asynchronous task
+ * notifies the object.
+ *
+ * If the asynchronous task was launched by IPreparedModel::execute, an empty vector
+ * will be returned.
+ *
+ * @return outputShapes 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 status is either
+ * NONE or OUTPUT_INSUFFICIENT_SIZE.
+ */
+ const std::vector<OutputShape>& getOutputShapes();
+
+ private:
ErrorStatus mErrorStatus;
+ std::vector<OutputShape> mOutputShapes;
};
diff --git a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
index 3b4eb21..b5a8607 100644
--- a/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/1.0/vts/functional/GeneratedTestHarness.cpp
@@ -89,13 +89,24 @@
sp<ExecutionCallback>& callback) {
return preparedModel->execute_1_2(request, callback);
}
-static Return<ErrorStatus> ExecutePreparedModel(sp<V1_0::IPreparedModel>&, const Request&) {
+static Return<ErrorStatus> ExecutePreparedModel(sp<V1_0::IPreparedModel>&, const Request&,
+ hidl_vec<OutputShape>*) {
ADD_FAILURE() << "asking for synchronous execution at V1_0";
return ErrorStatus::GENERAL_FAILURE;
}
static Return<ErrorStatus> ExecutePreparedModel(sp<V1_2::IPreparedModel>& preparedModel,
- const Request& request) {
- return preparedModel->executeSynchronously(request);
+ const Request& request,
+ hidl_vec<OutputShape>* outputShapes) {
+ ErrorStatus result;
+ Return<void> ret = preparedModel->executeSynchronously(
+ request, [&result, &outputShapes](ErrorStatus error, const hidl_vec<OutputShape>& shapes) {
+ result = error;
+ *outputShapes = shapes;
+ });
+ if (!ret.isOk()) {
+ return ErrorStatus::GENERAL_FAILURE;
+ }
+ return result;
}
enum class Synchronously { NO, YES };
const float kDefaultAtol = 1e-5f;
@@ -197,6 +208,8 @@
inputMemory->commit();
outputMemory->commit();
+ ErrorStatus executionStatus;
+ hidl_vec<OutputShape> outputShapes;
if (sync == Synchronously::NO) {
SCOPED_TRACE("asynchronous");
@@ -211,18 +224,24 @@
// retrieve execution status
executionCallback->wait();
- ErrorStatus executionReturnStatus = executionCallback->getStatus();
- EXPECT_EQ(ErrorStatus::NONE, executionReturnStatus);
+ executionStatus = executionCallback->getStatus();
+ outputShapes = executionCallback->getOutputShapes();
} else {
SCOPED_TRACE("synchronous");
// execute
- Return<ErrorStatus> executionStatus = ExecutePreparedModel(
- preparedModel, {.inputs = inputs_info, .outputs = outputs_info, .pools = pools});
- ASSERT_TRUE(executionStatus.isOk());
- EXPECT_EQ(ErrorStatus::NONE, static_cast<ErrorStatus>(executionStatus));
+ Return<ErrorStatus> executionReturnStatus = ExecutePreparedModel(
+ preparedModel, {.inputs = inputs_info, .outputs = outputs_info, .pools = pools},
+ &outputShapes);
+ ASSERT_TRUE(executionReturnStatus.isOk());
+ executionStatus = static_cast<ErrorStatus>(executionReturnStatus);
}
+ ASSERT_EQ(ErrorStatus::NONE, executionStatus);
+ // TODO(xusongw): Check if the returned output shapes match with expectation once the
+ // sample driver implementation of dynamic output shape is finished.
+ ASSERT_EQ(outputShapes.size(), 0);
+
// validate results
outputMemory->read();
copy_back(&test, outputs_info, outputPtr);
diff --git a/neuralnetworks/1.2/Android.bp b/neuralnetworks/1.2/Android.bp
index 7d13104..d8762b0 100644
--- a/neuralnetworks/1.2/Android.bp
+++ b/neuralnetworks/1.2/Android.bp
@@ -27,6 +27,7 @@
"Operation",
"OperationType",
"OperationTypeRange",
+ "OutputShape",
],
gen_java: false,
}
diff --git a/neuralnetworks/1.2/IExecutionCallback.hal b/neuralnetworks/1.2/IExecutionCallback.hal
index 667e0d6..47de1b6 100644
--- a/neuralnetworks/1.2/IExecutionCallback.hal
+++ b/neuralnetworks/1.2/IExecutionCallback.hal
@@ -18,6 +18,7 @@
import @1.0::ErrorStatus;
import @1.0::IExecutionCallback;
+import OutputShape;
/**
* IExecutionCallback must be used to return the error status result from an
@@ -39,10 +40,16 @@
* - DEVICE_UNAVAILABLE if driver is offline or busy
* - GENERAL_FAILURE if the asynchronous task resulted in an
* unspecified error
- * - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is
- * not large enough to store the resultant values
+ * - OUTPUT_INSUFFICIENT_SIZE if at least one output
+ * operand buffer is not large enough to store the
+ * corresponding output
* - INVALID_ARGUMENT if one of the input arguments to
* prepareModel is invalid
+ * @param outputShapes A list of shape information of model output operands.
+ * The index into "outputShapes" corresponds with to index
+ * of the output operand in the Request outputs vector.
+ * outputShapes must be empty unless the status is either
+ * NONE or OUTPUT_INSUFFICIENT_SIZE.
*/
- oneway notify_1_2(ErrorStatus status);
+ oneway notify_1_2(ErrorStatus status, vec<OutputShape> outputShapes);
};
diff --git a/neuralnetworks/1.2/IPreparedModel.hal b/neuralnetworks/1.2/IPreparedModel.hal
index 4e91c67..044ca28 100644
--- a/neuralnetworks/1.2/IPreparedModel.hal
+++ b/neuralnetworks/1.2/IPreparedModel.hal
@@ -100,11 +100,17 @@
* - NONE if execution is performed successfully
* - DEVICE_UNAVAILABLE if driver is offline or busy
* - GENERAL_FAILURE if there is an unspecified error
- * - OUTPUT_INSUFFICIENT_SIZE if provided output buffer is
- * not large enough to store the resultant values
+ * - OUTPUT_INSUFFICIENT_SIZE if at least one output
+ * operand buffer is not large enough to store the
+ * corresponding output
* - INVALID_ARGUMENT if one of the input arguments is
* invalid
+ * @return outputShapes 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 status is either
+ * NONE or OUTPUT_INSUFFICIENT_SIZE.
*/
executeSynchronously(Request request)
- generates (ErrorStatus status);
+ generates (ErrorStatus status, vec<OutputShape> outputShapes);
};
diff --git a/neuralnetworks/1.2/types.hal b/neuralnetworks/1.2/types.hal
index 40c07e7..4738cc3 100644
--- a/neuralnetworks/1.2/types.hal
+++ b/neuralnetworks/1.2/types.hal
@@ -234,9 +234,6 @@
*
* For a scalar operand, dimensions.size() must be 0.
*
- * For a tensor operand, dimensions.size() must be at least 1;
- * however, any of the dimensions may be unspecified.
- *
* A tensor operand with all dimensions specified has "fully
* specified" dimensions. Whenever possible (i.e., whenever the
* dimensions are known at model construction time), a tensor
@@ -255,17 +252,20 @@
* . The operand has lifetime CONSTANT_COPY or
* CONSTANT_REFERENCE.
*
- * . The operand has lifetime MODEL_INPUT or MODEL_OUTPUT. Fully
+ * . The operand has lifetime MODEL_INPUT. Fully
* specified dimensions must either be present in the
* Operand or they must be provided in the corresponding
* RequestArgument.
- * EXCEPTION: If the input or output is optional and omitted
+ * EXCEPTION: If the input is optional and omitted
* (by setting the hasNoValue field of the corresponding
* RequestArgument to true) then it need not have fully
* specified dimensions.
*
* A tensor operand with some number of unspecified dimensions is
* represented by setting each unspecified dimension to 0.
+ *
+ * A tensor operand with unspecified rank is represented by providing
+ * an empty dimensions vector.
*/
vec<uint32_t> dimensions;
@@ -397,3 +397,18 @@
*/
bool relaxComputationFloat32toFloat16;
};
+
+/**
+ * Describes the shape information of an output operand after execution.
+ */
+struct OutputShape {
+ /**
+ * Dimensions of the operand.
+ */
+ vec<uint32_t> dimensions;
+
+ /**
+ * Whether the provided buffer size is sufficient for the output.
+ */
+ bool isSufficient;
+};
diff --git a/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp b/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
index d80fbcf..1eaea4b 100644
--- a/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
+++ b/neuralnetworks/1.2/vts/functional/ValidateRequest.cpp
@@ -110,15 +110,20 @@
executionCallback->wait();
ErrorStatus executionReturnStatus = executionCallback->getStatus();
+ const auto& outputShapes = executionCallback->getOutputShapes();
ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, executionReturnStatus);
+ ASSERT_EQ(outputShapes.size(), 0);
}
{
SCOPED_TRACE(message + " [executeSynchronously]");
- Return<ErrorStatus> executeStatus = preparedModel->executeSynchronously(request);
+ Return<void> executeStatus = preparedModel->executeSynchronously(
+ request, [](ErrorStatus error, const hidl_vec<OutputShape>& outputShapes) {
+ ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, error);
+ EXPECT_EQ(outputShapes.size(), 0);
+ });
ASSERT_TRUE(executeStatus.isOk());
- ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, static_cast<ErrorStatus>(executeStatus));
}
}