Introduce reusable execution to canonical interface -- HAL.
This CL modifies the canonical interface for reusable executions:
- Add new interface: IExecution with compute and computeFenced methods
- Add new method IPreparedModel::createExecution
In NNAPI runtime, the new interface IExecution is used to
memoize request-specific execution resources (e.g. converted HAL
request). The expected usage is that, IPreparedModel::createExecution
will be invoked in the first computation of a reusable NDK ANNExecution
object, and IExecution::compute* will be invoked repeatedly.
The IPreparedModel::execute* methods are preserved to avoid redundant
object creation and memoization overhead for a single-time
(non-reusable) execution.
For a vendor implementing the canonical interfaces, only the
IPreparedModel::execute* methods will be called because there is
currently no reusable execution at HAL interface. A DefaultExecution
implementation is provided to reduce the work needed on the vendor side.
Bug: 184073769
Test: NNT_static
Test: neuralnetworks_utils_hal_1_0_test
Test: neuralnetworks_utils_hal_1_1_test
Test: neuralnetworks_utils_hal_1_2_test
Test: neuralnetworks_utils_hal_1_3_test
Test: neuralnetworks_utils_hal_common_test
Test: neuralnetworks_utils_hal_aidl_test
Change-Id: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
Merged-In: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
(cherry picked from commit 727a7b2104b0962509fedffe720eec508b2ee6de)
diff --git a/neuralnetworks/1.2/utils/src/Execution.cpp b/neuralnetworks/1.2/utils/src/Execution.cpp
new file mode 100644
index 0000000..18d1c90
--- /dev/null
+++ b/neuralnetworks/1.2/utils/src/Execution.cpp
@@ -0,0 +1,74 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "Execution.h"
+
+#include "Callbacks.h"
+#include "Conversions.h"
+#include "Utils.h"
+
+#include <android/hardware/neuralnetworks/1.0/types.h>
+#include <android/hardware/neuralnetworks/1.1/types.h>
+#include <android/hardware/neuralnetworks/1.2/IPreparedModel.h>
+#include <android/hardware/neuralnetworks/1.2/types.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Result.h>
+#include <nnapi/TypeUtils.h>
+#include <nnapi/Types.h>
+#include <nnapi/hal/CommonUtils.h>
+#include <nnapi/hal/HandleError.h>
+
+#include <memory>
+#include <utility>
+#include <vector>
+
+// See hardware/interfaces/neuralnetworks/utils/README.md for more information on HIDL interface
+// lifetimes across processes and for protecting asynchronous calls across HIDL.
+
+namespace android::hardware::neuralnetworks::V1_2::utils {
+
+nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
+ std::shared_ptr<const PreparedModel> preparedModel, V1_0::Request request,
+ hal::utils::RequestRelocation relocation, V1_2::MeasureTiming measure) {
+ if (preparedModel == nullptr) {
+ return NN_ERROR() << "V1_2::utils::Execution::create must have non-null preparedModel";
+ }
+
+ return std::make_shared<const Execution>(PrivateConstructorTag{}, std::move(preparedModel),
+ std::move(request), std::move(relocation), measure);
+}
+
+Execution::Execution(PrivateConstructorTag /*tag*/,
+ std::shared_ptr<const PreparedModel> preparedModel, V1_0::Request request,
+ hal::utils::RequestRelocation relocation, V1_2::MeasureTiming measure)
+ : kPreparedModel(std::move(preparedModel)),
+ kRequest(std::move(request)),
+ kRelocation(std::move(relocation)),
+ kMeasure(measure) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
+ const nn::OptionalTimePoint& /*deadline*/) const {
+ return kPreparedModel->executeInternal(kRequest, kMeasure, kRelocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> Execution::computeFenced(
+ const std::vector<nn::SyncFence>& /*waitFor*/, const nn::OptionalTimePoint& /*deadline*/,
+ const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+ return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
+ << "IExecution::computeFenced is not supported on 1.2 HAL service";
+}
+
+} // namespace android::hardware::neuralnetworks::V1_2::utils