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/1.3/utils/src/Execution.cpp b/neuralnetworks/1.3/utils/src/Execution.cpp
new file mode 100644
index 0000000..3d17cc3
--- /dev/null
+++ b/neuralnetworks/1.3/utils/src/Execution.cpp
@@ -0,0 +1,84 @@
+/*
+ * 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 "Conversions.h"
+#include "PreparedModel.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/types.h>
+#include <android/hardware/neuralnetworks/1.3/IPreparedModel.h>
+#include <android/hardware/neuralnetworks/1.3/types.h>
+#include <nnapi/IExecution.h>
+#include <nnapi/Result.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_3::utils {
+
+nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
+ std::shared_ptr<const PreparedModel> preparedModel, Request request,
+ hal::utils::RequestRelocation relocation, V1_2::MeasureTiming measure,
+ OptionalTimeoutDuration loopTimeoutDuration) {
+ if (preparedModel == nullptr) {
+ return NN_ERROR() << "V1_3::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,
+ std::move(loopTimeoutDuration));
+}
+
+Execution::Execution(PrivateConstructorTag /*tag*/,
+ std::shared_ptr<const PreparedModel> preparedModel, Request request,
+ hal::utils::RequestRelocation relocation, V1_2::MeasureTiming measure,
+ OptionalTimeoutDuration loopTimeoutDuration)
+ : kPreparedModel(std::move(preparedModel)),
+ kRequest(std::move(request)),
+ kRelocation(std::move(relocation)),
+ kMeasure(measure),
+ kLoopTimeoutDuration(std::move(loopTimeoutDuration)) {}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Execution::compute(
+ const nn::OptionalTimePoint& deadline) const {
+ const auto hidlDeadline = NN_TRY(hal::utils::makeExecutionFailure(convert(deadline)));
+ return kPreparedModel->executeInternal(kRequest, kMeasure, hidlDeadline, kLoopTimeoutDuration,
+ 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 {
+ const auto hidlWaitFor = NN_TRY(hal::utils::convertSyncFences(waitFor));
+ const auto hidlDeadline = NN_TRY(convert(deadline));
+ const auto hidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
+ return kPreparedModel->executeFencedInternal(kRequest, hidlWaitFor, kMeasure, hidlDeadline,
+ kLoopTimeoutDuration,
+ hidlTimeoutDurationAfterFence, kRelocation);
+}
+
+} // namespace android::hardware::neuralnetworks::V1_3::utils
diff --git a/neuralnetworks/1.3/utils/src/PreparedModel.cpp b/neuralnetworks/1.3/utils/src/PreparedModel.cpp
index fd7f8f2..5163e8a 100644
--- a/neuralnetworks/1.3/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.3/utils/src/PreparedModel.cpp
@@ -18,6 +18,7 @@
#include "Callbacks.h"
#include "Conversions.h"
+#include "Execution.h"
#include "Utils.h"
#include <android/hardware/neuralnetworks/1.0/types.h>
@@ -139,8 +140,10 @@
const nn::OptionalDuration& loopTimeoutDuration) const {
// Ensure that request is ready for IPC.
std::optional<nn::Request> maybeRequestInShared;
- const nn::Request& requestInShared = NN_TRY(hal::utils::makeExecutionFailure(
- hal::utils::flushDataFromPointerToShared(&request, &maybeRequestInShared)));
+ hal::utils::RequestRelocation relocation;
+ const nn::Request& requestInShared =
+ NN_TRY(hal::utils::makeExecutionFailure(hal::utils::convertRequestFromPointerToShared(
+ &request, &maybeRequestInShared, &relocation)));
const auto hidlRequest = NN_TRY(hal::utils::makeExecutionFailure(convert(requestInShared)));
const auto hidlMeasure = NN_TRY(hal::utils::makeExecutionFailure(convert(measure)));
@@ -148,16 +151,27 @@
const auto hidlLoopTimeoutDuration =
NN_TRY(hal::utils::makeExecutionFailure(convert(loopTimeoutDuration)));
+ return executeInternal(hidlRequest, hidlMeasure, hidlDeadline, hidlLoopTimeoutDuration,
+ relocation);
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+PreparedModel::executeInternal(const Request& request, V1_2::MeasureTiming measure,
+ const OptionalTimePoint& deadline,
+ const OptionalTimeoutDuration& loopTimeoutDuration,
+ const hal::utils::RequestRelocation& relocation) const {
+ if (relocation.input) {
+ relocation.input->flush();
+ }
+
auto result = kExecuteSynchronously
- ? executeSynchronously(hidlRequest, hidlMeasure, hidlDeadline,
- hidlLoopTimeoutDuration)
- : executeAsynchronously(hidlRequest, hidlMeasure, hidlDeadline,
- hidlLoopTimeoutDuration);
+ ? executeSynchronously(request, measure, deadline, loopTimeoutDuration)
+ : executeAsynchronously(request, measure, deadline, loopTimeoutDuration);
auto [outputShapes, timing] = NN_TRY(std::move(result));
- NN_TRY(hal::utils::makeExecutionFailure(
- hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared)));
-
+ if (relocation.output) {
+ relocation.output->flush();
+ }
return std::make_pair(std::move(outputShapes), timing);
}
@@ -168,8 +182,9 @@
const nn::OptionalDuration& timeoutDurationAfterFence) const {
// Ensure that request is ready for IPC.
std::optional<nn::Request> maybeRequestInShared;
- const nn::Request& requestInShared =
- NN_TRY(hal::utils::flushDataFromPointerToShared(&request, &maybeRequestInShared));
+ hal::utils::RequestRelocation relocation;
+ const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+ &request, &maybeRequestInShared, &relocation));
const auto hidlRequest = NN_TRY(convert(requestInShared));
const auto hidlWaitFor = NN_TRY(hal::utils::convertSyncFences(waitFor));
@@ -178,27 +193,58 @@
const auto hidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
const auto hidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
+ return executeFencedInternal(hidlRequest, hidlWaitFor, hidlMeasure, hidlDeadline,
+ hidlLoopTimeoutDuration, hidlTimeoutDurationAfterFence,
+ relocation);
+}
+
+nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
+PreparedModel::executeFencedInternal(const Request& request, const hidl_vec<hidl_handle>& waitFor,
+ V1_2::MeasureTiming measure, const OptionalTimePoint& deadline,
+ const OptionalTimeoutDuration& loopTimeoutDuration,
+ const OptionalTimeoutDuration& timeoutDurationAfterFence,
+ const hal::utils::RequestRelocation& relocation) const {
+ if (relocation.input) {
+ relocation.input->flush();
+ }
+
auto cb = hal::utils::CallbackValue(fencedExecutionCallback);
- const auto ret = kPreparedModel->executeFenced(hidlRequest, hidlWaitFor, hidlMeasure,
- hidlDeadline, hidlLoopTimeoutDuration,
- hidlTimeoutDurationAfterFence, cb);
+ const auto ret =
+ kPreparedModel->executeFenced(request, waitFor, measure, deadline, loopTimeoutDuration,
+ timeoutDurationAfterFence, cb);
HANDLE_TRANSPORT_FAILURE(ret);
auto [syncFence, callback] = NN_TRY(cb.take());
// If executeFenced required the request memory to be moved into shared memory, block here until
// the fenced execution has completed and flush the memory back.
- if (maybeRequestInShared.has_value()) {
+ if (relocation.output) {
const auto state = syncFence.syncWait({});
if (state != nn::SyncFence::FenceState::SIGNALED) {
return NN_ERROR() << "syncWait failed with " << state;
}
- NN_TRY(hal::utils::unflushDataFromSharedToPointer(request, maybeRequestInShared));
+ relocation.output->flush();
}
return std::make_pair(std::move(syncFence), std::move(callback));
}
+nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
+ const nn::Request& request, nn::MeasureTiming measure,
+ const nn::OptionalDuration& loopTimeoutDuration) const {
+ // Ensure that request is ready for IPC.
+ std::optional<nn::Request> maybeRequestInShared;
+ hal::utils::RequestRelocation relocation;
+ const nn::Request& requestInShared = NN_TRY(hal::utils::convertRequestFromPointerToShared(
+ &request, &maybeRequestInShared, &relocation));
+
+ auto hidlRequest = NN_TRY(convert(requestInShared));
+ auto hidlMeasure = NN_TRY(convert(measure));
+ auto hidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
+ return Execution::create(shared_from_this(), std::move(hidlRequest), std::move(relocation),
+ hidlMeasure, std::move(hidlLoopTimeoutDuration));
+}
+
nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() const {
auto self = shared_from_this();
auto fallback = [preparedModel = std::move(self)](