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/PreparedModel.cpp b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
index b209a44..e01401b 100644
--- a/neuralnetworks/1.2/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
@@ -18,6 +18,7 @@
#include "Callbacks.h"
#include "Conversions.h"
+#include "Execution.h"
#include "ExecutionBurstController.h"
#include "ExecutionBurstUtils.h"
#include "Utils.h"
@@ -93,19 +94,31 @@
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)));
- auto result = kExecuteSynchronously ? executeSynchronously(hidlRequest, hidlMeasure)
- : executeAsynchronously(hidlRequest, hidlMeasure);
+ return executeInternal(hidlRequest, hidlMeasure, relocation);
+}
+
+nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
+PreparedModel::executeInternal(const V1_0::Request& request, MeasureTiming measure,
+ const hal::utils::RequestRelocation& relocation) const {
+ if (relocation.input) {
+ relocation.input->flush();
+ }
+
+ auto result = kExecuteSynchronously ? executeSynchronously(request, measure)
+ : executeAsynchronously(request, measure);
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);
}
@@ -120,6 +133,21 @@
<< "IPreparedModel::executeFenced is not supported on 1.2 HAL service";
}
+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));
+ return Execution::create(shared_from_this(), std::move(hidlRequest), std::move(relocation),
+ hidlMeasure);
+}
+
nn::GeneralResult<nn::SharedBurst> PreparedModel::configureExecutionBurst() const {
auto self = shared_from_this();
auto fallback = [preparedModel = std::move(self)](