HAL interface for compilation and execution hints
The following AIDL types are added:
- TokenValuePair
- PrepareModelConfig
- ExecutionConfig
The following AIDL methods are added:
- IDevice::prepareModelWithConfig
- IPreparedModel::executeSynchronouslyWithConfig
- IPreparedModel::executeFencedWithConfig
- IBurst::executeSynchronouslyWithConfig
The compilation and execution hints are being stored as a list of
token-value pairs as part of the PrepareModelConfig / ExecutionConfig.
And the PrepareModelConfig / ExecutionConfig parcelables are created in
order to make future extensions to the execution related interfaces
easier.
It is the drivers responsibility to verify the hints, and it is allowed
for the driver to ignore them.
Bug: 203248587
Test: neuralnetworks_utils_hal_aidl_test
Change-Id: I98240fd75089fc85cdfcaa0be28aab8a6f0dfca5
Merged-In: I98240fd75089fc85cdfcaa0be28aab8a6f0dfca5
(cherry picked from commit 0e671f3edb9d2c78658a4ef4169e3211e3f9bb00)
diff --git a/neuralnetworks/1.2/utils/src/Burst.cpp b/neuralnetworks/1.2/utils/src/Burst.cpp
index 911fbfa..23e8070 100644
--- a/neuralnetworks/1.2/utils/src/Burst.cpp
+++ b/neuralnetworks/1.2/utils/src/Burst.cpp
@@ -305,8 +305,9 @@
nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::execute(
const nn::Request& request, nn::MeasureTiming measure,
- const nn::OptionalTimePoint& deadline,
- const nn::OptionalDuration& loopTimeoutDuration) const {
+ const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
// This is the first point when we know an execution is occurring, so begin to collect
// systraces. Note that the first point we can begin collecting systraces in
// ExecutionBurstServer is when the RequestChannelReceiver realizes there is data in the FMQ, so
@@ -317,7 +318,7 @@
// fall back to another execution path
if (!compliantVersion(request).ok()) {
// fallback to another execution path if the packet could not be sent
- return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+ return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration, {}, {});
}
// ensure that request is ready for IPC
@@ -346,7 +347,7 @@
// send request packet
const auto requestPacket = serialize(hidlRequest, hidlMeasure, slots);
const auto fallback = [this, &request, measure, &deadline, &loopTimeoutDuration] {
- return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+ return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration, {}, {});
};
return executeInternal(requestPacket, relocation, fallback);
}
@@ -354,14 +355,17 @@
// See IBurst::createReusableExecution for information on this method.
nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
const nn::Request& request, nn::MeasureTiming measure,
- const nn::OptionalDuration& loopTimeoutDuration) const {
+ const nn::OptionalDuration& loopTimeoutDuration,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
NNTRACE_RT(NNTRACE_PHASE_EXECUTION, "Burst::createReusableExecution");
// if the request is valid but of a higher version than what's supported in burst execution,
// fall back to another execution path
if (!compliantVersion(request).ok()) {
// fallback to another execution path if the packet could not be sent
- return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration);
+ return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration, {},
+ {});
}
// ensure that request is ready for IPC
diff --git a/neuralnetworks/1.2/utils/src/Conversions.cpp b/neuralnetworks/1.2/utils/src/Conversions.cpp
index 838d9c4..62ec2ed 100644
--- a/neuralnetworks/1.2/utils/src/Conversions.cpp
+++ b/neuralnetworks/1.2/utils/src/Conversions.cpp
@@ -212,9 +212,9 @@
};
}
-GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
+GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
const hal::V1_2::Model::ExtensionNameAndPrefix& extensionNameAndPrefix) {
- return Model::ExtensionNameAndPrefix{
+ return ExtensionNameAndPrefix{
.name = extensionNameAndPrefix.name,
.prefix = extensionNameAndPrefix.prefix,
};
@@ -495,7 +495,7 @@
}
nn::GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
- const nn::Model::ExtensionNameAndPrefix& extensionNameAndPrefix) {
+ const nn::ExtensionNameAndPrefix& extensionNameAndPrefix) {
return Model::ExtensionNameAndPrefix{
.name = extensionNameAndPrefix.name,
.prefix = extensionNameAndPrefix.prefix,
diff --git a/neuralnetworks/1.2/utils/src/Device.cpp b/neuralnetworks/1.2/utils/src/Device.cpp
index e7acecd..3a58d2c 100644
--- a/neuralnetworks/1.2/utils/src/Device.cpp
+++ b/neuralnetworks/1.2/utils/src/Device.cpp
@@ -236,7 +236,9 @@
nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
const nn::Model& model, nn::ExecutionPreference preference, nn::Priority /*priority*/,
nn::OptionalTimePoint /*deadline*/, const std::vector<nn::SharedHandle>& modelCache,
- const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+ const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
// Ensure that model is ready for IPC.
std::optional<nn::Model> maybeModelInShared;
const nn::Model& modelInShared =
diff --git a/neuralnetworks/1.2/utils/src/PreparedModel.cpp b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
index 6df3df3..feb3951 100644
--- a/neuralnetworks/1.2/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
@@ -91,7 +91,9 @@
nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
const nn::Request& request, nn::MeasureTiming measure,
const nn::OptionalTimePoint& /*deadline*/,
- const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+ const nn::OptionalDuration& /*loopTimeoutDuration*/,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
// Ensure that request is ready for IPC.
std::optional<nn::Request> maybeRequestInShared;
hal::utils::RequestRelocation relocation;
@@ -123,19 +125,22 @@
}
nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFenced(const nn::Request& /*request*/,
- const std::vector<nn::SyncFence>& /*waitFor*/,
- nn::MeasureTiming /*measure*/,
- const nn::OptionalTimePoint& /*deadline*/,
- const nn::OptionalDuration& /*loopTimeoutDuration*/,
- const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+PreparedModel::executeFenced(
+ const nn::Request& /*request*/, const std::vector<nn::SyncFence>& /*waitFor*/,
+ nn::MeasureTiming /*measure*/, const nn::OptionalTimePoint& /*deadline*/,
+ const nn::OptionalDuration& /*loopTimeoutDuration*/,
+ const nn::OptionalDuration& /*timeoutDurationAfterFence*/,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
<< "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 {
+ const nn::OptionalDuration& /*loopTimeoutDuration*/,
+ const std::vector<nn::TokenValuePair>& /*hints*/,
+ const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
// Ensure that request is ready for IPC.
std::optional<nn::Request> maybeRequestInShared;
hal::utils::RequestRelocation relocation;