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/aidl/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
index 2460fba..8c8a87a 100644
--- a/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
@@ -63,6 +63,8 @@
// it is skipped. The field is set to true by default and is set to false in
// quantization coupling tests to suppress skipping a test
bool reportSkipping;
+ // `useConfig` indicates if a test should use execute*WithConfig functions for the execution.
+ bool useConfig;
TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
bool reusable)
: executor(executor),
@@ -70,7 +72,8 @@
outputType(outputType),
memoryType(memoryType),
reusable(reusable),
- reportSkipping(true) {}
+ reportSkipping(true),
+ useConfig(false) {}
TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
bool reusable, bool reportSkipping)
: executor(executor),
@@ -78,7 +81,17 @@
outputType(outputType),
memoryType(memoryType),
reusable(reusable),
- reportSkipping(reportSkipping) {}
+ reportSkipping(reportSkipping),
+ useConfig(false) {}
+ TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
+ bool reusable, bool reportSkipping, bool useConfig)
+ : executor(executor),
+ measureTiming(measureTiming),
+ outputType(outputType),
+ memoryType(memoryType),
+ reusable(reusable),
+ reportSkipping(reportSkipping),
+ useConfig(useConfig) {}
};
std::string toString(OutputType type) {
@@ -100,7 +113,8 @@
<< ", .measureTiming=" << (config.measureTiming ? "true" : "false")
<< ", .outputType=" << toString(config.outputType)
<< ", .memoryType=" << toString(config.memoryType)
- << ", .reusable=" << (config.reusable ? "true" : "false") << "}";
+ << ", .reusable=" << (config.reusable ? "true" : "false")
+ << ", .useConfig=" << (config.useConfig ? "true" : "false") << "}";
return ss.str();
}
@@ -587,8 +601,8 @@
std::shared_ptr<IExecution> execution;
if (testConfig.reusable) {
- const auto ret = preparedModel->createReusableExecution(request, testConfig.measureTiming,
- loopTimeoutDurationNs, &execution);
+ const auto ret = preparedModel->createReusableExecution(
+ request, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}}, &execution);
ASSERT_TRUE(ret.isOk()) << static_cast<nn::ErrorStatus>(ret.getServiceSpecificError());
ASSERT_NE(nullptr, execution.get());
}
@@ -607,6 +621,10 @@
::ndk::ScopedAStatus ret;
if (testConfig.reusable) {
ret = execution->executeSynchronously(kNoDeadline, &executionResult);
+ } else if (testConfig.useConfig) {
+ ret = preparedModel->executeSynchronouslyWithConfig(
+ request, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+ kNoDeadline, &executionResult);
} else {
ret = preparedModel->executeSynchronously(request, testConfig.measureTiming,
kNoDeadline, loopTimeoutDurationNs,
@@ -649,9 +667,16 @@
ExecutionResult executionResult;
// execute
- ret = burst->executeSynchronously(request, slots, testConfig.measureTiming,
- kNoDeadline, loopTimeoutDurationNs,
- &executionResult);
+ if (testConfig.useConfig) {
+ ret = burst->executeSynchronouslyWithConfig(
+ request, slots,
+ {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}}, kNoDeadline,
+ &executionResult);
+ } else {
+ ret = burst->executeSynchronously(request, slots, testConfig.measureTiming,
+ kNoDeadline, loopTimeoutDurationNs,
+ &executionResult);
+ }
ASSERT_TRUE(ret.isOk() || ret.getExceptionCode() == EX_SERVICE_SPECIFIC)
<< ret.getDescription();
if (ret.isOk()) {
@@ -680,6 +705,10 @@
::ndk::ScopedAStatus ret;
if (testConfig.reusable) {
ret = execution->executeFenced({}, kNoDeadline, kNoDuration, &executionResult);
+ } else if (testConfig.useConfig) {
+ ret = preparedModel->executeFencedWithConfig(
+ request, {}, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+ kNoDeadline, kNoDuration, &executionResult);
} else {
ret = preparedModel->executeFenced(request, {}, testConfig.measureTiming,
kNoDeadline, loopTimeoutDurationNs,
@@ -697,9 +726,19 @@
waitFor.emplace_back(dupFd);
// If a sync fence is returned, try start another run waiting for the sync
// fence.
- ret = preparedModel->executeFenced(request, waitFor, testConfig.measureTiming,
- kNoDeadline, loopTimeoutDurationNs,
- kNoDuration, &executionResult);
+ if (testConfig.reusable) {
+ ret = execution->executeFenced(waitFor, kNoDeadline, kNoDuration,
+ &executionResult);
+ } else if (testConfig.useConfig) {
+ ret = preparedModel->executeFencedWithConfig(
+ request, waitFor,
+ {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+ kNoDeadline, kNoDuration, &executionResult);
+ } else {
+ ret = preparedModel->executeFenced(
+ request, waitFor, testConfig.measureTiming, kNoDeadline,
+ loopTimeoutDurationNs, kNoDuration, &executionResult);
+ }
ASSERT_TRUE(ret.isOk());
waitForSyncFence(executionResult.syncFence.get());
}
@@ -830,11 +869,13 @@
std::vector<Executor> executorList;
std::vector<MemoryType> memoryTypeList;
std::vector<bool> reusableList = {false};
+ std::vector<bool> useConfigList = {false};
int deviceVersion;
ASSERT_TRUE(device->getInterfaceVersion(&deviceVersion).isOk());
if (deviceVersion >= kMinAidlLevelForFL8) {
reusableList.push_back(true);
+ useConfigList.push_back(true);
}
switch (testKind) {
@@ -879,11 +920,14 @@
for (const Executor executor : executorList) {
for (const MemoryType memoryType : memoryTypeList) {
for (const bool reusable : reusableList) {
- if (executor == Executor::BURST && reusable) continue;
- const TestConfig testConfig(executor, measureTiming, outputType, memoryType,
- reusable);
- SCOPED_TRACE(toString(testConfig));
- EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+ for (const bool useConfig : useConfigList) {
+ if ((useConfig || executor == Executor::BURST) && reusable) continue;
+ const TestConfig testConfig(executor, measureTiming, outputType,
+ memoryType, reusable,
+ /*reportSkipping=*/true, useConfig);
+ SCOPED_TRACE(toString(testConfig));
+ EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+ }
}
}
}
@@ -942,6 +986,13 @@
createPreparedModel(device, model, &preparedModel);
if (preparedModel == nullptr) return;
EvaluatePreparedModel(device, preparedModel, testModel, testKind);
+ int32_t deviceVersion;
+ ASSERT_TRUE(device->getInterfaceVersion(&deviceVersion).isOk());
+ if (deviceVersion >= kMinAidlLevelForFL8) {
+ createPreparedModel(device, model, &preparedModel, /*reportSkipping*/ true,
+ /*useConfig*/ true);
+ EvaluatePreparedModel(device, preparedModel, testModel, testKind);
+ }
} break;
case TestKind::QUANTIZATION_COUPLING: {
ASSERT_TRUE(testModel.hasQuant8CoupledOperands());