Certain mutation testing -- mutateOperandLifeTimeTest and
mutateOperandInputOutputTest -- can introduce potentially very large
CONSTANT_COPY operands, which can in turn create potentially very
large Models which must be passed across binder. To avoid overflowing
the binder buffer, we estimate the size of the mutated Model, and skip
the test if that size is too high. The old logic recognizes that our
tests only have a single active binder transaction at a time, and
assumes that there are no other clients using the same service at the
same time, and so we should have the binder buffer to ourselves; to be
conservative, we reject any Model whose estimated size exceeds half
the binder buffer size. Unfortunately, sometimes the binder buffer
still overflows, because it unexpectedly contains an allocation from
some other transaction: It appears that binder buffer memory
management is not serialized with respect to transactions from our
tests, and therefore depending on scheduler behavior, there may be a
sizeable allocation still in the buffer when we attempt to pass the
large Model. To fix this problem we become even more conservative,
and instead of limiting the Model to half the binder buffer size, we
limit it to half IBinder.MAX_IPC_SIZE (the recommended transaction
size limit). To confirm that this change does not exclude too many
tests, I checked how may times the size filter function
exceedsBinderSizeLimit is called, how many times it rejects a model
under the new logic (modelsExceedHalfMaxIPCSize), and how many times
it rejects a model under the old logic (modelsExceedHalfMaxIPCSize).
Test: VtsHalNeuralnetworksV1_0TargetTest --gtest_filter=TestGenerated/ValidationTest.Test/*-*dsp*
Test: # models = 3592, modelsExceedHalfMaxIPCSize = 212, modelsExceedHalfBufferSize = 18
Test: VtsHalNeuralnetworksV1_1TargetTest --gtest_filter=TestGenerated/ValidationTest.Test/*-*dsp*
Test: # models = 7228, modelsExceedHalfMaxIPCSize = 330, modelsExceedHalfBufferSize = 28
Test: VtsHalNeuralnetworksV1_2TargetTest --gtest_filter=TestGenerated/ValidationTest.Test/*-*dsp*
Test: # models = 52072, modelsExceedHalfMaxIPCSize = 506, modelsExceedHalfBufferSize = 28
Test: VtsHalNeuralnetworksV1_3TargetTest --gtest_filter=TestGenerated/ValidationTest.Test/*-*dsp*
Test: # models = 73342, modelsExceedHalfMaxIPCSize = 568, modelsExceedHalfBufferSize = 28
Test: VtsHalNeuralnetworksTargetTest
Bug: 227719657
Bug: 227719752
Bug: 231928847
Bug: 238777741
Bug: 242271308
Change-Id: I3f81d71ca3c0ad4c639096b1dc034a8909bc8971
- detect cycle (CycleTest)
- detect bad execution order (mutateExecutionOrderTest)
- detect lifetime inconsistent with whether operand is written (mutateOperandLifeTimeTest)
- detect lifetime inconsistent with Model inputIndexes/outputIndexes (mutateOperandInputOutputTest)
- detect incorrect number of consumers (mutateOperandNumberOfConsumersTest)
- detect operand written multiple times (mutateOperandAddWriterTest)
- detect operand never written (mutateOperationRemoveWriteTest)
Bug: 66478689
Test: VtsHalNeuralnetworksV1_*TargetTest
Change-Id: Id4ba19660bbd31a16f8a675f7b6437f4d779e8da
This CL also cleans up the validation framework code.
Bug: 67828197
Test: mma
Test: VtsHalNeuralnetworksV1_*TargetTest
Change-Id: I84661fb2b8204148788d10425ca0ac986158b15f
This CL does additional NNAPI VTS test cleanup, including consolidating
duplicate functionality. Specifically, this CL:
* consolidates the createPreparedModel function, removing the duplicate
* consolidates the std::out ErrorStatus and DeviceStatus code into Utils
* changes non-null constant pointers to constant references
* removes redudant leading namespace specifiers (V1_0::, ::testing, etc.)
* makes the Valdiation tests free functions
* renames device to kDevice and mTestModel to kTestModel
Bug: N/A
Test: mma
Test: VtsHalNeuralnetworksV1_*TargetTest (with sample-all)
Change-Id: Ic401bb1f1760cc10384ac0d30c0c93409b63a9c7
To make it easier to create the next version of NNAPI, this change
removes the following nonsensical dependence:
- NNAPI 1.0 VTS depends on NNAPI 1.1 and 1.2
- NNAPI 1.1 VTS depends on NNAPI 1.2
In particular, I made the following changes:
- split GeneratedTestHarness.cpp into three separate implementations,
- created a restricted version of Callbacks.h for 1.0 and 1.1,
- removed the dependency on frameworks/ml/nn/HalInterfaces.h,
- refactored Android.bp files for more autonomy between 1.0, 1.1, and 1.2,
- consolidated some common code into Utils.h,
- created structure for sharing code between VTS versions (VtsHalNeuralNetworksV1_0_utils).
Bug: 74827824
Bug: 124462414
Test: VtsHalNeuralnetworksV1_0TargetTest
Test: VtsHalNeuralnetworksV1_1TargetTest
Test: VtsHalNeuralnetworksV1_1CompatV1_0TargetTest
Test: VtsHalNeuralnetworksV1_2TargetTest
Test: VtsHalNeuralnetworksV1_2CompatV1_0TargetTest
Test: VtsHalNeuralnetworksV1_2CompatV1_1TargetTest
Change-Id: I4243d0b5e574255cef1070850f4d0a284f65f54e
- Instead of isCachingSupport returning a single boolean, switch to
getNumberOfCacheFilesNeeded returning the number of cache files. This
is to support use cases when driver needs more than one cache file for
each type, or when driver does not need data cache.
- Instead of a separate saveToCache, pass cache info along with
prepareModel_1_2 to save into cache as well as perform compilation.
This is to avoid a potential additional copy of cache files.
Bug: 123780248
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Test: VtsHalNeuralnetworksV1_xTargetTest with a test driver that can
read and write cache entries
Change-Id: I921b7b8ccc3c66af19f6589f7213c6870d6f07bf
The filter tensor is not always operand 1 if there are multiple
operations in the model.
Test: 1.2 VTS tests with sample driver
Change-Id: I8925dfd18072ae5579657e8aa13128e7d9f58324
* Adds a specification of invalid scale and zero point for TENSOR_BOOL8.
This fixes vts failures for comparison ops.
* Removes (FUNDAMENTAL_MIN - 1) from invalid OperationTypes.
FUNDAMENTAL_MIN is equal to 0 and resulting -1 was statically casted to
uint32_t and passed 4294967295 as an invalid OperationType. However, our
validateOperation function interpreted this ID as an extension ID and
didn't fail.
* Adds mutateOperationOperandTypeSkip for QUANTIZE and DEQUANTIZE.
* Adds removeOperandSkip for BIDIRECTIONAL_SEQUENCE_RNN.
Fix: 121130841
Fix: 123247345
Test: VtsHalNeuralnetworksV1_2TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Change-Id: Iefb502c6b9301d5470eb4cdaa46d398f1a0e512a
Enable VTS unit test for dynamic output shape deduction.
Only test dynamic output shape for V1_2::IDevice with V1_2::Model.
Bug: 73506513
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Change-Id: I4134e1ec54a15554eb8533134897279651b57da3
Create 1.2 version IPreparedModel, IPreparedModelCallback, and
IExecutionCallback.
Currently the new interfaces are created the same as 1.0 version,
but will have more methods introduced in later CLs.
Bug: 73506513
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Change-Id: Icf4d04c22f88e825d87562f1489377fdf6bf585d
Added ExtraParams union for extra Operand parameters.
It's a more sensible approach than adding more fields
to the Operand struct.
Bug: 119249581
Test: NeuralNetworksTest_static
Test: VtsHalNeuralnetworksV1_0TargetTest
Test: VtsHalNeuralnetworksV1_1TargetTest
Test: VtsHalNeuralnetworksV1_2TargetTest
Change-Id: I59731134cf0ea34cf9e10342686d331da9e9c3b3
Moves comments for type range values into their enums to make it harder
to forget to update them.
Adds missing float16 types for ARGMAX/ARGMIN/CAST.
Test: VtsHalNeuralnetworksV1_2TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Test: [ PASSED ] 2519 tests.
Change-Id: Ic7c3df8c8fbff45fe497f304b6b2c7a09e7dc5a6
* Update doc string
* Update zero point mutation to check for symmetric quantization
Fix: 118671831
Test: VtsHalNeuralnetworksV1_2TargetTest
Change-Id: Id1999c793c839b892cfe45cbb245611b12db2a72
This abstracts the boundary values for OperandType and
OperationType to avoid the need to update them in the
model validation functions.
Test: VtsHalNeuralnetworksV1_2TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Test: VtsHalNeuralnetworksV1_2CompatV1_1TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Test: VtsHalNeuralnetworksV1_2CompatV1_0TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Change-Id: I39155148d67215e32b4eb1991b885f65d5eeaca8
Skip L2_NORMALIZATION, LOCAL_RESPONSE_NORMALIZATION, and SOFTMAX for
addOperationInputTest.
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Change-Id: I16de09e2eef7610e564f1f49ff067af7efca3784
This is a copy the v1.1 tests since we don't have any new ops
implemented in v1.2 yet.
Bug: 114365802
Test: mm
Test: NNAPI VTS
Change-Id: Ida7525fcd3ae0fd6f88ff9591e06aba922bdae64
