The documentation said that cell-to-input weights are required to be
present when input-to-input weights, recurrent-to-input weights and
input gate bias are present. This was incorrect since this weights can
be omitted if peephole connections are not used even if all the other
tensors are present.
Another bug that is fixed in this change is that for output #0 the docs
said "of shape [batch_size, num_units * 4] with CIFG, or [batch_size,
num_units * 3] without CIFG" when in fact it is the opposite, i.e. "of
shape [batch_size, num_units * 3] with CIFG, or [batch_size, num_units *
4] without CIFG."
Existing CTS/VTS tests expect behaviour described in the fixed documentation.
Existing CPU implementation is also compliant with the fixed documentation.
Fix: 111842951
Test: mma
Change-Id: Id011783e33672ae65dc6fe3784cb26feb832acf9
Merged-In: Id011783e33672ae65dc6fe3784cb26feb832acf9
(cherry picked from commit e0537f09fb)
Prior to this CL, the VTS compilation caching tests are using float32
models. However, there exists several drivers that only support QUANT8
operations, and most of the tests are simply skipped.
This CL makes the CompilationCachingTest fixture parameterized that
every test will run twice, with the first pass running with float32
models and the second pass running with quant8 models.
Bug: 132112463
Test: 1.2 VTS with sample drivers
Test: 1.2 VTS with test drivers that can read and write cache entries
Change-Id: Iaf28065e46c2d90ab8b347b9848be1e6546cce86
Merged-In: Iaf28065e46c2d90ab8b347b9848be1e6546cce86
(cherry picked from commit 35e7599eb2)
Before this change, getSupportedOperations_1_2 was invoked in
saveModelToCache prior to prepareModel_1_2, which delayed the model
preparation and cache overwriting always ran faster in
SaveToCache_TOCTOU test.
This change increases the chance that the model is overwritten after it
is saved in SaveToCache_TOCTOU test by moving getSupportedOperations_1_2
out of saveToCache.
Bug: 123433989
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Test: VtsHalNeuralnetworksV1_xTargetTest with a test driver that can
read and write cache entries
Test: TOCTOU tests with a sample driver that is deliberately broken
Change-Id: Ieba6a82252623ac429d5fcb6950b99db949f9886
Merged-In: Ieba6a82252623ac429d5fcb6950b99db949f9886
(cherry picked from commit f8c5987bbd)
We only expect the driver to not crash.
Bug: 123433989
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Test: VtsHalNeuralnetworksV1_xTargetTest with a test driver that can
read and write cache entries
Change-Id: Ic9bd7ad6e42d77d505955cb9dda597a39e95cdb6
Merged-In: Ic9bd7ad6e42d77d505955cb9dda597a39e95cdb6
(cherry picked from commit 83ab17f224)
Two tests are added into VTS:
1. The TOCTOU test, for a fixed number of iterations.
We try to attack the driver by replacing the model cache with another
cache entry from a similar model while the driver is saving to or
preparing from cache, and see if the driver crashes or falsely prepares
an unexpected model (by checking the execution result). Either of the
following results are acceptable
* Fails -> the driver detects this corruption.
* Succeeds and produces correct execution result -> the corruption
happens before saving to cache or after preparing from cache.
Due to the racy nature, this test is probabilistic and we run it several
times.
2. Similar to the TOCTOU test but replace the content between
compile-to-cache and compile-from-cache (once only).
Additionally, remove tmp cache directory if test succeeds.
Bug: 123433989
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Test: VtsHalNeuralnetworksV1_xTargetTest with a test driver that can
read and write cache entries
Test: Check /data/local/tmp and confirm that
* dirs are preserved on failure
* dirs are removed on success
Change-Id: Ie04fa905f465c3775979f0ca74359d185dcacea9
Merged-In: Ie04fa905f465c3775979f0ca74359d185dcacea9
(cherry picked from commit a44e130a92)
This CL adds the following two types of validation tests on the NNAPI
Burst serialized format:
(1) it directly modifies the serialized data (invalidating it) to ensure
that vendor driver services properly validates the serialized
request
(2) it ensures that vendor driver services properly fail when the result
channel is not large enough to return the data
This CL additionally includes miscellaneous cleanups:
(1) having a generic "validateEverything" function
(2) moving the "prepareModel" function that's common across
validateRequest and validateBurst to a common area
Fixes: 129779280
Bug: 129157135
Test: mma
Test: VtsHalNeuralnetworksV1_2TargetTest (with sample-all)
Change-Id: Ib90fe7f662824de17db5a254a8c501855e45f6bd
Merged-In: Ib90fe7f662824de17db5a254a8c501855e45f6bd
(cherry picked from commit 20f28a24e9)
We had no tests for quantized PAD in NNAPI 1.1 and think that vendors might have implemented different behaviors.
Bug: 122243484
Test: N/A
Change-Id: Ibfc0801ab746fc271dc5f8efc764b818c6d49df4
Merged-In: Ibfc0801ab746fc271dc5f8efc764b818c6d49df4
(cherry picked from commit b01ce9644e)
RESIZE_NEAREST_NEIGHBOR
- The CPU implementation always had the order of {width, height}.
- In P, the documentation was incorrectly changed to {height, width}.
Bug: 131623949
Bug: 130035110
Test: mm
Change-Id: I6c79459fa73347fb51fc34a76ad78d5ac207f210
Merged-In: I6c79459fa73347fb51fc34a76ad78d5ac207f210
(cherry picked from commit 286339b4c8)
Update the docs for ops:
- LOG_SOFTMAX
- LOCAL_RESPONSE_NORMALIZATION
The docs mentioned possibility to use float16 as an input type but
didn't specify that other inputs must be float16 as well in that case.
Fix: 123500722
Test: mma
Change-Id: I9028c4a109c80f0b8571fab45555818e9e4bc783
Merged-In: I9028c4a109c80f0b8571fab45555818e9e4bc783
(cherry picked from commit 6d13ba258b)
hidl-generated makefiles are now generated such that bpfmt(file) == file.
Bug: 67417008
Test: enable bpfmt hook
Change-Id: I1f69d292bc23a7cc293a66110cb02d597e1019ad
In case of CIFG LSTM, input layer norm weights are not used in the
computation.
Bug: 129126572
Test: mma
Change-Id: I57bc578606132a2c44c71ab63dd7645dcc001302
Merged-In: I57bc578606132a2c44c71ab63dd7645dcc001302
(cherry picked from commit 0480af9aa8)
This CL adapts the VTS code to the corresponding changes made in the NN
utility library.
Bug: 128319484
Test: mma
Test: atest VtsHalNeuralnetworksV1_0Target
Test: atest VtsHalNeuralnetworksV1_1Target
Test: atest VtsHalNeuralnetworksV1_2Target
Change-Id: I470e8228cde2b75620ad851e4fe408f8e8329e7c
Merged-In: I470e8228cde2b75620ad851e4fe408f8e8329e7c
(cherry picked from commit 102e0442d8)
This CL adapts the VTS code to the corresponding changes made in the NN
utility library.
Bug: 119570067
Test: mma
Test: atest VtsHalNeuralnetworksV1_0Target
Test: atest VtsHalNeuralnetworksV1_1Target
Test: atest VtsHalNeuralnetworksV1_2Target
Change-Id: I7cbc1d7025c0352aa1ed29d71dc84c2fcfc20a4f
Merged-In: I7cbc1d7025c0352aa1ed29d71dc84c2fcfc20a4f
(cherry picked from commit e68668f65b)
- 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
Merged-In: I921b7b8ccc3c66af19f6589f7213c6870d6f07bf
(cherry picked from commit b61ba1ed0b)
Also updates documentation for this op and UnidirectionalSequenceLSTM
op.
Bug: 123644584
Test: in ag/6758764
Change-Id: I72d029fef6d890eb1771c21814b028b09af280c7
Merged-In: I72d029fef6d890eb1771c21814b028b09af280c7
(cherry picked from commit f404a1e894)
Performance information in Capabilities is used by the runtime when
it selects the appropriate processor to distribute work to. Prior to
this CL, Capabilities can only distinguish between float and non-float
data types -- so, for example, float16 and float32 performance is
considered to be the same, and performance for all non-float data types is
considered to be the same.
Bug: 124041010
Test: NeuralNetworksTest_static
Test: VtsHalNeuralnetworksV1_2TargetTest --hal_service_instance=android.hardware.neuralnetworks@1.2::IDevice/sample-all
Change-Id: I83fb5920c1c75afbd7750d793a0b8f3e72a0552c
Merged-In: I83fb5920c1c75afbd7750d793a0b8f3e72a0552c
(cherry picked from commit 632b4bd9b0)
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
Merged-In: I8925dfd18072ae5579657e8aa13128e7d9f58324
(cherry picked from commit 8804423ca1)
This CL creates a new suite of tests to enable presubmit tests:
* PresubmitHalNeuralnetworksV1_0TargetTest
* PresubmitHalNeuralnetworksV1_1TargetTest
* PresubmitHalNeuralnetworksV1_2TargetTest
These tests are the same as the VTS tests, with the exception that they
will skip running all tests (but still pass) if the service cannot be
found and its name starts with "service-".
This change does not affect the existing NNAPI VTS tests.
Test: mma
Test: atest
Bug: 124040554
Change-Id: I36a38b66b21fd51d0ca381bb4e05a39266dd353f
(cherry picked from commit ed68233697)
Argument-dependent lookup will only work for operator>> if the operator
is in one of the argument's namespaces. This caused the enumerations to
pretty-print for V1_0, but not for V1_1 or V1_2. This change ensures the
V1_0 namespace is used.
Test: mma
Test: atest VtsHalNeuralnetworksV1_0TargetTest (verified the test output "OFFLINE" for DeviceStatus and "DEVICE_UNAVAILABLE" for ErrorStatus instead of raw byte value representation)
Test: atest VtsHalNeuralnetworksV1_1TargetTest (verified the test output "OFFLINE" for DeviceStatus and "DEVICE_UNAVAILABLE" for ErrorStatus instead of raw byte value representation)
Test: atest VtsHalNeuralnetworksV1_2TargetTest (verified ran and passed tests)
Fixes: 124316129
Change-Id: I764a46e2d87615b1f3da0ab0e6edb134bb533887
(cherry picked from commit 42a35bee10)
The tmp directory is only removed when the driver reports caching not
supported, otherwise it is kept for debugging purpose.
Test: 1.2 VTS tests with sample driver
Change-Id: I5969beb1ec365c992765f40d7693630606f16e18
Merged-In: I5969beb1ec365c992765f40d7693630606f16e18
(cherry picked from commit 350d91b1df)
Test was failing with segmentation fault (and crashing whole VTS) when
device was not available.
Change-Id: Id0f28d061dc5858fa00ef1bac5f7aa467d860864
Merged-In: Id0f28d061dc5858fa00ef1bac5f7aa467d860864
(cherry picked from commit 70d25b813e)
* 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
Merged-In: Iefb502c6b9301d5470eb4cdaa46d398f1a0e512a
(cherry picked from commit 923b8c5842)
This is a follow-up to change Ia9b99015eec7a48bbf969cbe503862271f09adca
Bug: 118605927
Test: mma
Change-Id: I7ddafca04bce6fd37a9c0877270cee325111d833
Merged-In: I7ddafca04bce6fd37a9c0877270cee325111d833
(cherry picked from commit 1ffe69a8e9)
Add the following tests for compilation caching:
- validation tests
- Test isCachingSupported
- Test prepareModelFromCache with invalid numFd and invalid access mode
- Test saveToCache with invalid numFd, invalid access mode,
invalid file size, and invalid fd offset
- execution test
- Save a mobilenet model to cache and then retrieve and run accuracy
evaluation.
- The same test but the file offsets for prepareModelFromCache is not at zero.
- security test
- CompilationCachingSecurityTest.CorruptedSecuritySensitiveCache
Randomly flip one bit of security-sensitive cache.
- CompilationCachingSecurityTest.WrongLengthSecuritySensitiveCache
Randomly append bytes to security-sensitive cache.
- CompilationCachingSecurityTest.WrongToken
Randomly flip one bit of cache token.
Bug: 119616526
Test: VtsHalNeuralnetworksV1_xTargetTest with 1.2 sample driver
Test: VtsHalNeuralnetworksV1_xTargetTest with a test driver that can
read and write cache entries
Change-Id: Iae9211cb28ce972b29572dfedd45d1ade4dfdaf5
Merged-In: Iae9211cb28ce972b29572dfedd45d1ade4dfdaf5
(cherry picked from commit 3405878e5e)
Machine Learning is a fast moving domain: new operations and data types
are introduced at a staggering speed. The Android API, on the other
hand, evolves on a yearly cycle. Many application developers, OEMs, and
SoC vendors would like to be able to define new operations at a faster
cycle.
In OC-MR1, NNAPI provided a simple mechanism to mitigate this problem:
two OEM data types and an OEM operation. The downside of this mechanism is that
it is simplistic: There’s no guarantee of consistency between vendors.
Examples of features that are required by the first parties' use cases
that we would like to be able to define outside of the normal Android
release cycle:
- New data types:
- Sparse tensor (used by some speech generation models).
- 16 bit int tensor (needed by OCR).
- 16 bit float tensor.
- New operations:
- 16 bit quantized LSTM (needed by OCR).
- Basic primitives like sqrt and floor.
- Logical operations.
- Complex neural network layers.
- Control flow.
- Enhancement to existing operations:
- Concatenate supporting different scales/zeroPoints for arguments.
- High-dimensional tensors.
- Ordering of dimensions.
We are going to provide support for two types of extensions:
- NNAPI Platform extension. This is functionality that will become part of
future releases of NNAPI. These extensions provide generic,
non-vendor-specific functionality. Only the Android team can define new
platform extensions. These extensions will be defined in the master
branch of AOSP. Each extension comes with:
- Documentation defining the extension,
- A header file for the new constants,
- A parameter validation library to be used by drivers, and
- Validation tests akin to the CTS and VTS tests.
- Vendor extension. A vendor-specific extension is an alternative to OEM
operation and data types. Its usage will be limited only to first party apps
preinstalled on the /vendor or /odm partition.
Each vendor will be identified by a specific value to
prevent collisions between multiple IPs found on the same SoC. This
effectively creates a vendor-specific namespace. These identifiers are
assigned by Google.
This change only defines the new interface. The implementation follows
in changes Ie4e2530e4c81fabe4eb59b7a6ba4a3b4bb483bd1,
I02aa12f4a8444012ddf3b20c2bfbba6a21ce9ce9, and
Icf59ed04e602aa7a730eb1eb45e5f6b1204fafb3.
Bug: 118605927
Test: mma
Change-Id: Ia9b99015eec7a48bbf969cbe503862271f09adca
Merged-In: Ia9b99015eec7a48bbf969cbe503862271f09adca
(cherry picked from commit 9212018558)
