Age | Commit message (Collapse) | Author |
|
Squeeze, when used on a single-value tensor with 1s in
each dimension, will result in zero-sized tensor without dimensions.
This change fixes it, by making its output a tensor of size [1].
Bug: 155238914
Test: squeeze_b155238914.mod.py run on coral.
Change-Id: Ia5bf655c647084d9fa094ae8a2048917a8237b81
|
|
This CL fixes the ordering of sequences in the test. Specifically:
1) Causes the spin-loop in waitForExecutionToBegin to spin while the
execution has not yet started
2) Adds dummyExecution() to the driver implementations of FAIL_LAUNCH
3) Skips getDurationWhileRunning for all DriverKind::CPU test cases
Further, this CL adds dummyExecution() to
TestPreparedModelLatest::executeFenced to ensure the call to
executeFenced will not complete until after the test has called
getDuration. This brings executeFenced in line with the other execution
methods which call dummyExecution().
Finally, this CL has the following two minor changes:
1) code that was previously running from a detached thread in
executeFenced now instead runs synchronously
2) a for-loop that iterated by constant reference now instead iterates
by constant value
Fixes: 155215099
Test: mma
Test: NeuralNetworksTest_static --gtest_filter=Fenced/TimingTest*
Change-Id: I69a7448367ce8353791c9462d144122b6c913802
|
|
|
|
Bug: 140641425
Test: NNT_static
Change-Id: I3e037df067125b04efbd73f6d84da228d5e24f7a
|
|
Bug: 140641425
Bug: 151328024
Bug: 152446228
Bug: 152445711
Bug: 152446298
Fixes: 152446228
Test: NNT_static
Test: NNT_static_fuzzing
Change-Id: Ifd1f45a2b058fe9996dca817a067ce2ad6d47b46
|
|
This CL also re-enables TestUnspecifiedDimensions, which will trigger
the bug in validateRequest when sample driver is deployed.
Bug: 141294076
Fixes: 155228680
Test: NNT_static with sample driver
Change-Id: I268eb67bda8933be180bd58b604ee76b72577697
|
|
rvc-dev
|
|
Fixes: 154639297
Test: NNT_static
Test: NNT_static_fuzzing
Change-Id: I2b37dd7fe07f3239682b5bc7f7cef0ae76134ad0
|
|
- The tests require mTestDevice being valid to perform the tests, while
useCpuOnly will only allow CPU fallback path.
Fixes: 155229031
Test: mm
Test: NeuralNetworksTest_static
Change-Id: Ic866d14de39d0c8f52648ad319bac14bc8fdd666
|
|
This CL introduces the following kinds of tests:
* device creation and initialization errors
* "normal" functional positive and negative tests
* HIDL transport error tests
* mock service crash tests
Fixes: 139189546
Bug: 129572750
Test: mma
Test: atest NeuralNetworksTest_static
Change-Id: I3e1eb5433732695572044a74f1529ffc9c3a1982
|
|
Test: Build NeuralNetworksTest_static both normally and with
#define VERBOSE uncommented in TestPartitioningRandom.cpp
Bug: 136735929
Change-Id: I79e33239b63885b050d77b5a819b7be82a80528e
|
|
|
|
* changes:
Add STRIDED_SLICE with shrinkMask in RGG tests.
Increase the fp precision in dumped spec.
|
|
Fixes: 132322549
Test: NNT_static_fuzzing
Change-Id: I85577481125e131741992e9d896d12320af67e1a
|
|
|
|
In this test, the RGG produces a non-sensible graph with extreme large
output gain and highly clamped output range.
Fixes: 153690038
Test: NNT_static_fuzzing
Change-Id: I7063730e04ea8a18aecd20029eb6e726ac780c2a
|
|
These tests use a customized IDevice to test the device memory
allocation and fallback logic.
This CL also allows the runtime to dispatch device memory allocation
with dynamic shape to drivers.
Additionally, this CL fixes a bug that a failed device memory allocation
will return BAD_DATA -- it should return OP_FAILED instead.
Bug: 152209365
Test: NNT_static
Change-Id: I1facb2dad345958c3b9b1bab4a9564085c382c4a
|
|
|
|
|
|
Fix sporadic failures in CtsNNAPITestCases.TensorRankConstraint tests.
Improves description of expectation failures when comparing
OperandTypeWithExtraParams instances.
Test: CtsNNAPITestCases
Bug: 154093710
Change-Id: Ia2b8bc914892f2ff7908d247fa26743fa242f783
|
|
Fix: 148208229
Test: NNT_static
Change-Id: I853f1f294cedacb0cbb810042ef2a3466a072ffe
|
|
Bug: 149199424
Test: NNT_static
Test: VTS 1.3
Change-Id: I887be1b8530aad4568db9e75d3260b6b78884d23
|
|
|
|
|
|
|
|
Before this change, we only had tests with CONSTANT_COPY constant inputs
(the default). This change adds CONSTANT_REFERENCE counterparts of those
tests.
Bug: 137836124
Test: NNT_static
Change-Id: If6a0b897636f49b19ac7ccc1bc7ded3eee764744
|
|
Note that this change potentially changes EVERY pseudorandom test case
in this particular test set.
Test: NeuralNetworksTest_static --gtest_filter=Seed/RandomPartitioningTest.*
Also verified that some randomly generated graphs include the
newly-added V1_3 operations.
Bug: 141704808
Change-Id: Ib762c5fe8f7c42825025a6e57b61631040e0911a
|
|
* changes:
Prevent RGG from generating NaN values.
Add broadcast with different ranks in RGG.
|
|
|
|
|
|
|
|
Before this CL, the golden results in RGG tests are computed from
nnapi-reference with the actual data type. This CL changes the logic to
compute the golden results from an equivalent float32 model to avoid
bias if possible.
Currently only the golden results of single-operation models are
computed from fp32 models. This is because there is no multi-operation
quant model in RGG tests yet.
Fixes: 150805665
Test: NNT_static
Change-Id: I4e268d3aadeebdcdae1d7010a4d0564f45b11b82
|
|
This CL introduces "ValueProperty" to RGG, which defines the required
properties of input values as well as the guaranteed properties of the
output values. During the graph generation process, an operation
input may only be wired to an operation output with compatible
properties.
E.g., the input of SQRT will not come from CONV_2D because
CONV_2D cannot guarantee non-negative output values; while it could come
from RELU because RELU guarantees that the output values are always
non-negative.
Additionally, fixes the issue that REDUCE_PROD may generate NaN values
because 0 * inf evaluates to NaN in C++.
Bug: 145188288
Bug: 150805665
Test: NNT_static_fuzzing
Change-Id: I5464233839a91afbe2022b8fe603e4296825a7a4
|
|
Fixes: 132323699
Test: NNT_static_fuzzing
Change-Id: Ica52243ad5fda8d290289febea8ab4f1206dab36
|
|
Fix: 152963626
Test: NNT_static
Change-Id: Ie4d003d071533479719f7b73f792d00eb62de145
|
|
Bug: 149199424
Test: NNT_static
Change-Id: If6ae373876fa309580198526d44018576a4ebdd1
|
|
The following construct is used to specify subgraph conversions:
converter = DataTypeConverter().Identify({
op1: (target_type, target_scale, target_zero_point),
...
subgraphOp: DataTypeConverter().Identify({
...
}),
})
Implicit conversion is supported as well.
Bug: 149199424
Test: NNT_static
Change-Id: I3f28d36159f19a50e50bc6f438bb98331f77a5f9
|
|
execution" into rvc-dev
|
|
Test that getOutputOperand*() fails with ANEURALNETWORKS_BAD_STATE if called
on a running Execution.
Bring burstCompute() and startComputeWithDependencies() testing to
parity with compute() and startCompute() testing: Ensure that driver
errors are handled properly, and that getOutputOperand*() is handled
properly.
Also add NeuralNetworksWrapper::startComputeWithDependencies().
Bug: 148804027
Test: NeuralNetworksTest_static --gtest_filter=*ExecutionTest*Wait*
Change-Id: I7cca4d43d7f6369845d7fedac4e37ca2cc8b7afe
|
|
* changes:
Add rank check to LOCAL_RESPONSE_NORMALIZATION validation
Add rank check to FLOOR validation
Add rank checks to validation functions
|
|
into rvc-dev
|
|
Before this change, ExecutionPlan::Controller::mLastStepIndex is only
updated at the beginning of the ExecutionPlan::next() call. When
interpreted control flow is involved, this call might result in several
logical steps being processed until an ExecutionStep is reached.
When an interpreted CF loop terminates successfully but the next step is
an ExecutionStep that fails, ExecutionPlan::fallback() restarts from
mLastStepIndex, which happens to be the index of the WhileStep. This
results in the loop being executed again. This happens forever.
This change fixes the bug by updating mLastStepIndex when
nextCompound(const ExecutionStep* ...) is called.
This change renames mLastStepIndex to mFallbackNextStepIndex to reflect
its usage.
Fix: 152623151
Test: NNT_static
Change-Id: Iff3b68f541adb8facf078e968abac2a6ae8e74c1
|
|
Also add rank check to HARD_SWISH validation.
This commit removes nnAssert from the utility function that converts
NNAPI Shape to TF Lite's RuntimeShape. The assert would cause a crash if
the function was passed a shape with rank > 4 even though RuntimeShape
supports such shapes.
Bug: 152945527
Bug: 139957496
Test: NNTest_static
Change-Id: I69147915669bee525c8883e48fd53ff0f899a501
|
|
Also move LOCAL_RESPONSE_NORMALIZATION operation to OperationResolver
Bug: 124041202
Bug: 139957496
Test: NNTest_static
Change-Id: I4226bdb101d6ee0a744448e61357344db019dd70
|
|
Also move FLOOR operation to OperationResolver
Bug: 124041202
Bug: 139957496
Test: NNTest_static
Change-Id: I59070aff36ddfa57085258974168a314176ed27b
|
|
The change adds rank checks to validation of operations that only
support tensors of rank 4 or less. This requirement comes from legacy TF
Lite code and is likely to be relaxed in the future to be on par with TF
Lite.
Adding the checks to validation is benefitial for the TF Lite delegate
since in case of a validation error NNAPI node will be fully rejected by
the delegation but execution error will cause TF Lite to run NNAPI node
during every invocation only to receive an error and do the calculation
using CPU implementation.
Bug: 139957496
Test: NNTest_static
Change-Id: I5cc4c48e775826a237d5ac54c3d2078254bd17a2
|
|
Fix: 129571542
Test: NNTest_static
Change-Id: Ie94a0a36b1df10afff6a5269f9375f99bc6ffba7
|
|
The tests have been added in change I15ad10dd.
Bug: 152405977
Test: N/A
Change-Id: I477e50025e88cf1d04952746a47a4b0f0923d2d7
|
|
|
|
|