Prior to this CL, the NNAPI represented SharedHandle as a collection of
file descriptors and a vector of ints to act as metadata. This struct
mirrored Android's native_handle_t. However, the NNAPI only uses
SharedHandle to hold a single file descriptor to hold a single cache
file.
A sibling CL simplifies the SharedHandle by changing it to be a single
ref-counted file descriptor. This CL makes changes to the HAL utility
code in response to the SharedHandle change.
Bug: N/A
Test: mma
Test: NeuralNetworksTest_static
Change-Id: Ib390ad76bcdadffd85f8c2c502189aaa7da483e4
This CL modifies the shared memory allocation for pointer arguments to
use proper alignment and padding. We use default alignment (64) and
min padding (1) for HIDL drivers, and default alignment (64) and default
padding (64) for sAIDL drivers.
Bug: 184164929
Test: NNT_static
Change-Id: I22591640fa047d5f75d437edac1a7645d3b05526
Merged-In: I22591640fa047d5f75d437edac1a7645d3b05526
(cherry picked from commit 8bfa243972)
This CL modifies the canonical interface for reusable burst executions:
- Add new method IBurst::createExecution
The reusable burst execution will not fallback to another execution path
if sending request packet fails. The behavior of single-time burst
execution remains unchanged.
Additionally, this CL enables pointer -> shared memory conversion in
1.2/1.3 burst implementation.
Bug: 184073769
Test: NNT_static
Test: neuralnetworks_utils_hal_1_0_test
Test: neuralnetworks_utils_hal_1_1_test
Test: neuralnetworks_utils_hal_1_2_test
Test: neuralnetworks_utils_hal_1_3_test
Test: neuralnetworks_utils_hal_common_test
Change-Id: Iaac81668d247c2cb76d70e6abbd10f00b397b19f
Merged-In: Iaac81668d247c2cb76d70e6abbd10f00b397b19f
(cherry picked from commit ead6d37ae9)
This CL modifies the canonical interface for reusable executions:
- Add new interface: IExecution with compute and computeFenced methods
- Add new method IPreparedModel::createExecution
In NNAPI runtime, the new interface IExecution is used to
memoize request-specific execution resources (e.g. converted HAL
request). The expected usage is that, IPreparedModel::createExecution
will be invoked in the first computation of a reusable NDK ANNExecution
object, and IExecution::compute* will be invoked repeatedly.
The IPreparedModel::execute* methods are preserved to avoid redundant
object creation and memoization overhead for a single-time
(non-reusable) execution.
For a vendor implementing the canonical interfaces, only the
IPreparedModel::execute* methods will be called because there is
currently no reusable execution at HAL interface. A DefaultExecution
implementation is provided to reduce the work needed on the vendor side.
Bug: 184073769
Test: NNT_static
Test: neuralnetworks_utils_hal_1_0_test
Test: neuralnetworks_utils_hal_1_1_test
Test: neuralnetworks_utils_hal_1_2_test
Test: neuralnetworks_utils_hal_1_3_test
Test: neuralnetworks_utils_hal_common_test
Test: neuralnetworks_utils_hal_aidl_test
Change-Id: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
Merged-In: I91790bb5ccf5ae648687fe603f88ffda2c9fd2b2
(cherry picked from commit 727a7b2104)
This CL alters the AIDL utility "getDevices" to return whether or not a
service is updatable. This CL additional removes IDevice::isUpdatable
because it is now redundant with the changes to "getDevices".
Bug: 170696350
Test: mma
Test: NeuralNetworksTest_static
Change-Id: I86228763937c0dd2ee5b1d9d547c367cd7504ed4
Use __ANDROID_API_FUTURE__ until the S version is finalized.
Bug: 183657864
Test: Build & flashed CF, run TFLite benchmark util
Change-Id: I1352aa33866487f955136c6fb8e4d0d279225f13
This CL fixes the compiler errors that arose of changing the Memory
representation of the NN AIDL HAL, and updates the conversion and
utility code to work with the new Memory type.
This change also makes libaidlcommonsupport available to apex modules at
min sdk level 29.
Bug: 183118727
Test: mma
Test: VtsHalNeuralnetworksTargetTest
Change-Id: Ief565473b4d82e0bb43785fc3b8275b16bd26cf6
Merged-In: Ief565473b4d82e0bb43785fc3b8275b16bd26cf6
(cherry picked from commit b0fcb3927d)
Adds deadline and loopTimeoutDuration to nn::IBurst::execute.
This CL additionally adds a test case for
IPreparedModel::configureExecutionBurst and adds the suffix "_H" to
the header guards that were missing "_H".
Bug: 180492058
Bug: 177267324
Test: mma
Test: NeuralNetworksTest_static
Change-Id: Ic9ba2fb9b4dee4e40d99ae91bb9555a58d0508d2
Merged-In: Ic9ba2fb9b4dee4e40d99ae91bb9555a58d0508d2
(cherry picked from commit bbbdb0d5f1)
Added SPDX-license-identifier-Apache-2.0 to:
neuralnetworks/utils/adapter/Android.bp
security/keymint/aidl/vts/performance/Android.bp
Bug: 68860345
Bug: 151177513
Bug: 151953481
Test: m all
Exempt-From-Owner-Approval: janitorial work
Change-Id: I139497d894d893a405d9b16df5e23b69dc82674f
This CL creates Device, PreparedModel, and Buffer classes which adapt a
canonical interface to a NN 1.3 HIDL interface.
Bug: 160668438
Test: mma
Change-Id: Ifee27858efd280307a299b3fe2e1a4bd9240e9de
Merged-In: Ifee27858efd280307a299b3fe2e1a4bd9240e9de
(cherry picked from commit 323dd0c2dc)
Previously, countNumberOfConsumers would trigger a CHECK if the input
was invalid. This CL makes countNumberOfConsumers gracefully fail on
errors, instead returning the error through the GeneralResult.
Bug: N/A
Test: mma
Change-Id: Iee54f87768e52fdf701c22d94083c053b881733d
Merged-In: Iee54f87768e52fdf701c22d94083c053b881733d
(cherry picked from commit c4d98007fd)
Prior to this CL, the canonical memory type only held a SharedHandle,
which mirrors the behavior of native_handle_t/hidl_handle. This means
memory types including AHardwareBuffer were stored as this SharedHandle
type. With this CL, the canonical memory type is stored directly as
AHardwareBuffer to avoid using non-NDK AHardwareBuffer calls in the NN
runtime.
Bug: 179906132
Test: mma
Test: NeuralNetworksTest_static
Test: presubmit
Change-Id: I394071c193d15ac0c90ac47e5a2a9a79c635db6c
Merged-In: I394071c193d15ac0c90ac47e5a2a9a79c635db6c
(cherry picked from commit bbe43d950e)
This CL adds a simple implementation of IBurst that dispatches calls to
an IPreparedModel object and changes
IPreparedModel::configureExecutionBurst to return this new object
(instead of returning an error).
This CL additionally defines an InvalidBurst class that returns errors
whenever it is used and a ResilientBurst class to recover an IBurst
object when it has died.
Bug: 177267324
Test: mma
Change-Id: I4c7e7ff4e6559aeb5e62c4fa02f2e751fef9d87d
Merged-In: I4c7e7ff4e6559aeb5e62c4fa02f2e751fef9d87d
(cherry picked from commit 44f324fb0d)
A sibling change to this change introduces a canonical IBurst object and
a new method IPreparedModel::configureExecutionBurst. This CL performs
minimal changes to the NN hal utility code to make the code still
compile.
Bug: 177267324
Test: mma
Change-Id: I076067289dde7def5622a6cb25233619f80efa41
Merged-In: I076067289dde7def5622a6cb25233619f80efa41
(cherry picked from commit b6a7ed5d5f)
This CL introduces unit tests to validate the V1_X::utils::Device,
*PreparedModel, and *Buffer adapter classes. It does so by mocking the
underlying HIDL interface in order to simulate a driver returning bad
data, HIDL transport failures, and service crashes.
Note that the purpose of these new tests is to validate the adapter
classes themselves, not the HIDL interfaces they use. For example,
because nn::IPreparedModel does not currently define a method for
configuring a burst execution, V1_[23]::utils::PreparedModel similarly
does not use hardware::neuralnetworks::V1_[23]::IPreparedModel's
configureExecutionBurst method.
This CL also introduces unit tests to validate the utils::Resilient*
adapter classes, and mocks DEAD_OBJECT failures to ensure that the
underyling object can be recovered appropriately.
Bug: 163801800
Test: mma
Test: atest neuralnetworks_utils_hal_common_test
Test: atest neuralnetworks_utils_hal_1_[0-3]_test
Change-Id: I2c79865bf666d3f4bf53061ff5090746403583e9
Merged-In: I2c79865bf666d3f4bf53061ff5090746403583e9
(cherry picked from commit afc4d7cfe7)
Prior to this CL, ResilientPreparedModel and ResilientBuffer were
passthrough interfaces that just forwarded calls to the underlying
interface object. This CL implements the full recovery mechanism for
these two classes. However, because we do not want to enable this
functionality in the NN runtime yet, ResilientDevice hides the paths
that create ResilientPreparedModel and ResilientBuffer behind an #if
until we want to enable those paths.
Bug: N/A
Test: mma
Change-Id: Idfe8093c63c7ba2f16c995eec872d150696e7a08
Merged-In: Idfe8093c63c7ba2f16c995eec872d150696e7a08
(cherry picked from commit 667dc2dcac)
This change removes the 'blocking' parameter for the
ResilientPreparedModel::Factory and ResilientBuffer::Factory. The
'blocking' parameter is only useful for ResilientDevice::Factory, which
behind the scenes chooses between the HIDL calls IDevice::getService and
IDevice::tryGetService. The equivalent calls for IPreparedModel and
IBuffer are not used, as both are created from the IDevice object.
This change also modifies the ResilientDevice's device recovery
behavior. Prior to this change, ResilientDevice's recovery mechanism had
the following behavior:
* attempt to call a function
* if the function did not return a DEAD_OBJECT error, return
* if the function returned a DEAD_OBJECT error, attempt to recover the
device
* whether or not the recovery succeeded, call the function again
This CL changes the behavior so that if device recovery fails,
ResilientDevice will not call the function the second time.
Bug: N/A
Test: mma
Change-Id: Icf37d05c884c740178324fcd046ea56914ef7d44
Merged-In: Icf37d05c884c740178324fcd046ea56914ef7d44
(cherry picked from commit 11761e37a8)
This CL introduces a new templated class CallbackValue to handle HIDL
"return value" callbacks in a terser and more readable way.
This CL also introduces a new macro HANDLE_HAL_STATUS to return from the
current function when an error is present with the ability to append a
more descriptive error message.
Finally, this CL changes the behavior of synchronous executions. Prior
to this CL, IPreparedModel fell back to an asynchronous execution if the
synchronous execution was allowed and failed. This change instead
returns a failure if synchronous execution is allowed and fails.
Bug: 173084343
Test: mma
Change-Id: I62714a932e71dfc77401bbcb9eaaaf3d94fb9707
Merged-In: I62714a932e71dfc77401bbcb9eaaaf3d94fb9707
(cherry picked from commit 98ed9baf5d)
This CL copies information from
packages/modules/NeuralNetworks/runtime/VersionedInterfaces.cpp and
modifies the description to be more appropriate for the NN HAL utility
code.
Specific sections added to the README:
* "HIDL Interface Lifetimes across Processes"
* "Protecting Asynchronous Calls across HIDL"
Bug: 170289677
Test: mma
Change-Id: Id381895535d708b627f4746687b4d12e16560639
Merged-In: Id381895535d708b627f4746687b4d12e16560639
(cherry picked from commit 7a655bb3d4)
A sibling CL to this CL changes the definition of nn::TimePoint to the
same type as std::chrono::steady_clock::time_point but has changed the
underlying duration representation to use uint64_t. That sibling CL also
renames nn::OptionalTimeoutDuration to nn::OptionalDuration, and changes
the definition to the same type as std::nanoseconds except the
underlying duration representation now uses uint64_t.
This CL makes changes to the NN HAL utility code in response to the
changes in the sibling CL.
Bug: 174297663
Test: mma
Test: NeuralNetworksTest_static
Change-Id: If44d9aefadb2c78b632ff289b5ff5a49f766525c
Merged-In: If44d9aefadb2c78b632ff289b5ff5a49f766525c
(cherry picked from commit ca11420785)
Prior to this change, whenever the NN utility code encountered a HIDL
transport error, the error message would display the file and line
number of the "handleTransportError" function itself. This change
introduces a new macro "HANDLE_TRANSPORT_FAILURE" that handles the
transport error in a similar way but now the error message displays
the file and line number of where the macro is called.
Bug: N/A
Test: mma
Change-Id: I35b34f8f5be52b7fcff0fbb58a37ab2b8c7dd8bb
Merged-In: I35b34f8f5be52b7fcff0fbb58a37ab2b8c7dd8bb
(cherry picked from commit 61f508e018)
This change renames all `convert` functions to `unvalidatedConvert`.
This change also introduces new `convert` functions that act only on the
types that appear in the NN HIDL methods directly. These new `convert`
functions perform validation. Specifically, if either the source or
destination value is invalid, then the conversion fails.
Bug: 160667419
Test: mma
Test: NeuralNetworksTest_static
Change-Id: I492956ff60ad1466c67893993d28cdd6f3860708
Merged-In: I492956ff60ad1466c67893993d28cdd6f3860708
(cherry picked from commit 32acc06144)
Currently, if an IDevice object is a DEAD_OBJECT, the runtime attempts
to re-retrieve the handle to the rebooted IDevice service. If an update
occurs after the IDevice was originally created, the rebooted IDevice
object may have different metadata and behavior. This is problematic
because the original metadata is cached in the runtime. Further, an
application might have made decisions based on that metadata and
behavior. (Note that a driver service that is functionally the same but
has a different underlying implementation such as having more optimized
code will have different `getVersionString` metadata.) Instead, this CL
invalidates the IDevice object on cache mismatch, and always returns an
error if it is used.
Bug: 173081926
Test: mma
Change-Id: I805987361c627c32d45e1b7c7aed230376fc66ad
Merged-In: I805987361c627c32d45e1b7c7aed230376fc66ad
(cherry picked from commit 5a74c0fb0f)
This CL implements the canonical IDevice, IPreparedModel, and IBuffer
interfaces for the 1.0, 1.1, 1.2, and 1.3 NN HIDL HAL interfaces.
Further, it introduces "Resilient" adapter interfaces to automatically
retrieve a handle to a recovered interface object after it has died and
rebooted.
This CL also updates the conversion code from returning nn::Result to
nn::GeneralResult, which includes a ErrorStatus code in the case of an
error.
Finally, this CL introduces a new static library
neuralnetworks_utils_hal_service which consists of a single function
::android::nn::hal::getDevices which can be used by the NNAPI runtime to
retrieve the HIDL services without knowing the underlying HIDL types.
Bug: 160668438
Test: mma
Test: NeuralNetworksTest_static
Change-Id: Iec6ae739df196b4034ffb35ea76781fd541ffec3
Merged-In: Iec6ae739df196b4034ffb35ea76781fd541ffec3
(cherry picked from commit 3670c385c4)
This CL creates the following primary sets of functions:
* V1_X::utils::convert(<canonical_type>) -- Converts a canonical type
to the corresponding HAL version type.
* nn::convert(<V1_X_HAL_type>) -- Converts a HAL version type to the
corresponding canonical type.
* neuralnetworks::utils::hasNoPointerData -- Indicates if the object
contains no pointer-based data that could be relocated to shared
memory.
* neuralnetworks::utils::flushDataFromPointerToShared -- Relocate
pointer-based data to shared memory.
* neuralnetworks::utils::unflushDataFromSharedToPointer -- Undoes
`flushDataFromPointerToShared` on a Request object. More
specifically, `unflushDataFromSharedToPointer` copies the output
shared memory data from the transformed Request object back to the
output pointer-based memory in the original Request object.
It also introduces some other minor utility code, including
makeQuantized8PerformanceConsistentWithP, countNumberOfConsumers,
validate, valid, and validatedConvertToCanonical.
Bug: 160667419
Test: mma
Change-Id: I0732e658c1f4ed40cd122f1ca8581fb40b056757
Merged-In: I0732e658c1f4ed40cd122f1ca8581fb40b056757
(cherry picked from commit a685c3dbf4)
