|
24.02
|
|
24.02
|
Arm NN allows adding new backends through the 'Pluggable Backend' mechanism.
Backends reside under src/backends, in separate subfolders. For Linux builds they must have a backend.cmake file, which is read automatically by src/backends/backends.cmake. The backend.cmake file under the backend-specific folder is then included by the main CMakeLists.txt file at the root of the Arm NN source tree.
The backend.cmake has three main purposes:
armnnLibraries list.armnnUnitTestLibraries list.Example backend.cmake file taken from reference/backend.cmake:
As described in the previous section, adding a new backend will require creating a CMakeLists.txt in the backend folder. This follows the standard cmake conventions, and is required to build a static cmake OBJECT library to be linked into the Arm NN shared library. As with any cmake build, the code can be structured into subfolders and modules as the developer sees fit.
Example can be found under reference/CMakeLists.txt.
Arm NN on Android uses the native Android build system. New backends are integrated by creating a
For multiple backends that need common code, there is support for including them in the build similarly to the backend code. This requires adding three files under a subfolder at the same level as the backends folders. These are:
They work the same way as the backend files. The only difference between them is that common code is built first, so the backend code can depend on them.
[aclCommon](aclCommon) is an example for this concept and you can find the corresponding files:
Backends are identified by a string that must be unique across backends. This string is wrapped in the BackendId object for backward compatibility with previous Arm NN versions.
All backends need to implement the IBackendInternal interface. The interface functions to be implemented are:
Note that GetOptimizations() and SubgraphViewUniquePtr OptimizeSubgraphView(const SubgraphView& subgraph, bool& optimizationAttempted) have been deprecated. The method OptimizationViews OptimizeSubgraph(const SubgraphView& subgraph) should be used instead to apply specific optimizations to a given sub-graph.
The Arm NN framework then creates instances of the IBackendInternal interface with the help of the BackendRegistry singleton.
Important: the IBackendInternal object is not guaranteed to have a longer lifetime than the objects it creates. It is only intended to be a single entry point for the factory functions it has. The best use of this is to be a lightweight, stateless object and make no assumptions between its lifetime and the lifetime of the objects it creates.
For each backend one needs to register a factory function that can be retrieved using a BackendId. The Arm NN framework creates the backend interfaces dynamically when it sees fit and it keeps these objects for a short period of time. Examples:
GetLayerSupport() function and creates an ILayerSupport object to help deciding this.OptimizeSubgraphView() function on each of them and, if possible, substitutes the corresponding sub-graph in the original graph with its optimized version.IBackendContext object and keeps this context alive for the Runtime's lifetime. It notifies this context object before and after a network is loaded or unloaded.As mentioned above, all backends need to be registered through the BackendRegistry so Arm NN knows about them. Registration requires a unique backend ID string and a lambda function that returns a unique pointer to the IBackendInternal interface.
For registering a backend only this lambda function needs to exist, not the actual backend. This allows dynamically creating the backend objects when they are needed.
The BackendRegistry has a few convenience functions, like we can query the registered backends and are able to tell if a given backend is registered or not.
Dynamic backends are registered during the runtime creation.
Arm NN uses the ILayerSupport interface to decide if a layer with a set of parameters (i.e. input and output tensors, descriptor, weights, filter, kernel if any) are supported on a given backend. The backends need a way to communicate this information by implementing the GetLayerSupport() function on the IBackendInternal interface.
Examples of this can be found in the RefLayerSupport header and the RefLayerSupport implementation.
The IWorkloadFactory interface is used for creating the backend specific workloads. The factory function that creates the IWorkloadFactory object in the IBackendInterface takes an IMemoryManager object.
To create a workload object the IWorkloadFactory takes a WorkloadInfo object that holds the input and output tensor information and a workload specific queue descriptor.
Backends may choose to implement custom memory management. Arm NN supports this concept through the following mechanism:
IBackendInternal interface has a CreateMemoryManager() function, which is called before creating the workload factoryCreateWorkloadFactory(...) function so the workload factory can use it for creating the backend-specific workloadsAcquire() on the memory manager before it starts executing the network and it calls Release() in its destructorThe backends may choose to implement backend-specific optimizations. This is supported through the method OptimizationViews OptimizeSubgraph(const SubgraphView& subgraph) of the backend interface that allows the backends to apply their specific optimizations to a given sub-graph.
The OptimizeSubgraph(...) method returns an OptimizationViews object containing three lists:
The previous way backends had to provide a list optimizations to the Optimizer (through the GetOptimizations() method) is still in place for backward compatibility, but it's now considered deprecated and will be remove in a future release.
Backends may need to be notified whenever a network is loaded or unloaded. To support that, one can implement the optional IBackendContext interface. The framework calls the CreateBackendContext(...) method for each backend in the Runtime. If the backend returns a valid unique pointer to a backend context, then the runtime will hold this for its entire lifetime. It then calls the following interface functions for each stored context:
BeforeLoadNetwork(NetworkId networkId)AfterLoadNetwork(NetworkId networkId)BeforeUnloadNetwork(NetworkId networkId)AfterUnloadNetwork(NetworkId networkId)Backends can also have an associated CustomMemoryAllocator registered with them that ArmNN will use to allocate intra/inter-layer memory. This particular feature is required if you want a backend to use ProtectedContentAllocation. To support this on your own backend you must implement the UseCustomMemoryAllocator interface.
This interface returns a boolean value which indicates if the provided allocator is supported by the backend. This interface is also used by the lambda function returned by the Backend Registry to configure the CustomMemoryAllocator. Within the backend itself there should be a wrapper class to convert the generic CustomMemoryAllocator provided by the interface into something that is more suitable for your own backend.
Examples of how this can be done are in the ClBackend header and the ClRegistryInitializer header
This is a list of BackendCapabilities currently supported by the backend. It consists of a constant list of Name/Value pairs, each containing a string name, and a boolean value to indicate support. For example to indicate support for ProtectedContentAllocation you would return {"ProtectedContentAllocation", true}
An example can be found at the top of ClBackend header
Backends can also be loaded by Arm NN dynamically at runtime. To be properly loaded and used, the backend instances must comply to the standard interface for dynamic backends and to the versioning rules that enforce ABI compatibility.
The dynamic backend shared object must expose the following interface for Arm NN to handle it correctly:
Interface details:
extern "C" is needed to use avoid C++ name mangling, necessary to allow Arm NN to dynamically load the symbols.GetBackendId(): must return the unique id of the dynamic backends. If at the time of the loading the id already exists in the internal Arm NN's backend registry, the backend will be skipped and not loaded in Arm NNGetVersion(): must return the version of the dynamic backend. The version must indicate the version of the Backend API the dynamic backend has been built with. The current Backend API version can be found by inspecting the IBackendInternal interface. At the time of loading, the version of the backend will be checked against the version of the Backend API Arm NN is built with. If the backend version is not compatible with the current Backend API, the backend will not be loaded as it will be assumed that it is not ABI compatible with the current Arm NN build.BackendFactory(): must return a valid instance of the backend. The backend instance is an object that must inherit from the version of the IBackendInternal interface declared by GetVersion(). It is the backend developer's responsibility to ensure that the backend implementation correctly reflects the version declared by GetVersion(), and that the object returned by the BackendFactory() function is a valid and well-formed instance of the IBackendInternal interface.Dynamic backend versioning policy:
Updates to Arm NN's Backend API follow these rules: changes to the Backend API (the IBackendInternal interface) that break ABI compatibility with the previous API version will be indicated by a change of the API's major version, while changes that guarantee ABI compatibility with the previous API version will be indicated by a change in API's the minor version.
For example:
During the creation of the Runtime, Arm NN will scan a given set of paths searching for suitable dynamic backend objects to load. A list of (absolute) paths can be specified at compile-time by setting a define named DYNAMIC_BACKEND_PATHS in the form of a colon-separated list of strings.
The paths will be processed in the same order as they are indicated in the macro.
It is also possible to override those paths at runtime when creating the Runtime object by setting the value of the m_DynamicBackendsPath member in the CreationOptions class. Only one path is allowed for the override via the CreationOptions class. By setting the value of the m_DynamicBackendsPath to a path in the filesystem, Arm NN will entirely ignore the list of paths passed via the
Only alphanumeric characters are allowed for both the <vendor> and the <name> fields, namely lowercase and/or uppercase characters, and/or numerical digits (see the table below for examples). Only dots and numbers are allowed for the optional <version> field.
Symlinks to other files are allowed to support the standard linux shared object versioning:
Files are identified by their full canonical path, so it is allowed to have files with the same name in different directories. However, if those are actually the same dynamic backend, only the first in order of parsing will be loaded.
Examples:
| Filename | Description |
|---|---|
| Arm_GpuAcc_backend.so | valid: basic backend name |
| Arm_GpuAcc_backend.so.1 | valid: single field version number |
| Arm_GpuAcc_backend.so.1.2 | valid: multiple field version number |
| Arm_GpuAcc_backend.so.1.2.3 | valid: multiple field version number |
| Arm_GpuAcc_backend.so.10.1.27 | valid: Multiple digit version |
| Arm_GpuAcc_backend.so.10.1.33. | not valid: dot not followed by version number |
| Arm_GpuAcc_backend.so.3.4..5 | not valid: dot not followed by version number |
| Arm_GpuAcc_backend.so.1,1.1 | not valid: comma instead of dot in the version |
| Arm123_GpuAcc_backend.so | valid: digits in vendor name are allowed |
| Arm_GpuAcc456_backend.so | valid: digits in backend id are allowed |
| ArmCo_GpuAcc_backend.so | not valid: invalid character in vendor name |
| Arm_Gpu.Acc_backend.so | not valid: invalid character in backend id |
| GpuAcc_backend.so | not valid: missing vendor name |
| _GpuAcc_backend.so | not valid: missing vendor name |
| Arm__backend.so | not valid: missing backend id |
| Arm_GpuAcc.so | not valid: missing "backend" at the end |
| __backend.so | not valid: missing vendor name and backend id |
| __.so | not valid: missing all fields |
| Arm_GpuAcc_backend | not valid: missing at least ".so" at the end |
| Arm_GpuAcc_backend_v1.2.so | not valid: extra version info at the end |
| Arm_CpuAcc_backend.so | valid: basic backend name |
| Arm_CpuAcc_backend.so.1 -> Arm_CpuAcc_backend.so | valid: symlink to valid backend file |
| Arm_CpuAcc_backend.so.1.2 -> Arm_CpuAcc_backend.so.1 | valid: symlink to valid symlink |
| Arm_CpuAcc_backend.so.1.2.3 -> Arm_CpuAcc_backend.so.1.2 | valid: symlink to valid symlink |
| Arm_no_backend.so -> nothing | not valid: symlink resolves to non-existent file |
| pathA/Arm_GpuAcc_backend.so | valid: basic backend name |
| pathB/Arm_GpuAcc_backend.so | valid: but duplicated from pathA/ |
Arm NN will try to load the dynamic backends in the same order as they are parsed from the filesystem.
The source code includes an example that is used to generate some mock dynamic backends for testing purposes. The source files are:
TestDynamicBackend.hpp TestDynamicBackend.cpp
This example is useful for going through all the use cases that constitute an invalid dynamic backend object, such as an invalid/malformed implementation of the shared object interface, or an invalid value returned by any of the interface methods that would prevent Arm NN from making use of the dynamic backend.
A dynamic implementation of the reference backend is also provided. The source files are:
RefDynamicBackend.hpp RefDynamicBackend.cpp
The implementation itself is quite simple and straightforward. Since an implementation of this particular backend was already available, the dynamic version is just a wrapper around the original code that simply returns the backend id, version and an instance of the backend itself via the factory function. For the sake of the example, the source code of the reference backend is used to build the dynamic version (as you would for any new dynamic backend), while all the other symbols needed are provided by linking the dynamic backend against Arm NN.
The makefile used for building the reference dynamic backend is also provided: CMakeLists.txt
A unit test that loads the reference backend dynamically and that exercises it is also included in the file DynamicBackendTests.cpp, by the test case CreateReferenceDynamicBackend. In the test, a path on the filesystem is scanned for valid dynamic backend files (using the override option in CreationOptions) where only the reference dynamic backend is. In this example the file is named Arm_CpuRef_backend.so, which is compliant with the expected naming scheme for dynamic backends. A DynamicBackend is created in the runtime to represent the newly loaded backend, then the backend is registered in the Backend Registry with the id "CpuRef" (returned by GetBackendId()). The unit test makes sure that the backend is actually registered in Arm NN, before trying to create an instance of the backend by calling the factory function provided through the shared object interface (BackendFactory()). The backend instance is used to verify that everything is in order, testing basic 2D convolution support by making use of the Layer Support API and the Workload Factory. At the end of test, the runtime object goes out of scope and the dynamic backend instance is automatically destroyed, and the handle to the shared object is closed.
1.8.17