@spec available() :: [atom()]

Which delegate kinds this build of the library can construct.

This answers "was it compiled in", not "is a device present" -- the two have different answers on the same binary. Whether a device is there is discovered by trying to create the delegate and getting {:error, reason} back.

Note that an interpreter, and any delegate attached to it, belongs to one process at a time.

@spec external(binary() | list()) :: nif_resource_ok() | nif_error()

Load a delegate from a shared library implementing TfLite's delegate plugin interface -- tflite_plugin_create_delegate and tflite_plugin_destroy_delegate.

That covers Edge TPU, a GPU delegate built elsewhere, and any vendor delegate, without this library having to know anything about them.

The path is resolved to an absolute one before loading, because the loader is asked for exactly the file named: a bare libfoo.so would otherwise be looked up along the system search path, which is not where anyone means.

Options are handed to the plugin as strings, since that is the whole of the plugin ABI -- atoms and integers are converted, and at most 256 pairs fit. What the keys mean is the plugin's business.

{:ok, delegate} =
  TFLiteElixir.Delegate.external("/opt/lib/libvendor_delegate.so",
    device: 0, precision: :fp16)

The library is never unloaded, which is what TfLite does too.

@spec external(binary() | list(), Keyword.t() | map()) ::
  nif_resource_ok() | nif_error()

Load a delegate from a shared library. See external/1 for what the options mean.

Raising version of external/1.

Raising version of external/2.

@spec xnnpack() :: nif_resource_ok() | nif_error()

Create an XNNPACK delegate.

XNNPACK is compiled into every target except armv6 and armv7l; available/0 is what says so, and this returns {:error, reason} where it is not.

Options

• :num_threads. Size of the delegate's thread pool. Zero or less means no thread pool at all, which is XNNPACK's own default and therefore this one. Note that this is not the same knob as TFLiteElixir.InterpreterBuilder.set_num_threads/2, which drives TfLite's CPU backend: a delegate created here carries its own pool.
• :flags. A list of atoms, added to XNNPACK's defaults rather than replacing them -- TfLite spells turning a default off as its own flag, such as :disable_subgraph_reshaping. One of :qs8, :qu8, :force_fp16, :dynamic_fully_connected, :variable_operators, :transient_indirection_buffer, :enable_latest_operators, :enable_subgraph_reshaping, :slow_consistent_arithmetic, :disable_subgraph_reshaping or :disable_dynamically_quantized_ops.
• :weight_cache_file_path. Where to keep XNNPACK's cache of packed weights, which is read if it exists and written if it does not.

@spec xnnpack(Keyword.t() | map()) :: nif_resource_ok() | nif_error()

Create an XNNPACK delegate. See xnnpack/0 for the options.

Raising version of xnnpack/0.

Raising version of xnnpack/1.