Crate executorch_sys

Expand description

Unsafe bindings for ExecuTorch - On-device AI across mobile, embedded and edge for PyTorch.

Provides a low level Rust bindings for the ExecuTorch library. For the common use case, it is recommended to use the high-level API provided by the executorch crate, where a more detailed documentation can be found.

To build the library, you need to build the C++ library yourself first. Currently the supported Cpp executorch version is 1.0.1 (or 1.0.0). The C++ library allow for great flexibility with many flags, customizing which modules, kernels, and extensions are built. Multiple static libraries are built, and the Rust library links to them. In the following example we build the C++ library with the necessary flags to run example hello_world:

# Clone the C++ library
cd ${EXECUTORCH_CPP_DIR}
git clone --depth 1 --branch v1.0.1 https://github.com/pytorch/executorch.git .
git submodule sync --recursive
git submodule update --init --recursive

# Install requirements
./install_requirements.sh

# Build C++ library
mkdir cmake-out && cd cmake-out
cmake \
    -DDEXECUTORCH_SELECT_OPS_LIST=aten::add.out \
    -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
    -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=OFF \
    -DEXECUTORCH_BUILD_PORTABLE_OPS=ON \
    -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON \
    -DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON \
    -DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON \
    -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON \
    -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON \
    -DEXECUTORCH_ENABLE_PROGRAM_VERIFICATION=ON \
    -DEXECUTORCH_ENABLE_LOGGING=ON \
    ..
make -j

# Run example
# We set EXECUTORCH_RS_EXECUTORCH_LIB_DIR to the path of the C++ build output
cd ${EXECUTORCH_RS_DIR}/examples/hello_world
python export_model.py
EXECUTORCH_RS_EXECUTORCH_LIB_DIR=${EXECUTORCH_CPP_DIR}/cmake-out cargo run

The executorch crate will always look for the following static libraries:

libexecutorch.a
libexecutorch_core.a

Additional libs are required if feature flags are enabled. For example the libextension_data_loader.a is required if the data-loader feature is enabled, and libextension_tensor.a is required if the tensor-ptr feature is enabled. See the feature flags section for more info.

The static libraries of the kernels implementations are required only if your model uses them, and they should be linked manually by the binary that uses the executorch crate. For example, the hello_world example uses a model with a single addition operation, so it compile the C++ library with DEXECUTORCH_SELECT_OPS_LIST=aten::add.out and contain the following lines in its build.rs:

println!("cargo::rustc-link-lib=static:+whole-archive=portable_kernels");
println!("cargo::rustc-link-lib=static:+whole-archive=portable_ops_lib");

let libs_dir = std::env::var("EXECUTORCH_RS_EXECUTORCH_LIB_DIR").unwrap();
println!("cargo::rustc-link-search=native={libs_dir}/kernels/portable/");

Note that the ops and kernels libs are linked with +whole-archive to ensure that all symbols are included in the binary.

The EXECUTORCH_RS_EXECUTORCH_LIB_DIR environment variable should be set to the path of the C++ build output. If its not provided, its the responsibility of the binary to add the libs directories to the linker search path, and the crate will just link to the static libraries using cargo::rustc-link-lib=....

If you want to link to executorch libs yourself, set the environment variable EXECUTORCH_RS_LINK to 0, and the crate will not link to any library and not modify the linker search path.

The crate contains a small C/C++ bridge that uses the headers of the C++ library, and it is compiled using the cc crate (and the cxx crate, that uses cc under the hood). If custom compiler flags (for example -DET_MIN_LOG_LEVEL=Debug) are used when compiling the C++ library, you should set the matching environment variables that cc reads during cargo build (for example CFLAGS=-DET_MIN_LOG_LEVEL=Debug CXXFLAGS=-DET_MIN_LOG_LEVEL=Debug), see the cc docs.

§Cargo Features

By default the std feature is enabled.

data-loader: Includes the FileDataLoader and MmapDataLoader structs. Without this feature the only available data loader is BufferDataLoader. The libextension_data_loader.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON.
module: Includes the Module struct, a high-level API for loading and executing PyTorch models. It is an alternative to the lower-level Program API, which is more suitable for embedded systems. The libextension_module_static.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_MODULE=ON. Also includes the std, data-loader and flat-tensor features.
tensor-ptr: Includes a few functions creating cxx::SharedPtr<Tensor> pointers, that manage the lifetime of the tensor object alongside the lifetimes of the data buffer and additional metadata. The libextension_tensor.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_TENSOR=ON. Also includes the std feature.
flat-tensor: Includes the FlatTensorDataMap struct that can read .ptd files with external tensors for models. The libextension_flat_tensor.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON.
etdump: Includes the ETDumpGen struct, an implementation of an EventTracer, used for debugging and profiling. The libetdump.a static library is required, compile C++ executorch with EXECUTORCH_BUILD_DEVTOOLS=ON and EXECUTORCH_ENABLE_EVENT_TRACER=ON. In addition, the flatcc (or flatcc_d) library is required, available at {CMAKE_DIR}/third-party/flatcc_ep/lib/, and should be linked by the user.
std: Enable the standard library. This feature is enabled by default, but can be disabled to build executorch in a no_std environment. NOTE: no_std is still WIP, see https://github.com/pytorch/executorch/issues/4561

Re-exports§

pub use cxx;std

Modules§

util: Utility functions and structs.

Structs§

ArrayRefBool
ArrayRefChar
ArrayRefDimOrderType
ArrayRefEValue
ArrayRefEValuePtr
ArrayRefF64
ArrayRefI32
ArrayRefI64
ArrayRefOptionalTensor
ArrayRefSizesType
ArrayRefStridesType
ArrayRefTensor
ArrayRefU8
ArrayRefUsizeType
BoxedEvalueListI64
BoxedEvalueListOptionalTensor
BoxedEvalueListTensor
BufferDataLoader
DataLoaderRefMut
ETDumpGen
EValueRef
EValueRefMut
EValueStorage
EventTracermodule: EventTracer is a class that users can inherit and implement to log/serialize/stream etc.
EventTracerRefMut
ExecutorchPalImpl
FileDataLoader
FlatTensorDataMap
HierarchicalAllocator
MallocMemoryAllocator: Dynamically allocates memory using malloc() and frees all pointers at destruction time.
MemoryAllocator
MemoryManager
Method
MethodMeta
MmapDataLoader
Modulemodule: A facade class for loading programs and executing methods within them.
NamedDataMapRef
NamedDataMapRefMut
OptionalTensorRef
OptionalTensorRefMut
OptionalTensorStorage
Program
SpanI64
SpanOptionalTensor
SpanSpanU8
SpanTensor
SpanU8
Tensortensor-ptr: A minimal Tensor type whose API is a source compatible subset of at::Tensor.
TensorImpl
TensorInfo
TensorLayout
TensorRef
TensorRefMut
TensorStorage
VecChar
VecEValue
VecVecChar
executorch_tick_ratio: Represents the conversion ratio from system ticks to nanoseconds. To convert, use nanoseconds = ticks * numerator / denominator.

Enums§

Error: ExecuTorch Error type.
MmapDataLoaderMlockConfig: Describes how and whether to lock loaded pages with mlock().
ModuleLoadMode: Enum to define loading behavior.
ProgramHeaderStatus: Describes the presence of an ExecuTorch program header.
ProgramVerification: Types of validation that the Program can do before parsing the data.
ScalarType
Tag
TensorShapeDynamism: The resizing capabilities of a Tensor.
executorch_pal_log_level: Severity level of a log message. Values must map to printable 7-bit ASCII uppercase letters.

Constants§

EXECUTORCH_CPP_VERSION: The version of the ExecuTorch C++ library that this crate is compatible and linked with.

Functions§

MallocMemoryAllocator_as_memory_allocator^⚠: Get a pointer to the base class MemoryAllocator.
MallocMemoryAllocator_new: Construct a new Malloc memory allocator.
Module_execute^⚠module: Execute a specific method with the given input values and retrieve the output values. Loads the program and method before executing if needed.
Module_is_loadedmodule: Checks if the program is loaded.
Module_is_method_loadedmodule: Checks if a specific method is loaded.
Module_loadmodule: Load the program if needed.
Module_load_method^⚠module: Load a specific method from the program and set up memory management if needed.
Module_method_meta^⚠module: Get a method metadata struct by method name.
Module_method_names^⚠module: Get a list of method names available in the loaded program.
Module_newmodule: Constructs an instance by loading a program from a file with specified memory locking behavior.
Module_num_methods^⚠module: Get the number of methods available in the loaded program.
Module_unload_method^⚠module: Unload a specific method from the program.
TensorPtr_new^⚠tensor-ptr: Create a new tensor pointer.
executorch_BufferDataLoader_as_data_loader_mut^⚠
executorch_BufferDataLoader_new^⚠
executorch_ETDumpGen_as_event_tracer_mut^⚠
executorch_ETDumpGen_get_etdump_data^⚠
executorch_ETDumpGen_new^⚠
executorch_EValue_as_bool^⚠
executorch_EValue_as_bool_list^⚠
executorch_EValue_as_f64^⚠
executorch_EValue_as_f64_list^⚠
executorch_EValue_as_i64^⚠
executorch_EValue_as_i64_list^⚠
executorch_EValue_as_optional_tensor_list^⚠
executorch_EValue_as_string^⚠
executorch_EValue_as_tensor^⚠
executorch_EValue_as_tensor_list^⚠
executorch_EValue_copy^⚠
executorch_EValue_destructor^⚠
executorch_EValue_move^⚠
executorch_EValue_new_from_bool^⚠
executorch_EValue_new_from_bool_list^⚠
executorch_EValue_new_from_f64^⚠
executorch_EValue_new_from_f64_list^⚠
executorch_EValue_new_from_i64^⚠
executorch_EValue_new_from_i64_list^⚠
executorch_EValue_new_from_optional_tensor_list^⚠
executorch_EValue_new_from_string^⚠
executorch_EValue_new_from_tensor^⚠
executorch_EValue_new_from_tensor_list^⚠
executorch_EValue_new_none^⚠
executorch_EValue_tag^⚠
executorch_FileDataLoader_as_data_loader_mut^⚠
executorch_FileDataLoader_destructor^⚠
executorch_FileDataLoader_new^⚠
executorch_FlatTensorDataMap_as_named_data_map_mut^⚠
executorch_FlatTensorDataMap_load^⚠
executorch_HierarchicalAllocator_destructor^⚠
executorch_HierarchicalAllocator_new^⚠
executorch_MemoryAllocator_allocate^⚠
executorch_MemoryAllocator_new^⚠
executorch_MemoryManager_new^⚠
executorch_MethodMeta_attribute_tensor_meta^⚠
executorch_MethodMeta_get_backend_name^⚠
executorch_MethodMeta_input_tag^⚠
executorch_MethodMeta_input_tensor_meta^⚠
executorch_MethodMeta_memory_planned_buffer_size^⚠
executorch_MethodMeta_name^⚠
executorch_MethodMeta_num_attributes^⚠
executorch_MethodMeta_num_backends^⚠
executorch_MethodMeta_num_inputs^⚠
executorch_MethodMeta_num_memory_planned_buffers^⚠
executorch_MethodMeta_num_outputs^⚠
executorch_MethodMeta_output_tag^⚠
executorch_MethodMeta_output_tensor_meta^⚠
executorch_MethodMeta_uses_backend^⚠
executorch_Method_destructor^⚠
executorch_Method_execute^⚠
executorch_Method_get_attribute^⚠
executorch_Method_get_output^⚠
executorch_Method_inputs_size^⚠
executorch_Method_outputs_size^⚠
executorch_Method_set_input^⚠
executorch_MmapDataLoader_as_data_loader_mut^⚠
executorch_MmapDataLoader_destructor^⚠
executorch_MmapDataLoader_new^⚠
executorch_NamedDataMap_get_key^⚠
executorch_NamedDataMap_get_num_keys^⚠
executorch_NamedDataMap_get_tensor_layout^⚠
executorch_OptionalTensor_get^⚠
executorch_Program_check_header^⚠
executorch_Program_destructor^⚠
executorch_Program_get_method_name^⚠
executorch_Program_get_named_data_map^⚠
executorch_Program_load^⚠
executorch_Program_load_method^⚠
executorch_Program_method_meta^⚠
executorch_Program_num_methods^⚠
executorch_TensorImpl_new^⚠
executorch_TensorInfo_dim_order^⚠
executorch_TensorInfo_is_memory_planned^⚠
executorch_TensorInfo_name^⚠
executorch_TensorInfo_nbytes^⚠
executorch_TensorInfo_scalar_type^⚠
executorch_TensorInfo_sizes^⚠
executorch_TensorLayout_dim_order^⚠
executorch_TensorLayout_nbytes^⚠
executorch_TensorLayout_scalar_type^⚠
executorch_TensorLayout_sizes^⚠
executorch_Tensor_const_data_ptr^⚠
executorch_Tensor_coordinate_to_index^⚠
executorch_Tensor_coordinate_to_index_unchecked^⚠
executorch_Tensor_destructor^⚠
executorch_Tensor_dim^⚠
executorch_Tensor_dim_order^⚠
executorch_Tensor_element_size^⚠
executorch_Tensor_mutable_data_ptr^⚠
executorch_Tensor_nbytes^⚠
executorch_Tensor_new^⚠
executorch_Tensor_numel^⚠
executorch_Tensor_scalar_type^⚠
executorch_Tensor_size^⚠
executorch_Tensor_sizes^⚠
executorch_Tensor_strides^⚠
executorch_VecChar_destructor^⚠
executorch_VecEValue_destructor^⚠
executorch_VecVecChar_destructor^⚠
executorch_get_pal_impl^⚠: Returns the PAL function table, which contains function pointers to the active implementation of each PAL function.
executorch_is_valid_dim_order_and_strides^⚠
executorch_pal_abort^⚠: Immediately abort execution, setting the device into an error state, if available.
executorch_pal_allocate^⚠: NOTE: Core runtime code must not call this directly. It may only be called by a MemoryAllocator wrapper.
executorch_pal_current_ticks^⚠: Return a monotonically non-decreasing timestamp in system ticks.
executorch_pal_emit_log_message^⚠: Severity level of a log message. Values must map to printable 7-bit ASCII uppercase letters.
executorch_pal_free^⚠: Frees memory allocated by et_pal_allocate().
executorch_pal_init^⚠: Initialize the platform abstraction layer.
executorch_pal_ticks_to_ns_multiplier^⚠: Return the conversion rate from system ticks to nanoseconds as a fraction. To convert a system ticks to nanoseconds, multiply the tick count by the numerator and then divide by the denominator: nanoseconds = ticks * numerator / denominator
executorch_register_pal^⚠: Override the PAL functions with user implementations. Any null entries in the table are unchanged and will keep the default implementation.
executorch_stride_to_dim_order^⚠

Type Aliases§

DimOrderType: The type used for elements of Tensor.dim_order().
SizesType: The type used for elements of Tensor.sizes().
StridesType: The type used for elements of Tensor.strides().
executorch_timestamp_t: Platform timestamp in system ticks.