# Kernel Registration ## Overview At the last stage of [ExecuTorch model exporting](./export-overview.md), we lower the operators in the dialect to the _out variants_ of the [core ATen operators](./ir-ops-set-definition.md). Then we serialize these operator names into the model artifact. During runtime execution, for each operator name we will need to find the actual _kernels_, i.e., the C++ functions that do the heavy-lifting calculations and return results. Portable kernel library is the in-house default kernel library, it’s easy to use and portable for most of the target backends. However it’s not optimized for performance, because it’s not specialized for any certain target. Therefore we provide kernel registration APIs for ExecuTorch users to easily register their own optimized kernels. ## Design Principles **What do we support?** On the operator coverage side, the kernel registration APIs allow users to register kernels for all core ATen ops as well as custom ops, as long as the custom ops schemas are specified. Notice that we also support _partial kernels, _for example the kernel only supports a subset of tensor dtypes and/or dim orders. **Kernel contract**: kernels need to comply with the following requirements: * Match the calling convention derived from operator schema. The kernel registration API will generate headers for the custom kernels as references. * Satisfy the dtype constraints defined in edge dialect. For tensors with certain dtypes as arguments, the result of a custom kernel needs to match the expected dtypes. The constraints are available in edge dialect ops. * Gives correct result. We will provide a testing framework to automatically test the custom kernels. ## High Level Architecture ![](./_static/img/kernel-library-custom-aten-kernel.png) ExecuTorch users are asked to provide: 1. the custom kernel library with C++ implementations 2. a yaml file associated with the library that describes what operators are being implemented by this library. For partial kernels, the yaml file also contains information on the dtypes and dim orders supported by the kernel. More details in the API section. ### Workflow At build time, the yaml files associated with kernel libraries will be passed to the _kernel resolver_ along with the model op info (see selective build doc) and the outcome is a mapping between a combination of operator names and tensor metadata, to kernel symbols. Then codegen tools will use this mapping to generate C++ bindings that connect the kernels to ExecuTorch runtime. ExecuTorch users need to link this generated library into their application to use these kernels. At static object initialization time, kernels will be registered into the ExecuTorch kernel registry. At runtime initialization stage, ExecuTorch will use the operator name and argument metadata as a key to lookup for the kernels. For example, with “aten::add.out” and inputs being float tensors with dim order (0, 1, 2, 3), ExecuTorch will go into the kernel registry and lookup for a kernel that matches the name and the input metadata. ## APIs There are two sets of APIs: yaml files that describe kernel - operator mappings and codegen tools to consume these mappings. ### Yaml Entry for Core ATen Op Out Variant Top level attributes: * `op` (if the operator appears in `native_functions.yaml`) or `func` for custom operator. The value for this key needs to be the full operator name (including overload name) for `op` key, or a full operator schema (namespace, operator name, operator overload name and schema string), if we are describing a custom operator. For schema syntax please refer to this [instruction](https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md). * `kernels`: defines kernel information. It consists of `arg_meta` and `kernel_name`, which are bound together to describe "for input tensors with these metadata, use this kernel". * `type_alias`(optional): we are giving aliases to possible dtype options. `T0: [Double, Float]` means `T0` can be one of `Double` or `Float`. * `dim_order_alias`(optional): similar to `type_alias`, we are giving names to possible dim order options. Attributes under `kernels`: * `arg_meta`: a list of "tensor arg name" entries. The values for these keys are dtypes and dim orders aliases, that are implemented by the corresponding `kernel_name`. This being `null` means the kernel will be used for all types of input. * `kernel_name`: the expected name of the C++ function that will implement this operator. You can put whatever you want to here, but you should follow the convention of replacing the `.` in the overload name with an underscore, and lowercasing all characters. In this example, `add.out` uses the C++ function named `add_out`. `add.Scalar_out` would become `add_scalar_out`, with a lowercase `S`. We support namespace for kernels, but note that we will be inserting a `native::` to the last level of namespace. So `custom::add_out` in the `kernel_name` will point to `custom::native::add_out`. Some examples of operator entry: ```yaml - op: add.out kernels: - arg_meta: null kernel_name: torch::executor::add_out ``` An out variant of a core ATen operator with a default kernel ATen operator with a dtype/dim order specialized kernel (works for `Double` dtype and dim order needs to be (0, 1, 2, 3)) ```yaml - op: add.out type_alias: T0: [Double] dim_order_alias: D0: [[0, 1, 2, 3]] kernels: - arg_meta: self: [T0, D0] other: [T0 , D0] out: [T0, D0] kernel_name: torch::executor::add_out ``` ### Custom Ops C++ API For a custom kernel that implements a custom operator, we provides 2 ways to register it into ExecuTorch runtime: 1. Using `EXECUTORCH_LIBRARY` and `WRAP_TO_ATEN` C++ macros. 2. Using `functions.yaml` and codegen'd C++ libraries. The first option requires C++17 and doesn't have selective build support yet, but it's faster than the second option where we have to go through yaml authoring and build system tweaking. The first option is particularly suitable for fast prototyping but can also be used in production. Similar to `TORCH_LIBRARY`, `EXECUTORCH_LIBRARY` takes the operator name and the C++ function name and register them into ExecuTorch runtime. #### Prepare custom kernel implementation Define your custom operator schema for both functional variant (used in AOT compilation) and out variant (used in ExecuTorch runtime). The schema needs to follow PyTorch ATen convention (see native_functions.yaml). For example: ```yaml custom_linear(Tensor weight, Tensor input, Tensor(?) bias) -> Tensor custom_linear.out(Tensor weight, Tensor input, Tensor(?) bias, *, Tensor(a!) out) -> Tensor(a!) ``` Then write your custom kernel according to the schema using ExecuTorch types, along with APIs to register to ExecuTorch runtime: ```c++ // custom_linear.h/custom_linear.cpp #include Tensor& custom_linear_out(const Tensor& weight, const Tensor& input, optional bias, Tensor& out) { // calculation return out; } ``` #### Use a C++ macro to register it into PyTorch & ExecuTorch Append the following line in the example above: ```c++ // custom_linear.h/custom_linear.cpp // opset namespace myop EXECUTORCH_LIBRARY(myop, "custom_linear.out", custom_linear_out); ``` Now we need to write some wrapper for this op to show up in PyTorch, but don’t worry we don’t need to rewrite the kernel. Create a separate .cpp for this purpose: ```c++ // custom_linear_pytorch.cpp #include "custom_linear.h" #include at::Tensor custom_linear(const at::Tensor& weight, const at::Tensor& input, std::optional bias) { // initialize out at::Tensor out = at::empty({weight.size(1), input.size(1)}); // wrap kernel in custom_linear.cpp into ATen kernel WRAP_TO_ATEN(custom_linear_out, 3)(weight, input, bias, out); return out; } // standard API to register ops into PyTorch TORCH_LIBRARY(myop, m) { m.def("custom_linear(Tensor weight, Tensor input, Tensor(?) bias) -> Tensor", custom_linear); m.def("custom_linear.out(Tensor weight, Tensor input, Tensor(?) bias, *, Tensor(a!) out) -> Tensor(a!)", WRAP_TO_ATEN(custom_linear_out, 3)); } ``` #### Compile and link the custom kernel Link it into ExecuTorch runtime: In our `CMakeLists.txt`` that builds the binary/application, we just need to add custom_linear.h/cpp into the binary target. We can build a dynamically loaded library (.so or .dylib) and link it as well. Link it into PyTorch runtime: We need to package custom_linear.h, custom_linear.cpp and custom_linear_pytorch.cpp into a dynamically loaded library (.so or .dylib) and load it into our python environment. One way of doing this is: ```python import torch torch.ops.load_library("libcustom_linear.so/dylib") # Now we have access to the custom op, backed by kernel implemented in custom_linear.cpp. op = torch.ops.myop.custom_linear.default ``` ### Custom Ops Yaml Entry As mentioned above, this option provides more support in terms of selective build and features such as merging operator libraries. First we need to specify the operator schema as well as a `kernel` section. So instead of `op` we use `func` with the operator schema. As an example, here’s a yaml entry for a custom op: ```yaml - func: allclose.out(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False, bool dummy_param=False, *, Tensor(a!) out) -> Tensor(a!) kernels: - arg_meta: null kernel_name: torch::executor::allclose_out ``` The `kernel` section is the same as the one defined in core ATen ops. For operator schema, we are reusing the DSL defined in this [README.md](https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md), with a few differences: #### Out variants only ExecuTorch only supports out-style operators, where: * The caller provides the output Tensor or Tensor list in the final position with the name `out`. * The C++ function modifies and returns the same `out` argument. * If the return type in the YAML file is `()` (which maps to void), the C++ function should still modify `out` but does not need to return anything. * The `out` argument must be keyword-only, which means it needs to follow an argument named `*` like in the `add.out` example below. * Conventionally, these out operators are named using the pattern `.out` or `._out`. Since all output values are returned via an `out` parameter, ExecuTorch ignores the actual C++ function return value. But, to be consistent, functions should always return `out` when the return type is non-`void`. #### Can only return `Tensor` or `()` ExecuTorch only supports operators that return a single `Tensor`, or the unit type `()` (which maps to `void`). It does not support returning any other types, including lists, optionals, tuples, or scalars like `bool`. #### Supported argument types ExecuTorch does not support all of the argument types that core PyTorch supports. Here's a list of the argument types we currently support: * Tensor * int * bool * float * str * Scalar * ScalarType * MemoryFormat * Device * Optional * List * List> * Optional> ### Build Tool Macros We provide build time macros to help users to build their kernel registration library. The macro takes the yaml file describing the kernel library as well as model operator metadata, and packages the generated C++ bindings into a C++ library. The macro is available on both CMake and Buck2. #### CMake `generate_bindings_for_kernels(FUNCTIONS_YAML functions_yaml CUSTOM_OPS_YAML custom_ops_yaml)` takes a yaml file for core ATen op out variants and also a yaml file for custom ops, generate C++ bindings for kernel registration. It also depends on the selective build artifact generated by `gen_selected_ops()`, see selective build doc for more information. Then `gen_operators_lib` will package those bindings to be a C++ library. As an example: ```cmake # SELECT_OPS_LIST: aten::add.out,aten::mm.out gen_selected_ops("" "${SELECT_OPS_LIST}" "") # Look for functions.yaml associated with portable libs and generate C++ bindings generate_bindings_for_kernels(FUNCTIONS_YAML ${EXECUTORCH_ROOT}/kernels/portable/functions.yaml) # Prepare a C++ library called "generated_lib" with _kernel_lib being the portable library, executorch is a dependency of it. gen_operators_lib("generated_lib" KERNEL_LIBS ${_kernel_lib} DEPS executorch) # Link "generated_lib" into the application: target_link_libraries(executorch_binary generated_lib) ``` We also provide the ability to merge two yaml files, given a precedence. `merge_yaml(FUNCTIONS_YAML functions_yaml FALLBACK_YAML fallback_yaml OUTPUT_DIR out_dir)` merges functions_yaml and fallback_yaml into a single yaml, if there's duplicate entries in functions_yaml and fallback_yaml, this macro will always take the one in functions_yaml. Example: ```yaml # functions.yaml - op: add.out kernels: - arg_meta: null kernel_name: torch::executor::opt_add_out ``` And out fallback: ```yaml # fallback.yaml - op: add.out kernels: - arg_meta: null kernel_name: torch::executor::add_out ``` The merged yaml will have the entry in functions.yaml. #### Buck2 `executorch_generated_lib` is the macro that takes the yaml files and depends on the selective build macro `et_operator_library`. For an example: ```python # Yaml file for kernel library export_file( name = "functions.yaml" ) # Kernel library cxx_library( name = "add_kernel", srcs = ["add.cpp"], ) # Selective build artifact, it allows all operators to be registered et_operator_library( name = "all_ops", include_all_ops = True, # Select all ops in functions.yaml ) # Prepare a generated_lib executorch_generated_lib( name = "generated_lib", functions_yaml_target = ":functions.yaml", deps = [ ":all_ops", ":add_kernel", ], ) # Link generated_lib to ExecuTorch binary cxx_binary( name = "executorch_bin", deps = [ ":generated_lib", ], ) ```