Stub 注册流程
所有的 stub 定义几乎都在 aten/src/ATen/native/DispatchStub.h 文件，可以慢慢看。里面有段注释：
// Implements instruction set specific function dispatch.
//
// Kernels that may make use of specialized instruction sets (e.g. AVX2) are
// compiled multiple times with different compiler flags (e.g. -mavx2). A
// DispatchStub contains a table of function pointers for a kernel. At runtime,
// the fastest available kernel is chosen based on the features reported by
// cpuinfo.
//
// Example:
//
// In native/MyKernel.h:
// using fn_type = void(*)(const Tensor& x);
// DECLARE_DISPATCH(fn_type, stub)
//
// In native/MyKernel.cpp
// DEFINE_DISPATCH(stub);
//
// In native/cpu/MyKernel.cpp:
// namespace {
// // use anonymous namespace so that different cpu versions won't conflict
// void kernel(const Tensor& x) { ... }
// }
// REGISTER_DISPATCH(stub, &kernel);
//
// To call:
// stub(kCPU, tensor);
//
// TODO: CPU instruction set selection should be folded into whatever
// the main dispatch mechanism is.
//
// Supported device types for registration:
// - CPU: Central Processing Unit
// - CUDA: NVIDIA GPUs
// - HIP: AMD GPUs
// - MPS: Apple Silicon GPUs (Metal Performance Shaders)
// - MTIA: Meta Training and Inference Devices
// - XPU: Intel GPUs
// - HPU: Reserved for HPU (Intel Gaudi) device types
// - PrivateUse1: Reserved for private/custom device types
//
// If you want to update the list of supported devices, add a new dispatch_ptr
// member in DispatchStubImpl.h and update the get_call_ptr switch.
// As well you will need to update the inlined list in 'is_device_supported`
//
//
// ignore warnings about DispatchStub::DEFAULT, AVX, AVX2 defined elsewhere
DispatchStub 模板基类定义
见 aten/src/ATen/native/DispatchStub.h。DispatchStub 类型为：
template <typename rT, typename T, typename... Args>
struct DispatchStub<rT (*)(Args...), T>;
其中主要包含几类方法，一是调用，会根据设备类型来选择函数指针，强制转换后调用：
