Torch Lua中addmm函数定义位置及Lua与C代码关联疑问

Let's break down how the addmm method you're seeing in Linear:updateGradInput connects to the TH library's C implementation—this is a common point of confusion when diving into Torch's internals.

First: addmm is a Tensor Method, Not a Lua Function Defined in the Codebase #

When you see self.gradInput:addmm(0, 1, gradOutput, self.weight) in Linear.lua, you're calling a member method of the Torch Tensor object. This method isn't written in Lua—it's a wrapper around a C function from the TH library, exposed to Lua via Torch's Lua/C binding system.

How the Lua-C Binding Works #

Torch7 is built on LuaJIT, which has robust support for calling C functions. Here's the high-level flow:

• TH Library Implements Core Logic: The actual matrix multiplication logic for addmm (short for "add matrix-matrix product") lives in TH's C code. There are type-specific versions (like THFloatTensor_addmm or THDoubleTensor_addmm) in THTensorMath.c to handle different tensor types (float, double, etc.).
• Torch Registers C Functions as Lua Tensor Methods: During Torch's initialization, the core library registers these C functions as methods on the torch.Tensor metatable. So when you call tensor:addmm(...), Lua looks up the method in the tensor's metatable and invokes the corresponding C function.
• Automatic Type Dispatch: The Lua method automatically routes your call to the correct C function based on the tensor's underlying type. For example, a float tensor will use THFloatTensor_addmm, while a double tensor uses THDoubleTensor_addmm.

Verifying This Yourself #

If you want to confirm this, fire up a Torch Lua shell and run:

print(torch.Tensor.addmm)

You'll see output like function: 0x..., which signals this is a C function (not a Lua-defined function) exposed to the Lua environment.

Why You Can't Find a Lua Definition for addmm #

Simply put: there isn't one. All low-level tensor operations in Torch are implemented in C for performance, and only exposed to Lua via bindings. The Linear.lua code just uses these pre-registered methods to handle the computations needed for backpropagation.

内容的提问来源于stack exchange，提问作者krngrvr09