relay-aot

An experimental ahead of time compiler for Relay.

The ahead of time compiler enables the execution of Relay code without requiring a framework's interpreter written in C++ or Python.

The removal of framework and interpretation overhead combined with optimized operators produced by TVM's operators dramatically reduces execution time. Additionally, the compiler produces a single binary which only depends on TVM, simplifying the deployment story. This repository contains an external library which implements ahead of time compilation for Relay. The current approach is a proof of concept which lowers Relay to C++, and relies on a C++ compiler such as gcc or clang to produce an executable.

The ahead of time compiler comes as a standalone library which exposes a primitive compile function which compiles a relay.Function into a Python closure which wraps the compiled native code.

The compiler's design is straight forward. It lowers functions into a small C++-like IR, and generates a C++ program which can be compiled and dynamically linked. We extract the corresponding symbol from the dynamic library, and wrap it as a Python closure.

You can see an example below:

import numpy as np
import tvm
from tvm.relay import Module, GlobalVar, Function
from aot import compile

def double_example():
    # Declare a Relay module.
    mod = Module()

    # Implement the double function.
    x = var('x', shape=())
    double = GlobalVar('double')
    mod[double] = Function([x], x + x)

    # Generate a function which calls double twice.
    x = var('x', shape=())
    f = Function([x], double(double(x)))
    # Compile the function.
    cfunc = compile(f, mod)

    a = tvm.nd.array(np.array(1.5, dtype='float32'))
    output = cfunc(a).asnumpy() # array(6.)

Currently there is no Python package due to the lack of a package for tvm itself. You can use the ahead of time compiler by adding it to your PYTHONPATH.

export PYTHONPATH="THE_AOT_PATH:${PYTHONPATH}"

You must set the variable TVM_HOME in order to use the native compiler currently.

You can test to ensure you setup the ahead of time correctly by running:

TVM_HOME=~/Git/tvm python3 examples/readme_ex.py

Note: this only sets the TVM_HOME for this command you must reexport in your shell.