Need a `mutex` op in Fluid

[kavyasrinet]

During the discussion of designing Go op in Fluid, we discovered that we might need a synchronized op in fluid that acquires a mutex on shared variables, if the Go op is used with other operators in Fluid. For example, to execute the following:

import fluid

W = fluid.tensor(...)
X = fluid.tensor(...)

with fluid.go():
    mul_op(W, X)

# main thread
mul_op(W, X)

We need to acquire a mutex on the variables W and X in case other Go routines are trying to access them. This is how Go performs it as well (reference: https://golang.org/ref/mem#tmp_5). So one proposal is:

import fluid

W = fluid.tensor(...)
X = fluid.tensor(...)

mutex = fluid.mutex(...)

with fluid.go():
    with fluid.synchronize(mutex):
        mul_op(W, X)

# main thread
with fluid.synchronize(mutex):
        mul_op(W, X)

We need to figure out how to share the lock between threads.
This work would require us to create a new variable that holds a mutex and a new operator synchronize that takes in the mutex and the block to protect.

[wangkuiyi]

I agree that in some use cases, mutex would be easier than CSP, like the above example.

A few suggestions on naming:

mu = fluid.make_mutex()
with fluid.lock_mutex(mu):
  mul_op(W, X)