Problem with equality operator of expr.Mod?

[wweic]

I'm looking at unit test of ir builder, looks like the == operator does not work for Mod as expected:

    ib = tvm.ir_builder.create()
    n = tvm.var("n")
    A = ib.pointer("float32", name="A")
    with ib.for_range(0, n, name="i") as i:
        with ib.if_scope((i % 2) == 0):
            A[i] = A[i] + 1
        with ib.else_scope():
            A[0] = A[i] + 2

The generated IR is:

➜  ~ ipython
Python 2.7.10 (default, Jul 14 2015, 19:46:27)
Type "copyright", "credits" or "license" for more information.

IPython 5.1.0 -- An enhanced Interactive Python.
?         -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help      -> Python's own help system.
object?   -> Details about 'object', use 'object??' for extra details.

In [1]: import tvm
   ...:

In [2]:     ib = tvm.ir_builder.create()
   ...:     n = tvm.var("n")
   ...:     A = ib.pointer("float32", name="A")
   ...:     with ib.for_range(0, n, name="i") as i:
   ...:         with ib.if_scope((i % 2) == 0):
   ...:             A[i] = A[i] + 1
   ...:         with ib.else_scope():
   ...:             A[0] = A[i] + 2
   ...:
   ...:     body = ib.get()
   ...:

In [3]: body
Out[3]:
for (i, 0, n) {
  if (0) {
    A[i] = (A[i] + 1.000000f)
  } else {
    A[0] = (A[i] + 2.000000f)
  }
}

Note that the condition expression is constant 0 instead of an equality check expression. The __eq__ operator for Mod comes from tvm/_ffi/node.py

In [4]: i = ib.for_range(0, n, name="i")

In [5]: a = i._enter_value % 1

In [6]: a.__eq__.im_func
Out[6]: <function tvm._ffi.node.__eq__>

Which I assume what we want is tvm/expr.py?

I modify the unit test to use self.equal directly, and the code now makes sense to me:

In [7]: %paste
    ib = tvm.ir_builder.create()
    n = tvm.var("n")
    A = ib.pointer("float32", name="A")
    with ib.for_range(0, n, name="i") as i:
        with ib.if_scope((i % 2).equal(0)):
            A[i] = A[i] + 1
        with ib.else_scope():
            A[0] = A[i] + 2

    body = ib.get()
## -- End pasted text --

In [8]: body
Out[8]:
for (i, 0, n) {
  if (((i % 2) == 0)) {
    A[i] = (A[i] + 1.000000f)
  } else {
    A[0] = (A[i] + 2.000000f)
  }
}

Let me know if you think this is a problem. I'd be interested to help figure out why.

[tqchen]

this is mainly because python do not allow simple overload of equal operator, we need to use tvm.equal instead

[wweic]

Thanks for looking. It should be fine to override equal operator in python. I think what's happening here is __eq__ in NodeBase override __eq__ in ExprOp. For instance I directly instantiate an object of ExprOp(so the object does not inherit from NodeBase), == works fine:

In [50]: a = tvm.expr.ExprOp()

In [51]: a == 1
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-51-8ff30fa1c86a> in <module>()
----> 1 a == 1

/Library/Python/2.7/lib/python/site-packages/tvm-0.1.0-py2.7-macosx-10.10-intel.egg/tvm/expr.pyc in __eq__(self, other)
     88
     89     def __eq__(self, other):
---> 90         return self.equal(other)
     91
     92     def __ne__(self, other):

Since I'm quite new to the codebase, not sure who is the consumer of the interface. If it's a piece of code that generate IR with the language, should be easy to just let that code stick to equal instead of ==.

[tqchen]

I see, can you submit a patch to ExprOp to add this change? I agree that it is indeed helpful

[wweic]

I'll take a closer look and prepare a patch. thanks.

[wweic]

Turns out larger portion of tvm codebase assumes object identity semantics of == operator, instead of logical equality operator. After I force __eq__ to the one in ExprOp, a bunch of unit tests failed since they assumes object identity semantics(as defined in NodeBase.__eq__).

======================================================================
ERROR: test_arith_detect_linear_equation.test_basic
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/python2.7/site-packages/nose/case.py", line 197, in runTest
    self.test(*self.arg)
  File "//tvm/tests/python/unittest/test_arith_detect_linear_equation.py", line 18, in test_basic

    assert m[0] == b  <-------  logical equality operator will fail here since assert will call bool() on the expression, which will cause failure due to ExprOp.__bool__().

  File "/tvm-0.1.0-py2.7-macosx-10.12-intel.egg/tvm/expr.py", line 103, in __nonzero__
    "use tvm.all / tvm.any instead")
ValueError: Cannot use and / or / not operator to Expr, hint: use tvm.all / tvm.any instead

I'll look if we can replace the usage of object identity with is operator, and use == for logical equality. This should make it clearer about the meaning of ==.

For now I'll post a quick fix for the unit test first.

[tqchen]

get it, thanks! We cannot use is operator either, because the underlying object are different(they share the same handle). What we could do is to introduce a same_as function similar to C++ end that does the same testing

[wweic]

That sounds good. I also prefer leaving the syntax sugar == for logical equality operator. Will follow up this later.

[wweic]

resolved in #550