[BUG] iadd does not do in-place add

[suranap]

Software versions

On Sapling2 at Stanford

$ legate-issue
Python : 3.10.14 | packaged by conda-forge | (main, Mar 20 2024, 12:45:18) [GCC 12.3.0]
Platform : Linux-5.4.0-169-generic-x86_64-with-glibc2.31
Legion : legion-24.03.0
Legate : 24.01.00.dev+38.g90944d7
Cunumeric : 24.01.00.dev+32.g364e95dc.dirty
Numpy : 1.26.4
Scipy : 1.13.1
Numba : 0.59.1
CTK package : cuda-version-11.7-h67201e3_3 (conda-forge)
GPU driver : 535.54.03
GPU devices :
GPU 0: Tesla P100-SXM2-16GB
GPU 1: Tesla P100-SXM2-16GB
GPU 2: Tesla P100-SXM2-16GB
GPU 3: Tesla P100-SXM2-16GB

Jupyter notebook / Jupyter Lab version

No response

Expected behavior

Both iadd and add(x,y,out=x) should reuse the space occupied by a and not make a new copy. In the stack trace below, it appears the code is doing this:

t = copy(a)
t = task(a + b)

Observed behavior

This code gets to DeferredArray.binary_op. Here's a code snippet:

        lhs = self.base
        src1 = src1._copy_if_overlapping(self).  # Is this line right? 
        rhs1 = src1._broadcast(lhs.shape)
        src2 = src2._copy_if_overlapping(self)
        rhs2 = src2._broadcast(lhs.shape)

The 2nd line compares a to a, confirms they are overlapping and copies the array into a new thunk.

Example code or instructions

import cunumeric as np
a = np.random.randint(0, 1000, size=(50257, 768))
b = np.random.randint(0, 1000, size=(50257, 768))
a += b

Stack traceback or browser console output

>>> pdb.run("np.add(a, b, out=a)", globals=globals())
> <string>(1)<module>()
(Pdb) c
> ~/projects/legate/cunumeric/cunumeric/deferred.py(119)wrapper()
-> self = args[0]
(Pdb) n
> ~/projects/legate/cunumeric/cunumeric/deferred.py(120)wrapper()
-> args = tuple(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(124)wrapper()
-> for (idx, arg) in enumerate(args)
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(120)wrapper()
-> args = tuple(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(126)wrapper()
-> for k, v in kwargs.items():
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(130)wrapper()
-> return func(*args, **kwargs)
(Pdb) p args
(<cunumeric.deferred.DeferredArray object at 0x7f94556d4820>, <BinaryOpCode.ADD: 1>, <cunumeric.deferred.DeferredArray object at 0x7f94556d4820>, <cunumeric.deferred.DeferredArray object at 0x7f94556d5750>, True, ())
(Pdb) s
--Call--
> ~/projects/legate/cunumeric/cunumeric/deferred.py(3301)binary_op()
-> @auto_convert("src1", "src2")
(Pdb) n
> ~/projects/legate/cunumeric/cunumeric/deferred.py(3310)binary_op()
-> lhs = self.base
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(3311)binary_op()
-> src1 = src1._copy_if_overlapping(self)
(Pdb) s
--Call--
> ~/projects/legate/cunumeric/cunumeric/deferred.py(291)_copy_if_overlapping()
-> def _copy_if_overlapping(self, other: DeferredArray) -> DeferredArray:
(Pdb) n
> ~/projects/legate/cunumeric/cunumeric/deferred.py(292)_copy_if_overlapping()
-> if not self.base.overlaps(other.base):
(Pdb) self.base.overlaps(other.base)
True
(Pdb) n
> ~/projects/legate/cunumeric/cunumeric/deferred.py(294)_copy_if_overlapping()
-> copy = cast(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(295)_copy_if_overlapping()
-> DeferredArray,
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(296)_copy_if_overlapping()
-> self.runtime.create_empty_thunk(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(297)_copy_if_overlapping()
-> self.shape,
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(298)_copy_if_overlapping()
-> self.base.type,
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(299)_copy_if_overlapping()
-> inputs=[self],
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(296)_copy_if_overlapping()
-> self.runtime.create_empty_thunk(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(294)_copy_if_overlapping()
-> copy = cast(
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(302)_copy_if_overlapping()
-> copy.copy(self, deep=True)
(Pdb)
> ~/projects/legate/cunumeric/cunumeric/deferred.py(119)wrapper()
-> self = args[0]
(Pdb)

It's the create_empty_thunk that looks strange to me. And the copy.copy is just more unneeded work.

[manopapad]

I was able to reproduce this issue, and @bryevdv will be investigating.