[ONNX] only broadcast matmul if the shape has changed

[mbrookhart]

cc @AndrewZhaoLuo @jwfromm

#9911 introduced some extra broadcasting to handle edge cases in shapes. In cases where the broadcasted shape matched the input shape, this ended up being a null op that caused some issues with a model I've been testing.

I think the issue is deeper in how broadcast_to is lowered, but removing the unnecessary broadcast also fixes my bug, so I'm punting and adding the deeper issue to my backlog.

[AndrewZhaoLuo]

LGTM

[jwfromm]

Lowering to a null op is very interesting behavior. We should make sure to fix that eventually.. This work around in the meantime makes sense to me.

[AndrewZhaoLuo]

Can we make a tracking issue for this?

[jwfromm]

We might have to poke around a little more to figure out the issue. I wasn't able to replicate a failure using a dedicated script with useless broadcast_to calls inserted. This seems to run fine for example:

import numpy as np
import tvm
from tvm import relay
from tvm.contrib import graph_executor

x = relay.var('x', shape=[10], dtype='float32')
y = relay.var('y', shape=[10], dtype='float32')
val = relay.const(5, dtype='float32')
x2 = relay.broadcast_to(x, relay.shape_of(y))
out = x2 + val
mod = tvm.IRModule.from_expr(out)

with relay.build_config(opt_level=3):
    lib = relay.build(mod, target="llvm")

gmod = graph_executor.GraphModule(lib["default"](tvm.cpu()))
x_np = np.random.normal(size=[10]).astype('float32')
y_np = np.random.normal(size=[10]).astype('float32')
gmod.set_input('x', x_np)
gmod.set_input('y', y_np)
gmod.run()
print(gmod.get_output(0))

Do we expect that to trigger the issue or is it more nuanced?

[mbrookhart]

I haven't been able to isolate the problem from the model yet, but we'll work to do that offline, reproduce as a test, and then file the issue

[mbrookhart]

Thanks @AndrewZhaoLuo @jwfromm