[Broadcasting] Concat inputs scale with the broadcast size. about glow

[jfix71]

@rdzhabarov Good point, I actually don't think we actually need the flexibility as far as I've seen. I think we will generally see either all inputs to the concat with count = 1, or a single repeated input (through a Broadcast or Tile).

An alternative that is less complex and doesn't provide this unnecessary flexibility is to add count and axis to the InsertTensorNode. The InsertTensorInst currently already has count and axis, so this is a somewhat natural change. Then we could move createTile() (which createBroadcast() uses) to create an InsertTensorNode with a count instead of ConcatNode with repeated input as it currently does.

@qcolombet I would imagine the input data for a Splat with a repeat count needs to be statically known, similar to how current splats only take a statically known float value? If so, I don't think it is applicable in this case, because the data here that is to be concatenated isn't statically known.

[qcolombet]

Sounds sensible.
One thing I would suggest is to tie the input and number of times more closely if possible: i.e., (A, B) (2048, 10) vs. (A, 2048) (B, 10).

[rdzhabarov]

Sounds like a good idea to me. Do we anticipate that we'd need the flexibility to use: A,5 times; B,4 times; A,10 times again etc? Maybe not, but I think what @qcolombet suggests will help us to support this case pretty easy if needed.

[qcolombet]

One thing that came to mind.
We could also introduce a new splat node that would repeat the input x time.
E.g.,
instead of doing
concat A, 2040, B 10

We would have, (concat (splat A, 2048), (splat B, 10)).

I am just suggesting that because I think we were thinking of adding such a "splat" node.

[jfix71]

Done by #1332