aten.constant_pad_nd.default not support rank > 4

[swimdi]

aten.constant_pad_nd.default currently not support lowering to ttnn.pad if its input rank > 4

or it will TT_THROW Tensor rank is greater than 4

btw, there's no rank > 4 case in models, so I think can lower its priority

[ayerofieiev-tt]

@swimdi can you explain what this op does? do we lower it to some TT-NN operation?

[swimdi]

aten.constant_pad_nd.default is do the padding on the op and it current partially lowering to ttnn.pad

pytorch2.0_ttnn/torch_ttnn/passes/lowering/to_tt_pass.py

Lines 732 to 765 in fa5218c

	if node.target == torch.ops.aten.constant_pad_nd.default:
	input, pad = args[0], args[1]
	if len(args) > 2:
	value = args[2]
	else:
	value = 0
	# TODO(#516)
	if any(p < 0 for p in pad):
	return None
	input_shape = input.meta["val"].size()
	output_shape = node.meta["val"].size()
	rank = len(input_shape)
	full_pad = [(0, 0)] * (rank - len(pad))
	# The order of pad from pytorch is reversed
	full_pad += [(pad[i], pad[i + 1]) for i in range(0, len(pad), 2)][::-1]
	# TODO(#514)
	if rank > 4:
	return None
	# TODO(#192): Front padding isn't well supported so skip for now
	if not all(f == 0 for f, _ in full_pad):
	return None
	# Change layout to row-major for non-tile-size-aligned tensor
	if (
	rank < 2
	or input_shape[-1] % ttnn.TILE_SIZE != 0
	or input_shape[-2] % ttnn.TILE_SIZE != 0
	or full_pad[-1][1] % ttnn.TILE_SIZE != 0
	or full_pad[-2][1] % ttnn.TILE_SIZE != 0
	):
	input = g.call_function(ttnn.to_layout, args=(input, TtnnRowMajorLayout()))
	# TODO(#515)
	if output_shape[-1] % 2 != 0:
	return None
	return g.call_function(ttnn.pad, args=(input, full_pad, value))