ttnn.pad does not when support output_shape[-1] is odd

[swimdi]

We convert aten.constant_pad_nd.default to ttnn.pad and it currently does not support cases wheb output_shape[-1] is odd

or it will TT_THROW pad_op.cpp:34: this->output_tensor_shape[3] % 2 == 0 RM padding requires output X dim to be a multiple of 2

These input varation are appeared in models and is related to this constraint

["Tensor<[1, 3, 224, 224]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 96, 112, 112]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 144, 56, 56]> self = ?", "List[int] pad = [1, 2, 1, 2]", "number value = 0.0"],
["Tensor<[1, 240, 28, 28]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 672, 14, 14]> self = ?", "List[int] pad = [1, 2, 1, 2]", "number value = 0.0"],
["Tensor<[1, 3, 240, 240]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 96, 120, 120]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 144, 60, 60]> self = ?", "List[int] pad = [1, 2, 1, 2]", "number value = 0.0"],
["Tensor<[1, 240, 30, 30]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 672, 15, 15]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 3, 260, 260]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 96, 130, 130]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 144, 65, 65]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 288, 33, 33]> self = ?", "List[int] pad = [1, 1, 1, 1]", "number value = 0.0"],
["Tensor<[1, 720, 17, 17]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 3, 300, 300]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 144, 150, 150]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 192, 75, 75]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 288, 38, 38]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 816, 19, 19]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 3, 380, 380]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 144, 190, 190]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 192, 95, 95]> self = ?", "List[int] pad = [2, 2, 2, 2]", "number value = 0.0"],
["Tensor<[1, 336, 48, 48]> self = ?", "List[int] pad = [0, 1, 0, 1]", "number value = 0.0"],
["Tensor<[1, 960, 24, 24]> self = ?", "List[int] pad = [1, 2, 1, 2]", "number value = 0.0"],

[jaykru-tt]

I think I have a fix for this that should be landing soon with the fixes for #12896 and #15511. I'll give you a heads up when it's merged.

[jaykru-tt]

I didn't manage to land a fix for this in my PR, but we'll be tackling this as an onboarding exercise for a new teammate in the New Year :)