Unsupported cases for `aten.amin/amax` conversion

[jerrysky3]

When trying to convert aten.amin/amax to ttnn.min/max, we ran into the issues below with tt-metal and currently don't support them:

• [ ] keepdim = false isn't supported in tt-metal
• [ ] Reduction on dims < rank - 2 results in assertion failure Unsupported dim inside tt-metal
• [ ] Rank of input shape < 2 gives (1, 1) output shape instead of the expected (1)
• [ ] (..., x, y) input shape where x or y % ttnn.TILE_SIZE != 0. In this case TILE_LAYOUT pads the input tensor with 0 which gives out incorrect results in min/max operation. It should pads inf/-inf corresponding to ttnn.min/max
• [ ] When the output shape is (x, ttnn.TILE_SIZE * y) where x % ttnn.TILE_SIZE != 0, tt-metal tries to reshape the padded output back to the expected output shape and results in assertion failure Unable to reshape a tensor in TILE_LAYOUT to non-tile height and width! Please convert the tensor to ROW_MAJOR_LAYOUT first.

[jerrysky3]

We also found ttnn.max(input_tensor) always return 0 even if all values < 0

[bbradelTT]

@jerrysky3 could you please provide one or more concrete examples of calling ttnn.max for the last case ("When the output shape is (x, ttnn.TILE_SIZE * y) where x % ttnn.TILE_SIZE != 0")?

Do you mean something like

def test_max3(device): 
    x = -torch.ones((3, 1, 3, 64), dtype=torch.float32)
    x_tt = ttnn.from_torch(x, dtype=ttnn.bfloat16, layout=ttnn.TILE_LAYOUT, device=device)
    keepdim = False # True
    y_tt = ttnn.max(x_tt, dim=0, keepdim=keepdim)
    y = ttnn.to_torch(y_tt)
    z = torch.max(x, dim=0, keepdim=keepdim).values
    print(f'x {x}\ny_tt {y_tt}\nz {z}')
    assert_with_pcc(z, y)