CVS-175119-[OVEP] Fixed possibility of array index out of bounds in subgraph partitioning

[RajeevSekar]

Models with nodes like QLinear, DQLinear and similar ones were crashing while running perf test, due to an issue in the sub graph partitioning logic

[preetha-intel]

can we add the increment part of loop so that the additional check can be removed.

something like
while (++j < total_clusters && !append_node)

[RajeevSekar]

Hi Preetha, I have made that change, thanks for pointing it out

[RajeevSekar]

Hi Preetha, due to this change of adding the increment part in the loop, an issue was created.

Old logic (larger loop)
In the case where j < total_clusters is true but append node is also true (the loop is not entered), j is not incremented

New logic (shorter loop)
In the case ++j is still < total_clusters but append node is true, the loop is not entered but j is aldready incremented (unlike the first logic).

so, just after the loop, in this block

if (append_node) {
connected_clusters[j].emplace_back(index);
}

incorrect value of j is used which causes some model to fail.

So we cant use "while (++j < total_clusters && !append_node)"

Is there a way to optimize the loop logic but retain the j if we exit the loop due to append_node == true