This option enables parallel dispatches for recompute tangents 0: off (default), triangle pass is interleaved with vertex pass, requires resource barriers in between. 1: on, batch triangle passes together, resource barrier, followed by vertex passes together, cost more memory.
