network stack bypassing with hw queue classification and QOS?

I have hardware that can classify incoming IP-packets L2-L4, and on
the DS field.
This classification allows me to know what outgoing NIC the packets
should end up on, but I still want normal QOS functionality for this
traffic, so I can't push the packets across while staying in the
driver.

Is there any good way of making use of the availability of a
TCAM-alike packet rule classifier while still having the kernel
sort/drop/queue the packets according to various QOS-settings (with no
limitation on what king of policy that can be used for a certain flow)
before they end up in the outgoind NIC? The aim for this application
is to get good packet rate in a system that doesn't have much CPU to
spare.

/Daniel