Re: Re How many vector registers are useful?

ssimmons@convex.com (Steve Simmons) writes:


      No, I do not know of any published studies that state the optimal number
      of vector registers. However, there are costs as more vector operators
      are added. These costs are expansion of instruction format, more complex
      micro code within the vector unit, more complex chaining analysis in both
      the compilers and vector unit, etc... Ultimately, this cost is paid for by
      the custumer.


Santosh Abraham, Ed Davidson, and I had a paper two asplos's ago that
looked at the minimal number of vector registers required for specific
codes. The question of the optimal is really tough to address for all the
obvious reasons: the dependence on compiler technology, variability in
program demands, and the real hardware costs of various machine designs
(like the register-fu interconnect). Because of this we decided to focus
on the minimal number of registers required to achieve optimal
performance. Sure, in theory its an intractable problem for a compiler,
but my experience along the way suggests that in fact you can often find
these optimal performance schedules without killing yourself, and the
early discoveries are often fairly close to the minimal (in terms of
register demands).


Last summer we had a follow up paper at ICS that presented a theory for
why the asplos results looked they way they did. This paper extended the
work on register demands to vliw and single issue machines, though still
focused on vectorizable loop codes.


Bill


--

Return to the comp.compilers page.
Search the comp.compilers archives again.