🚀 go-pugleaf

In article <9106010224.AA28532@ucbvax.Berkeley.EDU> jbs@WATSON.IBM.COM writes:
   Path: ichips!iWarp.intel.com!ogicse!dali.cs.montana.edu!uakari.primate.wisc.edu!zaphod.mps.ohio-state.edu!cis.ohio-state.edu!ucbvax!WATSON.IBM.COM!jbs

	     Henry Spencer says:
   My understanding is that the funnier-looking features, in particular the
   infinities, NaNs, and signed zeros, mostly cost essentially nothing.

	     I believe this statement is incorrect.  Infs and nans compli-
   cate the design of the floating point unit.  Even if there is ultimately
   no performance impact, this added design effort represents a very real
   cost.  Perhaps some hardware designers could comment.

Ok, I designed the square root/divide hardware in the first generation of
Multiflow's machines.  I'm mostly on Henry's side -- you do need to detect
inputs that can cause wrong answers (square root of negative, 0/0, etc.),
and it does take extra hardware and design time, but it isn't that hard,
and it doesn't affect the cycle time.  Of the "nanocode" that controlled
this floating point unit, something like half dealt with getting
single/double divide/sqrt results, and the other half looked for
exceptions, rounding, and the like.

Getting the exceptions right was harder than getting the basic
functionality right, and was tougher to validate.  We found four bugs in
this functional unit over the two years it was in use, and three were in
exceptions, not basic functionality.  But it still isn't all that hard:  I
spent three months coming up with the basic design from a standing start.
And since there was a standard, I could run industry standard validation
suites against it to build my own confidence in it.  To me, that alone
proved the worth of the standard.

Getting the denorms right is not easy, and it does affect performance.
Especially in a statically scheduled machine, where the compiler was
expecting the result in cycle N, but your floating point unit just figured
out that it generated an underflow in cycle N-1 and must now do some
additional processing.  This means that the floating unit must be able to
stall the machine, something it previously didn't need, and stall is
inevitably on the critical timing path.

Bob Colwell  colwell@ichips.intel.com  503-696-4550
Intel Corp.  JF1-19
5200 NE Elam Young Parkway
Hillsboro, Oregon 97124