NetNews Usenet Archive 1992 #27/NN_1992_27.iso/spool/comp/arch/11035

Internet Message Format | 1992-11-19 | 2.5 KB

Xref: sparky comp.arch:11035 comp.lang.misc:3845
Newsgroups: comp.arch,comp.lang.misc
Path: sparky!uunet!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!hamblin.math.byu.edu!hellgate.utah.edu!lanl!cochiti.lanl.gov!jlg
From: [e-mail] (J. Giles)
Subject: Re: Hardware Support for Numeric Algorithms
Message-ID: <[e-mail]>
Sender: [e-mail]
Organization: Los Alamos National Laboratory
References: <[e-mail]> <[e-mail]> <[e-mail]>
Date: Mon, 23 Nov 1992 18:32:04 GMT
Lines: 37

In article <[e-mail]>, [e-mail] (Wayne Throop) writes:
|> [...]
|> But ultimately, when you are going after that last 1% or .1% of speed,
|> you will simply have to spell out the istream, operation by operation.

Unfortunately, the speedup from going to assembly is usually in the
5% to 10% area.  And yes, there are people to whom such an increase
is worth a man-year or two of effor to achieve.  They'd rather spend
less effort if the CS community would provide better integrated tools
for optimizing, and there are people who cannot afford the effort
these kinds of optimizations require - but would benefit also if the
tools were improved.

|> [...]
|> No amounts of hints to the optimizer will ever be enough, because
|> ultimately the optimizer can only deal with hints it is prepared to
|> accept, and if a programmer comes up with something unforseen by the
|> optimizer, there's no simple way to give a hint about it.

True.  However it's possible to have the amount of explicit
assembly kept to a minimum if the proper interfacing protocols
are present.  For example, neither C nor Fortran (nor most other
high-level languages) have a way for the user to get the leading
zero count of the vector mask on a Cray.  I have algorithms which
need that computation and which are time-critical.  The nearest
that the high-level languages can do is some clumsy loop.  All
I really need are some intrinsics corresponding to the instructions
which set the vector mask based on some array conditionals and an
intrinsic which does a leading zero count instruction.  In fact,
if  all the machine instructions were avalable as intrinsic functions
calls, then the compiler need only be able to do  adequate register
allocation between these intrinsics (and also into the code generated
from the higher level) and I could optimize any part of the code
without paying procedure-call overhead.

--
J. Giles