Re: Separating algorithms from implementations (long)

"Toby Sharp" <TSharp@Serif.com> writes:


<snip>


> Intel's C/C++ compiler claims automatic
> vectorisation for using their SIMD extenstions, but personally (having
> tried their compiler) I certainly wouldn't trust my inner loops with
> it.


<snip>


> Then a very, very clever compiler could spend a lot of time looking at
> the algorithm description, assembling instructions into
> implementations, testing sample data, and making changes as
> appropriate.


Do you see an inconsistency between those statements? On the one hand,
you "wouldn't trust" an existing attempt at compiler cleverness, but
you posit the development of a much more clever compiler. Why do you
think a more ambitious effort would succeed where previous efforts
have, in your view, failed?


> However, the algorithm doesn't change with the CPU, so neither does
> its description. It can be recompiled later to generate a new
> implementation that benefits from the features of a new CPU.


Except, of course, that algorithms do depend on the hardware they
use. An algorithm for a massively-parallel machine will look different
from one for a single processor. And one machine might have fast
access to memory (so lookup tables and storing intermediate results is
wise), while another has small slow memory (making recalculation the
efficient choice).


<snip>
--
-Stephen H. Westin

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