>
 >On Sun, Aug 15, 2010 at 7:23 PM, Tyler Tricker <[email protected]>
 >wrote:
 >> Instead of SIMD built into the compiler, what would be the downsides
 >in
 >> having a generic assembly function( like C's asm but with a type
 >signature)
 >> and putting SIMD into a support library (a support library that could
 >be
 >> made to support things like OpenCL)? A support library could also have
 >a
 >> reference implementation for vanilla instruction sets.
 >
 >My personal opinion as a language user (rather than a language
 >implementor) is the "use our partially reliable research prototype
 >syntax, or else do absolutely everything yourself in assembly" is a
 >dichotomy that new languages often seem to end up presentingthe user
 >with, and it's precisely because I don't want either that I still
 >program in C++. It's a shame, because I'd really like to use a
 >language with more programmer support and higher level features.
 >
 >The example routine that I posted earlier in the thread shows the kind
 >of things that a programmer might want to do: it combines intrinsics
 >with "scalar" bit-shifting and looping that I'd both prefer not to
 >write in raw assembly and which don't seem to need it.
 >

Completelely agree here , a language that can do 80% of this  SIMD in the
language makes a strong argument for adoption especially for cross platform.
Regarding the support lib what language would it be ? We want to write less
assembly not more.. Learning all the SIMD instructions for a number of
platforms and idiosyncrasies is not something devs should need to do.

Ben


_______________________________________________
bitc-dev mailing list
[email protected]
http://www.coyotos.org/mailman/listinfo/bitc-dev

Reply via email to