generic simd and c++

generic simd and c++

[Andrew Pinski]

Does generic simd works in c++?
When I try it I get the following error:
findPeaks.cc:47: error: invalid operands of types `vector int' 
and `vector int'
    to binary `operator-'

Thanks,
Andrew Pinski

Re: generic simd and c++

[Aldy Hernandez]

>>>>> "Andrew" == Andrew Pinski <pinskia@physics.uc.edu> writes:

 > Does generic simd works in c++?
 > When I try it I get the following error:
 > findPeaks.cc:47: error: invalid operands of types `vector int' and
 > `vector int'
 >     to binary `operator-'

vector doesn't exist as a keyword.  You're going to have to define a
macro defining it as we did for altivec:

        #define vector __attribute__((vector_size(16)))

This will give you 128 bit vectors.  Use 8 for 64-bit vectors, etc.

however, try to use something else but vector, since this will cause
conflicts with C++ libraries.

Aldy

Re: generic simd and c++

[Andrew Pinski]

no this is not the problem, if you copy gcc/testsuite/gcc.c-torture/execute/simd-1.c to simd.cc and add #include <stdlib.h> to the top of the file and try to compile it, you will get the same error as I reported.

Thanks,
Andrew Pinski

I will be filling a bug report soon for some of simd problems with -O1 and -msse also.


> 
> >>>>> "Andrew" == Andrew Pinski <pinskia@physics.uc.edu> writes:
> 
>  > Does generic simd works in c++?
>  > When I try it I get the following error:
>  > findPeaks.cc:47: error: invalid operands of types `vector int' and
>  > `vector int'
>  >     to binary `operator-'
> 
> vector doesn't exist as a keyword.  You're going to have to define a
> macro defining it as we did for altivec:
> 
>         #define vector __attribute__((vector_size(16)))
> 
> This will give you 128 bit vectors.  Use 8 for 64-bit vectors, etc.
> 
> however, try to use something else but vector, since this will cause
> conflicts with C++ libraries.
> 
> Aldy
> 
> 

Re: generic simd and c++

[Benjamin Kosnik]

Somebody should just make a template specialization of std::vector that
uses the simd extensions. See <vector> and perhaps the vector<bool>
specialization already in place if you're curious.

That would rule.

Such a specialization could easily be shipped with libstdc++ as a file
in the "include/ext" directory, once copyright assignments were taken
care of.

-benjamin

Re: generic simd and c++

[Daniel Egger]

[-- Attachment #1: Type: text/plain, Size: 687 bytes --]

Am Mit, 2002-07-31 um 21.59 schrieb Benjamin Kosnik:

> Somebody should just make a template specialization of std::vector that
> uses the simd extensions. See <vector> and perhaps the vector<bool>
> specialization already in place if you're curious.

I'm curious how you think this might work. What should the template
parameter be? SIMD instructions are somewhat different from typical
containter operations and would require special methods while some 
operations a vector specialisation needs would be hardly implementable
and probably not make much sense.

Maybe I'm just thinking in the wrong direction so please show me the
way.... :)

-- 
Servus,
       Daniel

[-- Attachment #2: Dies ist ein digital signierter Nachrichtenteil --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: generic simd and c++

[Gabriel Dos Reis]

Daniel Egger <degger@fhm.edu> writes:

| [1  <text/plain>]
| Am Mit, 2002-07-31 um 21.59 schrieb Benjamin Kosnik:
| 
| > Somebody should just make a template specialization of std::vector that
| > uses the simd extensions. See <vector> and perhaps the vector<bool>
| > specialization already in place if you're curious.
| 
| I'm curious how you think this might work. What should the template
| parameter be? SIMD instructions are somewhat different from typical
| containter operations and would require special methods while some 
| operations a vector specialisation needs would be hardly implementable
| and probably not make much sense.
| 
| Maybe I'm just thinking in the wrong direction so please show me the
| way.... :)

Actually, I thought the Altivec vector workd Aldy did may be extended
to support valarray<>.  I didn't reallywork hard; but I think it is
doable.  For fixed sized valarray<>.

-- Gaby