Re: how to enable vector intrinsic functions in fortran frontend

[Sa Liu <SALIU at de dot ibm dot com>]

Tobias Burnus <burnus@net-b.de> wrote on 15.10.2007 21:05:17:

> I also would like that the vector types compile on (almost) any system.
> The advantage of Fortran is that it is essentially system independent;
> there are only very few parts in Fortran programs which are system
> dependent (such as the arguments to the "SYSTEM" call).
> 
> VECTOR sounds like a concept very much suited to number crunching in
> general and thus it makes sense - in terms of portability - that a
> program targeted for a vector-supporting system also runs on systems
> which do not; and even if the generated code on non-PowerPC is clumsy
> and slow, having it supported there makes the feature more appealing to
> scientific users (which rarely target one system only).

Generally I think that the VECTOR type should be kept behind a flag, so 
that it only gets enabled on the systems supporting vector without 
bothering the other. As the scalar types of vector vary from system to 
system, idealy the checking routines should be kept in each backend. But 
the fortran parse should be manipulated to recognize the VECTOR syntax. In 
general we'd like to have a Fortran extension to support VECTOR and allow 
Fortran applications to benefit from vectorizable platforms. 

> > I think an intrinsic module would be the best place for these to go.
> The advantage of an intrinsic module would be that it keeps the
> namespace clean; the disadvantage is that it is not directly compatible
> with the XL Fortran compiler. 

Concretely why is it not compatible with XL Fortran compiler? Sorry I'm 
not quite familar with intrinsic module.

> One could combine these: A flag to enable the VECTOR(...) declaration
> and an intrinsic module to provide VEC_*. However, as VEC_* only takes
> vectors, one could directly enable it with the -fvector-type flag. I
> also do not see the problem name clashes for this reasons.
> 

I actually prefer the former option. As we try to enable vector intrinsics 
on more than one systems, the -fvector-type should only switch on data 
declaration VECTOR(...), and different intrinsic sets should be kept 
separately, say in different intrinsic modules (if the compatibility 
problem getting resolved). And each intrinsic module should only be 
activated on the target system.

To take PowerPC and SPU as an example, there are a couple of steps to go:

- allow declaration VECTOR(...) with -fvector-type

- check types/kinds on PowerPC
        VECTOR(INTEGER(1))              vector signed char 
        VECTOR(INTEGER(2))              vector signed short
        VECTOR(INTEGER(4))              vector signed int, vector signed 
long
        VECTOR(PIXEL)                   vector pixel
        VECTOR(REAL(4))         vector float
        VECTOR(UNSIGNED(1))     vector unsigned char 
        VECTOR(UNSIGNED(2))             vector unsigned short 
        VECTOR(UNSIGNED(4))             vector unsigned int, vector 
unsigned long

- check types/kinds on SPU 
        VECTOR(INTEGER(1))              vector signed char
        VECTOR(INTEGER(2))              vector signed short
        VECTOR(INTEGER(4))              vector signed int
        VECTOR(INTEGER(8))              vector signed long long
        VECTOR(REAL(4))         vector float
        VECTOR(REAL(8))         vector double
        VECTOR(UNSIGNED(1))     vector unsigned char
        VECTOR(UNSIGNED(2))     vector unsigned short
        VECTOR(UNSIGNED(4))             vector unsigned int
        VECTOR(UNSIGNED(8))             vector unsigned long long

- add intrinsic module ppc-altivec-intrinsics.def for PowerPC

- add intrinsic module spu-vector-intrinsics.def for SPU

I will think of detailed solutions when I dig a bit more in the code. 
Further suggestions/comments/criticism are alway welcome!


Thanks!
Sa