<cc'ing llvmdev, please don't email me directly>
On Sun, 27 May 2007, Christophe Avoinne wrote:
> On 23/05/2007 22:05:48, Chris Lattner (sabre at nondot.org) wrote:
> > LLVM 2.0 is done! Download it here: http://llvm.org/releases/ or view
> > the release notes:
http://llvm.org/releases/2.0/docs/ReleaseNotes.html
> >
>
> Hello Chris.
>
> Your project sounds quite exciting and I may be interested to use it
instead
> of gcc.
>
> Well, let's me explain why I'm interested in this project :
>
> I have a PlayStation Portable and like to develop several projects using
> all of its potential.
>
> PSP has two Allegrex cpus : SC (system control) and ME (Media Engine).
> Both are designed with a MIPS 2 core plus a dozen instructions borrowed
from
> MIPS4.
> Both also have a standard FPU coprocessor.
> But only SC processor has a second coprocessor called VFPU which allows
> matrix and vector computations.
>
> Actually, gcc doesn't handle VFPU at all (i'm speaking about the
compiler,
> not the assembler).
>
> I tried to add the necessary things to allow gcc to handle them but I have
> only a half success because of the way gcc handles registers and a few
other
> messing things.
>
> So i'm looking for an alternative : llvm-gcc.
Ok.
> First, what is VFPU :
>
> it is a great coprocessor which allows us to handle up to 128 32-bit float
> registers as a scalar or as a vector of two, three or four elements or as a
> square matrix of 4, 9 or 16 elements.
>
> - a scalar : S000 = [$0], S001, [$1], S002 = [$2], S003 = [$3], S100 =
[$4],
> ..., S010 = [$32], ..., S020 = [$64], ..., S733 = [$127]
> - a column : C000.p = [S000, S010], C020.p = [S020, S030], ..., C000.t =
> [S000, S010, S020], C010.t = [S010, S020, S030], ..., C000.q = [S000, S010,
> S020, S030]
> - a row : R000.p = [S000, S001], R002.p = [S002, S003], R000.t = [S000,
S001,
> S002], R001.t = [S001, S002, S003], R000.q = [S000, S001, S002, S003]
> - a matrix : M000.p = [R000.p, R010.p], M020.p = [R020.p, R030.p], ...,
> M000.t = [R000.t, R010.t, R020.t], ..., M000.q = [R000.q, R010.q, R020.q,
> R030q]
> - a transposed matrix : E000.p = [C000.p, C001.p], M020.p = [R020.p,
R030.p],
> ..., M000.t = [R000.t, R010.t, R020.t], ..., M000.q = [R000.q, R010.q,
> R020.q, R030q]
>
> (NOTE : <t><#m><#r><#c>.<s> : t = S
(scalar)/C (column)/R (row)/M (matrix)/E
> (transposed matrix), #m = matrix, #r = row, #c = column, s = p (2d vector),
t
> (3d vector), q (4d vector))
>
> now, if you want just a 3d dot product, just issue this instruction for
> instance : vdot.t S000.s, C100.t, C110.t
> or a matrix product : vmmul.q M000.q, M100.q, E200.q
>
> As you can see, that VFPU has a great potential but, alas, it is not very
> well exploited by the open source psp-gcc.
Ok.
> Now, lets's speak about llvm-gcc.
>
> 1) Does a (uncompleted) MIPS target exist for llvm-gcc ? or must I create
> one ?
Bruno just submitted one, he's the right person to talk to. I have not
had a chance to review the code, so it has not been committed yet, but it
should be soon.
> 2) Do you think that llvm-gcc can handle those combinaisons of registers :
> 2.1) Can llvm-gcc handle vector float as a super-set of scalar registers ?
> 2.2) Can llvm-gcc handle matrix float as a super-set of vector float ?
LLVM has good support for vectors, if you treat matrices as large vectors,
you should have no problem.
> 3) Can llvm-gcc vectorize float operations ?
llvm-gcc supports explicitly vectorized code, but we have no
autovectorization pass yet.
-Chris
--
http://nondot.org/sabre/
http://llvm.org/
