Re: [RFC] Vectorization of indexed elements

[Vidya Praveen <vidyapraveen@arm.com>]

On Tue, Oct 01, 2013 at 09:26:25AM +0100, Richard Biener wrote:
> On Mon, 30 Sep 2013, Vidya Praveen wrote:
> 
> > On Mon, Sep 30, 2013 at 02:19:32PM +0100, Richard Biener wrote:
> > > On Mon, 30 Sep 2013, Vidya Praveen wrote:
> > > 
> > > > On Fri, Sep 27, 2013 at 04:19:45PM +0100, Vidya Praveen wrote:
> > > > > On Fri, Sep 27, 2013 at 03:50:08PM +0100, Vidya Praveen wrote:
> > > > > [...]
> > > > > > > > I can't really insist on the single lane load.. something like:
> > > > > > > > 
> > > > > > > > vc:V4SI[0] = c
> > > > > > > > vt:V4SI = vec_duplicate:V4SI (vec_select:SI vc:V4SI 0)
> > > > > > > > va:V4SI = vb:V4SI <op> vt:V4SI
> > > > > > > > 
> > > > > > > > Or is there any other way to do this?
> > > > > > > 
> > > > > > > Can you elaborate on "I can't really insist on the single lane load"?
> > > > > > > What's the single lane load in your example? 
> > > > > > 
> > > > > > Loading just one lane of the vector like this:
> > > > > > 
> > > > > > vc:V4SI[0] = c // from the above scalar example
> > > > > > 
> > > > > > or 
> > > > > > 
> > > > > > vc:V4SI[0] = c[2] 
> > > > > > 
> > > > > > is what I meant by single lane load. In this example:
> > > > > > 
> > > > > > t = c[2] 
> > > > > > ...
> > > > > > vb:v4si = b[0:3] 
> > > > > > vc:v4si = { t, t, t, t }
> > > > > > va:v4si = vb:v4si <op> vc:v4si 
> > > > > > 
> > > > > > If we are expanding the CONSTRUCTOR as vec_duplicate at vec_init, I cannot
> > > > > > insist 't' to be vector and t = c[2] to be vect_t[0] = c[2] (which could be 
> > > > > > seen as vec_select:SI (vect_t 0) ). 
> > > > > > 
> > > > > > > I'd expect the instruction
> > > > > > > pattern as quoted to just work (and I hope we expand an uniform
> > > > > > > constructor { a, a, a, a } properly using vec_duplicate).
> > > > > > 
> > > > > > As much as I went through the code, this is only done using vect_init. It is
> > > > > > not expanded as vec_duplicate from, for example, store_constructor() of expr.c
> > > > > 
> > > > > Do you see any issues if we expand such constructor as vec_duplicate directly 
> > > > > instead of going through vect_init way? 
> > > > 
> > > > Sorry, that was a bad question.
> > > > 
> > > > But here's what I would like to propose as a first step. Please tell me if this
> > > > is acceptable or if it makes sense:
> > > > 
> > > > - Introduce standard pattern names 
> > > > 
> > > > "vmulim4" - vector muliply with second operand as indexed operand
> > > > 
> > > > Example:
> > > > 
> > > > (define_insn "vmuliv4si4"
> > > >    [set (match_operand:V4SI 0 "register_operand")
> > > >         (mul:V4SI (match_operand:V4SI 1 "register_operand")
> > > >                   (vec_duplicate:V4SI
> > > >                     (vec_select:SI
> > > >                       (match_operand:V4SI 2 "register_operand")
> > > >                       (match_operand:V4SI 3 "immediate_operand)))))]
> > > >  ...
> > > > )
> > > 
> > > We could factor this with providing a standard pattern name for
> > > 
> > > (define_insn "vdupi<mode>"
> > >   [set (match_operand:<mode> 0 "register_operand")
> > >        (vec_duplicate:<mode>
> > >           (vec_select:<scalarmode>
> > >              (match_operand:<mode> 1 "register_operand")
> > >              (match_operand:SI 2 "immediate_operand))))]
> > 
> > This is good. I did think about this but then I thought of avoiding the need
> > for combiner patterns :-) 
> > 
> > But do you find the lane specific mov pattern I proposed, acceptable? 
> 
> The specific mul pattern?  As said, consider factoring to vdupi to
> avoid an explosion in required special optabs.
> 
> > > (you use V4SI for the immediate?  
> > 
> > Sorry typo again!! It should've been SI.
> > 
> > > Ideally vdupi has another custom
> > > mode for the vector index).
> > > 
> > > Note that this factored pattern is already available as vec_perm_const!
> > > It is simply (vec_perm_const:V4SI <source> <source> <immediate-selector>).
> > > 
> > > Which means that on the GIMPLE level we should try to combine
> > > 
> > > el_4 = BIT_FIELD_REF <v_3, ...>;
> > > v_5 = { el_4, el_4, ... };
> > 
> > I don't think we reach this state at all for the scenarios in discussion.
> > what we generally have is:
> > 
> >  el_4 = MEM_REF < array + index*size >
> >  v_5 = { el_4, ... }
> > 
> > Or am I missing something?
> 
> Well, but in that case I doubt it is profitable (or even valid!) to
> turn this into a vector lane load from the array.  If it is profitable
> to perform a vector read (because we're going to use the other elements
> of the vector as well) then the vectorizer should produce a vector
> load and materialize the uniform vector from one of its elements.
> 
> Maybe at this point you should show us a compilable C testcase
> with a loop that should be vectorized using your instructions in
> the end?

Here's a compilable example:

void 
foo (int *__restrict__ a,
     int *__restrict__ b,
     int *__restrict__ c)
{
  int i;

  for (i = 0; i < 8; i++)
    a[i] = b[i] * c[2];
}

This is vectorized by duplicating c[2] now. But I'm trying to take advantage
of target instructions that can take a vector register as second argument but
use only one element (by using the same value for all the lanes) of the 
vector register.

Eg. mul <vec-reg>, <vec-reg>, <vec-reg>[index]
    mla <vec-reg>, <vec-reg>, <vec-reg>[index] // multiply and add

But for a loop like the one in the C example given, I will have to load the
c[2] in one element of the vector register (leaving the remaining unused)
rather. This is why I was proposing to load just one element in a vector 
register (what I meant as "lane specific load"). The benefit of doing this is
that we avoid explicit duplication, however such a simplification can only
be done where such support is available - the reason why I was thinking in
terms of optional standard pattern name. Another benefit is we will also be
able to support scalars in the expression like in the following example:

void
foo (int *__restrict__ a,
     int *__restrict__ b,
     int c)
{
  int i;

  for (i = 0; i < 8; i++)
    a[i] = b[i] * c;
}


Another example which can take advantage of the target feature:

void 
foo (int *__restrict__ a,
     int *__restrict__ b,
     int *__restrict__ c)
{
  int i,j;

  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++)
      a[j] += b[j] * c[i];
}

This scenario we discussed this earlier (you suggested handling this at TER).

Cheers
VP