Re: Re: [PATCH] VECT: Add WHILE_LEN pattern for decrement IV support for auto-vectorization

[钟居哲 <juzhe.zhong@rivai.ai>]

[-- Attachment #1: Type: text/plain, Size: 3843 bytes --]

Thanks Kewen.

Current flow in this patch like you said:
....
len = WHILE_LEN (n,vf);
...
v = len_load (addr,len);
..
addr = addr + vf (in byte align);
....

This patch is just keep adding address with a vector factor (adjust as byte align).
For example, if your vector length = 512bit. Then this patch is just updating address as
addr = addr + 64;

However, today after I read RVV ISA more deeply, it should be more appropriate that
the address should updated as : addr = addr + (len * 4) if len is element number of INT32.
the len is the result by WHILE_LEN which calculate the len. 

I assume for IBM target, it's better to just update address directly adding the whole register bytesize 
in address IV. Since I think the second way (address = addr + (len * 4)) is too RVV specific, and won't be suitable for IBM. Is that right?
If it is true, I will keep this patch flow (won't change to  address = addr + (len * 4)) to see what else I need to do for IBM.
I would rather do that in RISC-V backend port.

>> I tried
>>to compile the above source files on Power, the former can adopt doloop
>>optimization but the latter fails to. 
You mean GCC can not do hardward loop optimization when IV loop control is variable ? 


juzhe.zhong@rivai.ai
 
From: Kewen.Lin
Date: 2023-04-13 15:29
To: 钟居哲
CC: gcc-patches; Jeff Law; rdapp; richard.sandiford; rguenther
Subject: Re: [PATCH] VECT: Add WHILE_LEN pattern for decrement IV support for auto-vectorization
Hi Juzhe,
 
on 2023/4/12 21:22, 钟居哲 wrote:
> Thanks Kewen. 
> 
> It seems that this proposal WHILE_LEN can help s390 when using --param vect-partial-vector-usage=2 compile option.
> 
 
Yeah, IMHO, the previous sequence vs. the proposed sequence are like:
 
int
foo (int *__restrict a, int *__restrict b, int n)
{
  if (n <= 0)
    return 0;
 
  int iv = 0;
  int len = MIN (n, 16);
  int sum = 0;
  do
    {
      sum += a[len] + b[len];
      iv += 16;
      int n1 = MIN (n, iv);                   // line A
      int n2 = n - n1;
      len = MIN (n2, 16);
    }
  while (n > iv);
 
  return sum;
}
 
vs.
 
int
foo (int *__restrict a, int *__restrict b, int n)
{
  if (n <= 0)
    return 0;
 
  int len;
  int sum = 0;
  do
    {
      len = MIN (n, 16);
      sum += a[len] + b[len];
      n -= len;
    }
  while (n > 0);
 
  return sum;
}
 
it at least saves one MIN (at line A) and one length preparation in the
last iteration (it's useless since loop ends).  But I think the concern
that this proposed IV isn't recognized as simple iv may stay.  I tried
to compile the above source files on Power, the former can adopt doloop
optimization but the latter fails to.
 
> Would you mind apply this patch && support WHILE_LEN in s390 backend and test it to see the overal benefits for s390
> as well as the correctness of this sequence ? 
 
Sure, if all of you think this approach and this revision is good enough to go forward for this kind of evaluation,
I'm happy to give it a shot, but only for rs6000. ;-)  I noticed that there are some discussions on withdrawing this
WHILE_LEN by using MIN_EXPR instead, I'll stay tuned.
 
btw, now we only adopt vector with length on the epilogues rather than the main vectorized loops, because of the
non-trivial extra costs for length preparation than just using the normal vector load/store (all lanes), so we don't
care about the performance with --param vect-partial-vector-usage=2 much.  Even if this new proposal can optimize
the length preparation for --param vect-partial-vector-usage=2, the extra costs for length preparation is still
unavoidable (MIN, shifting, one more GPR used), we would still stay with default --param vect-partial-vector-usage=1
(which can't benefit from this new proposal).
 
BR,
Kewen