[llvm-dev] [Proposal][RFC] Epilog loop vectorization

[Nema, Ashutosh via llvm-dev]

From: anemet at apple.com [mailto:anemet at apple.com]
Sent: Wednesday, March 15, 2017 6:21 AM
To: Hal Finkel <hfinkel at anl.gov>
Cc: Nema, Ashutosh <Ashutosh.Nema at amd.com>; Zaks, Ayal <ayal.zaks at intel.com>; Renato Golin <renato.golin at linaro.org>; mkuper at google.com; Mehdi Amini <mehdi.amini at apple.com>; Daniel Berlin <dberlin at dberlin.org>; llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] [Proposal][RFC] Epilog loop vectorization


On Mar 14, 2017, at 11:33 AM, Hal Finkel <hfinkel at anl.gov<mailto:hfinkel at anl.gov>> wrote:



On 03/14/2017 12:11 PM, Adam Nemet wrote:

On Mar 14, 2017, at 9:49 AM, Hal Finkel <hfinkel at anl.gov<mailto:hfinkel at anl.gov>> wrote:


On 03/14/2017 11:21 AM, Adam Nemet wrote:

On Mar 14, 2017, at 6:00 AM, Nema, Ashutosh <Ashutosh.Nema at amd.com<mailto:Ashutosh.Nema at amd.com>> wrote:

Summarizing the discussion on the implementation approaches.

Discussed about two approaches, first running ‘InnerLoopVectorizer’ again on the epilog loop immediately after vectorizing the original loop within the same vectorization pass, the second approach where re-running vectorization pass and limiting vectorization factor of epilog loop by metadata.

<Approach-2>
Challenges with re-running the vectorizer pass:
1)      Reusing alias check result:
When vectorizer pass runs again it finds the epilog loop as a new loop and it may generates alias check, this new alias check may overkill the gains of epilog vectorization.
We should use the already computed alias check result instead of re computing again.
2)      Rerun the vectorizer and hoist the new alias check:
It’s not possible to hoist alias checks as its not fully redundant (not dominated by other checks), it’s not getting execute in all paths.

<Mail Attachment.jpeg>


I don’t understand. Looks like you have the same alias checks for the epilog loop too.  How is this CFG different from the re-vectorization of the scalar loop?

You're looking at the wrong thing. This *is* the image from re-vectorization. The other image (linked below in step (3)) shows the other option.

Ah ok, the numbering confused me here.




 Would be good to have both CFGs here and highlighting the difference.

I thought that the whole point was that *if* you reached the epilog vector loop via the first vector loop, you want to bypass the alias checks before the epilog vector.

Yes, but, that's not quite true now. You can also reach the epilogue loop if you fail the min-trip-count check, and so you don't know anything about the aliasing checks.

OK, so we want this loops to be handled specially.  We effectively say that we only vectorize this loop if it does not require any alias checks or if the alias checks can be predicate-forwarded to this loop from existing checks.

This still seems like an orthogonal issue that may be interesting to solve independently.  In other words this could be a nice feature in the vectorizer anyway: the loop is estimated to be low-trip count so feel free to predicate the new vector loop so that the alias check result could be reused from some other block.  We obviously don’t have this capability today but it’s something that could be nice aside from the vectorizer.

That sounds great. I'm not sure, however, exactly what this means. This "predicate forwarding" sounds like a non-local restructuring of the surrounding code (because the predicates aren't known to be true at that point, we need to search different predecessors to find one in which the conditions might be true and insert the vector loop across that predecessor edge instead). Maybe we could do this, for example, by calling SE->isKnownPredicate enhanced with some additional context sensitivity because we currently check dominating conditional branches for things that are AddRecs in some loop? Moreover, we then have the problem of restructuring the order of the trip-count checks (because we need to fast fail to the scalar loop for the smallest trip counts). Maybe we can do this the same way? This means finding a dominating check on the trip count that implies the check we're about to insert, change that check to the check we want, keeping the original only on the true side of the trip-count check (i.e. if the trip count is larger than the small threshold, then check the large threshold).

If we're going to do all of that, then I'd lean toward saying that this does not belong in the vectorizer at all. Rather, this seems like something we'd want in some general transformation (this seems somewhat akin to what JumpThreading does). The general transformation seems something like this; the basic vectorized loop looks like this:

int start = 0;
if (n >= vf) {
  if (check) {
    for (...; start += vf)
      ...
  }
}

for (i = start; i < n; ++i) {
  ...
}

and after we vectorize the epilogue loop we end up with this:

int start = 0;
if (n >= vf) {
  if (check) {
    for (...; start += vf)
      ...
  }
}

if (n >= vf2) {

This is (n - start >= vf2) which I think makes this a bit more complicated since now “check” is not always redundant if n >= vf2 so hoisting becomes a cost decision.


  if (check) {

But we can turn this into a predicate we already have: if (n >= vf && check).  BTW, this is I think how Ashutosh’ approach 1 works too; if the first vector loop does not iterate once the alias result will be false in the epilog vector loop.


    for (...; start += vf2)
      ...
  }
}

for (i = start; i < n; ++i) {
  ...
}

and we need to end up with this:

int start = 0;
if (n >= vf2) {
  if (check) {
    if (n >= vf) {
      for (...; start += vf)
        ...
    }

    for (...; start += vf2)
      ...
  }
}

for (i = start; i < n; ++i) {
  ...
}

where we've recognized here that 'check' is the same in both cases, and that because vf2 < vf, the one trip-count check implies the other. This latter part seems like the part that our existing passes might not know what to do with currently. Thoughts?


And then we get:

int start = 0;
bool check_result = false;
if (n >= vf) {
  if (check) {
    check_result = true;
    for (...; start += vf)
      ...
  }
}

if (n - start >= vf2) {
  if (check_result) {
    for (...; start += vf2)
      ...
  }
}

for (i = start; i < n; ++i) {
  …
}

What do you think?

This is exactly what my approach is, we do not want to execute alias check for epilog vector loop, even when “n < vf" as epilog loop has very limited trip count. We just check if first vector loop is not iterated once then the alias result check is ‘false’.


Adam



 -Hal



Adam



 -Hal



I still don’t understand why that’s not possible with some sophisticated predicate propagation independent from the vectorizer.  I am not saying it’s already possible but it should be.

Adam




NOTE: We cannot prepone alias check as its expensive compared to other checks.

<Approach-1>
1)      Current patch depends on the existing functionality of LoopVectorizer, it uses ‘InnerLoopVectorizer’ again to vectorize the epilog loop, as it happens in the same vectorization pass we have flexibility to reuse already computed alias result check & limit vectorization factor for the epilog loop.
2)      It does not generate the blocks for new block layout explicitly, rather it depends on ‘InnerLoopVectorizer::createEmptyLoop’ to generate new block layout. The new block layout get automatically generated by calling the ‘InnerLoopVectorizer:: vectorize’ again.
3)      Block layout description with epilog loop vectorization is available at
https://reviews.llvm.org/file/data/fxg5vx3capyj257rrn5j/PHID-FILE-x6thnbf6ub55ep5yhalu/LayoutDescription.png

Approach-1 looks feasible, please comment if any objections.

Regards,
Ashutosh


...



--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170315/eca62248/attachment.html>