[llvm-dev] [RFC] Adding thread group semantics to LangRef (motivated by GPUs)

Jan Sjodin via llvm-dev llvm-dev at lists.llvm.org
Thu Jan 31 06:59:40 PST 2019


 
> >    I think you misunderstand me if you think I was listing disallowed
> >    transforms. My question was if it is possible to have multi-threaded
> >    semantics in the source language, but in the compiler have a
> >    single-threaded view, where some properties of the CFG would determine
> >    what is legal and not for some functions with a special flag.
> 
>
> But that's practically the same as listing allowed and disallowed
> transforms -- you're defining what the final IR is allowed to look
> like, not how it's allowed to execute. If at all possible, the former
> should be derived from the latter.

It does at least say something about transforms that haven't been written
yet, but from that viewpoint, sure you can say that.

> > > The only way to
> > > conclusively prove that transforms will never break code that's
> > > supposed to work (except for bugs in the transforms) is to define the
> > > semantics of the IR, and then to make sure that it refines the
> > > semantics of the source language. This is why the LangRef almost never
> > > talks about allowed/disallowed transforms except as examples to
> > > explain some given semantics, and if you don't follow that rule, your
> > > patch will probably be rejected.
> >
> > I'm trying to figure out if we are in the "almost never" territory
> >  here or not.
> > 

> I don't think so at all. In addition to being much, much harder to
> reason about and prove that the approach is sound, I think it'll be
> more intrusive.

It is much harder to come up with rules that are correct, and will not
be too restrictive for sure. I am not convinced that it is more
intrusive since it will only affect some transforms.

> >  Any transform that re-arranges control flow would potentially have to
> >  know about the properties of ballot(), and the rules with respect to
> >  the CFG (and maybe consider the target) to know where to insert the
> >  intrinsics.

> But the same is true for basically any approach to handling this. In
> fact, adding the merge intrinsics makes this much easier. Beyond the
> usual problems with hoisting ballots, which passes currently avoid via
> the current convergent + speculatable attributes, we'd only have to
> add awareness to passes that they can't duplicate/combine merge
> intrinsics or move them past convergent intrinsics, which is a local
> property and hence easily checkable. In example I explained, without
> some kind of merge intrinsic, tail duplication has to look at the
> entire loop to know whether the transform is legal. Of course, you
> could have some kind of "no convergent calls" flag on a function, but
> that doesn't eliminate the nastyness when there are convergent calls.

We will have to determine if the intrinsics are worth more compared to
scanning the code.

> >   I had the impression that the control flow convergence was
> >   in part specified by what the target architecture can handle.

> This is true, although hopefully we can agree on something that
> everyone can implement.

Yes, hopefully there is an execution model that is works for everyone.

> >  One of
> >  the more complicated cases would be linearization where the control
> >  flow is completely rewritten, and is encoded in a variable that says
> >  which basic block is the next one to execute.

> It's certainly interesting to think about how to maintain correctness
> in the face of ballots() with such a pass, but a) it's certainly no
> harder with merge intrinsics than merges being implicit and b) I doubt
> that's useful for anything you'd want to do with a GPU.

Irreducible control flow has to be handled somehow, and linearization
is the only transform I know of, that will handle everything. I'm not
sure what the execution model says about irreducible control flow.

> >  Another case is DCE,
> >  where a ballot() could be eliminated, and it would potentially have to
> >  remove a number of intrinsics to enable later optimizations (unless it
> >  would affect performance?), so the intrinsics will require some
> >  non-local updates.

> Removing merge intrinsics if it's profitable and allowed is a
> separate optimization, that can be done relatively quickly in a single
> pass without impacting the performance of other optimizations. It's
> requiring expensive non-local checks in optimizations which modify
> control flow that's a problem.

There are certainly a lot of tradeoffs. My point was simply that the
intrinsics are not strictly local.

> >  So we might have them without a ballot(), which would seem to make it
> >  more difficult for structurizers or other transforms to maintain the
> >  semantics and insert intrinsics.

> It's not any more difficult code-wise, since the case where there is
> a ballot needs to be handled anyways. And while it might take longer
> to process stuff when this hypothetical pass encounters a merge
> intrinsic (I can't think of a real-world case where it would matter),
> it should result in better code generated.

I was thinking of linearization or something similar, where an
instrinsic by itself might be hard to preserve, compared to looking at
a ballot() and derive where intrinsics should be inserted.

> > > > How are uniform branches handled? Do they affect the convergence model?
> > > >
> > > We may be able to remove convergence points if branches are uniform.
> > > In Nicolai's proposal, I believe we'd want to remove a merge intrinsic
> > > when all the branches that it post-dominates that aren't also
> > > post-dominated by some other merge intrinsic are uniform.
> >
> > I couldn't quite understand the last sentence, but I assume the
> > conditions would prevent removing convergence points that help
> > performance. Post-domination might not be adequate if there are loops
> > btw.
> 
> It should be -- Nicolai's proposal is that a merge intrinsic merges
> all the divergences caused by branches post dominated by the
> intrinsic, so if all the divergent branches are merged by some other
> intrinsic earlier, then there's no divergence and the merge intrinsic
> is a no-op.

Makes sense.

- Jan
    On Wednesday, January 30, 2019, 11:41:29 AM EST, Connor Abbott <cwabbott0 at gmail.com> wrote:  
 
 On Wed, Jan 30, 2019 at 4:20 PM Jan Sjodin <jan_sjodin at yahoo.com> wrote:

 
>
> > > for (int i = 0; i < 2; i++) {
> > >  foo = ballot(true); // ballot 1
> > >
> > >    if (threadID /* ID of the thread within a wavefront/warp */ % 2 == 0) continue;
> > >
> > >    bar = ballot(true); // ballot 2
> > > }
> > >
> > > versus:
> > >
> > > int i = 0;
> > > while (true) {
> > >    do {
> > >        if (i == 2) break;
> > >        foo = ballot(true); // ballot 1
> > >        i++;
> > >    } while (threadID % 2 == 0);
> > >
> > >    if (i == 2) break;
> > >    bar = ballot(true); // ballot 2
> > >    i++;
> > > }
> >
> > I think you can remove the second "i++", otherwise we can increment "i" twice
> > if threadID % 2 != 0, but I see the issue. Using the equivalence classes would
> > prevent this transform, since we have re-arranged the control flow in that way,
>
> What do you mean? Note that the transforms that lead from the first
> example to the second are actually desirable if there aren't any
> ballots or other convergent operations, so banning them entirely is a
> no-go. Otherwise, note that the ballot here can be nested arbitrarily
> deep, which means that jump forwarding/tail duplication has to be
> aware of the entire loop unless we have some kind of merge intrinsic.

Yes, if there weren't any calls to a ballot, then the transform would
be legal. What I was saying in the beginning was that ballot() would
have some special rules attached to it. It is of course undesirable to
have flags to enforce correctness, but that is the point we are at
right now.

> Also, I should say that while interesting, control equivalence classes
> can be part of the *implementation*, but not the *specification*. We
> need
> to define the semantics of the IR -- that is, how a program
> compiled from any given IR is allowed to do when executed (and *not*
> what it looks like when compiled) -- *first*, and then which
> transforms are allowed/not allowed will fall out of that. We can't
> start by listing disallowed transforms, because then when someone
> comes along and writes a new optimization, it might be technically
> "allowed" even though it breaks something in practice.

I think you misunderstand me if you think I was listing disallowed
transforms. My question was if it is possible to have multi-threaded
semantics in the source language, but in the compiler have a
single-threaded view, where some properties of the CFG would determine
what is legal and not for some functions with a special flag.

But that's practically the same as listing allowed and disallowed transforms -- you're defining what the final IR is allowed to look like, not how it's allowed to execute. If at all possible, the former should be derived from the latter.
 
 I agree
it is more desirable to have the the semantics specified in the
IR. However, I am exploring this from a practical point of view, since
ballot() is very rare compared to all code that is being compiled for
all targets. These intrinsics would always have to be considered when
writing a pass. They seem to me harder to think about, and
test, for someone who is working on a single-threaded target, compared
to looking at a flag. If we allow ourselves to look at which
transforms that might violate these properties, we could would free up
the rest of the compiler (and developers) to not have to worry about
these things. Intrinsics would have to be maintained throughout the
entire compilation process in every pass.

> The only way to
> conclusively prove that transforms will never break code that's
> supposed to work (except for bugs in the transforms) is to define the
> semantics of the IR, and then to make sure that it refines the
> semantics of the source language. This is why the LangRef almost never
> talks about allowed/disallowed transforms except as examples to
> explain some given semantics, and if you don't follow that rule, your
> patch will probably be rejected.

I'm trying to figure out if we are in the "almost never" territory
here or not.


I don't think so at all. In addition to being much, much harder to reason about and prove that the approach is sound, I think it'll be more intrusive.
 

> Now, to define a semantics for ballot(), we need to define what it's
> allowed to return for any given execution of any given program, and to
> do that, we need to define which threads must be active together when
> it's reached, which in turns means we need to define when control flow
> can re-converge somehow. Any proposal for how to handle ballot() must
> start here.

> > I'm not sure if using these rules will be easier or harder than dealing with
> > intrinsics. One problem is that the rules for single-threaded code might seem
> > arbitrary, and it would be hard to reason about them in a larger context.

> > What happens to new control flow created by transforms, and what will guide
> > the insertion of intrinsics in the new code? Code structurization is one example
> > were this could happen.

> Hopefully, it's clear from the above how this all falls out. The
> frontend for e.g. GLSL or SPIR-V would have to insert the merge
> intrinsics to preserve the semantics of the source language. Any
> transforms must refine the semantics of the IR, although I can't think
> of a scenario where that would involve emitting any new merge
> intrinsics. Usually, structurized IR's have their own semantics about
> when control flow converges, so a code structurizer should respect the
> original semantics. AMDGPU has its own code structurizer that runs
> very late in the pipeline (although there are plans to remove it), so
> we might have to change that to make it respect the merge intrinsic
> intrinsics, and then we'll correctly implement them "for free".

Any transform that re-arranges control flow would potentially have to
know about the properties of ballot(), and the rules with respect to
the CFG (and maybe consider the target) to know where to insert the
intrinsics.

But the same is true for basically any approach to handling this. In fact, adding the merge intrinsics makes this much easier. Beyond the usual problems with hoisting ballots, which passes currently avoid via the current convergent + speculatable attributes, we'd only have to add awareness to passes that they can't duplicate/combine merge intrinsics or move them past convergent intrinsics, which is a local property and hence easily checkable. In example I explained, without some kind of merge intrinsic, tail duplication has to look at the entire loop to know whether the transform is legal. Of course, you could have some kind of "no convergent calls" flag on a function, but that doesn't eliminate the nastyness when there are convergent calls. 
 I had the impression that the control flow convergence was
in part specified by what the target architecture can handle.

This is true, although hopefully we can agree on something that everyone can implement.
 
 One of
the more complicated cases would be linearization where the control
flow is completely rewritten, and is encoded in a variable that says
which basic block is the next one to execute.

It's certainly interesting to think about how to maintain correctness in the face of ballots() with such a pass, but a) it's certainly no harder with merge intrinsics than merges being implicit and b) I doubt that's useful for anything you'd want to do with a GPU.
 
Another case is DCE,
where a ballot() could be eliminated, and it would potentially have to
remove a number of intrinsics to enable later optimizations (unless it
would affect performance?), so the intrinsics will require some
non-local updates.


Removing merge intrinsics if it's profitable and allowed is a separate optimization, that can be done relatively quickly in a single pass without impacting the performance of other optimizations. It's requiring expensive non-local checks in optimizations which modify control flow that's a problem.
 

> > > Would they only be present if ballot and similar functions are used, or do they
> > > have to be present everywhere?
> >
> > They'd only have to be present when ballot or other convergent
> > functions are called, since otherwise it doesn't matter when control
> flow re-converges. However, we may want to keep them around for
> performance reasons (usually destroying convergence points is bad for
> performance).

So we might have them without a ballot(), which would seem to make it
more difficult for structurizers or other transforms to maintain the
semantics and insert intrinsics.


It's not any more difficult code-wise, since the case where there is a ballot needs to be handled anyways. And while it might take longer to process stuff when this hypothetical pass encounters a merge intrinsic (I can't think of a real-world case where it would matter), it should result in better code generated.
 

> > How are uniform branches handled? Do they affect the convergence model?
> >
> We may be able to remove convergence points if branches are uniform.
> In Nicolai's proposal, I believe we'd want to remove a merge intrinsic
> when all the branches that it post-dominates that aren't also
> post-dominated by some other merge intrinsic are uniform.

I couldn't quite understand the last sentence, but I assume the
conditions would prevent removing convergence points that help
performance. Post-domination might not be adequate if there are loops
btw.


It should be -- Nicolai's proposal is that a merge intrinsic merges all the divergences caused by branches post dominated by the intrinsic, so if all the divergent branches are merged by some other intrinsic earlier, then there's no divergence and the merge intrinsic is a no-op.
 

- Jan



    On Wednesday, January 30, 2019, 6:29:52 AM EST, Connor Abbott <cwabbott0 at gmail.com> wrote:  
 
 On Mon, Jan 28, 2019 at 9:09 PM Jan Sjodin <jan_sjodin at yahoo.com> wrote:
>
> > for (int i = 0; i < 2; i++) {
> >  foo = ballot(true); // ballot 1
> >
> >    if (threadID /* ID of the thread within a wavefront/warp */ % 2 == 0) continue;
> >
> >    bar = ballot(true); // ballot 2
> > }
> >
> > versus:
> >
> > int i = 0;
> > while (true) {
> >    do {
> >        if (i == 2) break;
> >        foo = ballot(true); // ballot 1
> >        i++;
> >    } while (threadID % 2 == 0);
> >
> >    if (i == 2) break;
> >    bar = ballot(true); // ballot 2
> >    i++;
> > }
>
> I think you can remove the second "i++", otherwise we can increment "i" twice
> if threadID % 2 != 0, but I see the issue. Using the equivalence classes would
> prevent this transform, since we have re-arranged the control flow in that way,

What do you mean? Note that the transforms that lead from the first
example to the second are actually desirable if there aren't any
ballots or other convergent operations, so banning them entirely is a
no-go. Otherwise, note that the ballot here can be nested arbitrarily
deep, which means that jump forwarding/tail duplication has to be
aware of the entire loop unless we have some kind of merge intrinsic.

Also, I should say that while interesting, control equivalence classes
can be part of the *implementation*, but not the *specification*. We
need to define the semantics of the IR -- that is, how a program
compiled from any given IR is allowed to do when executed (and *not*
what it looks like when compiled) -- *first*, and then which
transforms are allowed/not allowed will fall out of that. We can't
start by listing disallowed transforms, because then when someone
comes along and writes a new optimization, it might be technically
"allowed" even though it breaks something in practice. The only way to
conclusively prove that transforms will never break code that's
supposed to work (except for bugs in the transforms) is to define the
semantics of the IR, and then to make sure that it refines the
semantics of the source language. This is why the LangRef almost never
talks about allowed/disallowed transforms except as examples to
explain some given semantics, and if you don't follow that rule, your
patch will probably be rejected.

Now, to define a semantics for ballot(), we need to define what it's
allowed to return for any given execution of any given program, and to
do that, we need to define which threads must be active together when
it's reached, which in turns means we need to define when control flow
can re-converge somehow. Any proposal for how to handle ballot() must
start here.

> I'm not sure if using these rules will be easier or harder than dealing with
> intrinsics. One problem is that the rules for single-threaded code might seem
> arbitrary, and it would be hard to reason about them in a larger context.
>
> > Nicolai's proposal solves this by having the frontend emit a merge intrinsic
> > before the i++ is emitted. This prevents the jump forwarding from occurring.
>
> I was thinking about getting through the single-thread view and the issues with
> that first, then I will think more about the multi-thread and explicit convergence.
>
> If we are done with the single-thread stuff for now these are the question that I
> have been thinking about with the multi-threaded view:
>
> What happens to new control flow created by transforms, and what will guide
> the insertion of intrinsics in the new code? Code structurization is one example
> were this could happen.

Hopefully, it's clear from the above how this all falls out. The
frontend for e.g. GLSL or SPIR-V would have to insert the merge
intrinsics to preserve the semantics of the source language. Any
transforms must refine the semantics of the IR, although I can't think
of a scenario where that would involve emitting any new merge
intrinsics. Usually, structurized IR's have their own semantics about
when control flow converges, so a code structurizer should respect the
original semantics. AMDGPU has its own code structurizer that runs
very late in the pipeline (although there are plans to remove it), so
we might have to change that to make it respect the merge intrinsic
intrinsics, and then we'll correctly implement them "for free".

>
> Would they only be present if ballot and similar functions are used, or do they
> have to be present everywhere?

They'd only have to be present when ballot or other convergent
functions are called, since otherwise it doesn't matter when control
flow re-converges. However, we may want to keep them around for
performance reasons (usually destroying convergence points is bad for
performance).

>
> How are uniform branches handled? Do they affect the convergence model?

We may be able to remove convergence points if branches are uniform.
In Nicolai's proposal, I believe we'd want to remove a merge intrinsic
when all the branches that it post-dominates that aren't also
post-dominated by some other merge intrinsic are uniform.


>
>
> - Jan
>
>
> On Monday, January 28, 2019, 11:16:36 AM EST, Connor Abbott <cwabbott0 at gmail.com> wrote:
>
>
>
> On Fri, Jan 25, 2019 at 3:05 AM Jan Sjodin <jan_sjodin at yahoo.com> wrote:
> >
> > > for (...) {
> > >    ballot();
> > >    if (... /* non-uniform */) continue;
> > > }
> > >
> > > into
> > >
> > > for (...) {
> > >    do {
> > >        ballot();
> > >    } while (... /* non-uniform */);
> > > }
> >
> > I'm not sure if I follow this example, could you and explain a bit more?
> > It looks to me that the condition in the "if" must be false (if the
> > same condition is used in the while), or we would
> > call ballot the wrong number of times.
>
> Yes, the idea is that the same condition is used in the if and the do-while. I think I messed up the example a little... in the second snippet, we're supposed to break out of the inner loop if the outer loop's exit condition is true. Here's a more concrete example:
>
> for (int i = 0; i < 2; i++) {
>    foo = ballot(true); // ballot 1
>
>    if (threadID /* ID of the thread within a wavefront/warp */ % 2 == 0) continue;
>
>    bar = ballot(true); // ballot 2
> }
>
> versus:
>
> int i = 0;
> while (true) {
>    do {
>        if (i == 2) break;
>        foo = ballot(true); // ballot 1
>        i++;
>    } while (threadID % 2 == 0);
>
>    if (i == 2) break;
>    bar = ballot(true); // ballot 2
>    i++;
> }
>
> From a single-threaded perspective, these two snippets are identical, even if ballot() writes to arbitrary memory. The first can easily get transformed to something like the second when LLVM decides to duplicate the final i++ through jump forwarding, and then re-interprets the loop as two nested loops and splits the loop header in two. This is what currently happens with DOOM when we try to enable subgroup operations with it. Let's say there are two threads in a wavefront. Then the execution trace mandated by SPIR-V for the first looks like:
>
> thread 0        | thread 1
> ballot 1 = 0b11 | ballot 1 = 0b11
> skipped        | ballot 2 = 0b10
> ballot 1 = 0b11 | ballot 1 = 0b11
> skipped        | ballot 2 = 0b10
>
> Now, contrast this with the execution trace that programmers would expect for the second example:
>
> thread 0        | thread 1
> ballot 1 = 0b11 | ballot 1 = 0b11
> ballot 1 = 0b01 | skipped
> skipped        | ballot 2 = 0b10
> skipped        | ballot 1 = 0b10
> skipped        | ballot 2 = 0b10
>
> Nicolai's proposal solves this by having the frontend emit a merge intrinsic before the i++ is emitted. This prevents the jump forwarding from occurring.
>
>
> >
> > About the CSE, when would that be legal? I can imagine with uniform
> > branches that it could work, but would like to see an example to
> > fully understand this.
> >
> > I agree that it would be more conservative than if we model the threading,
> > but I'm not sure about the cost/benefit. I am mostly curious if it is
> > possible to have a single-thread view or not. Then we would have to
> > see if it is adequate.
> >
> > - Jan
> >
> > On Thursday, January 24, 2019, 10:31:47 AM EST, Connor Abbott <cwabbott0 at gmail.com> wrote:
> >
> >
> > I don't see how this would fix the continue vs. nested loop problem I
> > explained earlier. That is, how would this prevent turning:
> >
> > for (...) {
> >    ballot();
> >    if (... /* non-uniform */) continue;
> > }
> >
> > into
> >
> > for (...) {
> >    do {
> >        ballot();
> >    } while (... /* non-uniform */);
> > }
> >
> > and vice versa? Note that there's no duplication going on here, and
> > the single-threaded flow of control is exactly the same.
> >
> > Another reason this isn't so great is that it prevents e.g. CSE on
> > ballots that actually should be combined, since you're modelling it as
> > a write. It seems like the optimizer is going to need some special
> > knowledge of convergent things that fake memory constraints can't give
> > us.
> >
> > On Thu, Jan 24, 2019 at 4:06 PM Jan Sjodin <jan_sjodin at yahoo.com> wrote:
> > >
> > >
> > > I was looking into ballot() and how if it is possible to keep a single-threaded
> > > view of the code, but add some extra conditions that must hold after the
> > > transformation. I had the initial idea that each call to ballot() in a
> > > single-threaded program can be seen as a partial write to a memory
> > > location, and each location memory location is unique for every call site,
> > > plus there some externally observable side effect. We can abstract this
> > > away by tagging the calls, e.g. by using aliases.
> > >
> > > For example:
> > >
> > > if (...) {
> > >      foo1 = ballot();
> > > } else {
> > >      foo2 = ballot();
> > > }
> > >
> > > simply becomes:
> > >
> > > if (...) {
> > >      foo1 = ballot_1();
> > > } else {
> > >      foo2 = ballot_2();
> > > }
> > >
> > >
> > > and
> > >
> > > if (...) {
> > > } else {
> > > }
> > > ballot();
> > >
> > > becomes
> > >
> > > if (...) {
> > > } else {
> > > }
> > > ballot_1();
> > >
> > > In the first case it would prevent combining the two calls into one
> > > after the if. In the second example there is generally nothing that
> > > says it could not be transformed into the first example with two
> > > calls to ballot_1(), which should not be allowed.
> > >
> > > Another form of duplication that we must allow are loop transforms,
> > > like unrolling or peeling. These might seem similar to the example
> > > above, since we are cloning code and with conditions etc. But
> > > they are different since they calls are in different loop iterations.
> > >
> > > The condition that needs to be met is that:
> > >
> > > There must be a single path between all cloned ballot_n() functions.
> > >
> > > The reason for this condition is that if we clone the same call, then
> > > the copies must be mutually exclusive, but if they are cloned from
> > > a loop, there must be a path, or we would skip iterations.
> > >
> > > If we want to be more sophisticated we can add:
> > >
> > > If there is no such path, the calls must be separated by uniform branches.
> > >
> > > After the transform everything should be re-tagged, since we already
> > > checked the calls and we don't want to check them again. Also, not all
> > > transforms need (or should) have the tagging checked. One example is
> > > inlining, where multiple copies are created, but they are clearly different
> > > calls. The tagging can be done temporarily for a single pass, and then
> > > eliminated. This could be a good tool for debugging as well, since it can
> > > detect if a transform is suspect.
> > >
> > > The code would of course have to make sense as far as control flow. If
> > > we have:
> > >
> > > for(;;) {
> > >    if(...) {
> > >      ballot();
> > >      break;
> > >    }
> > > }
> > >
> > > This would have to be translated to:
> > >
> > > for(;;) {
> > >    if(...) {
> > >      ballot();
> > >      if(UnknownConstant) {  // Not a uniform condition, but later on translated to "true"
> > >        break;
> > >      }
> > > }
> > >
> > > However, I think that this is the way the code is generated today anyway.
> > > There would have to be some attribute that indicate that these calls (or functions)
> > > contain ballot or other cross-lane operations so they could be tagged and
> > > checked. The attribute would be used by the passes to know that the special
> > > conditions exist for those calls.
> > >
> > > As far as what it means to have a path, it could be complicated.
> > > For example:
> > >
> > > x = ...
> > > ballot_1();
> > >
> > > could be transformed to:
> > >
> > > if (x < 4711) {
> > >  ballot_1();
> > >
> > > if(x >= 4711) {
> > >  ballot_1();
> > > }
> > >
> > > So a simple path check would say there is a path, and the transform is legal,
> > > but if we examine the conditions, there is no path, and the transform should not be legal.
> > > It could be made even more obscure of course, but I don't see any optimizations really
> > > doing this kind of thing,
> > >
> > > - Jan
> > >
> > > On Saturday, December 29, 2018, 11:32:25 AM EST, Nicolai Hähnle via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> > >
> > >
> > > On 20.12.18 18:03, Connor Abbott wrote:
> > > >    We already have the notion of "convergent" functions like
> > > >    syncthreads(), to which we cannot add control-flow dependencies.
> > > >    That is, it's legal to hoist syncthreads out of an "if", but it's
> > > >    not legal to sink it into an "if".  It's not clear to me why we
> > > >    can't have "anticonvergent" (terrible name) functions which cannot
> > > >    have control-flow dependencies removed from them?  ballot() would be
> > > >    both convergent and anticonvergent.
> > > >
> > > >    Would that solve your problem?
> > > >
> > > >
> > > > I think it's important to note that we already have such an attribute,
> > > > although with the opposite sense - it's impossible to remove control
> > > > flow dependencies from a call unless you mark it as "speculatable".
> > >
> > > This isn't actually true. If both sides of an if/else have the same
> > > non-speculative function call, it can still be moved out of control flow.
> > >
> > > That's because doing so doesn't change anything at all from a
> > > single-threaded perspective. Hence why I think we should model the
> > > communication between threads honestly.
> > >
> > >
> > > > However, this doesn't prevent
> > > >
> > > > if (...) {
> > > > } else {
> > > > }
> > > > foo = ballot();
> > > >
> > > > from being turned into
> > > >
> > > > if (...) {
> > > >      foo1 = ballot();
> > > > } else {
> > > >      foo2 = ballot();
> > > > }
> > > > foo = phi(foo1, foo2)
> > > >
> > > > and vice versa. We have a "noduplicate" attribute which prevents
> > > > transforming the first into the second, but not the other way around. Of
> > > > course we could keep going this way and add a "nocombine" attribute to
> > > > complement noduplicate. But even then, there are even still problematic
> > > > transforms. For example, take this program, which is simplified from a
> > > > real game that doesn't work with the AMDGPU backend:
> > > >
> > > > while (cond1 /* uniform */) {
> > > >      ballot();
> > > >      ...
> > > >      if (cond2 /* non-uniform */) continue;
> > > >      ...
> > > > }
> > > >
> > > > In SPIR-V, when using structured control flow, the semantics of this are
> > > > pretty clearly defined. In particular, there's a continue block after
> > > > the body of the loop where control flow re-converges, and the only back
> > > > edge is from the continue block, so the ballot is in uniform control
> > > > flow. But LLVM will get rid of the continue block since it's empty, and
> > > > re-analyze the loop as two nested loops, splitting the loop header in
> > > > two, producing a CFG which corresponds to this:
> > > >
> > > > while (cond1 /* uniform */) {
> > > >      do {
> > > >          ballot();
> > > >          ...
> > > >      } while (cond2 /* non-uniform */);
> > > >      ...
> > > > }
> > > >
> > > > Now, in an implementation where control flow re-converges at the
> > > > immediate post-dominator, this won't do the right thing anymore. In
> > > > order to handle it correctly, you'd effectively need to always flatten
> > > > nested loops, which will probably be really bad for performance if the
> > > > programmer actually wanted the second thing. It also makes it impossible
> > > > when translating a high-level language to LLVM to get the "natural"
> > > > behavior which game developers actually expect. This is exactly the sort
> > > > of "spooky action at a distance" which makes me think that everything
> > > > we've done so far is really insufficient, and we need to add an explicit
> > > > notion of control-flow divergence and reconvergence to the IR. We need a
> > > > way to say that control flow re-converges at the continue block, so that
> > > > LLVM won't eliminate it, and we can vectorize it correctly without
> > > > penalizing cases where it's better for control flow not to re-converge.
> > >
> > > Well said!
> > >
> > > Cheers,
> > >
> > > Nicolai
> > > --
> > > Lerne, wie die Welt wirklich ist,
> > > Aber vergiss niemals, wie sie sein sollte.
> > > _______________________________________________
> > > LLVM Developers mailing list
> > > llvm-dev at lists.llvm.org
> > > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev  
  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20190131/1fb7314d/attachment-0001.html>


More information about the llvm-dev mailing list