It sounds like LLVM IR is being considered for optimizations in OpenMP
constructs. There seems to be plans regarding improvement of LLVM IR
Framework for providing things required for OpenMP / flang(?)
LLVM has the OpenMPOpt pass now [0] in which we can put OpenMP specific
transformations. For now it is simple but we have some more downstream
patches, e.g., parallel region expansion [Section 5, 1]. Other
optimizations [Section 3 & 4, 1], will be performed by the Attributor
(see [4] after [2,3]) after one missing piece (basically [5] with some
more plumming) was put in place, see [2,3] for details on the idea.
Please feel free to ask questions on any of this.
[0] ⚙ D69930 [OpenMP] Introduce the OpenMPOpt transformation pass
[1] http://compilers.cs.uni-saarland.de/people/doerfert/par_opt18.pdf
[2] https://www.youtube.com/watch?v=zfiHaPaoQPc
[3] http://compilers.cs.uni-saarland.de/people/doerfert/par_opt_lcpc18.pdf
[4] https://youtu.be/CzWkc_JcfS0
[5] ⚙ D71505 [Utils] Provide a callback encapsulation utility for call sites
It might also worth looking into [6,7] mentioned below.
Are there any design considerations which contain pros and cons about using
the MLIR vs LLVM IR for various OpenMP related optimizations/
transformations?
The biggest pro for LLVM-IR is that it works for C/C++ right now. In
addition, as I mentioned before, LLVM-IR has mature analysis and
transformation passes for real world programs and support things like
LTO out of the box.
The latest RFC [ (3) in my original post ] mentions that:
> So there exist some questions regarding where the optimisations should be
carried out.
Could you please provide more details on this?
I would like to quote Chris here:
“if you ignore the engineering expense, it would clearly make sense to
reimplement the mid-level LLVM optimizers on top of MLIR and replace
include/llvm/IR with a dialect definition in MLIR instead.“ --
[llvm-dev] [RFC] Writing loop transformations on the right representation is more productive
*Rest of the comment are inlined.*
> Hi Vinay,
>
> Thanks for taking an interest and the detailed discussion.
>
> To start by picking a few paragraph from your email to clarify a couple
> of things that lead to the current design or that might otherwise need
> clarification. We can talk about other points later as well.
>
> [
> Site notes:
> 1) I'm not an MLIR person.
> 2) It seems unfortnuate that we do not have a mlir-dev list.
> ]
>
>
> > 1. With the current design, the number of transformations / optimizations
> > that one can write on OpenMP constructs would become limited as there can
> > be any custom loop structure with custom operations / types inside it.
>
> OpenMP, as an input language, does not make many assumptions about the
> code inside of constructs*.
This isn’t entirely correct because the current OpenMP API specification (
1 Introduction) assumes that the code
inside the constructs belong to C, C++ and Fortran programs.
(FWIW, my next sentence specified that I talk about the base language
but anyway.)
While technically true, I will recommend not to make that assumption. We
do already allow non-base language constructs, e.g., CUDA intrinsics in
target regions, and that will not go away because it is required to
maximize performance.
> So, inside a parallel can be almost anything
> the base language has to offer, both lexically and dynamically.
>
I am mostly concerned with the MLIR side of things for OpenMP
representation.
MLIR can not only support operations for General Purpose languages like
C,C++, Fortran, etc but also various Domain Specific Language
representations as dialects (Example, ML, etc.). Note that there is also
SPIR V dialect which is again meant for “Parallel Compute”.
It becomes important to define the scope of the dialects / operations /
types supported inside OpenMP operations in MLIR.
Arguably, the OpenMP dialect in MLIR should match the OpenMP directives
and clauses as defined by the standard. Anything else is "not OpenMP".
> Assuming otherwise is not going to work. Analyzing a "generic" OpenMP
> representation in order to determine if can be represented as a more
> restricted "op" seems at least plausible. You will run into various
> issue, some mentioned explicitly below.
Isn’t it the other way around? For example, it doesn’t make much sense to
wrap OpenMP operations for SPIR-V operations / types.
I maybe misunderstanding but I thought you want to use something like
the GPU / Affine dialect to represent an OpenMP target region / loop.
That is plausible if you analyze the target region / loop and verify it
fits into the more generic dialect semantics.
I think it is important to specify (in the design) which existing MLIR
dialects are supported in this effort and the various lowerings /
transformations / optimizations which are planned for them.
That I cannot really help you with. TBH, I don't even know what
transformations people plan to do on OpenMP MLIR (and why).
> For starters, you still have to
> generate proper OpenMP runtime calls, e.g., from your GPU dialect, even
> if it is "just" to make sure the OMPD/OMPT interfaces expose useful
> information.
>
>
You can have a well-defined call-like mlir::Operation which calls the GPU
kernel. Perform all cross-device transformations in an easier way.
Then, this operation can be lowered to OpenMP runtime calls during LLVM
dialect conversion.
You missed my point I made in the other email. An OpenMP target region
is statically not a GPU offload so you should not model it as such "for
some time".
I think this is much better than directly having calls
to the OpenMP runtime library based on a kernel name mentioned in
llvm::GlobalVariable.
(Current) implementation is not semantics. There is no reason not to
change the way we lower OpenMP, e.g., by getting rid of the global
variables. They are present for a reason but not intrinsically required.
See the TRegions for example [6,7], they totally change the GPU lowering,
making it device agnostic and easy to analyze and optimize in the middle
end. Arguing the current encoding of OpenMP in LLVM-IR is problematic is
the same as arguing MLIR's LLVM dialect doesn't support atomic_rmw, it
might be true but its changeable.
[6] The TRegion Interface and Compiler Optimizations for OpenMP Target Regions | SpringerLink
[7] http://parallel.auckland.ac.nz/iwomp2019/slides_TRegion.pdf
> * I preclude the `omp loop` construct here as it is not even implemented
> anywhere as far as I know.
>
>
> > 2. It would also be easier to transform the Loop nests containing OpenMP
> > constructs if the body of the OpenMP operations is well defined (i.e.,
> does
> > not accept arbitrary loop structures). Having nested redundant
> "parallel" ,
> > "target" and "do" regions seems unnecessary.
>
> As mentioned above, you cannot start with the assumption OpenMP input is
> structured this this way. You have to analyze it first. This is the same
> reason we cannot simply transform C/C++ `for loops` into `affine.for`
> without proper analysis of the loop body.
>
> Now, more concrete. Nested parallel and target regions are not
> necessarily redundant, nor can/should we require the user not to have
> them. Nested parallelism can easily make sense, depending on the problem
> decomposition. Nested target will make a lot of sense with reverse
> offload, which is already in the standard, and it also should be allowed
> for the sake of a modular (user) code base.
>
Just to be clear, having all three of “target”, “parallel” and “do” doesn’t
represent “Nested parallelism” at all in the proposed design! ( 2(d) ).
omp.target {
omp.parallel {
omp.do {
…...
}
}
}
Above invokes a call to the tgt_target() for the code inside omp.do as
mentioned in the proposal.
I do not follow. Just to make sure, the above should be roughly
equivalent to the code below, correct? There is no "nested"
parallelism, sure, but I thought you were talking about the case where
there is, e.g. add another `#pragma omp parallel` inside the one that
already is there. That is nested parallelism which can happen and make
total sense for the application.
#pragma omp target
{
#pragma omp parallel
{
#pragma omp for
for (...)
{
...
}
}
}
>
> > 3. There would also be new sets of loop structures in new dialects when
> > C/C++ is compiled to MLIR. It would complicate the number of possible
> > combinations inside the OpenMP region.
>
> Is anyone working on this? If so, what is the timeline? I personally was
> not expecting Clang to switch over to MLIR any time soon but I am happy
> if someone wants to correct me on this. I mention this only because it
> interacts with the arguments I will make below.
>
>
> > E. Lowering of target constructs mentioned in ( 2(d) ) specifies direct
> > lowering to LLVM IR ignoring all the advantages that MLIR provides. Being
> > able to compile the code for heterogeneous hardware is one of the biggest
> > advantages that MLIR brings to the table. That is being completely missed
> > here. This also requires solving the problem of handling target
> information
> > in MLIR. But that is a problem which needs to be solved anyway. Using GPU
> > dialect also gives us an opportunity to represent offloading semantics in
> > MLIR.
>
> I'm unsure what the problem with "handling target information in MLIR" is
> but
> whatever design we end up with, we need to know about the target
> (triple) in all stages of the pipeline, even if it is just to pass it
> down.
>
>
> > Given the ability to represent multiple ModuleOps and the existence of
> GPU
> > dialect, couldn't higher level optimizations on offloaded code be done at
> > MLIR level?. The proposed design would lead us to the same problems that
> we
> > are currently facing in LLVM IR.
> >
> > Also, OpenMP codegen will automatically benefit from the GPU dialect
> based
> > optimizations. For example, it would be way easier to hoist a memory
> > reference out of GPU kernel in MLIR than in LLVM IR.
>
> While I agree with the premise that you can potentially reuse MLIR
> transformations, it might not be as simple in practice.
>
> As mentioned above, you cannot assume much about OpenMP codes, almost
> nothing for a lot of application codes I have seen. Some examples:
>
> If you have a function call, or any synchronization event for that
> matter, located between two otherwise adjacent target regions (see
> below), you cannot assume the two target regions will be offloaded to
> the same device.
> ```
> #omp target
> {}
> foo();
> #omp target
> {}
> ```
>
These kinds of optimizations are much easier to write in MLIR:
LLVM IR for the above code would contain a series of instructions of OpenMP
runtime call setup and foo() in the middle followed by another set of
OpenMP runtime related instructions. The body of the two target constructs
would be in two different outlined functions (if not modules).
It takes quite a bit of code to do analysis / transformation to write any
optimization on the generated LLVM IR.
You are right about the module's being a problem. As I mentioned in my
last email, we are working on that by not having them in different ones
during the optimization pipeline. If we make the `target` `parallel`
instead we can simulate that right now. The bodies are in different
functions, sure, but does it matter? Let's walk through parallel region
expansion (see above [Section 5, 1]) so you can judge for yourself:
#omp parallel
{ body0 }
some_code
#omp parallel
{ body1 }
will become
__kmpc_fork_call(..., @body0_fn, ...)
some_code
__kmpc_fork_call(..., @body1_fn, ...)
in IR. Simplified, there are 3 cases here:
1) some_code is harmless, meaning all of it can be executed redundantly.
2) parts of some some_code need to be guarded to be sequential but
they can be executed in a parallel region otherwise, e.g., the code
will not observe the difference through runtime calls.
3) parts of some some_code cannot be executed in a parallel region as
they might observe the difference through runtime calls.
First note that you need to do the classification regardless of your
encoding (=IR). In case of 3) we are done and nothing is happening.
Let's consider case 2) as 1) is basically a special case of it. As shown
in the paper [1], you need to broadcast values created by some_code
across all threads and synchronize appropriately to preserve semantic.
Other than that, the transformation is straight forward:
A) Create a function "@body01_fn" that is basically the outlined region
in which code is then guarded and __kmpc_fork_call are replaced by
direct calls. It looks like this:
call @body0_fn(...)
#omp master
some_code
#omp barrier
call @body1_fn(...)
B) Replace the region you put in the new function with a
__kmpc_fork_call to it:
__kmpc_fork_call(..., @body01_fn, ...)
C) Done.
If you are interested in the implementation I'll add you as a reviewer
once I put it on Phab. I'm in the process of cleaning up my stand alone
pass and moving it into the OpenMPOpt pass instead.
vs.
MLIR provides a way to represent the operations closer to the source. It is
as simple as checking the next operation(s) in the mlir::Block. OpenMP
target operation contains an inlined region which can easily be fused/
split / or any other valid transformation for that matter.
Note that you can also perform various Control Structure Analysis /
Transformations much easier in MLIR. For example, you can decide to execute
foo() based on certain conditions, and you can merge the two target regions
in the else path.
At the end, it's an encoding difference. Sure, the handling might be
easier in certain situations but all the validity checks, hence code
analyses, are still required. The actual "rewrite" is usually not the
hard part.
> Similarly, you cannot assume a `omp parallel` is allowed to be executed
> with more than a single thread, or that a `omp [parallel] for` does not
> have loop carried data-dependences, ...
>
With multi-dimensional index support for arrays, wouldn’t it be better to
do the data dependence analysis in MLIR?
Yes, probably.
LLVM IR has linearized subscripts for multi-dimensional arrays.
llvm::DependenceAnalysis tries to “guess” the indices based on different
patterns in SCEV. It takes an intrinsic
<The LLVM Compiler Infrastructure Project; or metadata or
some other mechanism of communication from the front end (not the built-in
set of instructions) to solve this problem.
Not disagreeing with you on this one 
The only caveat is that we still live in a world in which C/C++ is a
thing.
> Data-sharing attributes are also something that has to be treated
> carefully:
> ```
> x = 5;
> #omp task
> x = 3;
> print(x);
> ```
> Should print 5, not 3.
>
You can have “x” as a locally defined variable inside the “task” contained
region in MLIR OR custom data-sharing attributes in OpenMP dialect.
I'm not saying it is impossible or even hard, but maybe not as straight
forward as one might think. Your encoding is for example very reasonable.
In the example below you need to print 3, not 5, e.g., constant prop on
the outer level should not happen.
x = 5;
#omp task shared(x)
{
x = 3;
some_form_of_sync();
...
}
some_form_of_sync();
print(x);
> I hope I convinced you that OpenMP is not trivially mappable to existing
> dialects without proper analysis. If not, please let me know why you
> expect it to be.
>
I do not see much reason why the issues you mentioned can’t trivially be
mapped to the MLIR infrastructure. There is an easy way to define custom
operations / types / attributes in OpenMP dialect and perform optimizations
based on the *IR that is created especially for OpenMP*. The analysis /
transformations required can be easily written on the custom operations
defined rather than having a lowered form in the LLVM IR.
You can totally define your OpenMP dialect and map it to that. Mapping
to other dialects is the problematic part. As I mentioned, `omp
parallel` does not mean "parallel" or "dependence-free".
Since you mention it, why do you think it is conceptually or practically
harder to write an analysis/transformations on IR? I mean, you teach
your analysis what the op "omp.parallel" means, right? Why not teach an
(interprocedural) analysis what __kmpc_fork_call() does (see [2,3] above)?
FWIW, there are LLVM analyses and transformations that already know
about the transitive call made by __kmpc_fork_call and pthread_create
(see [4] above). It is done in a way that you can easily annotate your
own C/C++ or IR to make use of it, e.g., for your own transitive
callbacks:
Attributes in Clang — Clang 18.0.0git documentation
LLVM Language Reference Manual — LLVM 18.0.0git documentation
The various dialects / transformations in MLIR are in development / early
phase (Example, GPU dialect) waiting to be improved with use cases such as
this!
Great! I am eagerly looking forward to this.
Cheers,
Johannes
