[cfe-dev] RFC: clacc: translating OpenACC to OpenMP in clang

[Hal Finkel via cfe-dev]

On 12/05/2017 01:06 PM, Joel E. Denny wrote:
> Hi,
>
> We are working on a new project, clacc, that extends clang with 
> OpenACC support.  Clacc's approach is to translate OpenACC (a 
> descriptive language) to OpenMP (a prescriptive language) and thus to 
> build on clang's existing OpenMP support.  While we plan to develop 
> clacc to support our own research, an important goal is to contribute 
> clacc as a production-quality component of upstream clang.

Great.

>
> We have begun implementing an early prototype of clacc. Before we get 
> too far into the implementation, we would like to get feedback from 
> the LLVM community to help ensure our design would ultimately be 
> acceptable for contribution.  For that purpose, below is an analysis 
> of several high-level design alternatives we have considered and their 
> various features.  We welcome any feedback.
>
> Thanks.
>
> Joel E. Denny
> Future Technologies Group
> Oak Ridge National Laboratory
>
>
> Design Alternatives
> -------------------
>
> We have considered three design alternatives for the clacc compiler:
>
> 1. acc src  --parser-->   omp AST  --codegen-->  LLVM IR + omp rt calls

I don't think that we want this option because, if nothing else, it will 
preclude builting source-level tooling for OpenACC.

> 2. acc src --parser-->  acc AST --codegen-->  LLVM IR + omp rt calls
> 3. acc src  --parser-->  acc AST  --ttx-->  omp AST  --codegen-->  
> LLVM IR + omp rt calls

My recommendation: We should think about the very best way we could 
refactor the code to implement (2), and if that is too ugly (or 
otherwise significantly degrades maintainability of the OpenMP code), 
then we should choose (3).

>
> In the above diagram:
>
> * acc src = C source code containing acc constructs.
> * acc AST = a clang AST in which acc constructs are represented by
>   nodes with acc node types.  Of course, such node types do not
>   already exist in clang's implementation.
> * omp AST = a clang AST in which acc constructs have been lowered
>   to omp constructs represented by nodes with omp node types.  Of
>   course, such node types do already exist in clang's
>   implementation.
> * parser = the existing clang parser and semantic analyzer,
>   extended to handle acc constructs.
> * codegen = the existing clang backend that translates a clang AST
>   to LLVM IR, extended if necessary (depending on which design is
>   chosen) to perform codegen from acc nodes.
> * ttx (tree transformer) = a new clang component that transforms
>   acc to omp in clang ASTs.
>
> Design Features
> ---------------
>
> There are several features to consider when choosing among the designs
> in the previous section:
>
> 1. acc AST as an artifact -- Because they create acc AST nodes,
>    designs 2 and 3 best facilitate the creation of additional acc
>    source-level tools (such as pretty printers, analyzers, lint-like
>    tools, and editor extensions).  Some of these tools, such as pretty
>    printing, would be available immediately or as minor extensions of
>    tools that already exist in clang's ecosystem.
>
> 2. omp AST/source as an artifact -- Because they create omp AST
>    nodes, designs 1 and 3 best facilitate the use of source-level
>    tools to help an application developer discover how clacc has
>    mapped his acc to omp, possibly in order to debug a mapping
>    specification he has supplied.  With design 2 instead, an
>    application developer has to examine low-level LLVM IR + omp rt
>    calls.  Moreover, with designs 1 and 3, permanently migrating an
>    application's acc source to omp source can be automated.
>
> 3. omp AST for mapping implementation -- Designs 1 and 3 might
>    also make it easier for the compiler developer to reason about and
>    implement mappings from acc to omp.  That is, because acc and omp
>    syntax is so similar, implementing the translation at the level of
>    a syntactic representation is probably easier than translating to
>    LLVM IR.
>
> 4. omp AST for codegen -- Designs 1 and 3 simplify the
>    compiler implementation by enabling reuse of clang's existing omp
>    support for codegen.  In contrast, design 2 requires at least some
>    extensions to clang codegen to support acc nodes.
>
> 5. Full acc AST for mapping -- Designs 2 and 3 potentially
>    enable the compiler to analyze the entire source (as opposed to
>    just the acc construct currently being parsed) while choosing the
>    mapping to omp.  It is not clear if this feature will prove useful,
>    but it might enable more optimizations and compiler research
>    opportunities.

We'll end up doing this, but most of this falls within the scope of the 
"parallel IR" designs that many of us are working on. Doing this kind of 
analysis in the frontend is hard (because it essentially requires it to 
do inlining, simplification, and analysis akin to what the optimizer 
itself does).

>
> 6. No acc node classes -- Design 1 simplifies the compiler
>    implementation by eliminating the need to implement many acc node
>    classes.  While we have so far found that implementing these
>    classes is mostly mechanical, it does take a non-trivial amount of
>    time.
>
> 7. No omp mapping -- Design 2 does not require acc to be mapped to
>    omp.  That is, it is conceivable that, for some acc constructs,
>    there will prove to be no omp syntax to capture the semantics we
>    wish to implement.

I'm fairly certain that not everything maps exactly. They'll be some 
things we need to deal with explicitly in CodeGen.

> It is also conceivable that we might one day
>    want to represent some acc constructs directly as extensions to
>    LLVM IR, where some acc analyses or optimizations might be more
>    feasible to implement.  This possibility dovetails with recent
>    discussions in the LLVM community about developing LLVM IR
>    extensions for various parallel programming models.

+1

>
> Because of features 4 and 6, design 1 is likely the fastest design to
> implement, at least at first while we focus on simple acc features and
> simple mappings to omp.  However, we have so far found no advantage
> that design 1 has but that design 3 does not have except for feature
> 6, which we see as the least important of the above features in the
> long term.
>
> The only advantage we have found that design 2 has but that design 3
> does not have is feature 7.  It should be possible to choose design 3
> as the default but, for certain acc constructs or scenarios where
> feature 7 proves important (if any), incorporate design 2.  In other
> words, if we decide not to map a particular acc construct to any omp
> construct, ttx would leave it alone, and we would extend codegen to
> handle it directly.

This makes sense to me, and I think is most likely to leave the CodeGen 
code easiest to maintain (and has good separation of concerns). 
Nevertheless, I think we should go through the mental refactoring 
exercise for (2) to decide on the value of (3).

Thanks again,
Hal

>
> Conclusions
> -----------
>
> For the above reasons, and because design 3 offers the cleanest
> separation of concerns, we have chosen design 3 with the possibility
> of incorporating design 2 where it proves useful.
>
> Because of the immutability of clang's AST, the design of our proposed
> ttx component requires careful consideration.  To shorten this initial
> email, we have omitted those details for now, but we will be happy to
> include them as the discussion progresses.

-- 
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/cfe-dev/attachments/20171206/e49bf7cc/attachment.html>