Johannes,
I definitely agree that LLVM needs better support for optimizing parallel programs. Comments inline:
Hi Vikram,
Thanks for the quick reply. I put some comments below.
Johannes,
I definitely agree that LLVM needs better support for optimizing parallel programs. Comments inline:
> From: Johannes Doerfert <jdoerfert@anl.gov>
> Date: Wednesday, June 6, 2018 at 10:52 AM
> To: LLVM-Dev <llvm-dev@lists.llvm.org>
> Cc: Hal Finkel <hfinkel@anl.gov>, TB Schardl <neboat@mit.edu>, "Adve, Vikram Sadanand" <vadve@illinois.edu>, "Tian, Xinmin" <xinmin.tian@intel.com>, "llvmpar@lists.cs.illinois.edu" <llvmpar@lists.cs.illinois.edu>
> Subject: [RFC] Abstract Parallel IR Optimizations
>
> This is an RFC to add analyses and transformation passes into LLVM to
> optimize programs based on an abstract notion of a parallel region.
>
> == this is _not_ a proposal to add a new encoding of parallelism ==Actually, I suspect that your three “abstract” interfaces below
(ParallelRegionInfo, ParallelCommunicationInfo, ParallelIR/Builder)
*are* a parallel IR definition, and what you call the
“implementations” essentially encapsulate *target-specific*
variations.
It is not an IR (encoding) in the sense that there are no new
instructions, intrinsics, or anything else that is materialized. The
interface provides a unified view of existing parallel representations
and it is as general as possible.
I am unsure what you mean by target-specific here but our
implementations are the actual parallel IR representations as we
discussed them several times on other occasions.
IOW, this looks like a perhaps-very-preliminary parallel IR, not some
generic abstract interface. I can only suspect this because I don’t
have any information about what info or operations these interfaces
provide.
I disagree. Please check out the operations and info for the interface
here
⚙ Changeset View
and for the builder here
⚙ Changeset View
so we talk about the same thing.
In fact, I think it *is* a good idea to have a proper parallel IR to
support the kinds of optimizations you’re describing. I just don’t
think you can define a universal parallel IR for all the parallel
languages and targets that LLVM may want to support. At a minimum, I
think we need to experiment more before committing to one (*if* one
proved enough).
I think you misunderstood our intention here. This patch is in support
of such experiments as it allows to apply the optimizations we have and
might come up with to different parallel IR representations. Not all
optimizations will be applicable to all representations but I don't see
a reason why this is a problem. I also don't think this is a universal
parallel IR of sorts but it is a way to extract critical information
necessary for optimizations out of parallel representations.
Instead of defining a specific parallel IR or interface, as you’re
trying to do, it would be much better to provide IR-independent hooks
in LLVM so that different parallel IRs can sit “above” the LLVM
representation and support these and other optimizations. Your IR (or
IR implementations, whatever you call them) could be layered in this
way.
I do not understand what you mean here. The "implementations" (as I
called them so recklessly) lift _some_ information from the actual IR
representation to the high level abstract parallel IR interface level
for consummation by the optimization passes. Different parallel IRs can
come with their own implementation and provide more or less information
to enable optimizations.
The earlier RFC that Hal Finkel and Xinmin Tian circulated were for
such a set of hooks. In fact, these hooks also work on regions, but
are much more limited and are *not* meant to support parallel passes
themselves: those are left to any IR built above them.As you know, we (Hal, Xinmin, TB Schardl, George Stelle, Hashim
Sharif, I, and a few others) are trying to refine that proposal to
provide examples of multiple IRs the hooks can support, and to make a
clearer argument for the soundness and correctness impact of the hooks
on LLVM passes. You’ve been invited to our conference calls and to
edit the Google Doc. It would be valuable if you could add examples
of how your IR information could be connected to LLVM via these hooks.
I'm still confused about the hooks.
More specific questions below.
> We currently perform poorly when it comes to optimizations for parallel
> codes. In fact, parallelizing your loops might actually prevent various
> optimizations that would have been applied otherwise. One solution to
> this problem is to teach the compiler about the semantics of the used
> parallel representation. While this sounds tedious at first, it turns
> out that we can perform key optimizations with reasonable implementation
> effort (and thereby also reasonable maintenance costs). However, we have
> various parallel representations that are already in use (KMPC,
> GOMP, CILK runtime, ...) or proposed (Tapir, IntelPIR, ...).
>
> Our proposal seeks to introduce parallelism specific optimizations for
> multiple representations while minimizing the implementation overhead.
> This is done through an abstract notion of a parallel region which hides
> the actual representation from the analysis and optimization passes.In general, for analysis and optimization passes to work with multiple
representations, you’d need to have a pretty rich abstract notion of a
parallel region. I suspect that your design is pretty
OpenMP-specific. Even within that context, it would have to capture
quite a rich set of parallel constructs. Can you describe what
exactly is your abstract notion of a parallel region, and how is it
different from a concrete representation?
Good question. The interface implementations will abstract away of the
representation details that the optimization should not worry about such
as:
- runtime library calls
- variable passing via structures
- intrinsic calls to encode information
- ...
If you look at our variable privatization pass (which was not included
in this commit) it does not need to know if variable capturing is done
through a runtime library call (OpenMP), if captured variables are
handed to a special intrinsic to mark them (IntelPIR), or if capturing
is just a def-use chain crossing the parallel region boundaries (Tapir).
The pass requires to know the mapping (variable in sequential code ->
variable in parallel code) as well as the start/end instruction of the
parallel region. Then it can test for conditions that would allow to
privatize the variable, pass it by-value instead of by-reference, etc.
Similarly, other passes do not need to know the details of the encoding.
You can probably see that for barrier elimination but also parallel
region expansion is actually only looking at the surrounding code to
determine if expansion (and region merging) might make sense.
> In the schemata below, our current five optimizations (described in
> detail here [0]) are shown on the left, the abstract parallel IR
> interface is is in the middle, and the representation specific
> implementations is on the right.
>
> Optimization (A)nalysis/(T)ransformation Impl.
> ---------------------------------------------------------------------------
> CodePlacementOpt \ /---> ParallelRegionInfo (A) ---------|-> KMPCImpl (A)
> RegionExpander -\ | | GOMPImpl (A)
> AttributeAnnotator -|-|---> ParallelCommunicationInfo (A) --/ ...
> BarrierElimination -/ |
> VariablePrivatization / \---> ParallelIR/Builder (T) -----------> KMPCImpl (T)I wasn’t able to understand parts of this figure. What info do
ParallelRegionInfo() and ParallelCommunicationInfo() provide? What
operations does ParallelIR/Builder provide?
So far not much. ParallelRegionInfo just collects and manages the
parallel regions. These know about their extend (start/end), provide a
way to visit the blocks and instructions and have some helper functions
to reason about (potential) barriers. Nothing is specific to OpenMP.
ParallelCommunicationInfo provides the mapping of variables outside the
parallel region to the ones inside and tells you about the way they are
used (e.g., the pointer is actually a read-only container). Finally, the
builder is the only way to modify the parallel regions (as we don't know
the actual representation in the optimization). Currently it can only
add/remove attributes but my local version can create parallel regions,
build barriers, etc.
For the current code check the two links from above.
> In our setting, a parallel region can be an outlined function called
> through a runtime library but also a fork-join/attach-reattach region
> embedded in an otherwise sequential code. The new optimizations will
> provide parallelism specific optimizations to all of them (if
> applicable). There are various reasons why we believe this is a
> worthwhile effort that belongs into the LLVM codebase, including:Before adding something this broad into the LLVM code base, I think we
need to understand a whole slew of things about it, starting with the
questions above.
I'm not sure about broad. Do you mean because of the commit size? Most
of it is boilerplate and it is a fully working horizontal prototype.
Actually, there are only 1091 lines (all types included) added to
llvm/lib/.
> 1. We improve the performance of parallel programs, today.
There are a number of important parallel languages and both
language-specific and target-specific parallel optimizations. Just
because these five optimizations improve OpenMP program performance
isn’t enough to justify adding a new parallel IR (or interface) to
LLVM, without knowing whether other languages, targets and
optimizations could be correctly and effectively implemented using
this approach.
This will not prevent us from testing other languages and optimizations,
actually it should enable us to do so. If it somehow turns out that this
abstraction makes it impossible to implement an optimization correctly
and effectively and that optimization is very important to us we can
still get rid of the abstraction and keep the rest of the code. The
other optimization we have will still work even without the abstraction.
The same holds true if we decide on a parallel IR representation and
do not need to support multiple ones anymore.
> 2) It serves as a meaningful baseline for future discussions on
> (optimized) parallel representations.We can’t just add a new parallel IR design (abstract or concrete) to
mainline just to serve as a baseline for future discussions.
Not _just_! Adding something beneficial to mainline that makes sense (at
least for me but let's see what others think) does not strike me as a
problem. If that would happen, we would finally have a baseline for
parallel optimizations instead of comparing unoptimized programs to
optimized ones encoded in the new IR we came up with.
> 3) It allows to determine the pros and cons of the different schemes
> when it comes to actual optimizations and inputs.Same problem.
This is not the same problem. When you design your IR and you have to
make trade offs you would probably perform experiments and decide.
However, for that you need a test bed to perform the experiments in, an
implementation of both choices, etc. If we combine optimizations we
think are meaningful with representations we think are reasonable we can
actually learn if and how they impact each other for benchmarks we care
about.
> 4) It helps to identify problems that might arise once we start to
> transform parallel programs but _before_ we commit to a specific
> representation.I suspect you *are* committing to a specific representation, although
we can’t be sure until you provide more details.
Please check out the links to the interface codes above.