[LLVMdev] [RFC] Parallelization metadata and intrinsics in LLVM (for OpenMP, etc.)

[Hal Finkel]

On Tue, 14 Aug 2012 10:22:35 +0300
Pekka Jääskeläinen <pekka.jaaskelainen at tut.fi> wrote:

> On 08/13/2012 10:54 PM, Hal Finkel wrote:
> > I had thought about uses for shared-memory OpenCL implementations,
> > but I don't know enough about the use cases to make a specific
> > proposal. Is your metadata documented anywhere?
> 
> It is now a quick "brute force hack", that's why I got interested in
> your proposal. We just wanted to communicate the OpenCL work item
> information further down in the compiler as easily as possible and
> didn't have time to beautify it.
> 
> Now all instructions of the "chained" OpenCL kernel instances
> (work items) are annotated with their work item ID, their "parallel
> region ID" (from which region between barriers the instruction
> originates from) and a sequence ID. So, lots of metadata bloat.
> 
> These annotations allow finding the matching instructions later on to
> vectorize multiple work items together by just combining the matching
> instructions from the different WIs. The alias analyzer uses this
> metadata to return NO_ALIAS for any memory access combination where
> the accesses are from different work items within the same parallel
> region (the specs say if they do alias, the results are undefined,
> thus a programmer's fault).
> 
> With your annotations this hack could be probably cleaned up by using
> the "parallel for loop" metadata which the vectorizer and/or "thread
> lib call injector" (or the static instruction scheduler for a
> VLIW/TTA) can then use to parallelize the kernel as desired.
> 
> I'd remind that its usefulness is not limited to a shared memory
> multicore (or even multicore) for the kernel execution device. All
> non-SIMT targets require laying out the code for all the work-items
> (like they were parallel for loops, unrolled or vectorized or not) for
> valid OpenCL kernel execution when there are more than 1 WI per
> work-group, thus potentially benefit from this.

Fair enough. My Thought process here was that, first, I was not going
to propose anything specifically for non-shared-memory systems (those
require data copying directives, and I'd want to let others who have
experience with those do the proposing), and second, I was not going to
propose anything specifically for multi-target (heterogeneous) systems.
I think that single-target shared-memory systems fall into the model
I've sketched, and support for anything else will require further
extension.

> 
> > I agree, and this is specifically why I don't want to support
> > OpenMP by lowering it into runtime calls in the frontend. I want to
> > allow for other optimizations (vectorization, etc.) in combination
> > with (or instead of) multi-threading. I think that my current
> > proposal allows for that.
> 
> Yes it should, as far as I can see. If the loop body is a function and
> the iteration count (or its multiple) is known, one should be able to
> (vectorize multiple copies of the function without dependence
> checking. In the multi-WI OpenCL C case this function would contain
> the code for a single work item between a region between barriers
> (implicit or not).
> 
> I'm unsure if forcing the function extraction of the parallel
> regions brings unnecessary problems or not. Another option would be to
> mark the basic blocks that form parallel regions. Maybe all of the BBs
> could be marked with a PR identifier MD? This would require BB
> metadata (are they supported?).

I thought about this. There had been some patches provided for BB
metadata (by Ralf Karrenberg back in May), I don't recall what happened
with those. BB metadata might work, but I worry about existing
optimization passes, which don't know about this metadata, moving
things in and out of parallel regions in illegal ways. For example,
moving a call to some get_number_of_threads() function, or some inline
assembly region, in or out of a parallel region. Putting things in
functions just seemed safer (and BB metadata is not upstream). Also, it
would require extra checking to keep the parallel basic blocks together.
Furthermore, in many cases, the parallel regions need to end up as
separate functions anyway (because their passed as callbacks to the
runtime library).

> 
> >> Also, one user of this metadata could be the alias analysis: it
> >> should be easy to write an AA that can exploit the parallelism
> >> information. Parallel regions by definition do not have (defined)
> >> dependencies between each other (between synchronization points)
> >> which should be useful information for optimization purposes even
> >> if parallel hardware was not targeted.
> >
> > I really like this idea! -- and it sounds like you may already have
> > something like this in POCL?
> 
> Yes, an OpenCL AA that exploits the work-item independence and address
> space independence. With your annotations there could be a generic
> AA for the "independence information from parallelism metadata" part
> and a separate OpenCL-specific AA for the rest.
> 
> > Regarding having scheduling be separate, care is required to ensure
> > correctness. A large constraint on the design of a metadata API is
> > that
> 
> OK, I see.
> 
> I suppose it's not a big deal to add the scheduling property. At
> least if one (later) allows adding scheduling modes supported by other
> standards than OpenMP as well. I.e., not modes like "static" but
> "openmp31_static" or similar. For OpenCL work item loops the
> scheduling mode could be "auto" or left empty.

I think that this makes sense. For some things, like 'static', we can
define backend-independent semantics. For other things, like OpenMP's
'runtime', which is tied to how the application calls OpenMP runtime
functions, I agree, we should probably call that 'openmp_runtime' (or
something like that).

> 
> > I agree. I think that vectorization is best done earlier in the
> > optimization schedule. Vectorization, however, should appropriately
> > update loop metadata to allow for proper integration with
> > parallelization, etc. Lowering to runtime libraries (for
> > multi-threading in whatever form) should be done relatively late in
> > the process (because further higher-level optimizations are often
> > not possible after that point).
> 
> Yes, to enable automatic mixing of vectorization and threading from
> the single (data parallel) kernel.

Yep, that is exactly what I want to be able to do.

Thanks again,
Hal

> 



-- 
Hal Finkel
Postdoctoral Appointee
Leadership Computing Facility
Argonne National Laboratory