[cfe-dev] [RFC][OpenMP][CUDA] Unified Offloading Support in Clang Driver

[Justin Lebar via cfe-dev]

> I don't have any other tool that would be compatible with a bundle of all four object files, doesn't matter how I do it,

I think we've given up on letting existing tools (e.g. objdump) work
well on our host+device*n object files, see discussions earlier in the
thread.  Even I, probably the loudest supporter of this on thread,
think that it's probably not an important case that cuobjdump work
sensibly when you are embedding two different nvptx bundles into one
object file, since this is an edge case that most people aren't going
to encounter.

> Creating linkable host objects shouldn't be that hard (clang-bundler -embed).

This seems to be basically the same as embedding the device code into
the object file -- just a different format.

> If the bundler uses a target agnostic format, my example above would just work regardless of the compiler phases -- -S, -emit-ast, -emit-llvm, -emit-ast would work just fine.

What I believe we do with CUDA compilation is say that all of these
output code just for the host, unless you specify a device
architecture, in which case you can only specify one.  It seems to
work OK.  I'll grant the "wouldn't it be nice if" we could output a
single file containing .s files for all archs, but I'm not convinced
goes beyond wouldn't-it-be-nice.

> - Extending the bundler to enable compatibility modes can then be discussed on a per-case basis if there are a lot of users asking for that feature.

There are, on this very thread.  :)  I don't think in any case that
these compatibility modes can be a second-thought -- they'll need to
be first-class citizens in the design.

The whole first half of this mail seems quite complicated to me,
frankly.  Even if it may be convenient for complex compilation modes
like the openmp + cuda one you described, it may not be convenient for
simpler modes, which I expect will be far more common.

-Justin

On Tue, Mar 15, 2016 at 2:37 PM, Samuel F Antao <sfantao at us.ibm.com> wrote:
> Let me provide a more concrete example of what I had in mind (sorry for the
> long email...):
>
> Lets assume I have two source files a.cu and b.cu, both with OpenMP and CUDA
> kernels and several device functions. Device functions defined in a.cu are
> used in b.cu and vice-versa.
>
> I have a system that consists of a ppc64le host, an nvptx64 device and a
> ppc64le accelerator. The exact details of this system are not relevant, I am
> just using this example because it is something supported by the prototype
> we have in github and exemplifies that the same target can be both host and
> device (in general we can have AMD GPUs, DSPs, FPGAs, each with its own
> binary formats).
>
> ###
> Case A: We pass all the source files and compile in a single run of clang.
>
> A1) clang invocation (the openmp options are just an example, *.cu extension
> tells the driver we should compile for CUDA as well):
>   clang -target ppc64le -fopenmp -fomptargets=nvptx64,ppc64le a.cu b.cu
>
> A2) driver commands:
>   // Obtain cuda obj for a.cu
>   clang -cc1 -fcuda-is-device -target nvptx64 -S a.cu -o
> a-device-cuda-nvptx64.s
>   ptxas -c a-device-cuda-nvptx64.s -o a-device-cuda-nvptx64.o
>
>   // Obtain openmp nvptx64 obj for a.cu
>   clang -cc1 -fopenmp-is-device  -target nvptx64 -S a.cu -o
> a-device-openmp-nvptx64.s
>   ptxas -c a-device-openmp-nvptx64.s -o a-device-openmp-nvptx64.o
>
>   // Obtain openmp ppc64le device obj for a.cu
>   clang -cc1 -fopenmp-is-device -target ppc64le -c a.cu -o
> a-device-openmp-ppc64le.o
>
>   // Obtain openmp ppc64le host obj for a.cu
>   clang -cc1 -fopenmp -target ppc64le -c a.cu -o a-host-cuda-openmp.o
>
>   // Repeat everything  for b.cu
>
>   // Obtain cuda device image:
>   nvlink a-device-cuda-nvptx64.o b-device-cuda-nvptx64.o -o
> a.out-device-cuda-nvptx64
>
>    // Obtain openmp nvptx device image:
>   nvlink a-device-openmp-nvptx64.o b-device-openmp-nvptx64.o -o
> a.out-device-openmp-nvptx64
>
>   // Obtain openmp ppc64le device image (making it a shared library is just
> easier to load, but not required):
>   ld -shared a-device-openmp-ppc64le.o b-device-openmp-ppc64le.o -o
> a.out-device-openmp-ppc64le
>
>   // Create a host compatible object for each device (this is done at link
> time only, the user is not expected to deal with this tool - the name of the
> tool and commands are just an example).
>   // I am calling this "embed-mode" but is not really embedding, is generate
> an host object that solely contains device code...
>   clang-bundler -embed -host ppc64le -models=cuda,openmp,openmp
> -targets=nvptx64,nvptx64,ppc64le
> -inputs=a.out-device-cuda-nvptx64,a.out-device-openmp-nvptx64,a.out-device-openmp-ppc64le
> -outputs=a.out-device-cuda-nvptx64.o,a.out-device-openmp-nvptx64.o,a.out-device-openmp-ppc64le.o
>
>   // Generate the host binary
>   ld a-host-cuda-openmp.o a.out-device-cuda-nvptx64.o
> a.out-device-openmp-nvptx64.o a.out-device-openmp-ppc64le.o -o a.out
>
> ###
> Case B: We do separate compilation.
> B1) clang invocation (the openmp options are just an example, *.cu extension
> tells the driver we should compile for CUDA as well):
>   clang -c -target ppc64le -fopenmp -fomptargets=nvptx64,ppc64le a.cu b.cu
>   clang -target ppc64le -fopenmp -fomptargets=nvptx64,ppc64le a.o b.o
>
> B2) driver commands - 1st run:
>
>    // Same thing as in A2) up to the nvlink command
>
>    // Create the bundles a.o and b.o so that the user sees a single file.
>    clang-bundler -bundle -models=host,cuda,openmp,openmp
> -targets=nvptx64,nvptx64,ppc64le
> -inputs=a-host-cuda-openmp.o,a-device-cuda-nvptx64.o,a-device-openmp-nvptx64.o,a-device-openmp-ppc64le.o
> -outputs=a.o
>    clang-bundler -bundle -models=host,cuda,openmp,openmp
> -targets=nvptx64,nvptx64,ppc64le
> -inputs=b-host-cuda-openmp.o,b-device-cuda-nvptx64.o,b-device-openmp-nvptx64.o,b-device-openmp-ppc64le.o
> -outputs=b.o
>
> B2) driver commands - 2nd run:
>    // Attempt to unbundle the inputs because the user is asking for
> offloading support and they are not source files. If the bundler understands
> the file is a bundle it creates the individual files, otherwise create empty
> files for the devices.
>    clang-bundler -unbundle -models=host,cuda,openmp,openmp
> -targets=nvptx64,nvptx64,ppc64le -outputs
> =a-host-cuda-openmp.o,a-device-cuda-nvptx64.o,a-device-openmp-nvptx64.o,a-device-openmp-ppc64le.o
> -inputs=a.o
>    clang-bundler -unbundle -models=host,cuda,openmp,openmp
> -targets=nvptx64,nvptx64,ppc64le -outputs
> =b-host-cuda-openmp.o,b-device-cuda-nvptx64.o,b-device-openmp-nvptx64.o,b-device-openmp-ppc64le.o
> -inputs=b.o
>
>   // Same thing in A2 starting from the nvlink commands.
>
> ###
> A few more comments/observations:
>
> - I don't have any other tool that would be compatible with a bundle of all
> four object files, doesn't matter how I do it, this is just not supported by
> any other compiler and I'm not doing anything wrong as per the programming
> models specification. But, I have tools compatible with each object. That's
> is way I agree we should give the user the ability to extract objects. I
> agree with Justin that extracting stuff from tar is more convenient that
> doing it from ELF (I don't think most users will now how to do it from ELF,
> and it is very likely they don't even know what ELF is, for them is just a
> binary, so not exposing these details is desirable). I am not trying to push
> for a specific format, I just think it should be target agnostic.
>
> - Creating linkable host objects shouldn't be that hard (clang-bundler
> -embed). For CUDA I can create an IR file that defines
> "__cuda_fatbin_wrapper" with the bytes coming from the fat binary (that
> global would have to be made external though). For OpenMP it would do
> something similar but for ".omp_offloading.img_start/end.<device>". The
> resulting IR could then be converted into an object by calling the host
> backend. The tool itself wouldn't care about obj formats.
>
> - If the bundler uses a target agnostic format, my example above would just
> work regardless of the compiler phases -- -S, -emit-ast, -emit-llvm,
> -emit-ast would work just fine.
>
> - Extending the bundler to enable compatibility modes can then be discussed
> on a per-case basis if there are a lot of users asking for that feature. If
> that is to be supported, someone with experience with that format could
> contribute it and there is nothing in the design that prevents it. Expecting
> to have something that works for all formats obj formats and device
> combinations from the very beginning seems not reasonable. My plan of action
> was to contribute something that works and can scale to as many use cases as
> possible, I can't however contribute specific code for every possible
> device, in particular those that are not supported in clang yet.
>
> - The user does not have to deal with the bundling tool at any time if we
> adopt a format that makes it easy to extract objects. I don't know if one
> should avoid completely having a tool (again the user doesn't have to use
> it, but it can look it up in the install tree of clang). An alternative is
> supporting something like clang -ccbundle. Maybe other people are aware of
> what are the precedents in clang?
>
> About shared libraries: I am not sure what is the concern with shared
> libraries. They would be created as binaries would, and would contain
> bundles on them. Nothing that is proposed here precludes that. Supporting
> device-side shared libraries depends on whether the runtime library
> implements a loader or not. It is very possible I am missing something
> here...
>
> Thanks!
> Samuel
>
> 2016-03-15 14:36 GMT-04:00 Eric Christopher via cfe-dev
> <cfe-dev at lists.llvm.org>:
>>
>>
>>
>> On Tue, Mar 15, 2016 at 12:25 AM C Bergström <cfe-dev at lists.llvm.org>
>> wrote:
>>>
>>> Afaik - The way that each vendor *embeds* is a little different and
>>> non-conflicting. (AMD vs NV vs Intel) Embedding for each vendor in a
>>> single object may work. Wouldn't this achieve the desired goal without
>>> the bundling? Bundling is a non-solution - it's just delays dealing
>>> with the offload target code, at which point you have to extract and
>>> jump around to make sure it all magically works (fragile - This side
>>> up)
>>>
>>
>> I'm somewhat in agreement here. An alternate idea is to just "do what
>> we've always done" and have compatibility modes where we try to do what's
>> expected for the particular end use that we're trying to copy. That would
>> get us maximum flexibility (at the cost of a number of implementations, but
>> abstraction hopefully ftw?).
>>
>>>
>>> With embedding approach we have all the pieces to at least read/embed
>>> - it's just dealing with elf/macho/pxe objects and data sections.
>>>
>>
>> Agreed.
>>
>>>
>>> Also there's the question of how to deal with shared libraries - Even
>>> if initially it's not a concern, eventually someone will request it.
>>> If this isn't ever going to be supported or a concern it's not valid
>>> of course..
>>>
>>
>> I'm sure someone is going to want it eventually.
>>
>> -eric
>>
>>>
>>> On Tue, Mar 15, 2016 at 2:54 PM, Justin Lebar via cfe-dev
>>> <cfe-dev at lists.llvm.org> wrote:
>>> >> I agree with Justin when he says that supporting all possible
>>> >> combinations of host-device bundling, each with a custom format adds a lot
>>> >> of complexity and does not necessarily helps the user.
>>> >
>>> > Maybe it would be helpful to speak concretely about this.  What is a
>>> > specific case you're concerned about?
>>> >
>>> > Frankly if we try to build a thing that does everything -- both embeds
>>> > and bundles -- I'm concerned that will be much more complex than
>>> > either approach by itself.  But maybe there's something about bundling
>>> > code for many different devices that's qualitatively different than
>>> > bundling code for one or just a few devices?  For example, are there
>>> > extant formats that we would like to be compatible with that are
>>> > incompatible with each other?  Again, I think it would be helpful to
>>> > be concrete.
>>> >
>>> >> Using a header compatible with tar is not something hard to do, having
>>> >> the bundler tool as part of clang ensures that it works across operating
>>> >> systems too.
>>> >
>>> > This is a minor point, but -- without getting myself back into the
>>> > minefield of discussing whether using the tar format is or is not good
>>> > -- if we do use said format, given what I understand at the moment,
>>> > I'm not in favor of building our own user-facing tool, if that's
>>> > possible.  If you're on Windows and want to inspect a tar file you
>>> > can, download 7z.  It's no different than downloading objdump.  But we
>>> > can postpone this discussion until later; I suspect it won't be
>>> > necessary, given the pushback to using tar in the first place.
>>> >
>>> >> this leaves room to have custom-format bundling activated with options
>>> >> for the cases (would have to be evaluated one by one) that would greatly
>>> >> benefit of interoperability.
>>> >
>>> > This is, as I read it, exactly what everyone else has been arguing so
>>> > strenuously against (and what I conceded to).
>>> >
>>> > On Mon, Mar 14, 2016 at 7:25 PM, Samuel F Antao <sfantao at us.ibm.com>
>>> > wrote:
>>> >> Hi all,
>>> >>
>>> >> I decided a take shot on a possible implementation for the part of
>>> >> this
>>> >> proposal that I think is more consensual (i.e. the part that does not
>>> >> relate
>>> >> with the bundling). I posted three patches
>>> >> (http://reviews.llvm.org/D18170,
>>> >> http://reviews.llvm.org/D18171, http://reviews.llvm.org/D18172) with a
>>> >> possible implementation, so that we have something more concrete to
>>> >> discuss.
>>> >> Let me know your thoughts.
>>> >>
>>> >> Going back to the bundling discussion:
>>> >>
>>> >> I agree with Justin when he says that supporting all possible
>>> >> combinations
>>> >> of host-device bundling, each with a custom format adds a lot of
>>> >> complexity
>>> >> and does not necessarily helps the user. Therefore, I think reasonable
>>> >> to
>>> >> have intermediate files bundled in some already existent format (say
>>> >> tar)
>>> >> that is agnostic of the programming model. Actually, that was my
>>> >> motivation
>>> >> when I proposed the custom format in the bundler.
>>> >>
>>> >> When I look at all the different concerns I think that a possible
>>> >> solution
>>> >> is to have a bundler with three operation modes:
>>> >>  i) "Embed": it generates a host object that contains the device image
>>> >> and
>>> >> properly defines the symbols a programming model requires. Therefore
>>> >> it can
>>> >> be linked with host objects successfully. This file is never exposed
>>> >> to the
>>> >> user unless save-temps is used.
>>> >> ii) "Bundle": Combines host and device object using, by default, a
>>> >> format
>>> >> easy to interact with that is agnostic of the programming model.
>>> >> iii) "Unbundle": The inverse of ii), it assumes the input uses that
>>> >> default
>>> >> format.
>>> >>
>>> >> Using a header compatible with tar is not something hard to do, having
>>> >> the
>>> >> bundler tool as part of clang ensures that it works across operating
>>> >> systems
>>> >> too. At the same time, this leaves room to have custom-format bundling
>>> >> activated with options for the cases (would have to be evaluated one
>>> >> by one)
>>> >> that would greatly benefit of interoperability.
>>> >>
>>> >> Does this sound reasonable?
>>> >>
>>> >> Thanks!
>>> >> Samuel
>>> >>
>>> >> 2016-03-10 13:59 GMT-05:00 Justin Lebar via cfe-dev
>>> >> <cfe-dev at lists.llvm.org>:
>>> >>>
>>> >>> > Justin, is this convincing enough?
>>> >>>
>>> >>> Okay, okay.  Uncle.
>>> >>>
>>> >>> There are two things here that I find convincing.
>>> >>>
>>> >>> 1) Although we're not going to be compatible with the details of
>>> >>> some,
>>> >>> if not all, other compilers' formats, we can at least be compatible
>>> >>> with the spirit by using object files as opposed to tar.
>>> >>>
>>> >>> 2) The postscript in Andrey's e-mail:
>>> >>>
>>> >>> > Re: objdump doesn't understand ELF format with code for multiple
>>> >>> > targets. The same is true for fat executable files as well, isn't?
>>> >>> > So if we
>>> >>> > need to teach objdump how to recognize fat files, we already have
>>> >>> > this
>>> >>> > problem.
>>> >>>
>>> >>> It's probably much more important that objdump work on executables
>>> >>> than on object files, since if you have object files, you can
>>> >>> probably
>>> >>> recompile with -save-temps, but if you only have an executable, you
>>> >>> don't necessarily have access to intermediate files, or even a
>>> >>> compiler for the relevant architecture, much less the specific
>>> >>> compiler which generated the executable.
>>> >>>
>>> >>> Shoving device code into the host *executable* seems unavoidable.
>>> >>> I'm
>>> >>> still not thrilled with doing the same for object files -- it still
>>> >>> feels like we're using ELF when we actually want an archive format --
>>> >>> but (1) makes it less bad.
>>> >>>
>>> >>> -Justin
>>> >>>
>>> >>> On Wed, Mar 9, 2016 at 5:59 AM, Andrey Bokhanko
>>> >>> <andreybokhanko at gmail.com> wrote:
>>> >>> > All,
>>> >>> >
>>> >>> > I asked Intel GCC guys who implemented OpenMP offloading support in
>>> >>> > GCC,
>>> >>> > and
>>> >>> > as they told me, GCC also employs option #4 from Hal's list -- it
>>> >>> > puts
>>> >>> > both
>>> >>> > host and target code in a single ELF file. "Code" in GCC case is
>>> >>> > always
>>> >>> > GCC's IR (Gimple), though -- they require GCC invocation from
>>> >>> > linker in
>>> >>> > order to produce a multi-target executable. This makes GCC
>>> >>> > non-interoperable
>>> >>> > with any other offloading compiler and effectively produces its own
>>> >>> > standard.
>>> >>> >
>>> >>> > Thus, prior art from:
>>> >>> > * nvcc
>>> >>> > * Pathscale
>>> >>> > * GCC
>>> >>> > * ICC
>>> >>> >
>>> >>> > indicates only one direction -- compiler driver produces a single
>>> >>> > object
>>> >>> > file with target code embedded in data section.
>>> >>> >
>>> >>> > Justin, is this convincing enough? I don't see any good reasons why
>>> >>> > clang
>>> >>> > should go against what every other compiler on the planet does.
>>> >>> >
>>> >>> > Re: objdump doesn't understand ELF format with code for multiple
>>> >>> > targets.
>>> >>> > The same is true for fat executable files as well, isn't? So if we
>>> >>> > need
>>> >>> > to
>>> >>> > teach objdump how to recognize fat files, we already have this
>>> >>> > problem.
>>> >>> >
>>> >>> > Yours,
>>> >>> > Andrey
>>> >>> > =====
>>> >>> > Software Engineer
>>> >>> > Intel Compiler Team
>>> >>> >
>>> >>> >
>>> >>> _______________________________________________
>>> >>> cfe-dev mailing list
>>> >>> cfe-dev at lists.llvm.org
>>> >>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>> >>
>>> >>
>>> > _______________________________________________
>>> > cfe-dev mailing list
>>> > cfe-dev at lists.llvm.org
>>> > http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>> _______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>
>>
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>
>