[llvm-dev] [RFC] Upstreaming a proper SPIR-V backend

[Mehdi AMINI via llvm-dev]

On Wed, Mar 3, 2021 at 7:05 AM James Y Knight <jyknight at google.com> wrote:

>
>
> On Tue, Mar 2, 2021 at 6:46 PM Mehdi AMINI <joker.eph at gmail.com> wrote:
>
>>
>>
>> On Tue, Mar 2, 2021 at 3:18 PM James Y Knight <jyknight at google.com>
>> wrote:
>>
>>> On Tue, Mar 2, 2021 at 4:40 PM Mehdi AMINI via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>>> On Tue, Mar 2, 2021 at 3:07 AM Trifunovic, Konrad via llvm-dev <
>>>> llvm-dev at lists.llvm.org> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> A very good question. I was actually expecting it 😊
>>>>>
>>>>> So, at the moment, it does not integrate into MLIR SPIRV backend and
>>>>> we have not thought about it. I guess You are referring to having a SPV
>>>>> dialect in MLIR and using a 'serialize' option to produce a SPIR-V binary?
>>>>>
>>>>> I agree that developing two backends in parallel is a bit redundant.
>>>>> If SPIR-V LLVM backend becomes a production quality it means actually it
>>>>> could consume any LLVM IR (provided it does conform to some SPIR-V
>>>>> restrictions).
>>>>> By any LLVM IR input I mean: it should be irrelevant whether it is
>>>>> produced by a clang, MLIR to LLVM IR lowering or just some other front-end
>>>>> that produces LLVM IR.
>>>>
>>>> The biggest 'impedance mismatch' that I currently see is that SPV MLIR
>>>>> dialect is now targeted mostly at Vulkan, while LLVM SPIR-V backend targets
>>>>> compute. Besides instruction set, the fundamental difference is a memory
>>>>> model.
>>>>> So if we want to unify those, we should actually make SPIR-V LLVM
>>>>> backend able to produce Vulkan dialect of SPIR-V as well.
>>>>>
>>>>> My answer is a bit elusive, but I totally agree with Your proposal: we
>>>>> should work towards having a one solution, and, LLVM SPIR-V backend seems
>>>>> like a more universal one (since it sits lower in the compiler stack).
>>>>> My proposal would be to include some MLIR -> LLVM-IR translated code
>>>>> in the testing so to have this final goal in mind.
>>>>>
>>>>
>>>> Something you're missing here, and maybe Lei clarified but I'll
>>>> reiterate: the SPIRV dialect in MLIR is equivalent to what your GlobalISel
>>>> pass will produce. It can actually round-trip to/from the SPIRV binary
>>>> format. So it is sitting lower than your backend in my view.
>>>> I can't figure out a situation where it would make sense to go from
>>>> MLIR SPIRV dialect to LLVM to use this new backend, but I may miss
>>>> something here...
>>>>
>>>> It would be really great to find a common path here before duplicating
>>>> a lot of the same thing in the lllvm-project monorepo, for example being
>>>> able to target the MLIR dialect from GlobalISel, or alternatively
>>>> converting the MIR to it right after would be an interesting thing to
>>>> explore.
>>>> I haven't seen it, but there was a talk last Sunday on this topic:
>>>> https://llvm.org/devmtg/2021-02-28/#vm1
>>>>
>>>
>>> This sort of problem seems like just one of those unfortunate
>>> consequences of MLIR being effectively an "LLVM IR 2.0 -- Generic Edition",
>>> but not yet actually layered underneath LLVM where it really wants to be.
>>>
>>
>> I don't understand what you mean here with "layered underneath LLVM"? Can
>> you elaborate on this?
>>
>
> That ultimately the goal should be for LLVM IR to be a dialect of MLIR,
> and for much of the optimization and codegen processes in LLVM to be
> implemented as MLIR dialect lowering. Then, MLIR is foundational --
> "layered" underneath LLVM's core -- LLVM would have a hard dependency on
> MLIR.
>

OK I see what you mean now, I didn't connect to this because I think it is
an open question whether we see this happen in this decade ;)

So my assumption coming here is that:
1) LLVM IR as it is now is "granted" (at least in the context of this
thread).
2) A SPIRV backend that takes LLVMIR and use GlobalISel is desirable.

Considering this, my angle is mainly one of library, software engineering,
and reuse / avoiding duplication.
So adding intrinsics to LLVM IR and improving the GPU support in LLVM IR is
something I see as "obviously good" and necessary for this project. The
only opportunity for sharing and avoiding duplication appears to me right
after GlobalISel for the rest of the pipeline.

At that point, SPIR-V as an MLIR dialect, and the SPIR-V backend doing MLIR
> dialect lowering would be effectively no different from how every target
> works -- just with a different output dialect.
>
> I think it doesn't really make sense to tie *this* project to those
>>> long-term goals of layering MLIR under LLVM-IR, given the extremely long
>>> timescale that is likely to occur in. The "proper" solution probably won't
>>> be possible any time soon.
>>>
>>
>> I'm not sure if we're talking about the same thing here: there is nothing
>> that I suggest that would operate at the level of LLVM IR. And nothing that
>> requires a "long timescale", it seems quite easily in scope to me here.
>>
>
> So, in the meantime, we could implement a special-case hack just for
>>> SPIRV, to enable lowering it to MLIR-SPIRV dialect. But, what's the
>>> purpose? It wouldn't really help move towards the longer term goal, I don't
>>> think? And if someone does need that at the moment, they can just feed the
>>> SPIRV binary format back into the existing MLIR SPIRV dialect, right?
>>>
>>
>> Do we want to maintain, in the LLVM monorepo, *two* different
>> implementations of a SPIRV IR and associated serialization (and potential
>> deserialization)? All the tools associated to manipulate it? I assume the
>> backend may even want to implement optimization passes, are we gonna
>> duplicate these as well?
>> (note that this isn't at the LLVM IR level, but post-instruction
>> selection, so very ad-hoc to the backend anyway).0
>>
>
> Quite possibly yes. It's unfortunate to have duplication, but given the
> current state of things, I think it should not be ruled out.
>
> My inclination is that the following factors are likely to be true:
> - The amount of code for SPIRV binary format serialization is not
> particularly large or tricky.
> - The work to emit SPIR-V MLIR dialect from the LLVM SPIR-V backend will
> not be simpler than serializing to SPIR-V directly.
> - Writing this custom code to emit SPIR-V MLIR dialect from the SPIR-V
> backend will not noticably further the longer-term goals of having LLVM
> core be implemented as MLIR dialect lowering.
>

These are great considerations, I subscribe entirely :)


>
> It seems to me that the choice here is either writing new code in LLVM to
> emit the SPIR-V MLIR dialect in the GlobalISel SPIR-V backend, or new code
> in LLVM to emit SPIR-V directly. And while I find the long-term prospects
> of MLIR integration into LLVM extremely promising, using MLIR just as
> step-stone to MLIR SPIR-V serialization does not seem particularly
> interesting.
>
> So, to me the interesting question is whether we'd expect to be doing
> something interesting after converting to the SPIR-V MLIR dialect form
> besides simply serializing to SPIR-V binary format. Something that would
> make the added complexity of serializing through MLIR seem more worthwhile.
> I guess I'm not immediately seeing this as likely to be the case, but it
> seems well worth further discussion.
>
> A possibility you've mentioned is post-instruction-selection
> optimizations. Do you have something in particular in mind there?
>

I suspect that post-Global ISel there is a bit more "than taking the MIR
as-is and emit the SPIRV serialization in a single traversal". So
converting MIR to MLIR means that everything that we want to happen at this
point will be shared.
Note that this is different from other backends, because I don't expect
SPIRV to share passes (RA, Scheduling, ...) on MIR or to take advantage of
the MC layer in the same way (if I'm wrong here then my point is less
strong though).

Cheers,

-- 
Mehdi
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210303/31b918e7/attachment-0001.html>