llvm dev - Mar 2021 - [RFC] Upstreaming a proper SPIR-V backend

Chiming in mostly from the perspective of MLIR SPIR-V support. More
comments inlined, but first some general comments. :)

As I understand it, SPIR-V is actually a mix of multiple things. It is
first and foremost 1) a binary format for encoding GPU executables that
cross the toolchain and hardware driver boundaries. Then it's 2) an
intermediate level language for expressing such GPU executables. It is also
3) a flexible and extensible spec with all sorts of capability and
extension mechanisms in order to support the needs of multiple APIs and
hardware features. It's unclear to me what a production-quality SPIR-V LLVM
backend would entail; but to actually support various use cases SPIR-V can
support (OpenCL, OpenGL, Vulkan; shader/kernel; various levels of
extensions; etc.), it looks to me that we need a story for all the above
points, where the IR aspect (2) is actually just facet. My understanding
over LLVM is it's mostly focusing on 2): we have a very coherent single IR
threading through the majority layers of the compiler stack and the IR
focuses very much as a means for compiler transformations (i.e., no
instruction versioning etc.). There isn't much native modelling for most
points for 1) and 3) (which makes sense as LLVM IR is a compiler IR). So to
make it work, one would need to shorehore through existing LLVM mechanisms
(e.g., using intrinsics for various GPU related builtins, using metadata
for SPIR-V decorations?, etc.), unless we want to evolve LLVM
infrastructure to have native support for the missing SPIR-V mechanisms,
which I think might be too much to take on. This is just general
mechanisms, not mentioning the different semantics between different SPIR-V
consumers (e.g., shader vs. kernel and what that means over
memory/execution model, etc.) that needs to be sorted out too.. Just
supporting a certain use case of what SPIR-V supports is certainly simpler
though as we can bake in assumptions and avoid some infrastructure needs
for the full generality.

That's why I think using MLIR as the infrastructure to build general
support for SPIR-V is more preferable as we control everything there and
can feel free to model all SPIR-V concepts in the most native way. For
example we can feel free to define all SPIR-V ops natively, including all
ops introduced by SPIR-V extensions and extended instruction sets. We can
support versions/extensions/capabilities natively and integrate it with the
target environment to automatically filter out
<https://github.com/llvm/llvm-project/blob/37eca08e5bcfbe926176412a4a0acf5b963da7e6/mlir/lib/Dialect/SPIRV/Transforms/SPIRVConversion.cpp#L690-L755>
CodeGen patterns generating ops not available on the target, etc. To me,
MLIR's open dialect/op/type/etc. system is a perfect fit for the open
SPIR-V spec with many capabilities/extensions/etc. For example we can even
make the SPIR-V dialect itself open to allow out-of-tree extensions and
development and such.

With that said, I understand that software development has many reality
concerns (like existing codebase, familiarity with different components,
etc.) and we have many different use cases, which may mean that different
paths make sense. So please don't take this as a negative feedback in
general. It's just that to me it's unclear how we can unify here right
now.
Even when the time arrives for unification, I'd believe going through MLIR
is better to have general SPIR-V support. :)

On Tue, Mar 2, 2021 at 6:07 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,
> e.g. in Vulkan we have logical adressing mode).
> 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.
>
I think we have an assumption here: LLVM itself should support all
mechanisms and use cases SPIR-V can support, if to make LLVM a layer before
SPIR-V. I think there is a huge gap here.

>
> 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.

To be specific, Vulkan compute is the most well supported use case right
now. But there are interests from the community to push on Vulkan graphics
and OpenCL.

> 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).
>
I actually believe the opposite, because of the reasons I listed at the
very beginning. To me SPIR-V also stays at a higher level than LLVM. (But
again, depending on what subset we are talking about.)

My proposal would be to include some MLIR -> LLVM-IR translated code in
the
> testing so to have this final goal in mind.
>
> PS: one more thought: SPIR-V does come with a set of builtin/intrinsic
> functions that expose the full capabilities of target architecture (mostly
> GPU). This set of intrinsics is actually a dialect in its own. So this is
> LLVM IR + SPIR-V specific intrinsics and their semantics that fully define
> the SPIR-V dialect at LLVM IR level. I believe this idea could be used in
> MLIR path: MLIR -> LLVM-IR with SPIR-V intrinsics (let's call it a
LLVM IR
> SPIR-V dialect) -> SPIR-V binary (generated by a backend). So the idea
of
> 'SPIR-V dialect' still exists, it is just now expressed at the LLVM
IR
> level.

Not sure this is the prefered way, given that we can define SPIR-V ops
easily in MLIR in its own dialect with native support for various aspects.

> regards,
> konrad
>
> > From: Renato Golin <rengolin at gmail.com>
> > Sent: Tuesday, March 2, 2021 11:12 AM
> > To: Trifunovic, Konrad <konrad.trifunovic at intel.com>
> > Cc: llvm-dev at lists.llvm.org; Paszkowski, Michal <
> michal.paszkowski at intel.com>; Bezzubikov, Aleksandr <
> aleksandr.bezzubikov at intel.com>; Tretyakov, Andrey1 <
> andrey1.tretyakov at intel.com>
> > Subject: Re: [llvm-dev] [RFC] Upstreaming a proper SPIR-V backend
> >
> > On Tue, 2 Mar 2021 at 09:36, Trifunovic, Konrad via llvm-dev
<mailto:
> llvm-dev at lists.llvm.org> wrote:
> > Hi all,
> >
> > We would like to propose this RFC for upstreaming a proper SPIR-V
> backend to LLVM:
> >
> > Hi,
> >
> > Perhaps a parallel question: how does that integrate with MLIR's
SPIRV
> back-end?
> >
> > If this proposal goes through and we have a production-quality SPIRV
> back-end in LLVM, do we remove MLIR's own version and lower to LLVM,
then
> to SPIRV? Or do we still need the MLIR version?
> >
> > In a perfect world, translating to LLVM IR then to SPIRV shouldn't
make
> a difference, but there could be some impedance mismatch between
MLIR->LLVM
> lowering that isn't compatible with SPIRV?
> >
> > But as a final goal, if SPIRV becomes an official LLVM target, it
would
> be better if we could iron out the impedance problems and keep only one
> SPIRV backend.
> >
> > cheers,
> > --renato
> >
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210302/f2504a49/attachment-0001.html>

On Tue, 2 Mar 2021 at 19:02, Lei Zhang <antiagainst at google.com> wrote:
> With that said, I understand that software development has many reality
> concerns (like existing codebase, familiarity with different components,
> etc.) and we have many different use cases, which may mean that different
> paths make sense. So please don't take this as a negative feedback in
> general. It's just that to me it's unclear how we can unify here
right now.
> Even when the time arrives for unification, I'd believe going through
MLIR
> is better to have general SPIR-V support. :)
>
Thank you for such a detailed response!

Honestly, I don't know much about SPIRV, so my comments were without
context. If there are reasons to keep the back-end on both sides, I'm not
against it.

I just proposed unifying things in case they're duplicated. If we can make
that case, then it should definitely be part of the plan.

cheers,
--renato
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210302/cc7c2499/attachment.html>

> As I understand it, SPIR-V is actually a mix of multiple things. It is
first and foremost 1) a binary format for encoding GPU executables that cross
the toolchain and hardware driver boundaries. Then it's 2) an intermediate
level language for expressing such GPU executables. It is also 3) a flexible and
extensible spec with all sorts of capability and extension mechanisms in order
to support the needs of multiple APIs and hardware features. It's unclear to
me what a production-quality SPIR-V LLVM backend would entail; but to actually
support various use cases SPIR-V can support (OpenCL, OpenGL, Vulkan;
shader/kernel; various levels of extensions; etc.), it looks to me that we need
a story for all the above points, where the IR aspect (2) is actually just
facet.
Agreed. Indeed, 'production quality SPIR-V backend' is vaguely defined
here and we proposed one discussion point on this. For this proposal needs, we
should focus on one subset of SPIR-V and one use-case (OpenCL). By production
quality I mean that we can correctly produce the code for that subset of SPIR-V.
I totally agree that having the 'Full SPIR-V' coverage is something very
broad and probably not achievable at all - but we are not aiming at that. I do
take the perspective of classical 'CPU backend' here: we do have to
generate the ISA code for the input LLVM-IR code. Now, besides instructions, our
backend needs to deduce proper capabilities and extensions, based on what subset
of instructions is selected. As You pointed out later, plain LLVM-IR is not
capable of describing the full SPIR-V. Some of
decorations/extensions/capabilities might be deduced by the backend, while some
need to be declared using various LLVM-IR concepts, such as metadata,
attributes, intrinsics - and that needs a clear definition.

 
>  My understanding over LLVM is it's mostly focusing on 2): we have a
very coherent single IR threading through the majority layers of the compiler
stack and the IR focuses very much as a means for compiler transformations
(i.e., no instruction versioning etc.). There isn't much native modelling
for most points for 1) and 3) (which makes sense as LLVM IR is a compiler IR).
So to make it work, one would need to shorehore through existing LLVM mechanisms
(e.g., using intrinsics for various GPU related builtins, using metadata for
SPIR-V decorations?, etc.), unless we want to evolve LLVM infrastructure to have
native support for the missing SPIR-V mechanisms, which I think might be too
much to take on.
Also agree. I do believe though, that LLVM-IR is still worth the effort, and we
can take an incremental approach into adopting some GPU concepts into LLVM-IR.
(e.g. 'convergent' attribute has been added mainly for GPU kind of
targets). The first step is defining the metadata and intrinsics that are target
specific for SPIR-V, but most of them could be generalized as GPU concepts and
even introduced into core LLVM-IR spec. I agree this is great effort - and not
really the main focus point of this proposal - yet, we should give LLVM-IR
it's own right into the world of GPUs.
> This is just general mechanisms, not mentioning the different semantics
between different SPIR-V consumers (e.g., shader vs. kernel and what that means
over memory/execution model, etc.) that needs to be sorted out too.. Just
supporting a certain use case of what SPIR-V supports is certainly simpler
though as we can bake in assumptions and avoid some infrastructure needs for the
full generality.
I would focus on just a subset and clearly define what input LLVM-IR GPU dialect
would look like for that subset.
> 
> That's why I think using MLIR as the infrastructure to build general
support for SPIR-V is more preferable as we control everything there and can
feel free to model all SPIR-V concepts in the most native way. For example we
can feel free to define all SPIR-V ops natively, including all ops introduced by
SPIR-V extensions and extended instruction sets. We can support
versions/extensions/capabilities natively and integrate it with the target
environment to automatically filter out CodeGen patterns generating ops not
available on the target, etc. To me, MLIR's open dialect/op/type/etc. system
is a perfect fit for the open SPIR-V spec with many capabilities/extensions/etc.
For example we can even make the SPIR-V dialect itself open to allow out-of-tree
extensions and development and such.
Right. For the 'General SPIR-V' support, MLIR is the right abstraction
level to use. And I would keep it that way. For the 'specific'/legacy
uses, backend is the way to fill that gap.
> With that said, I understand that software development has many reality
concerns (like existing codebase, familiarity with different components, etc.)
and we have many different use cases, which may mean that different paths make
sense. So please don't take this as a negative feedback in general. It's
just that to me it's unclear how we can unify here right now. Even when the
time arrives for unification, I'd believe going through MLIR is better to
have general SPIR-V support. :)
A very good discussion! I seem to be overly optimistic at the first place at
unifying those two approaches. Now I believe that we actually should have two
paths, for the reasons You have just explained and for the reasons of supporting
'legacy' paths/compilers that rely on a classical, years old approach:
Front-End -> LLVM-IR (opt) -> backend (llc). For that legacy path, a plain
old 'backend' approach is still (in my view) the way to go. On the other
hand, when MLIR evolves and gets wider adoption, it will be the way to go. From
the semantic point of view, MLIR is much better suited for representing
structured and extensible nature of SPIR-V. But for MLIR approach to be adopted,
new languages/front-ends need to be aware of that structure, so to take most of
the advantage of it. If Clang C/C++ start to use MLIR as its native generation
format - that would be a big case for MLIR approach, but until that happens, we
need to have some intermediate solution.

The most important use-case for the backend is the systems/languages that have
been targeting x86, ARM or NVPTX, AMDGPU. You want to replace that particular
backend with some other GPU backend (e.g. Intel GPU :) ). So the solution is to
use SPIR-V as the backend target, and then consume SPIR-V with a
proprietary/open-source GPU finalizer that eventually produces GPU assembly. (We
did not want to come up with yet another GPU intermediate language like PTX,
HSAIL etc. and want to use Khronos standard SPIR-V for that purpose). In this
use case, the stress is on points 1) and 2) and less on point 3) as You stated
earlier.

So my proposal is to keep two paths. They are complementary to each other. I
know of the maintenance cost concerns - but while there are use-cases for both,
it is still worth it. When the worlds stops using LLVM-IR, it will die silently,
so will die the SPIR-V backend - but that is a natural software lifecycle ;)

As it comes to unifying, there seems little implementation could be unified
(unless we make GlobalISel produce MLIR SPV dialect😊). Nevertheless, we might
collaborate on a conceptual level, especially on defining the subset of SPIR-V
that we want to support, what use cases (OpenCL, Vulkan compute, OGL?) are
relevant for LLVM community and to have a common vision there.

BTW: Intel is also interested in MLIR path and there is a group actively
contributing in that direction too.
...
>> 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).
> I actually believe the opposite, because of the reasons I listed at the
very beginning. To me SPIR-V also stays at a higher level than LLVM. (But again,
depending on what subset we are talking about.) 
Here by 'lower level in the compiler stack' I did not mean the
higher/lower semantics level, but the place in the compiler pipeline (front-end
-> optimizer -> back-end), where I assumed MLIR is at front-end level,
optimizer is LLVM-IR, and back-end comes last.
I agree that semantically, SPIR-V is higher level than LLVM-IR, especially when
it comes to other meta-data that LLVM-IR does not support natively
(extensions/extended instruction sets/capabilities/execution model etc.)
>> My proposal would be to include some MLIR -> LLVM-IR translated code
in the testing so to have this final goal in mind.
>> PS: one more thought: SPIR-V does come with a set of builtin/intrinsic
functions that expose the full capabilities of target architecture (mostly GPU).
This set of intrinsics is actually a dialect in its own. So this is LLVM IR +
SPIR-V specific intrinsics and their semantics that fully define the SPIR-V
dialect at LLVM IR level. I believe this idea could be used in MLIR path: MLIR
-> LLVM-IR with SPIR-V intrinsics (let's call it a LLVM IR SPIR-V
dialect) -> SPIR-V binary (generated by a backend). So the idea of
'SPIR-V dialect' still exists, it is just now expressed at the LLVM IR
level.
 
> Not sure this is the prefered way, given that we can define SPIR-V ops
easily in MLIR in its own dialect with native support for various aspects. 
Agree. Having in mind that we should actually keep both paths, I believe this
path of going MLIR -> LLVM-IR -> SPIR-V then does not make sense as it
might lose some information.

regards,
konrad
> From: Renato Golin <mailto:rengolin at gmail.com> 
> Sent: Tuesday, March 2, 2021 11:12 AM
> To: Trifunovic, Konrad <mailto:konrad.trifunovic at intel.com>
> Cc: mailto:llvm-dev at lists.llvm.org; Paszkowski, Michal
<mailto:michal.paszkowski at intel.com>; Bezzubikov, Aleksandr
<mailto:aleksandr.bezzubikov at intel.com>; Tretyakov, Andrey1
<mailto:andrey1.tretyakov at intel.com>
> Subject: Re: [llvm-dev] [RFC] Upstreaming a proper SPIR-V backend
> 
> On Tue, 2 Mar 2021 at 09:36, Trifunovic, Konrad via llvm-dev
<mailto:mailto:llvm-dev at lists.llvm.org> wrote:
> Hi all,
> 
> We would like to propose this RFC for upstreaming a proper SPIR-V backend
to LLVM:
> 
> Hi,
> 
> Perhaps a parallel question: how does that integrate with MLIR's SPIRV
back-end?
> 
> If this proposal goes through and we have a production-quality SPIRV
back-end in LLVM, do we remove MLIR's own version and lower to LLVM, then to
SPIRV? Or do we still need the MLIR version?
> 
> In a perfect world, translating to LLVM IR then to SPIRV shouldn't make
a difference, but there could be some impedance mismatch between MLIR->LLVM
lowering that isn't compatible with SPIRV?
> 
> But as a final goal, if SPIRV becomes an official LLVM target, it would be
better if we could iron out the impedance problems and keep only one SPIRV
backend.
> 
> cheers,
> --renato
>
_______________________________________________
LLVM Developers mailing list
mailto:llvm-dev at lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev