[RFC] Proposal for Adding SPIRV Target

Project Infrastructure LLVM Dev List Archives
1

Here is the revised proposal for the LLVM/SPIR-V converter. Please comment. Thanks.

Proposal of Adding SPIRV Target

Background

SPIR-V is a portable binary format for OpenCL kernels and GLSL shaders. A typical use case of SPIR-V is as follows:

  1. An application developer uses Clang to compile an OpenCL kernel source code to a SPIR-V binary which is common for all OpenCL platforms.

  2. The application developer ships the application containing the SPIR-V binary to customers.

  3. A customer runs the application on an OpenCL platform, which loads the SPIR-V binary through an OpenCL API function.

  4. The vendor-specific OpenCL runtime translates SPIR-V to LLVM IR, changes the target triple and data layout to suit the device which will execute the kernel, performs target specific optimizations, generates the ISA and executes the ISA on the device.

For OpenCL kernels, there is implicit data layout dependence when compiling the source to LLVM. Since SPIR-V is for common OpenCL platforms, a common data layout accepted by different OpenCL vendors is required. We choose the data layout which has been adopted by SPIR 1.2/2.0 for SPIR-V, since it has been successfully used for supporting consumption of SPIR 1.2/2.0 on various OpenCL platforms. For GLSL shaders, it is still under discussion whether to choose the same data layout as OpenCL, or a different data layout, or no data layout at all.

Location

From feedback of the previous version of the proposal, there are several suggestions about the location for the LLVM/SPIR-V converter:

  1. llvm/lib/SPIRV only, adding an option to Clang for outputting SPIR-V. The advantage is ease of use for bi-way translation. However it does not reflect the fact that only LLVM IR with specific target triple and data layout can be translated to SPIR-V.

  2. llvm/lib/SPIRV containing the main functionality of bi-way translation between LLVM IR and SPIR-V, llvm/lib/Target/SPIRV containing a thin wrapper as a target machine to allow Clang targeting SPIR-V. The advantage compared with 1 is that it allows a more conventional way of using Clang to produce SPIR-V. However it is subject to the same issue as 1 about not reflecting the requirement on the LLVM IR which can be translated to SPIR-V.

  3. llvm/lib/Target/SPIRV only. The advantage is that it reflects the requirement on the target triple and data layout for LLVM IR which can be translated to SPIR-V. However putting the SPIR-V to LLVM converter in the same directory is unconventional. Leaving the SPIR-V to LLVM converter out of LLVM source tree is also not desirable since OpenCL vendors need this functionality.

Our proposal is to take approach 3 and keep the bi-way converter in llvm/lib/Target/SPIRV. The functionality of the bi-way converter is exposed through llvm/include/Support/SPIRV.h. A thin wrapper as a target machine is also provided to allow Clang targeting SPIR-V. The rationale is that this directory structure better reflects the nature of SPIR-V. SPIR-V is not an alternative representation for arbitrary LLVM IR. Instead, it is an alternative representation for LLVM IR targeting generic OpenCL or Vulkan platforms. It has its own specific target triple and data layout. Therefore it makes sense for the functionality to be put under llvm/lib/Target. Also, as an alternative representation of LLVM IR, it makes sense to have a bi-way convertor.

Implementation

About the implementation of the converter, although there are suggestions to take the SelectionDAG/MC approach, it seems not a major concern in general. The current implementation uses a shared in-memory representation of SPIR-V, which facilitates supporting bi-way translation. The round-trip translated LLVM IR by the current implementation has passed OpenCL SPIR 1.2 conformance test, which proves the current implementation works. On the other hand, the SelectionDAG/MC approach would require significant tweaking compared to a conventional backend, since SPIR-V is not a low level machine ISA but a high level generic IR. Also, the SelectionDAG/MC approach only supports one-way translation from LLVM to SPIR-V. Therefore unless major concern arises, we will keep the current implementation approach.

Maintenance

The current implementation works by breaking down LLVM IR to instructions and re-constructing them in SPIR-V, and vice versa. Therefore the dependence of the current implementation on LLVM is mainly the C++ API in llvm/include/IR. As such, its dependence on LLVM is similar to typical module passes. Our experience with porting it among LLVM 3.2/3.4/3.6 is that the porting effort is moderate.

Milestones

Currently Clang can compile OpenCL 1.2/2.0 C kernel source to LLVM IR with spir/spir64 target triple, which is compatible with SPIR-V for the supported instructions, data types and builtin functions. Therefore the first milestone of the converter is to support compiling OpenCL 1.2/2.0 kernel to SPIR-V. This is also to lay the foundation for the upcoming OpenCL 2.1 C++ frontend development in Clang. The next milestone would be supporting OpenCL 2.1 C++, which hopefully would be in synch with the frontend development work. In the meantime, as the SPIR-V target becomes stable and is able to support the instructions common to OpenCL and GLSL, hopefully the GLSL frontend work and support for GLSL specific instructions would pick up and finally we would have a SPIR-V target supporting the complete SPIR-V spec.

Testing

We will add lit tests for SPIR-V. Also the converter would be tested by different OpenCL vendors for production quality through comprehensive conformance tests, since SPIR-V is required by OpenCL 2.1.

Logistics

AMD, Intel and some other SPIR WG members would join force for the development of the bi-way converter. Since supporting of SPIR-V is required by OpenCL 2.1, it is expected that OpenCL vendors would continue the maintenance efforts for supporting SPIRV target in LLVM.

Yaxun Liu

AMD

2

Hi Liu,

Thanks for the detailed proposal.

Here is the revised proposal for the LLVM/SPIR-V converter. Please comment. Thanks.

Proposal of Adding SPIRV Target

Background

SPIR-V is a portable binary format for OpenCL kernels and GLSL shaders. A typical use case of SPIR-V is as follows:

1. An application developer uses Clang to compile an OpenCL kernel source code to a SPIR-V binary which is common for all OpenCL platforms.
2. The application developer ships the application containing the SPIR-V binary to customers.
3. A customer runs the application on an OpenCL platform, which loads the SPIR-V binary through an OpenCL API function.
4. The vendor-specific OpenCL runtime translates SPIR-V to LLVM IR, changes the target triple and data layout to suit the device which will execute the kernel, performs target specific optimizations, generates the ISA and executes the ISA on the device.

Step 4 of your “typical use case” includes "changes the target triple and data layout to suit the device which will execute the kernel”. It implies that SPIR-V is data layout agnostic since you can load it with any data layout, or there are (to be specified) constraint on what a “compatible” data layout is, or you considered that it is up to the OpenCL vendor to figure out what will work or not, with the drawback that any LLVM update can break its use case.

For OpenCL kernels, there is implicit data layout dependence when compiling the source to LLVM. Since SPIR-V is for common OpenCL platforms, a common data layout accepted by different OpenCL vendors is required. We choose the data layout which has been adopted by SPIR 1.2/2.0 for SPIR-V, since it has been successfully used for supporting consumption of SPIR 1.2/2.0 on various OpenCL platforms. For GLSL shaders, it is still under discussion whether to choose the same data layout as OpenCL, or a different data layout, or no data layout at all.

Location

From feedback of the previous version of the proposal, there are several suggestions about the location for the LLVM/SPIR-V converter:

1. llvm/lib/SPIRV only, adding an option to Clang for outputting SPIR-V. The advantage is ease of use for bi-way translation. However it does not reflect the fact that only LLVM IR with specific target triple and data layout can be translated to SPIR-V.

How important is it to “reflect it”?
The SPIR-V emitter could just assert on the data layout matching what is expected.

2. llvm/lib/SPIRV containing the main functionality of bi-way translation between LLVM IR and SPIR-V, llvm/lib/Target/SPIRV containing a thin wrapper as a target machine to allow Clang targeting SPIR-V. The advantage compared with 1 is that it allows a more conventional way of using Clang to produce SPIR-V. However it is subject to the same issue as 1 about not reflecting the requirement on the LLVM IR which can be translated to SPIR-V.

I don’t see why you think there is an issue on the data layout in this case. Currently every target in LLVM has an expected data layout, and if the IR has been processed with a different data layout than what the target expect, it may or may not work.
I believe your “thin wrapper” in llvm/lib/Target/SPIRV would declare a triple and data layout and then makes this requirement explicit.
This looks like the cleaner solution to me, I’m not sure I understand why you are more in favor of option 3?

Thanks,

3

Here is the revised proposal for the LLVM/SPIR-V converter. Please comment. Thanks.

Proposal of Adding SPIRV Target

Background

SPIR-V is a portable binary format for OpenCL kernels and GLSL shaders. A typical use case of SPIR-V is as follows:

  1. An application developer uses Clang to compile an OpenCL kernel source code to a SPIR-V binary which is common for all OpenCL platforms.

  2. The application developer ships the application containing the SPIR-V binary to customers.

  3. A customer runs the application on an OpenCL platform, which loads the SPIR-V binary through an OpenCL API function.

  4. The vendor-specific OpenCL runtime translates SPIR-V to LLVM IR, changes the target triple and data layout to suit the device which will execute the kernel, performs target specific optimizations, generates the ISA and executes the ISA on the device.

For OpenCL kernels, there is implicit data layout dependence when compiling the source to LLVM. Since SPIR-V is for common OpenCL platforms, a common data layout accepted by different OpenCL vendors is required. We choose the data layout which has been adopted by SPIR 1.2/2.0 for SPIR-V, since it has been successfully used for supporting consumption of SPIR 1.2/2.0 on various OpenCL platforms. For GLSL shaders, it is still under discussion whether to choose the same data layout as OpenCL, or a different data layout, or no data layout at all.

Location

From feedback of the previous version of the proposal, there are several suggestions about the location for the LLVM/SPIR-V converter:

  1. llvm/lib/SPIRV only, adding an option to Clang for outputting SPIR-V. The advantage is ease of use for bi-way translation. However it does not reflect the fact that only LLVM IR with specific target triple and data layout can be translated to SPIR-V.

  2. llvm/lib/SPIRV containing the main functionality of bi-way translation between LLVM IR and SPIR-V, llvm/lib/Target/SPIRV containing a thin wrapper as a target machine to allow Clang targeting SPIR-V. The advantage compared with 1 is that it allows a more conventional way of using Clang to produce SPIR-V. However it is subject to the same issue as 1 about not reflecting the requirement on the LLVM IR which can be translated to SPIR-V.

  3. llvm/lib/Target/SPIRV only. The advantage is that it reflects the requirement on the target triple and data layout for LLVM IR which can be translated to SPIR-V. However putting the SPIR-V to LLVM converter in the same directory is unconventional. Leaving the SPIR-V to LLVM converter out of LLVM source tree is also not desirable since OpenCL vendors need this functionality.

Our proposal is to take approach 3 and keep the bi-way converter in llvm/lib/Target/SPIRV. The functionality of the bi-way converter is exposed through llvm/include/Support/SPIRV.h.

4

It is just a header file. All source code is under lib/Target/SPIRV since they really belong there. Those who want to use the conversion functionality just need to link with the SPIRV target library.

Any suggestion about a better location for the header file?

Thanks.

Sam

5

If all the code is in lib/Target/SPIRV, I think the header should be under include/Target/SPIRV even if it is bi-directional. It doesn’t make sense under Support.

-Chris

6

I think you failed to understand my email then.

For me, not using the common and shared legalization framework is a complete deal breaker. As you say, this is not a stable representation for any generic IR, only for the subset targeting a specific platform. But without a legalization layer that maps from generic IR to that specific platform’s representation, we would be unable to change the canonical form that the middle end optimizers produce from IR (including the IR generated by an OpenCL frontend) without updating your legalization layer. If that legalization layer in turn is a completely separate layer from the SelectionDAG legalization layer, you’ve added a whole new burden on the LLVM community that doesn’t seem reasonable.

I think it is important that this effort shift to thinking of SPIR-V as a virtual ISA (admittedly a very special purpose one) and use common infrastructure for lowering and targeting it.

At the same time, I freely acknowledge that the current infrastructure in LLVM may not be ideal for this purpose today. I think it is incumbent on your group[1] to undertake the effort of making LLVM’s infrastructure better suited to your usecase rather than introducing a new set of infrastructure that is not shared with any other targets.

-Chandler

[1]: A footnote that is really a meta point, and not specifically about this proposal. When I say “you will have to make significant changes to the LLVM infrastructure that you need”, I’m not saying you should go away and start writing patches to this effect. Changing the core code generation infrastructure of LLVM is a really huge undertaking. If you want to do this, you’ll need to first build up a reputation within the LLVM community, trust of the various developers, etc. Don’t dive into the infrastructure first, you need to start with bugs, fixes, and small improvements. I realize this is a huge challenge for a lot of contributors, but changing heavily used infrastructure in really invasive ways is an extremely high-risk endeavor and I think it is reasonable that the community has a relatively high bar for contributors proposing to do that.

7

I completely agree.

8

Thanks for the suggestion. It makes sense.

Sam

9

Hi Mehdi,

Thank you for your comments. My comments are below.

Sam

10

I don’t know how the discussion about SelectionDAG impact your plan.
If the IR -> SPIR-V path is implemented as a “regular” target using the legalization framework and so on, almost all the code will be in lib/Target/SPIRV.
However it is not clear to me why the SPIR-V -> IR path would benefit in any way to be there as well?

Conceptually I should be able to compile LLVM and disable the SPIR-V backend but still be able to read-in SPIR-V and target my fancy OpenCL compliant device with my backend.

11

Hi Chandler,

Thank you for your comments. We could upstream the current implementation as an experimental target, and then refactor it to use SelectionDAG/MC while already present in the LLVM tree. That way people who want to target SPIR-V now can use the experimental target, while we work on implementing the backend properly. Since the current implementation of ‘legalization’ is like a simple LLVM transformation pass, we do not expect any significant burden on the LLVM community during transition to SelectionDAG/MC.

Sam

12

Hi Liu,

Thanks for the detailed proposal.

13

It doesn’t not seem “natural” to me that a “target” backend would operate as a “source”.
It is not clear also what piece of the target infrastructure would be helpful to convert from SPIR-V to IR.
At some point there was a C target backend, it still would seem silly to implement clang as a target.

14

+1

It's highly unusual for lib/Target to implement X-to-LLVM-IR conversions.

Eli

15

Hi Liu,

Thanks for the detailed proposal.

16

So we have been in discussions within the Khronos SPIR-V work group on our push to get our SPIR-V code into tip LLVM and have drawn the following conclusions;

  • We absolutely must create a fully fledged backend that uses all the machinery that target backends are expected to use.
  • We probably have to split out the SPIR-V → LLVM IR into a separate project from LLVM ala Clang et al.

As we want to allow developers to use the SPIR-V production/consumption code now, and the time sink doing the above approach would incur, we are going to open source the current work on the Khronos GitHub page as a first step.

We intend to revisit introducing a SPIR-V backend to LLVM in the future.

Cheers,
-Neil.

17

So we have been in discussions within the Khronos SPIR-V work group on
our push to get our SPIR-V code into tip LLVM and have drawn the
following conclusions;

  * We absolutely must create a fully fledged backend that uses all the
    machinery that target backends are expected to use.

Can you give details on how you reached this conclusion? Why was this
determined to be better than the alternatives?

Thanks,
Tom

18

Hey Tom,

Really it was at the behest of the replies - we got a lot of feedback from the mailing list that indicated we'd be putting extra workload of people changing features of the IR if we didn't follow the same mechanisms of the other backends (mostly led by Chandler's very astute comments on the subject).

Cheers,
-Neil.

19

Sorry I’ve not updated this thread more often.

One thing I want to clarify something for other (possibly future) readers of the email thread to avoid confusion. I don’t think the SPIR folks have misunderstood, but context is always hard to convey. =]

I’m not pushing for using the machinery for the sake of using it. That wouldn’t make sense. I think there are specific reasons why the legalization machinery (in whatever form that takes) should be common across backends. This primarily to help ensure that a growing number of LLVM backends doesn’t make changing the core or common parts of LLVM intractably hard by consolidating and unifying the infrastructure which provides minimally correct functionality for a backend.

I worry there are similar maintenance concerns with not having an MC-based encoding layer, but I’m not as deeply familiar with MC so I would defer to others there.

I also want to make it clear that I’m not under any illusion that these layers are really in good shape to serve the needs of SPIR-V, or other virtual ISAs. I think that any of the virtual ISA backends currently being developed will end up needing to do non-trivial work to refactor and improve these layers of LLVM in order to make them a better fit.

As much as I’d like to have SPIR-V in the tree (and I would), I’m honestly happier to see appropriate planning for the amount of work that will likely be required. I think that means that when it does come into the tree, it will be a much healthier experience.

-Chandler

20

Hey Chandler!

I think one of the issues we have from a SPIR-V side is that we don’t have any of the usual LLVM community heavyweights on-board to work with us, which in turn means the community is more likely to be resistant to any of the changes we would need to make that would affect every backend, which is understandable.

Given that the WebAssembly effort seems to have more of the usual suspects in the LLVM community invested in working on it, and they are likely to encounter at least some of the same problems that an eventual SPIR-V backend would also have to deal with, I sincerely hope that when we do come to integrate a SPIR-V backend in the future some of the harder refactors are already done ;]

Cheers,
-Neil.