[PATCH] D112718: Add intrinsics and builtins for PTX atomics with semantic orders

Tue Jan 11 02:50:25 PST 2022

Naghasan added inline comments.

================
Comment at: clang/include/clang/Basic/BuiltinsNVPTX.def:1057
+
+BUILTIN(__nvvm_atom_xchg_global_i, "iiD*i", "n")
+TARGET_BUILTIN(__nvvm_atom_cta_xchg_global_i, "iiD*i", "n", SM_60)
----------------
tra wrote:
> t4c1 wrote:
> > tra wrote:
> > > t4c1 wrote:
> > > > tra wrote:
> > > > > We need to figure out how address-space-specific builtins are supposed to work.
> > > > > Right now two competing approaches.
> > > > > 
> > > > > This patch declares builtins with generic pointer as an argument, but, according to the test, expects to be used with the AS-specific pointer.
> > > > > It probably does not catch a wrong-AS pointer passed as an argument, either.
> > > > > It does happen to work, but I think it's mostly due to the fact that LLVM intrinsics are overloaded and we happen to end up addrspacecast'ing  things correctly if the builtin gets the right kind of pointer.
> > > > > 
> > > > > The other approach is to declare the pointer with the expected AS. E.g:
> > > > > > TARGET_BUILTIN(__nvvm_mbarrier_init_shared, "vWi*3i", "", AND(SM_80,PTX70))
> > > > > 
> > > > > IMO, this is the correct way to do it, but it is also rather inconvenient to use from CUDA code as it operates on generic pointers.
> > > > > 
> > > > > @jdoerfert - WDYT?
> > > > >TARGET_BUILTIN(__nvvm_mbarrier_init_shared, "vWi*3i", "", AND(SM_80,PTX70))
> > > > >IMO, this is the correct way to do it, but it is also rather inconvenient to use from CUDA code as it operates on generic pointers.
> > > > 
> > > > I tried doing this, however it is also completely unusable from OpenCL code (which is our use case). Trying to use it gives you errors about how casting pointers to different address spaces  - for example from local to AS3 is not allowed.
> > > Hmm. It should've worked. It would need the same explicit cast to a pointer in AS(3) as in your tests, but it would safeguard against attempts to pass it a generic pointer. E.g. https://godbolt.org/z/qE6oxzheM
> > > 
> > > Explicit casting to AS(X) for AS-specific variants is annoying, but it's probably unavoidable at the moment regardless of whether we declare the pointer argument to be AS-specific or not.  LLVM will not always be able to infer that a pointer is in particular AS.
> > > Using specific AS in the declaration has a minor benefit of safeguarding at compile time against unintentional use of generic pointers.
> > > 
> > > Ideally we may want to convert generic variant of the builtin to AS-specific one, if LLVM does know the AS. We currently do this for loads/stores, but not for other instructions.
> > > 
> > Well, it does not work. See: https://godbolt.org/z/vM6bKncc4. Declaring the pointer to be in generic AS is the only way I got it to work. If you know a way to call a builtin declared with AS numbers in OpenCL, I am happy to do that instead.
> Could you help me understand how different address spaces are handled by OpenCL in clang and LLVM?
> What exactly are `__local` or `__private` in OpenCL? Do they map to LLVM's address spaces? 
> Clang does complain that we're trying to change AS, but I do not see AS used in the IR: https://godbolt.org/z/soajoE991
> 
> AFAICT what happens is:
> * OpenCL does use explicit AS for pointers (`__local`, `__private` seem to pop up in the error messages)
> * `__attribute__((address_space(3)))` does not match the AS of `__local` and that leads to an error.
> * generic pointer argument works because we are allowed to cast any specific AS to generic one.
> 
> In other words, having specific-AS arguemt works as intended, we just have a mismatch between AS number used by OpenCL and AS number used by NVPTX and we're not allowed to cast between two specific AS.
> 
> If that's indeed the case, then what we appear to be missing is the mapping from OpenCL AS to the target-specific AS, which should, ideally, be done by clang itself. So, if NVPTX-specific builtin operating on shared pointer is called with a pointer in OpenCL's equivalent of CUDA's `__shared__`, it should be mapped to AS(3).
> 
> Could you help me understand how different address spaces are handled by OpenCL in clang and LLVM?

clang makes a strong distinction between language and target address spaces since ~3.6 (was more loose before). Each target in clang defines a map between language and target address space (e.g. https://github.com/llvm/llvm-project/blob/main/clang/lib/Basic/Targets/SPIR.h#L25). The map is used in clang codegen to lower `__local` to the right target address space.

> Clang does complain that we're trying to change AS, but I do not see AS used in the IR: https://godbolt.org/z/soajoE991

Same code but targeting SPIR: https://godbolt.org/z/4E3brzodq (ARM maps all languages address spaces to 0)

> having specific-AS arguemt works as intended, we just have a mismatch between AS number used by OpenCL and AS number used by NVPTX and we're not allowed to cast between two specific AS

OpenCL languages do map to nvptx target address space, it is just form clang's semantic a language asp is strongly distinct from target ones (lang asp simply isn't understood as a number). So you can't mix and match without an explicit cast, even if the lang asp eventually lowers to that target.

What is missing is the ability for the target builtin to accept a language asp if it lowers to the target asp.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D112718/new/

https://reviews.llvm.org/D112718