[PATCH] D99488: [SYCL][Doc] Add address space handling section to SYCL documentation

Mon Mar 29 12:09:29 PDT 2021

Anastasia added a comment.

Do you plan to implement any of the following restriction?

> To allocate local memory within a kernel, the user can either pass a sycl::local_accessor object as a argument to an ND-range kernel (that has a user-defined work-group size), or can define a variable in work-group scope inside sycl::parallel_for_work_group.

> Explicit pointer class values cannot be passed as arguments to kernels or stored in global memory.

Can I confirm that in your implementation any raw pointer obtained from `multi_ptr` will be in generic/default address space? This is something that might be worth adding in the documentation unless it is explained in the spec already?

================
Comment at: clang/docs/SYCLSupport.md:821
+SYCL compiler toolchain. Section 3.8.2 of SYCL 2020 specification defines
+[memory model](https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#_sycl_device_memory_model),
+section 4.7.7 - [address space classes](https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#_address_space_classes)
----------------
> The memory model for SYCL devices is based on the OpenCL 1.2 memory model.

Is this possibly a spec bug? OpenCL didn't have generic address space in v1.2, it has only been added in v2.0.

https://www.khronos.org/registry/OpenCL/specs/3.0-unified/html/OpenCL_C.html#the-generic-address-space

================
Comment at: clang/docs/SYCLSupport.md:830
+Similar to other single-source C++-based GPU programming modes like
+OpenMP/CUDA/HIP, SYCL uses clang's "default" address space for types with no
+address space attributes. This design has two important features: keeps the type system consistent with C++ on one hand and enable tools for emitting device code aligned with SPIR memory model (and other GPU targets).
----------------
Is this explained somewhere would you be able to add any reference?

================
Comment at: clang/docs/SYCLSupport.md:851
+
+Changing variable type has massive and destructive effect in C++. For instance
+this does not compile in C++ for OpenCL mode:
----------------
aaron.ballman wrote:
> 
> This example demonstrates the problem with compiling C++ code when address space type qualifiers are inferred.
> 
>     The example compiles in accordance with OpenCL language semantic...
> 
> https://godbolt.org/z/9jzxK5xc4 - ToT clang doesn't compile this example.

I am still not clear what message you are trying to convey here? In OpenCL kernel languages any object is always in some address space so if you write the following `decltype(p)`, it will always have address space attribute in a type. OpenCL spec is very explicit about this:

https://www.khronos.org/registry/OpenCL/specs/3.0-unified/html/OpenCL_C.html#addr-spaces-inference

So if you compare a type not attributed by an address space with an attributed one they will never compare as equal because according to C++ rules if the qualifiers differ the types will differ. You need to use a special type trait to remove an address space if you need to compare types not qualified by an address space. What is important to highlight however is that address space inference is where OpenCL differs to C or C++. But of course, neither C nor C++ have address spaces so it is hard to compare.

In relation to your documentation, it is not clear what you are trying to achieve with this paragraph?

================
Comment at: clang/docs/SYCLSupport.md:909
+| `__attribute__((opencl_local))` | local_space |
+| `__attribute__((opencl_private))` | private_space |
+
----------------
Since SYCL spec has constant AS you should explain whether it is going to be supported or not and if so then how.

================
Comment at: clang/docs/SYCLSupport.md:914-919
+Default address space represents "Generic-memory", which is a virtual address
+space which overlaps the global, local and private address spaces. SYCL mode
+enables conversion to/from default address space from/to address space
+attributed type.
+
+SPIR target allocates SYCL namespace scope variables in global address space.
----------------
Naghasan wrote:
> aaron.ballman wrote:
> > 
> I think this section should be extended.
> 
> Pointers to `Default` address space should get lowered into a pointer to a generic address space (or flat to reuse more general terminology).
> But depending on the allocation context, the `default` address space of a non-pointer type is assigned to a specific address space. This is described in https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#subsec:commonAddressSpace.
> 
> This is also in line with the behaviour of CUDA (small example https://godbolt.org/z/veqTfo9PK).
Ok, if the implementation plans to follow the spec precisely then just adding a reference should be sufficient. 

Do I understand it correctly that your implementation will use the first approach from the two described in:
> If the target of the SYCL backend can represent the generic address space, then the "common address space deduction rules" in Section 5.9.2 and the "generic as default address space rules" in Section 5.9.3 apply. If the target of the SYCL backend cannot represent the generic address space, then the "common address space deduction rules" in Section 5.9.2 and the "inferred address space rules" in Section 5.9.4 apply.

This should be added to the documentation btw.

Btw does this statement in any way relate to the following statement:

> Within kernels, the underlying C++ pointer types can be obtained from an accessor. The pointer types will contain a compile-time deduced address space. So, for example, if a C++ pointer is obtained from an accessor to global memory, the C++ pointer type will have a global address space attribute attached to it. The address space attribute will be compile-time propagated to other pointer values when one pointer is initialized to another pointer value using a defined algorithm.

from https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#_access_to_memory

Or if not where can I find the algorithm it refers to?

================
Comment at: clang/docs/SYCLSupport.md:915
+Default address space represents "Generic-memory", which is a virtual address
+space which overlaps the global, local and private address spaces. SYCL mode
+enables conversion to/from default address space from/to address space
----------------
aaron.ballman wrote:
> 
You should also explain what address spaces are super/sub-sets because this impacts implicit and explicit conversion behavior in an embedded C-like models. In relation to that, you should highlight that the private, local and global ASes are disjoint.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99488/new/

https://reviews.llvm.org/D99488