Add amdgpu intrinsics

[Flakebi]

Add intrinsics for the amdgpu architecture.

I’m not sure how to add/run CI (ci/run.sh fails for me e.g. for
nvptx because core cannot be found), but I checked that it compiles
without warnings with

$ CARGO_BUILD_RUSTFLAGS=-Ctarget-cpu=gfx900 cargo check --target amdgcn-amd-amdhsa -Zbuild-std=core

Tracking issue: rust-lang/rust#149988

[rustbot]

r? @sayantn

rustbot has assigned @sayantn.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[sayantn]

I think it can be "tested" the same way we do core::arch::nvptx (this is also a similar NO_RUN, NO_STD target). For those targets, we just test in CI that it builds. You can just look at how we handle nvptx64-nvidia-cuda in .github/workflows/main.yml. Also you would need to add this target to ci/dox.sh to check that the documentation builds

[Flakebi]

Thanks, I tried to replicate what’s there for nvptx + adding -Ctarget-cpu (as amdgpu doesn’t compile without, it has no generic cpu) + build-std=core.
I still can’t build ci/run-docker.sh or ci/run.sh locally (ci/dox.sh seems to work if I remove s390x), but I think I got it working now.
Sorry for the many force-pushes!

The diff to my first push from when opening this PR is here: https://github.com/rust-lang/stdarch/compare/b3f5bdae0efbdd5f7297d0225623bd31c7fe895b..e1028110e77561574bfb7ea349154d46b5ea7b86

[sayantn]

Thanks, I have put some comments

[sayantn]

Also I have noticed that Clang provides a different set of intrinsics than these (amdgpuintrin.h and gpuintrin.h). Is this divergence intentional? We normally stick to following the C conventions in stdarch

[Flakebi]

Thanks for the review!

Interesting, I thought core_arch is more akin to the __builtin_* family of functions in clang.
The clang __gpu intrinsics are definitely useful, though they are (to no surprise) missing low-level, hardware-specific intrinsics (e.g. wave_id, sethalt and buffer/image intrinsics, which I can’t add yet because they need simd type support in rustc).

I do plan to write a common gpu module for amdgpu and nvptx as well.
Maybe this one should follow gpuintrin.h?
We could e.g. have

mod amdgpu {
    mod gpu {
        use super::*;
        pub fn block_id_x() -> u32 {
            workitem_id_x()
        }
    }
}
// Same for nvptx

mod gpu {
    #[cfg(target_arch = "amdgpu")]
    pub use amdgpu::gpu::*;
    #[cfg(target_arch = "nvptx64")]
    pub use nvptx::gpu::*;

    // + more intrinsics as in gpuintrin.h
}

[sayantn]

The analogue to __builtin functions in C are #[rust_intrinsic] functions and linked LLVM intrinsics. They are generally not exposed to users (in stable) as they are tightly linked to the compiler implementation.

If there are some interesting platform-specific intrinsics, we can add them, but generally we follow GCC and Clang in regards to intrinsics.

Yea, a common gpu module can be added following gpuintrin.h.

[Flakebi]

Thanks for the review, that’s a nice case for const generics! I fixed all your inline comments.

Also thanks for linking the clang intrinsics, I tend to forget about those.
As far as I am aware, the moste used APIs for GPUs are CUDA/HIP/SYCL on the compute side and DirectX/Vulkan (through HLSL/GLSL) on the graphics side. Although clang is a building block for these, the clang intrinsics are most often not used directly. Therefore I’m rather hestitant to follow clang for GPUs.

As common GPU intrinsics is a larger topic, I started a Discourse post here (please add your thoughts!): https://internals.rust-lang.org/t/naming-gpu-things-in-the-rust-compiler-and-standard-library/23833
Though, I think progress on this doesn’t need to block amdgpu intrinsics (note that I don’t have a plan to stabilize amdgpu intrinsics, stabilizing generic GPU intrinsics seems more desirable).

My planned intrinsic “stack” is roughly as follows, from high-level to low-level:

1. Rust exposes (experimental) generic gpu intrinsics, e.g. at core::gpu, naming tbd, see the Discourse post
2. These are implemented for amdgpu and nvptx by calling target-specific intrinsics from core::arch::amdgpu/nvptx
3. core::arch::amdgpu/nvptx intrinsics expose (not on stable) their respective LLVM intrinsics, naming can follow the LLVM intrinsic names

We could also use clang names for core::arch::amdgpu/nvptx, then we would end up with three set of names:

1. Generic core::gpu intrinsics: Rust-specific names
2. core::arch::amdgpu/nvptx: clang names
3. LLVM intrinsics (used in core::arch): LLVM names

I would prefer following LLVM’s intrinsic names in core::arch::amdgpu/nvptx instead of clang’s names.
What do you think?

Note for my future self: I got run.sh to work with NORUN=1 NOSTD=1 TARGET=amdgcn-amd-amdhsa CARGO_UNSTABLE_BUILD_STD=core ci/run.sh.

[sayantn]

Thanks, left some more comments

[sayantn]

I don't have any strong preferences, if you feel that the lower-level LLVM intrinsics are more useful that just having the clang intrinsics, go ahead with that -- we can change it, remove it, or add the clang intrinsics too if required (as long as they are unstable)

[Flakebi]

Thanks for taking such a thorough look!
I fixed the (s.)wave.barrier and permlane64.i32 and answered your questions inline. I think the “works only on some target-cpus” is the most difficult one. I’m thinking, as long as it’s unstable, this state is fine. But that’s because I have no good idea to fix this, so if you or someone else has an idea, please let me know.

[sayantn]

I discovered some more problems with the intrinsics, wave.reduce.f{add,sub,max,min} and llvm.s.barrier.leave are not in the LLVM release branch yet, they were added like 1 month ago. So they will fail to compile once someone actually monomorphises the functions. We should add them when the changes actually propagates to the rust fork (~3 months probably), otherwise it will cause link failures

Also, you probably forgot to add intrinsics for the add and fadd ones.

The other changes LGTM

[Flakebi]

I fixed your comments and added a tests mod with instructions to check that generating the instructions works.
I changed the default target-cpu to gfx12, that seems like the one that supports most intrinsics (it doesn’t support s_memrealtime, s_get_waveid_in_workgroup, permlane16_swap and permlane32_swap).

[sayantn]

LGTM, just one nit

Edit: sorry, my mistake. No other problems. Thanks ❤️