Support checked physical and logical Warp collectives by tpn · Pull Request #1581 · NVIDIA/cuda-rust

tpn · 2026-10-21T00:46:42Z

Native collective calls can use controls and memory regions that agree within a physical warp or a power-of-two logical warp of 1–32 lanes. The compiler checks participation and conflicting accesses, recognizes containing barriers, and permits finite sibling loops to reconverge before a wider collective. Genuine group queries supply the identity facts; ordinary lane arithmetic does not.

Analysis runs in rustc-free prepared MIR after mem2reg. The existing commits keep the physical-Warp stage and logical-Warp and finite-loop stage separately reviewable. This incorporates the former #1584 scope; Warp scratch remains in #1586.

Recorded validation includes host and strict checks, 64 physical and 128 logical compiler cases, ordinary Git GPU examples and controls, sanitizers, expected traps and health replays. Logical-Warp Full-debug execution also passes. Fresh HTTPS builds match the scoped executable sections, ABI, resources and launch facts against their ordinary builds; those receipts do not claim HTTPS GPU execution. Whole artifacts and older-toolchain code differ, with the differences retained. The accepted source is unchanged by consolidation; current-head CI is checked separately.

Builds on #1575. Review the combined 16-file change (+1,494/−190). GitHub’s main-based Files tab includes prerequisites; a narrow stacked base requires a maintainer-created upstream branch or prerequisite landing. This PR remains draft.

With `lto = "fat"` or `lto = "thin"`, the build of a binary that uses a library with kernels stopped with an error: "failed to get bitcode from object file for LTO (Can't find section .llvmbc)". rustc keeps a list of the Rust objects in each rlib. This list contains the objects of all modules. With LTO, rustc reads bitcode from these objects and removes them from the link. The artifact objects have no bitcode. If a session writes an rlib, the backend now gives each artifact object to rustc as the DWARF object of its module. This object is in the rlib, but it is not in the list. LTO does not read this object, and the linker gets it from the rlib. The system linker does not get the DWARF objects. If the session also writes a binary or a dynamic library, the backend gives the artifact objects to this linker as link arguments. With `-Zlink-native-libraries=no`, rustc does not give link arguments to the linker. Then the artifact objects of this session are Rust objects, as before. The book said that `join_codegen()` and `link()` only call the LLVM backend. Now it tells what they add to the outputs. CI builds `cuda_module_in_lib` with fat LTO. The library is an rlib, and then an rlib and a cdylib. After each build, CI makes sure that the binary has the PTX bundles of the library and of the binary. It also makes sure that the cdylib has the PTX bundle of the library. Fixes NVIDIA#1495 Signed-off-by: Mikhail Borisov <mikhail.boriso2014@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Use Cargo target names for standalone emit-ltoir artifacts. Bind the compiler and reader to the same intermediate directory when invoked from a Cargo workspace member. Signed-off-by: Trent Nelson <trent@trent.me>

Accept ordered external LTO files when materializing embedded cubins. Include their content in the codegen fingerprint and verify owned bytes before linking through the existing finalizer. Preserve inputs that overlap generated output paths and diagnose unresolved device symbols. Add a standalone CUDA FFI example and a native regression script. Signed-off-by: Trent Nelson <trent@trent.me>

Reuse existing example dependency pins and classify external LTO as an FFI fixture. Run its linking regression in native sweeps and verify relocatable debug output in the existing full-debug route. Signed-off-by: Trent Nelson <trent@trent.me>

Preserve exact geometry and fixed dynamic shared-memory requirements through artifact loading so preparation can select and validate the compiled allocation. Keep artifact writers and readers on one in-tree implementation while preserving legacy bundles without launch facts. Signed-off-by: Trent Nelson <trent@trent.me>

Combine the reviewed compiler prerequisites for native library calls. Keep the external LTO fixture on the shared in-tree artifact implementation. Signed-off-by: Trent Nelson <trent@trent.me>

Use the reviewed Cargo target and output-directory correction in the combined compiler prerequisite tree. Signed-off-by: Trent Nelson <trent@trent.me>

Resolve concrete registration functions and scalar argument types during device compilation, then invoke the native provider and link its owned LTO output through the ordinary artifact finalizer. Signed-off-by: Trent Nelson <trent@trent.me>