You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Support numba 0.62+ versions and migrate to LLVM 20
Numba 0.62+ migration:
- Migrate to new pass manager and update IR codegen
- Add LLVM IR debug output for the outputs of the LLVM plugin pass
- Update libomptarget nvptx device RTL path for the new pass manager
- Enable and optimize toolchain binary generation by default for the CUDA target to match LLVM runtime ELF loader expectations
- Improve detection for openmp libraries to include system libraries and better error checking
LLVM 20 migration:
- Update to opaque pointer types and modernize LLVM API usage
- Use tgt_target_kernel interface for offloading codegen to match ABI conventions
- Update pass to use LLVM 20 APIs
- Replace deprecated CUDA nvvm.annotations with ptx_kernel calling convention
- Update tgt_register_lib and tgt_unregister_lib handling in ctor
Packaging improvements:
- Add ninja build system requirement for faster, more reliable builds
- Build openmp libraries conditionally via environment variables, to re-use existing packages and avoid conflicts
- Build manylinux wheels using native Clang/LLVM 20.1.8 version
- Update cibuildwheel to 3.3.1
- Ensure building for the macos 11.0 deployment target
Additional improvements:
- Set CUDA context flag CU_CTX_SCHED_BLOCKING_SYNC to match openmp runtime expectations
- Fix thread_limit semantics to align with omptarget defaults
- Improve build system with better CC/CXX flag handling
- Update Python version support and CI configuration
- Fix tests to be resilient to different thread configurations on test platforms
- Fix ploop.1.py example
0 commit comments