CuMetal v0.7.1

CuMetal v0.7.1

Release index

CuMetal 0.7.1 is a correctness release. CUDA event timing now measures GPU
execution rather than host submission. A possible hang in typed
grid.sync() is fenced. System atomics and roundf compile and keep their
semantics on both backends. The nvcc-compatible driver accepts the options
real CMake CUDA projects emit.

Highlights

• GPU-timed events. cudaEventElapsedTime reads the GPU end times of
  the Metal command buffers that cudaEventRecord commits. Previously it
  subtracted CPU timestamps from submission, so a benchmark with ~112 ms of
  GPU work reported 0.09 ms.
• Grid sync fence. The typed backend's grid.sync() leader snapshotted
  the barrier generation and published its arrival with unfenced relaxed
  atomics. A reordering could make it miss the release and spin forever,
  matching a hang reported on M1 Pro (#175). Validated on M4 Pro, where the
  race did not reproduce; M1 Pro confirmation is pending.
• System atomics and roundf. Clang 22's 32-bit NVVM system atomic
  intrinsics import into the existing coherent-UMA lowering, and unsigned
  system min/max keep unsigned comparisons. __nv_roundf no longer emits
  llvm.round.f32, which could fail Metal pipeline creation, and preserves
  signed zero on both backends.
• nvcc drop-in. --fp64=MODE, -isystem, -Xcompiler= and
  --compiler-options (also forwarded to linking), and
  --options-file/-optf response files are honoured. SYSTEM_INCLUDES
  is reported as -isystem, so find_package(CUDAToolkit) plus
  CUDA::cudart configures cleanly.
• Metal emission. Null-pointer comparisons and LLVM aggregates with
  template-like names, such as CUB BlockReduce, produce valid MSL.
• Install. The installed cumetal_bench keeps its signature and finds
  libcumetal.dylib without DYLD_LIBRARY_PATH.

Verification

On Apple M4 Pro, macOS 26, at 0271417 plus the version bump:

• Merged PR set, Debug with the binary shim ON: 450 of 454 passed in a
  -j 6 run. The four failures passed when rerun alone.
• Same set, Debug with the binary shim OFF: 448 of 453 passed in a
  -j 4 run. The five failures passed when rerun alone.
• The grid-sync fence: the cooperative-grid, cooperative-groups,
  backend-matrix and both corpus tests passed, and a unit assertion fails
  without the fix. A later full run lost the macOS Metal compiler service
  partway through; the 20 runtime-JIT tests that failed then do not use
  grid sync, and their generated code is unchanged by the fix.

These totals are evidence for the enrolled test surfaces, not a
percentage of the complete CUDA ecosystem.

Important limits

• Event intervals use Metal command-buffer timestamps, so they include any
  idle gap between command buffers and are not per-kernel counters.
• System-atomic import covers relaxed 32-bit integer forms on device
  storage. Wide and floating system atomics, other scopes and
  shared/private targets are not added.
• Linking existing objects with --fp64 does not change their arithmetic
  mode; select it when compiling each source.
• Static and anonymous-namespace kernels in native-AOT executables
  (#180) are not in this release.

Contributors

Thank you to the people who found, fixed and validated the work in this
release:

• Christopher Albert (@krystophny)
  reported #181–#189 and #198 while bringing the
  SIMPLE CUDA sources and a real CMake
  build to CuMetal, then fixed each one with a regression test:
  GPU-timed CUDA events (#199), the nvcc driver's --fp64, -isystem,
  -Xcompiler= and options-file handling and CMake system-include metadata
  (#192–#196), valid Metal emission for null pointers and template-like
  aggregates (#191, #197), the installed cumetal_bench (#190), and faster
  PTX import (#200).
• Harrison O'Connor-Hoover
  (@harrisonoconnorhover)
  brought Clang 22 system atomics to the source path and kept unsigned
  system min/max unsigned (#177), made roundf bit-exact on both
  backends (#178), and reported the cooperative grid-sync hang (#175)
  with the diagnostic that located it.

See the changelog for the complete list.