Codecov Report

❌ Patch coverage is 33.33333% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.01%. Comparing base (9528a33) to head (39cacc4).

@@             Coverage Diff             @@
##           master    #3017       +/-   ##
===========================================
- Coverage   89.31%   76.01%   -13.31%     
===========================================
  Files         148      148               
  Lines       12995    12918       -77     
===========================================
- Hits        11606     9819     -1787     
- Misses       1389     3099     +1710     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

For the record, this PR addresses #1989.

I'll have to clean up the patch, but apart from that it's finished.

• There is a way to launch kernels with non-trivial cluster sizes
• There are new kernel functions to query about clusters
• There is distributed shared memory
• There is a cluster_wait function.

Missing cleanups:

• Probably needs more device_capability >= v"9.0" checks
• Requires LLVM >= 20, i.e. Julia >= 1.13; will need to add @static if statements for this
• Need to make tests pass
• CUDA.jl is broken for Julia 1.13. I have a work-around, but I assume someone else will clean this up, and I'll then have to merge it