v21.11.00

This is the initial public alpha release of cuNumeric, an aspiring drop-in replacement for NumPy at scale.

Conda packages for this release are available at https://anaconda.org/legate/cunumeric.

What's Changed

• Refactoring for the broadcasting logic by @magnatelee in #18
• Improved partitioning and sharding for GEMV by @manopapad in #37
• Fix #16 by @manopapad in #38
• Add CI by @marcinz in #43
• Use a script on the runner to checkout CI repository by @marcinz in #44
• Fix CI by @marcinz in #45
• Extend tests with CPU/GPU/OMP testing by @marcinz in #48
• Remove accidental part of the job matrix from CI by @marcinz in #49
• Add missing alignment constraints for matrix-vector multiplication by @magnatelee in #58
• Force left alignment for pointers and references by @magnatelee in #59
• Don't alter the GC priority �for external instances by @magnatelee in #60
• Be strict when importing legate.numpy in examples by @manopapad in #61
• Fix for reinterpret casts that are actually unsafe in the modern c++ by @magnatelee in #62
• Remove the return type of the void-returning function in the mapper by @magnatelee in #63
• Remove dependency on numpy>=1.20 by @manopapad in #64
• Stop using looping templates by @magnatelee in #65
• Bug fix for release mode by @magnatelee in #66
• Port nozero to the new buffer API by @magnatelee in #68
• Missing constraint for bincount by @magnatelee in #69
• Clean up install script by @manopapad in #70
• Fixes to compile on MacOS by @manopapad in #71
• Disable absolute and allcose for complex types only with Clang by @magnatelee in #72
• Generalize the reshape operator by @magnatelee in #73
• Improve dot product for half precision floats by @magnatelee in #74
• Support for tensordot by @magnatelee in #75
• Bugfixes on operations by @manopapad in #76
• Add missing type casts for __half by @magnatelee in #77
• Pull the correct Core image by @marcinz in #78
• Port remaining fixes from old branch by @manopapad in #80
• Remove remaining conditional legate.numpy imports from examples by @manopapad in #81
• Always dump test output by @marcinz in #83
• Minor code cleanups by @manopapad in #85
• Attempt to address #84 by @manopapad in #86
• Always follow the core's choice regarding CUDA/OpenMP support by @manopapad in #88
• Fix legate data interface by @magnatelee in #92
• Handle overlapping stores correctly in dot by @magnatelee in #93
• Improvements to handling of scalar arrays by @manopapad in #90
• Port to the new calling convention by @magnatelee in #89
• Prevent CI on forks by @marcinz in #94
• Emptiness checks for matrix ops by @magnatelee in #95
• Mapper update by @magnatelee in #82
• Port to the new reduction op interface by @magnatelee in #96
• Stop using delinearization by @magnatelee in #97
• Dead code elimination by @magnatelee in #98
• Reorganizing source files by @magnatelee in #99
• Remove leftover requirements.txt by @manopapad in #100
• Update for build system changes by @manopapad in #101
• Updates for new attachment interface by @manopapad in #102
• Fix for matrix-vector multiplication by @magnatelee in #104
• Another attempt to fix degenerate cases by @magnatelee in #107
• Fix #111 by @magnatelee in #116
• Release 21.11.00 by @marcinz in #121

New Contributors

• @marcinz made their first contribution in #43

Full Changelog: https://github.com/nv-legate/cunumeric/commits/v21.11.00