• Add infix notation for comparators and math operations with scalars
  • A < B (creates expression; uses ewise_mult)
  • A + 1 (creates expression; uses apply)
  • A -= 1 (updates A)
  • Not valid: A + B; the user must choose ewise_add or ewise_mult
• Add matrix.ss.flatten() to flatten to a vector row-wise or column-wise
• Add vector.ss.reshape(nrows, ncols) to reshape to a matrix row-wise or column-wise
• Add “COOR” and “COOC” format options for importing and exporting matrices
  • COOR is coordinate format sorted row-wise
  • COOC is sorted column-wise
• Add configuration (via donfig) to enable/disable autocomputation of expressions
  • For example, with grblas.config.set(autocompute=True):
• Change Recorder() argument from record=False to start=True

Lots of big changes and new additions this release!

• Expressions can be used as values
  • expr.new() is no longer necessary (but is still recommended)
  • Methods like expr.apply(op) also work
  • Indexing expr[indices] does not yet work, nor do methods that mutate values
  • For example, A.ewise_mult(A.mxm(A.T).new()).new().reduce_scalar().new() can now be written as A.ewise_mult(A.mxm(A.T)).reduce_scalar()
• Add aggregators to grblas.agg that can be used in reductions!
  • For example, matrix.reduce_rows(agg.argmin)
  • Current aggregators: L0norm, L1norm, L2norm, Linfnorm, all, any, any_value, argmax, argmin, bitwise_all, bitwise_any, count, count_nonzero, count_zero, exists, first, first_index, geometric_mean, harmonic_mean, hypot, last, last_index, logaddexp, logaddexp2, max, mean, min, peak_to_peak, prod, root_mean_square, stdp, stds, sum, sum_of_inverses, sum_of_squares, varp, vars
• Add prefix scan (such as cumsum and cumprod) that works with any monoid
  • vector.scan(monoid)
  • matrix.scan_rows(monoid)
  • matrix.scan_columns(monoid)
• Add operator.get_semiring(monoid, binaryop) convenience function
• Add x.ss.build_scalar(...) for Vector and Matrix objects
• Add “coo” as import/export option
• Updated/fixed to_pygraphblas and from_pygraphblas

• Add Vector inner and outer products
  • v.inner(v, plus_times)
  • plus_times(v @ v)
• Fix computing A.ss.head() on iso-valued objects

• Update to SuiteSparse:GraphBLAS 5.1.3
• Add x.ss.is_iso to Vector and Matrix objects to indicate whether they are iso-valued

• Update to SuiteSparse:GraphBLAS 5.1.2.
• Add x.ss.pack and x.ss.unpack_* methods to Vector and Matrix objects that mirror import/export.
• Add x.ss.nbytes to get the number of bytes used by Vector and Matrix objects.

• Update to SuiteSparse:GraphBLAS 5.0.3
• Add new SuiteSparse GxB functions:
  • Matrix.ss.diag, Vector.ss.diag, and grblas.ss.diag
  • Matrix.ss.split
  • Matrix.ss.concat and grblas.ss.concat
• Add sane coercions to operators such as using exp on integer dtypes.
• Allow Scalars to be used in indexing
• Allow Scalars to be used when constructing numpy arrays
• Fix dtypes.unity for complex numbers (all pairs now match numpy behavior)
• Remove "pygraphblas" backend, as both libraries will use python-suitesparse-graphblas and be compatible

• Update grblas to be a pure Python package by depending on suitesparse-graphblas!
• Rename div to cdiv in semirings and add truediv and floordiv in semirings.
• Fix monoid attribute of semirings with land, lor, and lxor to coerce properly.
• Don’t allow expressions like v[w], which previously iterated over the values of Vector w.
• Don’t allow vectors or matrices to be directly cast to numpy arrays such as np.array(v).
  • Direct the users to e.g. v.to_values() instead.
• Support dup_op= keyword argument when converting from scipy.sparse and MatrixMarket formats.
• Improve recording by using name="" for scalars created during operations.
• Fix the recorder in Jupyter notebooks (again)

• Add "__weakref__" to __slots__ so objects are weakref-able.
• Test with Python 3.9

• Add infix sugar for ewise_plus, ewise_mult, mxm, apply, etc! For example:
  • C << op(A | B) # ewise_plus
  • C << op(A & B) # ewise_mult
  • C << op(A @ B) # mxm, vxm, mxv
  • C << op(A) # apply unary
  • C << op(A, 1) # apply binary, right=1
  • C << op(1, A) # apply binary, left=1
• Operators:
  • Add grblas.op namespace that has all unary, binary, and semiring objects.
  • Automatically convert binaryop to monoid or vice versa where appropriate if possible.
    • Hence, one can use a binaryop from grblas.op namespace as a monoid.
  • Add attributes such as binop.monoid, monoid.identity, and semiring.binaryop to operators.
  • Don’t allow a semiring where a monoid or binaryop is expected (do e.g. semiring.monoid instead).
  • Renamed grblas.ops to grblas.operator, which defines the base classes of operators.
• Add my_matrix.ss.head(n) and my_vector.ss.head(n) that returns the indices and values of n elements.
• Improve repr of very sparse Vector and Matrix objects by using the new C.ss.head() methods.
• Add grblas.replace singleton, so one can do e.g. C(mask, replace) << A.mxm(V).
• Add ability to read and write Matrix Market files with grblas.io.mmread and grblas.io.mmwrite.
• Allow Scalar, Vector, and Matrix objects to be serialized with pickle.
• Added __slots__ to all the things.

• Update to SuiteSparse 4.0.3 (new unary and binary ops, etc.)
• Added Windows support (without complex dtypes)
• Use NumPy’s malloc/free, so Python knows memory usage of GraphBLAS objects
• Many improvements to import and export with SuiteSparse
  • Renamed fast_export to export and fast_import to import_any
  • Added other import methods such as Matrix.ss.import_csr
  • support zero-copy import and export if possible (user must opt in)
  • formats: csr, csc, hypercsr, hypercsc, bitmapr, bitmapc, fullr, fullc
• Matrix and Vector
  • allow contains expressions
    • 1 in v, True if there is a value at v[1]
    • (1, 2) in A, True if there is a value at A[1, 2]
  • iteration
    • Yields the indices of the structure
  • Add .wait() method
  • Add format attribute as e.g. my_matrix.ss.format
    • One of “csr”, “hypercsc”, “bitmapr”, etc
• Scalar
  • Add .shape attribute (like Vector and Matrix, and like numpy scalars)
  • Able to cast to float and int
  • Allow things like __eq__ where .value is or computes a scalar
• Don’t allow __eq__ or __bool__ where they don’t make sense