sparsediff 0.6.1
- The bundled SparseDiffEngine moves from 0.3.0 to 0.6.1, the release
that ‘sparsediffpy’ 0.6.x and ‘CVXPY’ 1.9.3 use. From this release on,
the package version matches the bundled engine version.
- Fixed: a parametric problem re-evaluated after
sd_update_params() could return values computed from the
old parameter values (engine fix #107). For a modeling layer this meant
that re-solving a parametric nonlinear problem after changing a
parameter returned the previous solution, reported as optimal.
- Fixed: selecting entries with repeated indices in
sd_index() (for example a quadratic objective over a
symmetric variable) could crash R with a bus error when the Jacobian or
Hessian was evaluated (engine fix #105).
- New
sd_quad_form_dense(): the quadratic form with a
dense constant or parametric matrix. The parametric form follows
sd_update_params(). sd_quad_form() keeps its
signature and still takes a sparse constant matrix.
- Every ‘sparsediffpy’ 0.6.1 function now has an
sd_*
counterpart that calls the same engine routine with the same arguments;
the name map is on the package help page (?sparsediff).
New: sd_left_kron() and sd_right_kron()
(Kronecker products, used by the next CVXPY release), the
compressed-sparse-row derivative accessors sd_jacobian(),
sd_get_jacobian(), sd_init_hessian(),
sd_hessian() and sd_get_hessian(), and
sd_get_expr_dimensions() / sd_get_expr_size().
sd_rel_entr() now dispatches on operand size like
make_rel_entr.
- Fixed:
sd_init_hessian_coo() crashed R when called
before the Jacobian was initialized; it now initializes the Jacobian
first.
sd_rel_entr(), sd_rel_entr_first_scalar()
and sd_rel_entr_second_scalar() now check that their
arguments are two different variables, which the engine requires; other
arguments overflowed a heap buffer (found with AddressSanitizer).- The BLAS shim gains
cblas_ddot, which the 0.6.1 engine
calls.
- Fixed: the package failed to compile on Linux systems with glibc
2.36 or older (for example Debian 12), because
_GNU_SOURCE
was defined too late for fopencookie(). It is now set on
the compiler command line.
- Documentation: the sparse matrix arguments of
sd_quad_form() and sd_left_matmul() /
sd_right_matmul() are compressed-sparse-row arrays, and the
dense data arguments are row-major. Earlier documentation
said compressed-sparse-column and column-major. The grouped help pages
now have usage sections (flagged by current R-devel).
- The package now has a ‘testthat’ test suite.
sparsediff 0.4.0
- First CRAN release, bundling SparseDiffEngine 0.3.0.