Changelog

Changelog

This file only contains the currently unreleased functionality. For the complete Changelog of currently released functionality, please visit our CHANGELOG.md instead.

Unreleased

Version support changes

• Ginkgo now requires a compiler with C++ 17 support #1603

Interface changes

Behavior changes

Deprecations

• The Executor::run overload without a name as first parameter has been deprecated #1667

Summary of previous deprecations

• The device_reset parameter of CUDA and HIP executors no longer has an effect, and its allocation_mode parameters have been deprecated in favor of the Allocator interface.
• The CMake parameter GINKGO_BUILD_DPCPP has been deprecated in favor of GINKGO_BUILD_SYCL.
• The gko::reorder::Rcm interface has been deprecated in favor of gko::experimental::reorder::Rcm based on Permutation.
• The Permutation class' permute_mask functionality.
• Multiple functions with typos (set_complex_subpsace(), range functions such as conj_operaton etc).
• gko::lend() is not necessary anymore.
• The classes RelativeResidualNorm and AbsoluteResidualNorm are deprecated in favor of ResidualNorm.
• The class AmgxPgm is deprecated in favor of Pgm.
• Default constructors for the CSR load_balance and automatical strategies
• The PolymorphicObject's move-semantic copy_from variant
• The templated SolverBase class.
• The class MachineTopology is deprecated in favor of machine_topology.
• Logger constructors and create functions with the executor parameter.
• The virtual, protected, Dense functions compute_norm1_impl, add_scaled_impl, etc.
• Logger events for solvers and criterion without the additional implicit_tau_sq parameter.
• The global gko::solver::default_krylov_dim, use instead gko::solver::gmres_default_krylov_dim.
• array::get_num_elems() has been renamed to get_size()
• matrix_data::ensure_row_major_order() has been renamed to sort_row_major()
• device_matrix_data::get_num_elems() has been renamed to get_num_stored_elements()
• The CMake parameter GINKGO_COMPILER_FLAGS has been superseded by CMAKE_CXX_FLAGS, and GINKGO_CUDA_COMPILER_FLAGS has been superseded by CMAKE_CUDA_FLAGS
• The std::initializer_list overloads of matrix create methods and constructors are deprecated in favor of explicit array parameters

Added features

• Add Executor::get_description() for textual representation of the device #1615
• Add row and column scaling functionality to the distributed matrix #1640
• Add SolverProgress logger printing out or storing to disk the individual scalars (and vectors) of an iterative solvers after each iteration #1620
• Add new ortho_method parameter for GMRES, with classical Gram-Schmidt and classical Gram-Schmidt with re-orthogonalization options in addition to previously-available modified Gram-Schmidt #1646
• Add file config support for Schwarz #1658
• Add overload for Executor::run which accepts a name and a closure for the ReferenceExecutor as the first two arguments #1667
• Add function to fill device_matrix_data with zeros #1683
• Add (S)SOR and Gauss-Seidel preconditioner #1633, #1634
• Add support for additive read_distributed for the distriubted matrix #1650
• Add Ginkgo's own ILU and IC implementation #1684
• Add Nvidia ada architecture #1733
• Add half precision support #1706, #1708, #1711, #1712, #1713, #1716

Improvements

• Add workspace in residual norm check #1687, which reduces the alloc/free and corresponding overhead.
• Add distributed VectorCache and use it as workspace in Schwarz #1688.
• Add example to show the file config usage #1662
• Improve compile time for batched solvers #1629
• Reduce the thrust conflict possibility by adding custom thrust namespace #1730

Fixes

• Fix the algorithm usage of transposed triangular solver, which is the same as the original one not the default algorithm #1641
• Fix the inconsistent behavior on the zero diagonal value in scalar Jacobi #1642
• Fix an issue related to GCR and non-default strides in the rhs vector #1656
• Fix an issue related to triangular solvers with CUDA on Windows #1665
• Fix an issue where non-conforming MatrixMarket files were parsed without an error #1628
• Fix finding rocthrust if it's not installed paths included by default #1668
• Fix the casting issue in mixed-precision multigrid setup #1663
• Fix some test failures with ROCm 6.x #1670
• Fix the race condition in bicgstab #1676
• Fix an issue with MGS GMRES for complex numbers #1678
• Fix finding ROCm on recent ROCm version (5.0+) #1673
• Fix compiler error when using NVHPC with MPI enabled #1697
• Fix build issue of OMP backend when using HIPCC as C++ compiler #1695
• Fix build issue for Intel OneAPI 2025.0 #1718
• Fix inconsitent function and class/struct, which mainly fixes clang-cl #1725
• Fix undefined symbols in shared library in msys2/clang #1724
• Fix page fault issues when running on multiple Intel GPUs in parallel #1723