- Refactored system solver: Added helper functions to check and fix conflicts in configuration options and to reduce code duplication.
- Code returns errors instead of warnings when encountering ambiguous erronneus inputs from the user.
-
Exported HyKKT libraries and fixed solver reuse by refreshing numerical data, reusing allocations, resizing GPU transpose workspaces, and handling zero residuals.
-
Added classes and tests for permutation, Ruiz scaling, Cholesky factorization, Schur complement conjugate gradient and matrix multiplication and addition.
-
Changed random number generation int tests to be C++ style and fixed-seed, to avoid random failures.
-
Added Schur Complement Conjugate Gradient class.
-
Added cmake-format.
-
Reworked templates to include example tests.
-
Removed unnecessary full facotorization in the examples.
-
Update method names in
VectorHandlerto match BLAS naming. -
Added
conscounterparts toVector::getDatamethods. -
Made Vector::copyToExternal able to copy from device to host and vice versa.
-
Added
diagSolve,max, andabsvector operations. -
Added
allocateWithExternalSparsityPatternfor sparse matrices. -
Added initial guess support for FGMRES and randomized FGMRES iterative solvers.
-
Improved coding guidelines for developers on floating point conventions.
-
Made spack builds more robust.
-
Removed unnecessary device synchronization and added where needed (HIP only calls).
-
Changed variable and function names to be more explanatory.
-
Added Developer Guide documentation for vector and matrix classes and handlers.
-
Remove automatic memory allocation and memory syncing in various memory management functions and require the user to explicitly manage their own memory.
-
Set Re::Solve's real type and matrix/vector index type at CMake configuration time.
-
Added size assertions to
VectorHandler::gemvand fixed vector sizing inGramSchmidtaccordingly. -
Added line to testlog to tell the user to expect warnings as part of normal testing.
-
Fixed GPU refactorization allocation and synchronization errors, uninitialized GMRES example initial guesses, and CUDA ILU0 failures on stored numerical zero pivots.
-
Added cuDSS implementation of Cholesky solver (HyKKT) and refactorization
-
Optimized CGS2 coefficient accumulation in
GramSchmidtby eliminating an unnecessary device-to-host synchronization. -
Added examples/sysGmres.cpp to demonstrate how to use SystemSolver with GMRES.
-
Updated MatrixHandler::addConst to return integer error codes instead of void.
-
Added a preconditioner interface class so users can define their own preconditioners.
-
Added left preconditioning support for GMRES and a user-defined preconditioner class.
-
Added optional timing output to
gpuRefactor,kluRefactor,gluRefactor, andsysRefactorplus benchmark utilities for parsing logs and generating timing/residual plots. -
Added configurable zero-pivot and pivot-boost parameters for CUDA and HIP ILU0 and exposed the preconditioner solver through
SystemSolverfor configuring them.
-
Re::Solve now works reliably with asymmetric matrices, with no need for intermediate CSC storage. This requires switching
$L$ with$U$ and$P$ with$Q$ and reinterpretting them as CSR instead of CSC. It is seamless from the user perspective and fixed many bugs. -
Significant improvements to documentation and instructions inside and outside the code. Added general API description, including details on memory space synchronization.
-
Added more rigorous checks for PRs for clang formatting and to compile without warnings and memory leaks.
-
Updated pull request and issue templates.
-
Fixed a bug that produced inaccurate results for some asymmetric matrices with major feature 1.
-
Synchronized devices after HIP functions. HIP executes asynchronously, so bugs occured without synchronization.
-
Corrected the way cmake finds suitsparse.
-
Fixed various memory leaks and compiler warnings.
-
Changed all examples and tests to use Csr format, added uniform command line parsers (no longer hard-coded), and decluttered them.
-
Added asymmetric matrices and well-conditioned matrices to the test suite.
-
Removed RocSparse "fast mode" triangular solver and use RocSolver triangular solver only as it is now faster and removes dependencies.
-
Put sorting inside the KLU extraction because many solvers assume sorted factors and there's no need to reimplement sorting constantly.
-
Removed duplicate code, added code comments, corrected code to fit guidelines, removed magic numbers, and simplified code where possible.
-
Added the ability to reset a workspace without completely destroying it.
-
Improved testing and added tests where they were missing.
-
Added kernels for multiplying a vector by a diagonal matrix and a general matrix by a diagonal matrix (left and right).
-
Prohibitted sloppy memory syncing and added more descriptive error messages when a prohibited action is attempted.
-
The code now tracks the updated status for each vector in a multivector.
-
Added the ability to reuse a transpose allocation.
-
Added the ability to generically set solver parameters.
-
Added LUSOL direct solver, which can factorize matrices and extract factors independently of KLU.
-
Various Spack updates.