v3.41.0

LinearSolve v3.41.0

Diff since v3.40.3

Merged pull requests:

• Better safeguard on unloadable MKL (#793) (@ChrisRackauckas)

v3.40.3

LinearSolve v3.40.3

Diff since v3.40.2

Merged pull requests:

• Performance fixes for DualLinearProblems (#776) (@jClugstor)
• prepare to properly suppory lazy libraries (#789) (@vtjnash)
• Fix precompilation of usecuda, usemetal, and useblis (#792) (@ChrisRackauckas)

Closed issues:

• Precompilation with 3.40.2 fails (#790)

v3.40.2

LinearSolve v3.40.2

Diff since v3.40.1

Merged pull requests:

• Replace deprecated Libdl.dlsym_e/Libdl.dlopen_e with Libdl.dlsym/Libdl.dlopen (#782) (@devmotion)
• Backport #781 to 3.12 (#783) (@devmotion)
• Move use* extension functions to their respective extensions (#784) (@topolarity)
• Replace Krylov_{CG,GMRES} by KrylovJL_{CG,GMRES} (#786) (@jonas-schulze)
• Make MKL check to be fully static (#788) (@ChrisRackauckas)

Closed issues:

• usecuda() is not --trim compatible (#778)
• Solver hangs occasionally on __mkl_isavialable() (#787)

v3.40.1

LinearSolve v3.40.1

Diff since v3.40.0

Merged pull requests:

• Disable use of older MKL_jll versions (#781) (@devmotion)

v3.12.1

LinearSolve v3.12.1

Diff since v3.12.0

This release has been identified as a backport.
Automated changelogs for backports tend to be wildly incorrect.
Therefore, the list of issues and pull requests is hidden.

v3.40.0

LinearSolve v3.40.0

Diff since v3.39.2

Merged pull requests:

• Make length of Partials known at compile time in ForwardDiff overloads (#727) (@jClugstor)
• Hotfix typo in linearsolveautotune (#774) (@ChrisRackauckas)

Closed issues:

• Add a way to forward diff a function that uses linsolve inside (#646)
• LU algorithms that work when solving a linear system fail in autotune_setup (#771)

v3.39.2

LinearSolve v3.39.2

Diff since v3.39.1

Merged pull requests:

• Fix CUDA functional check in init_cacheval functions (#770) (@ChrisRackauckas-Claude)

Closed issues:

• Don't use cuda just because cuda is loaded (#761)

v3.39.1

LinearSolve v3.39.1

Diff since v3.39.0

Merged pull requests:

• Fix Float16 segfault with Metal algorithms (#743) (#764) (@ChrisRackauckas-Claude)
• Fix method overwriting error during precompilation on Apple Silicon (#769) (@ChrisRackauckas)

Closed issues:

• LinearSolveAutotune segfaults Julia when benchmarking Float16 on Apple Silicon (#743)
• Installation of LinearSolveAutotune fails on Apple Silicon (#768)

v3.39.0

LinearSolve v3.39.0

Diff since v3.38.0

Merged pull requests:

• Remove deepcopy operations from __init path (fixes #747) (#763) (@ChrisRackauckas-Claude)
• Fix BenchmarkTools.DEFAULT_PARAMETERS const assignment error in Julia 1.12 (#765) (@ChrisRackauckas-Claude)

Closed issues:

• Use of deepcopy doesn't trim well (#648)
• [LinearSolveAutotune] Setting BenchmarkTools.DEFAULT_PARAMETERS on Julia 1.12 fails since it is const (#744)
• Remove deepcopy from __init path. (#747)

v3.38.0

LinearSolve v3.38.0

Diff since v3.37.0

Merged pull requests:

• Fix allocations in 32Mixed precision methods by pre-allocating temporaries (#758) (@ChrisRackauckas-Claude)
• Simplify mixed precision: compute types on demand instead of caching (#759) (@ChrisRackauckas-Claude)
• fix sparspak solves in multithreaded loop (#760) (@j-fu)
• Improve usecuda function to verify CUDA availability (#762) (@hexaeder)

Closed issues:

• Don't use cuda just because cuda is loaded (#761)