Releases · mmottl/lacaml

WARNINIG WARNING WARNING

User code changes in existing code required! User code may still compile,
but can behave differently!

The following functions are affected:
- potrf
- potri
- potrs
The above functions now do not support the jitter argument anymore.
Users should remove the flag from calls to the above functions and
call the new Mat.add_const_diag function if they need to add jitter.
This call should happen right before the (now required) call to potrf.

More importantly, potri and potrs now do not support the factorize
flag anymore, which would call potrf automatically beforehand. This was
the (ill-conceived) default, which makes it harder to port LAPACK code
to Lacaml. In order to upgrade your code, please do the following:
- If potri or potrs were passed ~factorize:false, just remove
  the flag.
- If potri or potrs were passed ~factorize:true or were called
  without the factorize flag, remove the flag if necessary and call
  potrf with the exactly corresponding arguments before.
Luckily, these functions are typically used rarely, and the changes
are trivial. Apologies anyway for the churn!
New functions
- Mat.add_const_diag for adding a constant to the diagonal of a
  (sub-)matrix.
orgqr now detects if m < n and raises an exception instead of printing
a Fortran error message and continuing.
Eigenvalue offsets should now work correctly.
Improved documentation.

Switched to jbuilder and topkg
API changes
- trmm and trsm now do not label argument a anymore
- Many matrix functions now support an optional patt argument,
  which can be used to specify rectangular, triagonal, trapezoidal, and
  pentagonal patterns on which to perform an operation.
- New functions
  - Mat.sum_prod computes the sum of elemenet-wise products of
    two matrices. Some use cases are already covered by Mat.gemm_trace,
    but the latter does not support patterns.
Improved C-code to better support SIMD compiler optimizations
Many internal improvements, including untagged and unboxed passing of
parameters to and from external functions.
Compilation now uses -march=native -O3 -ffast-math by default, which
should be safe and exploit SIMD on platforms that support it to greatly
improve performance of some operations.
Improved documentation
Improved configuration and build process

Provide feedback