• entirely rewritten forward and backward substitution for better parallelism and use of GPUs (useful in case of multiple RHS)
• new CUDA kernels for panel reduction which drastically improve performance
• added a routine to compute A^TA and benchmark for normal equations
• multiple bugfixes and minor improvements

• improved cmake install scripts and procedure

• minor bugfixes: incoherent passing of arguments in block_axpy and block_copy tasks, in spmat_mv and in FindAMD

• Raise an error in potrf if matrix is indefinite
• better handling of transposition
• various bug fixes

• Fix in the cmake files which resulted in faulty behavior of the install step in Windows systems

• Support for Nvidia GPUs through the StarPU runtime system.
• Cholesky factorization for solving symmetric positive definite systems.
• Dynamic, hierarchical partitioning for the QR factorization.
• Switched to cmake for the build process.
• switched to fstarpu_mod StarPU module instead of hand-made interfaces and wrappers.
• Environment variables to set default values for all control parameters.

Version 2.0 is an almost complete rewrite of the qr_mumps package. Here are some of the main changes wrt previous versions

• Parallelism is now achieved using the StarPU runtime engine.
• 2D block partitioning can be used for frontal matrices in combination with communication avoiding dense factorization algorithms
• it is possible to bound the memory consumption of the parallel factorization phase
• pipelining of operation can be achieved through the asynchronous API
• the error handling has been deeply modified to make it thread-safe

• Added a method to extract the R factor once the factorization is computed.

• There is no limit to the number of concurrent instances of qrm_spmatc in the C interface
• A number of minor bugfixes