Changes:

• mixed-precision: introduce descriptor with precision adapted to local norms - Add CHAMELEON_[dz]gered... functions to reduce the precision of the tiles based on a requested accuracy - Add CHAMELEON_[dz]gerst... functions to restore the original numerical precision of the tiles in a descriptor
• types: add support for half precision arithmetic into the data descriptors
• cuda: add half precision conversion kernels, and variants of the gemm kernels (hgemm, and gemmex)
• cuda: Check error after lauching kernels
• descriptors: Add the possibility to pass arguments to the rankof function. This is used to provide custom distribuitions through a given file. WARNING: It changes the interface of CHAMELEON_Desc_Create_User that requires an additional , NULL parameters in the general case.
• control: Defined the default parameters through environment variables first and make sure the testings use the default value instead of overwritting them.
• control: Make the CHAM_context_t structure public, and provide a function to the user to access the pointer in case of the development of its own functionalities using the RUNTIME API.
• compute: Refactor the code that compute the kernel dimension to potentially enable variadic tile sizes
• compute/map: Rework the map functions family to be able to pass multiple descriptor with parameterized access types.
• compute/getrf: Add a basic LU factorization with partial pivoting (WARNING: this functionnality is still under development and does not provide full performance yet)
• compute/poinv: Add the possibility to use an intermediate distribution for the TRTRI operation
• compute/getrf_nopiv: Add lookahead through temporary buffers to better regulate the communication allocations
• runtime/starpu: Whenever possible replace the lacpy codelet by a direct memory copy from the input handler to the output one
• runtime/starpu: better separation of the public interface from the internal interface for code reusing the RUNTIME API
• testings: Display in help message the option possible values when possible
• bug: fix issue with undefined vasprintf
• bug: Make sure generic algorithms are used when at least one of the data descriptor is not 2D block cyclic and might cause issues.
• bug/starpu: Fix the --forcegpu option to integrate HIP devices withing the option and make sure it's applied only when possible
• Fix issue 124: RP_CHAMELEON_PRECISION is the set of supported precisions, while CHAMELEON_PRECISION is the set of enabled precisions
• Fix the trsm flops issue that was miscalculated.
• Fix integer overflow in malloc where size_t was not used
• ci/docker: provide a simpler docker image dedicated to the project

WARNING: Download the source archive by clicking on the link Download release above, please do not consider the automatic Source code links as they are missing the submodules. Visit the documentation to see how to install Chameleon.