ipiv: Rename ipiv_init and ipiv_reducek functions as INSERT_TASK  and make them match the INSERT_TASK API

PRUVOST Florent authored 1 year ago and Mathieu Faverge committed 1 year ago

Fix compilation problem with starpu versions < 1.4. STARPU_CUDA and STARPU_HIP macros may not be defined in starpu

Fix --forcegpu option to be able to force starpu to execute kernels on gpus, RUNTIME_CUDA was not well defined and introduce RUNTIME_HIP.

Matthieu Kuhn authored 1 year ago and Mathieu Faverge committed 1 year ago

runtime: Add struct fields and functions to manipulate the permutation and the inverse permutation arrays

Added permutation on top of partial pivoting, panel only.
First version of LU with pivoting.

runtime/starpu: extract the wontuse and mpi_data_register functions into header files to share them into multiple C files

runtime/starpu/control/runtime_descriptor.c

control: store the ncuda field in the options to know at insertion time if cuda kernels are enabled or not

hip: replace deprecated hipMallocHost by hipHostMalloc

runtime/starpu: add a warning if parallel workers are initialized without multi-threaded BLAS library

COJEAN Terry authored 8 years ago and Mathieu Faverge committed 2 years ago

Integrate the use of parallel tasks for chameleon within POTRF algorithm. Note that this requires StarPU
New cluster interface based on the hwloc vocabulary : CHAMELEON_CLUSTER_LEVEL=hardware-level:number-of-clusters and CHAMELEON_CLUSTER_SHOW to print the clusters