Add support for registering MPI data types

Closes #97

See merge request !204

THIBAULT Samuel authored 5 years ago and Mathieu Faverge committed 4 years ago

Fixes #97

Reduce the spectrum of the data accesses with Quark

See merge request !216

Reduce data access type whenever possible

This PR aims at reducing the data access type to enable more cases with the allocation on the fly.

For example, with a gemm: `alpha A *B  + beta * C`, and beta = 0. `C` is used only as output, and not as inout. This changes all the codelets to STARPU_VARIABLE_NBUFFERS.

See merge request !215

Reduce the data access types whenever it's possible to get more flexibility with the allocation onf the fly

StarPU - Add copy method for starpu <= 1.3.3

See merge request !214

specify a specific time limit that can exceed the project-level timeout, jobs...

See merge request solverstack/chameleon!213

specify a specific time limit that can exceed the project-level timeout, jobs submitted on a supercomputer like plafrim can be delayed due to the job scheduler, one week seems fair because we want this profiling done once a week

starpu: use starpu_mpi_wait_for_all when available

See merge request !210

THIBAULT Samuel authored 4 years ago and Mathieu Faverge committed 4 years ago

starpu_mpi_barrier is not really supposed to properly synchronize with STF.

Add execute_on_worker feature of starpu thanks to a new field `schedopt` in the runtime_request_s struct

See merge request solverstack/chameleon!211

PRUVOST Florent authored 4 years ago and Mathieu Faverge committed 4 years ago

Add execute_on_worker feature of starpu thanks to a new field `schedopt` in the runtime_request_s struct

starpu: Call starpu_fxt_autostart_profiling in non-MPI case too

See merge request !209

fix int overflow in core gram

See merge request solverstack/chameleon!208

Ci/kibana

See merge request solverstack/chameleon!205

Feature/issue94

* The API of Lapack_to_Tile/Tile_to_Lapack family function is modified to integrate the uplo parameter (This is done to enable the following points)
* A new option `mtxfmt` is introduced to change the allocation of the matrix.
  * 0: The matrices are allocated as a single big allocation
  * 1: The matrices are allocated tile by tile (Only with StarPU)
  * 2: The matrices are out-of-core (Only with StarPU)
* Fix issue #96 by adding an example at the end of the usage function
* Fix an issue with the descriptor copy functions that was using the incorrect get_blkldd/get_blkaddr functions.


See merge request !199

Make lapack_to_tile/tile_to_lapack interfaces deprecated and add Lap2Desc/Desc2Lap interfaces that takes into account an uplo parameter to limit the copy.

Set correct values for footprint in perfmodel files

See merge request !202

Testings cleanup

See merge request !200

Tune openmpi mapping option and adapt some parameters to improve performances

See merge request solverstack/chameleon!198