THIBAULT Samuel authored 8 years ago and Mathieu Faverge committed 7 years ago

Add MORSE_Desc_Create_OOC, which is like MORSE_Desc_Create, but does not
actually allocate a matrix, thus letting the runtime allocate on-demand the
tiles, possibly pushing them to the disk.

Add a --ooc option to tests to enable this.

    - use (starpu_cpu_func_t) 1 trick, same as cuda_func
    - cpu funtions are not defined anymore avoiding the dependency to coreblas
    - add #if !defined(CHAMELEON_SIMULATION) where it is needed
    - remove dependency to the coreblas library (become useless)
    - remove useless simucblas, simulapacke libraries
    - remove CHAMELEON_SIMULATION_MAGMA cmake variable and definition
      - keep using CHAMELEON_USE_CUDA and CHAMELEON_USE_MAGMA to consider CUDA kernels
      - this avoid to introduce useless new variables
    - work on messages

timing: add option --bigmat to choose if we allocate one big 'mat' array or if the runtime allocates the tile one by one

This routine, available in MKL, does a product in 6n^3 ops instead of 8n^3
but is interesting only for "large enough" matrices (to be tested...)
Potentially, we gain 25 % in all complex computations.
It could be interesting to look for it / implement it in cuda.

!!! Note that the flop counters are not updated         !!!
!!! In C/Z accuracy, most flops counter should be x0.75 !!!

IT is OFF by default
It is activated with MORSE_Enable(MORSE_GEMM3M)
In the timing routines, it is activated with --gemm3m

IT is OFF by default
It is activated with MORSE_Enable(MORSE_PROGRESS)
In the timing routines, it is activated with --progress
No progress is printed for tasks faster than 10 seconds

add a priority scheme for getrf_nopiv_tile (with dmdas) and an efficient restriction optimization (tested on sirocco)

instead of introducing RUNTIME_distributed_barrier

add cudablas library to make calls to cuda kernels (magma here, cublas will follow), no more calls to magma in runtime/starpu codelets

MORSE_Distributed_size, MORSE_Distributed_rank so that applications do not
hardcode the use of MPI.

Introduce RUNTIME_distributed_rank, RUNTIME_distributed_size,
RUNTIME_distributed_barrier, so that MORSE does not hardcode the use of MPI
either.

This allows to use simgrid-mpi.