compute/map: Improve the map function to take into account multiple descriptor and multiple implementations for the kernel to apply.

Abel Calluaud authored 1 year ago and Mathieu Faverge committed 1 year ago

Remove the use of CHAMELEON_PRECISIONS that is an upper duplicate of CHAMELEON_PRECISION provided by cmake_morse (Fix #124)

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

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

Change name of internal libraries cudablas and hipblas to gpucublas and gpuhipblas to avoid confusion with external libraries that can have the same name

CMake modifications to avoid cmake projects dependents to rediscover all our dependencies at configure time

Create a new library for chameleon blas/lapack api symbols to get things separated and clean. This avoids breaking users code using classical chameleon and cblas/lapacke beside (e.g. can cause redefinition of some cblas objects).

PRUVOST Florent authored 2 years ago and Mathieu Faverge committed 2 years ago

cmake: avoid to force cblas and lapacke mt to false because the application may want to use the multithreaded version and handle at its level the number of threads to use

Upgrade the version number to make the distinction between the release files, and the files from the master

coreblas/trace: add CHAMELEON_KERNELS_TRACE option to name the tiles and print information about the exectued kernels

SYLVAND Guillaume authored 4 years ago and Mathieu Faverge committed 3 years ago

lib is now called libhmat and target is well defined.