Aurelien Bouteiller

**Original comment by Omri Mor (Bitbucket: [omrimor2](https://bitbucket.org/omrimor2), GitHub: [omor1](https://github.com/omor1)).** ---------------------------------------- This is in `parsec/data_dist/matrix/matrixtypes.c`.

**Original comment by Omri Mor (Bitbucket: [omrimor2](https://bitbucket.org/omrimor2), GitHub: [omor1](https://github.com/omor1)).** ---------------------------------------- Actually, looking into the code of DPLASMA, it seems like CBLAS is called with `CblasColMajor`, so I suppose that column-major...

**Original comment by George Bosilca (Bitbucket: [bosilca](https://bitbucket.org/bosilca), GitHub: [bosilca](https://github.com/bosilca)).** ---------------------------------------- The default storage type is Lapack, and it is defined in matrix.h as column major.

**Original comment by George Bosilca (Bitbucket: [bosilca](https://bitbucket.org/bosilca), GitHub: [bosilca](https://github.com/bosilca)).** ---------------------------------------- We need documentation on the storage type used in the base matrix types.

**Original comment by George Bosilca (Bitbucket: [bosilca](https://bitbucket.org/bosilca), GitHub: [bosilca](https://github.com/bosilca)).** ---------------------------------------- This was designed in purpose. If we reset the GPU then each taskpool will have to push all the data...

**Original comment by Qinglei Cao (Bitbucket: [Qinglei_Cao](https://bitbucket.org/Qinglei_Cao), ).** ---------------------------------------- I made a simple example. Actually, it cannot run the second time when calling dplasma\_dpotrf. It shows “device\_cuda\_module.c:1223: parsec\_gpu\_data\_stage\_in: Assertion \`\(gpu\_elem->version...

**Original comment by Nuria Losada (Bitbucket: [nuriallv](https://bitbucket.org/nuriallv), GitHub: [nuriallv](https://github.com/nuriallv)).** ---------------------------------------- I believe this issue is related to the coherence state when we marked a copy shared between CPU and GPU....

This should be restudied after 3935f31 as it may now behave correctly

> I will take a look this weekend. Thanks. > […](#) We never documented what was the outcome of this study @QingleiCao

Snapshot of investigation: During parsec_redistribute_dtd_New we create an `adt` that we cleanup later in that same function. ``` 376 adt = parsec_dtd_create_arena_datatype(parsec, &TARGET); ... (gdb) n 426 adt = parsec_dtd_get_arena_datatype(parsec,...