 |
GHOST
1.1.2
General, Hybrid, and Optimized Sparse Toolkit
|
All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
|
GHOST
1.1.2
General, Hybrid, and Optimized Sparse Toolkit
|
A dense matrix in GHOST (ghost_densemat) can be either an actual dense matrix or a view of either another dense matrix or plain data. A view is a dense matrix which does not have its data stored in memory but only views data stored elsewhere Views can be identified by the flag GHOST_DENSEMAT_VIEW in the the ghost_densemat::traits.
A view can be either a compact view or a scattered (in which case the one of the flags GHOST_DENSEMAT_SCATTERED_LD or GHOST_DENSEMAT_SCATTERED_TR is set) view. A compact view is a view in which the data is contiguous with a certain stride and leading dimension. Otherwise, the view is scattered, where we distinguish between scattered views in leading (GHOST_DENSEMAT_SCATTERED_LD) or trailing (GHOST_DENSEMAT_SCATTERED_TR) dimension. Some function are not implemented natively for scattered views. If this is the case, a warning will be printed and a temporary densemat will be created for this function. Afterwards, the scattered view will be updated with the result data and the temporary densemat gets destroyed.
The dense matrix data is stored as a one-dimensional array ghost_densemat::val (or ghost_densemat::cu_val for GPU densemats). For scattered views, the according bits in ghost_densemat::rowmask or ghost_densemat::colmask have to be manipulated.
Analog for column-major densemats. We have a 4x4 source matrix A:
Create a view of A which skips row 0:
Create a view of A which skips row 0 and 3:
Create a view of A which skips row 1:
Create a view of A which skips column 3:
Create a view of A which skips column 3 and row 0:
Create a view of A which skips column 1 and row 0:
Create a view of A which skips column 1 and rows 0 and 2:
1.8.6