FM

Configuration options for the full matrices. [Edit on GitHub]

Keywords

• FORCE_BLOCK_SIZE

• NCOL_BLOCKS

• NROW_BLOCKS

• TYPE_OF_MATRIX_MULTIPLICATION

Keyword descriptions

FORCE_BLOCK_SIZE

Type: logical
Default: F
Lone keyword: T
Usage: FORCE_BLOCK_SIZE

Description: Ensure for small matrices that the layout is compatible with bigger ones, i.e. no subdivision is performed (can break LAPACK).

[Edit on GitHub]

NCOL_BLOCKS

Type: integer
Default: 64

Description: Defines the number of columns per scalapack block in the creation of vlock cyclic dense matrices. Use an internal default if zero or negative.

[Edit on GitHub]

NROW_BLOCKS

Type: integer
Default: 64

Description: Defines the number of rows per scalapack block in the creation of block cyclic dense matrices. Use an internal default if zero or negative.

[Edit on GitHub]

TYPE_OF_MATRIX_MULTIPLICATION

Type: enum
Default: COSMA

Description: Allows to switch between scalapack pxgemm and COSMA pxgemm. COSMA reduces the communication costs but increases the memory demands. The performance of Scalapack’s pxgemm on GPU’s depends crucially on the BLOCK_SIZES. Make sure optimized kernels are available.

Valid values:

• SCALAPACK Standard ScaLAPACK pdgemm

• PDGEMM Alias for ScaLAPACK

• COSMA COSMA is employed. See https://github.com/eth-cscs/COSMA.

[Edit on GitHub]