libflame
revision_anchor
|
Functions | |
FLA_Error | FLASH_UDdate_UT_inc_create_hier_matrices (FLA_Obj R_flat, FLA_Obj C_flat, FLA_Obj D_flat, dim_t depth, dim_t *b_flash, dim_t b_alg, FLA_Obj *R, FLA_Obj *C, FLA_Obj *D, FLA_Obj *T, FLA_Obj *W) |
dim_t | FLASH_UDdate_UT_inc_determine_alg_blocksize (FLA_Obj R) |
FLA_Error FLASH_UDdate_UT_inc_create_hier_matrices | ( | FLA_Obj | R_flat, |
FLA_Obj | C_flat, | ||
FLA_Obj | D_flat, | ||
dim_t | depth, | ||
dim_t * | b_flash, | ||
dim_t | b_alg, | ||
FLA_Obj * | R, | ||
FLA_Obj * | C, | ||
FLA_Obj * | D, | ||
FLA_Obj * | T, | ||
FLA_Obj * | W | ||
) |
References FLA_Abort(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_width(), FLA_Print_message(), FLASH_Obj_create_ext(), FLASH_Obj_create_hier_copy_of_flat(), and FLASH_UDdate_UT_inc_determine_alg_blocksize().
{ FLA_Datatype datatype; dim_t m_T, n_T; dim_t m_W, n_W; dim_t m_C; dim_t m_D; // *** The current UDdate_UT_inc algorithm implemented assumes that // the matrix has a hierarchical depth of 1. We check for that here // because we anticipate that we'll use a more general algorithm in the // future, and we don't want to forget to remove the constraint. *** if ( depth != 1 ) { FLA_Print_message( "FLASH_UDdate_UT_inc() currently only supports matrices of depth 1", __FILE__, __LINE__ ); FLA_Abort(); } // Create hierarchical copy of matrices R_flat, C_flat, and D_flat. FLASH_Obj_create_hier_copy_of_flat( R_flat, depth, b_flash, R ); FLASH_Obj_create_hier_copy_of_flat( C_flat, depth, b_flash, C ); FLASH_Obj_create_hier_copy_of_flat( D_flat, depth, b_flash, D ); // Query the datatype of matrix R_flat. datatype = FLA_Obj_datatype( R_flat ); // If the user passed in zero for b_alg, then we need to set the // algorithmic (inner) blocksize to a reasonable default value. if ( b_alg == 0 ) { b_alg = FLASH_UDdate_UT_inc_determine_alg_blocksize( *R ); } // Determine the element (not scalar) dimensions of the new hierarchical // matrix T. By using the element dimensions, we will probably allocate // more storage than we actually need (at the bottom and right edge cases) // but this is simpler than computing the exact amount and the excess // storage is usually small in practice. n_T = FLA_Obj_width( *R ); m_C = FLA_Obj_length( *C ); m_D = FLA_Obj_length( *D ); m_T = max( m_C, m_D ); // Create hierarchical matrix T, with element dimensions conformal to the // the larger of C and D, where each block is b_alg-by-b_flash. FLASH_Obj_create_ext( datatype, m_T * b_alg, n_T * b_flash[0], depth, &b_alg, b_flash, T ); // Determine the element (not scalar) dimensions of the new hierarchical // matrix W. The element length and width will be identical to that of R. // Once again, we will probably allocate excess storage, but we consider // this to be small. m_W = FLA_Obj_length( *R ); n_W = FLA_Obj_width( *R ); // Create hierarchical matrix W, with element dimensions conformal to R, // where each block is b_alg-by-b_flash. FLASH_Obj_create_ext( datatype, m_W * b_alg, n_W * b_flash[0], depth, &b_alg, b_flash, W ); return FLA_SUCCESS; }
References FLA_Obj_length().
Referenced by FLASH_UDdate_UT_inc_create_hier_matrices().
{ dim_t b_alg; dim_t b_flash; // Acquire the storage blocksize. b_flash = FLA_Obj_length( *FLASH_OBJ_PTR_AT( R ) ); // Scale the storage blocksize by a pre-defined scalar to arrive at a // reasonable algorithmic blocksize, but make sure it's at least 1. b_alg = ( dim_t ) max( ( double ) b_flash * FLA_UDDATE_INNER_TO_OUTER_B_RATIO, 1 ); return b_alg; }