libflame
revision_anchor
|
Functions | |
FLA_Error | FLASH_Apply_QUD_UT_inc_create_workspace (FLA_Obj T, FLA_Obj R, FLA_Obj *W) |
FLA_Error FLASH_Apply_QUD_UT_inc_create_workspace | ( | FLA_Obj | T, |
FLA_Obj | R, | ||
FLA_Obj * | W | ||
) |
References FLA_Abort(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_width(), FLA_Print_message(), FLASH_Obj_create_ext(), FLASH_Obj_depth(), FLASH_Obj_scalar_length_tl(), and FLASH_Obj_scalar_width_tl().
Referenced by FLASH_UDdate_UT_inc_update_rhs().
{ FLA_Datatype datatype; dim_t depth; dim_t b_alg; dim_t b_flash; dim_t m, n; // Query the depth. depth = FLASH_Obj_depth( T ); // *** The current Apply_QUD_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_Apply_QUD_UT_inc() currently only supports matrices of depth 1", __FILE__, __LINE__ ); FLA_Abort(); } // Query the datatype of matrix T. datatype = FLA_Obj_datatype( T ); // Inspect the length of a the top-left element of T to get the // algorithmic blocksize we'll use throughout the Apply_QUD_UT_inc // algorithm. b_alg = FLASH_Obj_scalar_length_tl( T ); // The width of the top-left element gives us the storage blocksize. b_flash = FLASH_Obj_scalar_width_tl( T ); // Determine the element (not scalar) dimensions of the new hierarchical // matrix W. 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. m = FLA_Obj_length( R ); n = 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 * b_alg, n * b_flash, depth, &b_alg, &b_flash, W ); return FLA_SUCCESS; }