libflame revision_anchor
|
Functions | |
FLA_Error | FLASH_Apply_Q_UT_create_workspace (FLA_Obj TW, FLA_Obj B, FLA_Obj *W) |
References FLA_Abort(), FLA_Obj_datatype(), FLA_Print_message(), FLASH_Obj_create_ext(), FLASH_Obj_depth(), FLASH_Obj_scalar_length_tl(), FLASH_Obj_scalar_width(), and FLASH_Obj_scalar_width_tl().
Referenced by FLASH_LQ_UT_solve(), and FLASH_QR_UT_solve().
{ 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( TW ); // *** The current Apply_Q_UT 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_Q_UT() currently only supports matrices of depth 1", __FILE__, __LINE__ ); FLA_Abort(); } // Query the datatype of matrix TW. datatype = FLA_Obj_datatype( TW ); // Inspect the dimensions of a the top-left element of TW to get the // algorithmic/storage blocksize we'll use throughout the Apply_Q_UT // algorithm. b_alg = FLASH_Obj_scalar_length_tl( TW ); b_flash = FLASH_Obj_scalar_width_tl( TW ); // The traditional (non-incremental) Apply_Q_UT algorithm-by-blocks // requires that the algorithmic blocksize be equal to the storage // blocksize. if ( b_alg != b_flash ) { FLA_Print_message( "FLASH_Apply_Q_UT() requires that b_alg == b_store", __FILE__, __LINE__ ); FLA_Abort(); } // The scalar length of W should be the algorithmc/storage blocksize // encoded in TW. m = b_alg; // Query the scalar (not element) width of the right-hand side // matrix B. n = FLASH_Obj_scalar_width( B ); // Create hierarchical matrix W. FLASH_Obj_create_ext( datatype, m, n, depth, &b_alg, &b_flash, W ); return FLA_SUCCESS; }