FLA_LU_incpiv.h File Reference

(r)

Go to the source code of this file.

Functions
FLA_Error	FLASH_LU_incpiv_create_hier_matrices (FLA_Obj A_flat, dim_t depth, dim_t *b_flash, dim_t b_alg, FLA_Obj *A, FLA_Obj *p, FLA_Obj *L)
dim_t	FLASH_LU_incpiv_determine_alg_blocksize (FLA_Obj A)
FLA_Error	FLASH_LU_incpiv_noopt (FLA_Obj A, FLA_Obj p, FLA_Obj L)
FLA_Error	FLASH_LU_incpiv_opt1 (FLA_Obj A, FLA_Obj p, FLA_Obj L)
FLA_Error	FLASH_LU_incpiv_solve (FLA_Obj A, FLA_Obj p, FLA_Obj L, FLA_Obj B, FLA_Obj X)

Function Documentation

FLASH_LU_incpiv_create_hier_matrices()

FLA_Error FLASH_LU_incpiv_create_hier_matrices	(	FLA_Obj	A_flat,
		dim_t	depth,
		dim_t *	b_flash,
		dim_t	b_alg,
		FLA_Obj *	A,
		FLA_Obj *	p,
		FLA_Obj *	L
	)

{
   FLA_Datatype datatype;
   dim_t        m, n;
   dim_t        one = 1;
   
   // *** The current LU_incpiv 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_LU_incpiv() currently only supports matrices of depth 1",
                         __FILE__, __LINE__ );
      FLA_Abort();
   }
 
   // Create hierarchical copy of matrix A_flat.
   FLASH_Obj_create_hier_copy_of_flat( A_flat, depth, b_flash, A );
 
   // Query the datatype of matrix A_flat.
   datatype = FLA_Obj_datatype( A_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_LU_incpiv_determine_alg_blocksize( *A );
   }
 
   // Query the element (not scalar) dimensions of the new hierarchical matrix.
   // This is done so we can create p and L with full blocks for the bottom
   // and right "edge cases" of A.
   m = FLA_Obj_length( *A );
   n = FLA_Obj_width ( *A );
 
   // Create hierarchical matrices p and L.
   FLASH_Obj_create_ext( FLA_INT,  m * b_flash[0], n, 
                         depth, b_flash, &one, 
                         p );
   
   FLASH_Obj_create_ext( datatype, m * b_flash[0], n * b_alg, 
                         depth, b_flash, &b_alg, 
                         L );
      
   return FLA_SUCCESS;
}

References FLA_Abort(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_width(), FLA_Print_message(), FLASH_LU_incpiv_determine_alg_blocksize(), FLASH_Obj_create_ext(), FLASH_Obj_create_hier_copy_of_flat(), and i.

FLASH_LU_incpiv_determine_alg_blocksize()

dim_t FLASH_LU_incpiv_determine_alg_blocksize

(

FLA_Obj

A

)

{
   dim_t b_alg;
   dim_t b_flash;
 
   // Acquire the storage blocksize.
   b_flash = FLA_Obj_length( *FLASH_OBJ_PTR_AT( A ) );
 
   // 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_LU_INNER_TO_OUTER_B_RATIO, 1 );
 
   return b_alg;
}

References FLA_Obj_length(), and i.

Referenced by FLASH_LU_incpiv_create_hier_matrices().

FLASH_LU_incpiv_noopt()

FLA_Error FLASH_LU_incpiv_noopt	(	FLA_Obj	A,
		FLA_Obj	p,
		FLA_Obj	L
	)

{
  dim_t     nb_alg;
  FLA_Error r_val;
  
  // Inspect the width of a the top-left element of L to get the algorithmic
  // blocksize we'll use throughout the LU_incpiv algorithm.
  nb_alg = FLASH_Obj_scalar_width_tl( L );
 
  // Begin a parallel region.
  FLASH_Queue_begin();
  
  // Enqueue tasks via a SuperMatrix-aware control tree.
  r_val = FLASH_LU_incpiv_var1( A, p, L, nb_alg, flash_lu_incpiv_cntl );
  
  // End the parallel region.
  FLASH_Queue_end();
 
  return r_val;
}

References flash_lu_incpiv_cntl, FLASH_LU_incpiv_var1(), FLASH_Obj_scalar_width_tl(), FLASH_Queue_begin(), FLASH_Queue_end(), and i.

Referenced by FLASH_LU_incpiv().

FLASH_LU_incpiv_opt1()

FLA_Error FLASH_LU_incpiv_opt1	(	FLA_Obj	A,
		FLA_Obj	p,
		FLA_Obj	L
	)

{
  dim_t     nb_alg;
  FLA_Error r_val;
  FLA_Obj   U;
 
  // Inspect the width of a the top-left element of L to get the algorithmic
  // blocksize we'll use throughout the LU_incpiv algorithm.
  nb_alg = FLASH_Obj_scalar_width_tl( L );
 
  // Create a temporary matrix to hold copies of all of the blocks along the
  // diagonal of A.
  FLASH_Obj_create_diag_panel( A, &U );
 
  // Begin a parallel region.
  FLASH_Queue_begin();
  
  // Enqueue tasks via a SuperMatrix-aware control tree.
  r_val = FLASH_LU_incpiv_var2( A, p, L, U, nb_alg, flash_lu_incpiv_cntl );
  
  // End the parallel region.
  FLASH_Queue_end();
 
  // Free the temporary matrix.
  FLASH_Obj_free( &U );
 
  return r_val;
}

References flash_lu_incpiv_cntl, FLASH_LU_incpiv_var2(), FLASH_Obj_create_diag_panel(), FLASH_Obj_free(), FLASH_Obj_scalar_width_tl(), FLASH_Queue_begin(), FLASH_Queue_end(), and i.

Referenced by FLASH_LU_incpiv().

FLASH_LU_incpiv_solve()

FLA_Error FLASH_LU_incpiv_solve	(	FLA_Obj	A,
		FLA_Obj	p,
		FLA_Obj	L,
		FLA_Obj	B,
		FLA_Obj	X
	)

{
  // Check parameters.
  if ( FLA_Check_error_level() >= FLA_MIN_ERROR_CHECKING )
    FLA_LU_incpiv_solve_check( A, p, L, B, X );
 
  FLASH_Copy( B, X );
 
  FLASH_FS_incpiv( A, p, L, X );
  FLASH_Trsm( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE,
              FLA_NONUNIT_DIAG, FLA_ONE, A, X );
 
  return FLA_SUCCESS;
}

References FLA_Check_error_level(), FLA_LU_incpiv_solve_check(), FLA_ONE, FLASH_Copy(), FLASH_FS_incpiv(), FLASH_Trsm(), and i.