libflame  revision_anchor
Functions
FLASH_LU_incpiv_var2.c File Reference

(r)

Functions

FLA_Error FLASH_LU_incpiv_var2 (FLA_Obj A, FLA_Obj p, FLA_Obj L, FLA_Obj U, dim_t nb_alg, fla_lu_t *cntl)
 

Function Documentation

◆ FLASH_LU_incpiv_var2()

FLA_Error FLASH_LU_incpiv_var2 ( FLA_Obj  A,
FLA_Obj  p,
FLA_Obj  L,
FLA_Obj  U,
dim_t  nb_alg,
fla_lu_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_LU_piv_copy_task(), FLA_Obj_length(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLASH_Queue_get_enabled(), FLASH_SA_LU(), and FLASH_Trsm_piv().

Referenced by FLASH_LU_incpiv_opt1().

14 {
15  FLA_Obj ATL, ATR, A00, A01, A02,
16  ABL, ABR, A10, A11, A12,
17  A20, A21, A22;
18 
19  FLA_Obj pTL, pTR, p00, p01, p02,
20  pBL, pBR, p10, p11, p12,
21  p20, p21, p22;
22 
23  FLA_Obj LTL, LTR, L00, L01, L02,
24  LBL, LBR, L10, L11, L12,
25  L20, L21, L22;
26 
27  FLA_Obj UL, UR, U0, U1, U2;
28 
29  FLA_Part_2x2( A, &ATL, &ATR,
30  &ABL, &ABR, 0, 0, FLA_TL );
31 
32  FLA_Part_2x2( p, &pTL, &pTR,
33  &pBL, &pBR, 0, 0, FLA_TL );
34 
35  FLA_Part_2x2( L, &LTL, &LTR,
36  &LBL, &LBR, 0, 0, FLA_TL );
37 
38  FLA_Part_1x2( U, &UL, &UR, 0, FLA_LEFT );
39 
40  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
41  FLA_Obj_width ( ATL ) < FLA_Obj_width ( A ) )
42  {
43  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
44  /* ************* */ /* ******************** */
45  &A10, /**/ &A11, &A12,
46  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
47  1, 1, FLA_BR );
48 
49  FLA_Repart_2x2_to_3x3( pTL, /**/ pTR, &p00, /**/ &p01, &p02,
50  /* ************* */ /* ******************** */
51  &p10, /**/ &p11, &p12,
52  pBL, /**/ pBR, &p20, /**/ &p21, &p22,
53  1, 1, FLA_BR );
54 
55  FLA_Repart_2x2_to_3x3( LTL, /**/ LTR, &L00, /**/ &L01, &L02,
56  /* ************* */ /* ******************** */
57  &L10, /**/ &L11, &L12,
58  LBL, /**/ LBR, &L20, /**/ &L21, &L22,
59  1, 1, FLA_BR );
60 
61  FLA_Repart_1x2_to_1x3( UL, /**/ UR, &U0, /**/ &U1, &U2,
62  1, FLA_RIGHT );
63 
64  /*------------------------------------------------------------*/
65 
66  if ( FLASH_Queue_get_enabled( ) )
67  {
68  // Enqueue
69  ENQUEUE_FLASH_LU_piv_copy( *FLASH_OBJ_PTR_AT( A11 ),
70  *FLASH_OBJ_PTR_AT( p11 ),
71  *FLASH_OBJ_PTR_AT( U1 ),
72  FLA_Cntl_sub_lu( cntl ) );
73  }
74  else
75  {
76  // Execute leaf
77  FLA_LU_piv_copy_task( *FLASH_OBJ_PTR_AT( A11 ),
78  *FLASH_OBJ_PTR_AT( p11 ),
79  *FLASH_OBJ_PTR_AT( U1 ),
80  FLA_Cntl_sub_lu( cntl ) );
81  }
82 
83  FLASH_Trsm_piv( U1, A12, p11,
84  FLA_Cntl_sub_trsm1( cntl ) );
85 
86  FLASH_SA_LU( A11, A12,
87  A21, A22, p21, L21, nb_alg, cntl );
88 
89  /*------------------------------------------------------------*/
90 
91  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
92  A10, A11, /**/ A12,
93  /* ************** */ /* ****************** */
94  &ABL, /**/ &ABR, A20, A21, /**/ A22,
95  FLA_TL );
96 
97  FLA_Cont_with_3x3_to_2x2( &pTL, /**/ &pTR, p00, p01, /**/ p02,
98  p10, p11, /**/ p12,
99  /* ************** */ /* ****************** */
100  &pBL, /**/ &pBR, p20, p21, /**/ p22,
101  FLA_TL );
102 
103  FLA_Cont_with_3x3_to_2x2( &LTL, /**/ &LTR, L00, L01, /**/ L02,
104  L10, L11, /**/ L12,
105  /* ************** */ /* ****************** */
106  &LBL, /**/ &LBR, L20, L21, /**/ L22,
107  FLA_TL );
108 
109  FLA_Cont_with_1x3_to_1x2( &UL, /**/ &UR, U0, U1, /**/ U2,
110  FLA_LEFT );
111  }
112 
113  return FLA_SUCCESS;
114 }
FLA_Error FLA_Repart_1x2_to_1x3(FLA_Obj AL, FLA_Obj AR, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:267
FLA_Error FLA_Repart_2x2_to_3x3(FLA_Obj ATL, FLA_Obj ATR, FLA_Obj *A00, FLA_Obj *A01, FLA_Obj *A02, FLA_Obj *A10, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj ABL, FLA_Obj ABR, FLA_Obj *A20, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition: FLA_View.c:142
FLA_Error FLA_LU_piv_copy_task(FLA_Obj A, FLA_Obj p, FLA_Obj U, fla_lu_t *cntl)
Definition: FLA_LU_piv_copy_task.c:13
FLA_Error FLA_Part_2x2(FLA_Obj A, FLA_Obj *A11, FLA_Obj *A12, FLA_Obj *A21, FLA_Obj *A22, dim_t mb, dim_t nb, FLA_Quadrant quadrant)
Definition: FLA_View.c:17
FLA_Bool FLASH_Queue_get_enabled(void)
Definition: FLASH_Queue.c:171
Definition: FLA_type_defs.h:158
dim_t FLA_Obj_width(FLA_Obj obj)
Definition: FLA_Query.c:123
FLA_Error FLA_Cont_with_1x3_to_1x2(FLA_Obj *AL, FLA_Obj *AR, FLA_Obj A0, FLA_Obj A1, FLA_Obj A2, FLA_Side side)
Definition: FLA_View.c:475
FLA_Error FLASH_SA_LU(FLA_Obj B, FLA_Obj C, FLA_Obj D, FLA_Obj E, FLA_Obj p, FLA_Obj L, dim_t nb_alg, fla_lu_t *cntl)
Definition: FLASH_SA_LU.c:13
FLA_Error FLA_Cont_with_3x3_to_2x2(FLA_Obj *ATL, FLA_Obj *ATR, FLA_Obj A00, FLA_Obj A01, FLA_Obj A02, FLA_Obj A10, FLA_Obj A11, FLA_Obj A12, FLA_Obj *ABL, FLA_Obj *ABR, FLA_Obj A20, FLA_Obj A21, FLA_Obj A22, FLA_Quadrant quadrant)
Definition: FLA_View.c:304
FLA_Error FLASH_Trsm_piv(FLA_Obj A, FLA_Obj B, FLA_Obj p, fla_trsm_t *cntl)
Definition: FLASH_Trsm_piv.c:13
FLA_Error FLA_Part_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:110
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116