libflame  revision_anchor
Functions
FLASH_LU_incpiv_var1.c File Reference

(r)

Functions

FLA_Error FLASH_LU_incpiv_var1 (FLA_Obj A, FLA_Obj p, FLA_Obj L, dim_t nb_alg, fla_lu_t *cntl)
 

Function Documentation

◆ FLASH_LU_incpiv_var1()

FLA_Error FLASH_LU_incpiv_var1 ( FLA_Obj  A,
FLA_Obj  p,
FLA_Obj  L,
dim_t  nb_alg,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_LU_piv_task(), FLA_Obj_length(), FLA_Obj_width(), FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLASH_Queue_get_enabled(), FLASH_SA_LU(), and FLASH_Trsm_piv().

Referenced by FLASH_LU_incpiv_noopt().

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_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_TL );
29 
30  FLA_Part_2x2( p, &pTL, &pTR,
31  &pBL, &pBR, 0, 0, FLA_TL );
32 
33  FLA_Part_2x2( L, &LTL, &LTR,
34  &LBL, &LBR, 0, 0, FLA_TL );
35 
36  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
37  FLA_Obj_width ( ATL ) < FLA_Obj_width ( A ) )
38  {
39  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
40  /* ************* */ /* ******************** */
41  &A10, /**/ &A11, &A12,
42  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
43  1, 1, FLA_BR );
44 
45  FLA_Repart_2x2_to_3x3( pTL, /**/ pTR, &p00, /**/ &p01, &p02,
46  /* ************* */ /* ******************** */
47  &p10, /**/ &p11, &p12,
48  pBL, /**/ pBR, &p20, /**/ &p21, &p22,
49  1, 1, FLA_BR );
50 
51  FLA_Repart_2x2_to_3x3( LTL, /**/ LTR, &L00, /**/ &L01, &L02,
52  /* ************* */ /* ******************** */
53  &L10, /**/ &L11, &L12,
54  LBL, /**/ LBR, &L20, /**/ &L21, &L22,
55  1, 1, FLA_BR );
56 
57  /*------------------------------------------------------------*/
58 
59  if ( FLASH_Queue_get_enabled( ) )
60  {
61  // Enqueue
62  ENQUEUE_FLASH_LU_piv( *FLASH_OBJ_PTR_AT( A11 ),
63  *FLASH_OBJ_PTR_AT( p11 ),
64  FLA_Cntl_sub_lu( cntl ) );
65  }
66  else
67  {
68  // Execute leaf
69  FLA_LU_piv_task( *FLASH_OBJ_PTR_AT( A11 ),
70  *FLASH_OBJ_PTR_AT( p11 ),
71  FLA_Cntl_sub_lu( cntl ) );
72  }
73 
74  FLASH_Trsm_piv( A11, A12, p11,
75  FLA_Cntl_sub_trsm1( cntl ) );
76 
77  FLASH_SA_LU( A11, A12,
78  A21, A22, p21, L21, nb_alg, cntl );
79 
80  /*------------------------------------------------------------*/
81 
82  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
83  A10, A11, /**/ A12,
84  /* ************** */ /* ****************** */
85  &ABL, /**/ &ABR, A20, A21, /**/ A22,
86  FLA_TL );
87 
88  FLA_Cont_with_3x3_to_2x2( &pTL, /**/ &pTR, p00, p01, /**/ p02,
89  p10, p11, /**/ p12,
90  /* ************** */ /* ****************** */
91  &pBL, /**/ &pBR, p20, p21, /**/ p22,
92  FLA_TL );
93 
94  FLA_Cont_with_3x3_to_2x2( &LTL, /**/ &LTR, L00, L01, /**/ L02,
95  L10, L11, /**/ L12,
96  /* ************** */ /* ****************** */
97  &LBL, /**/ &LBR, L20, L21, /**/ L22,
98  FLA_TL );
99  }
100 
101  return FLA_SUCCESS;
102 }
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_task(FLA_Obj A, FLA_Obj p, fla_lu_t *cntl)
Definition: FLA_LU_piv_task.c:15
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 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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116