libflame  revision_anchor
Functions
FLA_LU_piv_blk_var4.c File Reference

(r)

Functions

FLA_Error FLA_LU_piv_blk_var4 (FLA_Obj A, FLA_Obj p, fla_lu_t *cntl)
 

Function Documentation

◆ FLA_LU_piv_blk_var4()

FLA_Error FLA_LU_piv_blk_var4 ( FLA_Obj  A,
FLA_Obj  p,
fla_lu_t cntl 
)
16 {
17  FLA_Error r_val = FLA_SUCCESS, r_val_sub = FLA_SUCCESS;
18  FLA_Obj ATL, ATR, A00, A01, A02,
19  ABL, ABR, A10, A11, A12,
20  A20, A21, A22;
21 
22  FLA_Obj pT, p0,
23  pB, p1,
24  p2;
25 
26  FLA_Obj AB0, AB1, AB2;
27 
28  dim_t b;
29 
30 
31  FLA_Part_2x2( A, &ATL, &ATR,
32  &ABL, &ABR, 0, 0, FLA_TL );
33 
34  FLA_Part_2x1( p, &pT,
35  &pB, 0, FLA_TOP );
36 
37  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
38  FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
39 
40  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
41 
42  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
43  /* ************* */ /* ******************** */
44  &A10, /**/ &A11, &A12,
45  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
46  b, b, FLA_BR );
47 
48  FLA_Repart_2x1_to_3x1( pT, &p0,
49  /* ** */ /* ** */
50  &p1,
51  pB, &p2, b, FLA_BOTTOM );
52 
53  /*------------------------------------------------------------*/
54 
55  // A11 = A11 - A10 * A0
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
57  FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  // A21 = A21 - A20 * A01
61  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
62  FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
63  FLA_Cntl_sub_gemm3( cntl ) );
64 
65  // AB1 = / A11 \
66  // \ A21 /
67  FLA_Merge_2x1( A11,
68  A21, &AB1 );
69 
70  // AB1, p1 = LU_piv( AB1 )
71  FLA_LU_piv_internal( AB1, p1,
72  FLA_Cntl_sub_lu( cntl ) );
73 
74  // If the unblocked algorithm returns a null pivot,
75  // update the pivot index and return it.
76  if ( r_val == FLA_SUCCESS && r_val_sub >= 0 )
77  {
78  r_val = FLA_Obj_length( A01 ) + r_val_sub;
79  }
80 
81  // AB0 = / A10 \
82  // \ A20 /
83  FLA_Merge_2x1( A10,
84  A20, &AB0 );
85 
86  // AB2 = / A12 \
87  // \ A22 /
88  FLA_Merge_2x1( A12,
89  A22, &AB2 );
90 
91  // Apply pivots to remaining columns
92  FLA_Apply_pivots_internal( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB0,
93  FLA_Cntl_sub_appiv1( cntl ) );
94  FLA_Apply_pivots_internal( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB2,
95  FLA_Cntl_sub_appiv1( cntl ) );
96 
97  // A12 = A12 - A10 * A02
98  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
99  FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
100  FLA_Cntl_sub_gemm2( cntl ) );
101 
102  // A12 = trilu( A11 ) \ A12
103  FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
104  FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
105  FLA_ONE, A11, A12,
106  FLA_Cntl_sub_trsm1( cntl ) );
107 
108  /*------------------------------------------------------------*/
109 
110  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
111  A10, A11, /**/ A12,
112  /* ************** */ /* ****************** */
113  &ABL, /**/ &ABR, A20, A21, /**/ A22,
114  FLA_TL );
115 
116  FLA_Cont_with_3x1_to_2x1( &pT, p0,
117  p1,
118  /* ** */ /* ** */
119  &pB, p2, FLA_TOP );
120 
121  }
122 
123  return r_val;
124 }
FLA_Error FLA_Apply_pivots_internal(FLA_Side side, FLA_Trans trans, FLA_Obj p, FLA_Obj A, fla_appiv_t *cntl)
Definition: FLA_Apply_pivots_internal.c:13
FLA_Error FLA_Gemm_internal(FLA_Trans transa, FLA_Trans transb, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
Definition: FLA_Gemm_internal.c:16
FLA_Error FLA_LU_piv_internal(FLA_Obj A, FLA_Obj p, fla_lu_t *cntl)
Definition: FLA_LU_piv_internal.c:15
FLA_Error FLA_Trsm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl)
Definition: FLA_Trsm_internal.c:16
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
FLA_Obj FLA_ONE
Definition: FLA_Init.c:18
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 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_Error FLA_Cont_with_3x1_to_2x1(FLA_Obj *AT, FLA_Obj A0, FLA_Obj A1, FLA_Obj *AB, FLA_Obj A2, FLA_Side side)
Definition: FLA_View.c:428
FLA_Error FLA_Repart_2x1_to_3x1(FLA_Obj AT, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj AB, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:226
dim_t FLA_Obj_width(FLA_Obj obj)
Definition: FLA_Query.c:123
FLA_Error FLA_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:76
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
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
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
FLA_Error FLA_Merge_2x1(FLA_Obj AT, FLA_Obj AB, FLA_Obj *A)
Definition: FLA_View.c:541
int FLA_Error
Definition: FLA_type_defs.h:47
unsigned long dim_t
Definition: FLA_type_defs.h:71
Definition: FLA_type_defs.h:159

References FLA_Apply_pivots_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_piv_internal(), FLA_Merge_2x1(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_LU_piv_internal().