libflame  revision_anchor
Functions
FLA_LU_piv_blk_var3.c File Reference

(r)

Functions

FLA_Error FLA_LU_piv_blk_var3 (FLA_Obj A, FLA_Obj p, fla_lu_t *cntl)
 

Function Documentation

◆ FLA_LU_piv_blk_var3()

FLA_Error FLA_LU_piv_blk_var3 ( 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 AL, AR, A0, A1, A2;
23 
24  FLA_Obj pT, p0,
25  pB, p1,
26  p2;
27 
28  FLA_Obj AB0, AB1;
29 
30  dim_t b;
31 
32 
33  FLA_Part_2x2( A, &ATL, &ATR,
34  &ABL, &ABR, 0, 0, FLA_TL );
35 
36  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
37 
38  FLA_Part_2x1( p, &pT,
39  &pB, 0, FLA_TOP );
40 
41  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
42  FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
43 
44  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
45 
46  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
47  /* ************* */ /* ******************** */
48  &A10, /**/ &A11, &A12,
49  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
50  b, b, FLA_BR );
51 
52  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
53  b, FLA_RIGHT );
54 
55  FLA_Repart_2x1_to_3x1( pT, &p0,
56  /* ** */ /* ** */
57  &p1,
58  pB, &p2, b, FLA_BOTTOM );
59 
60  /*------------------------------------------------------------*/
61 
62  // Apply previously computed pivots
63  FLA_Apply_pivots_internal( FLA_LEFT, FLA_NO_TRANSPOSE, p0, A1,
64  FLA_Cntl_sub_appiv1( cntl ) );
65 
66  // A01 = trilu( A00 ) \ A10
67  FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
68  FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
69  FLA_ONE, A00, A01,
70  FLA_Cntl_sub_trsm1( cntl ) );
71 
72  // A11 = A11 - A10 * A01
73  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
74  FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
75  FLA_Cntl_sub_gemm1( cntl ) );
76 
77  // A21 = A21 - A20 * A01
78  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
79  FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
80  FLA_Cntl_sub_gemm2( cntl ) );
81 
82  // AB1 = / A11 \
83  // \ A21 /
84  FLA_Merge_2x1( A11,
85  A21, &AB1 );
86 
87  // AB1, p1 = LU_piv( AB1 )
88  r_val = FLA_LU_piv_internal( AB1, p1,
89  FLA_Cntl_sub_lu( cntl ) );
90 
91  // If the unblocked algorithm returns a null pivot,
92  // update the pivot index and return it.
93  if ( r_val == FLA_SUCCESS && r_val_sub >= 0 )
94  {
95  r_val = FLA_Obj_length( A01 ) + r_val_sub;
96  }
97 
98  // AB0 = / A10 \
99  // \ A20 /
100  FLA_Merge_2x1( A10,
101  A20, &AB0 );
102 
103  // Apply pivots to previous columns
104  FLA_Apply_pivots_internal( FLA_LEFT, FLA_NO_TRANSPOSE, p1, AB0,
105  FLA_Cntl_sub_appiv2( cntl ) );
106 
107  /*------------------------------------------------------------*/
108 
109  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
110  A10, A11, /**/ A12,
111  /* ************** */ /* ****************** */
112  &ABL, /**/ &ABR, A20, A21, /**/ A22,
113  FLA_TL );
114 
115  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
116  FLA_LEFT );
117 
118  FLA_Cont_with_3x1_to_2x1( &pT, p0,
119  p1,
120  /* ** */ /* ** */
121  &pB, p2, FLA_TOP );
122 
123  }
124 
125  if ( FLA_Obj_width( ATR ) > 0 )
126  {
127  /* Apply pivots to untouched columns */
128  FLA_Apply_pivots_internal( FLA_LEFT, FLA_NO_TRANSPOSE, p, ATR,
129  FLA_Cntl_sub_appiv1( cntl ) );
130 
131  /* ATR = trilu( ATL ) \ ATR */
132  FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
133  FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
134  FLA_ONE, ATL, ATR );
135  }
136 
137  return r_val;
138 }
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_Error FLA_Trsm_external(FLA_Side side, FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition: FLA_Trsm_external.c:13
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
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
dim_t FLA_Obj_width(FLA_Obj obj)
Definition: FLA_Query.c:123
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
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
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
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_1x3_to_1x2(), 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_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_piv_internal().