libflame  revision_anchor
Functions
FLA_LU_piv_unb_var3.c File Reference

(r)

Functions

FLA_Error FLA_LU_piv_unb_var3 (FLA_Obj A, FLA_Obj p)
 

Function Documentation

◆ FLA_LU_piv_unb_var3()

FLA_Error FLA_LU_piv_unb_var3 ( FLA_Obj  A,
FLA_Obj  p 
)

References FLA_Amax_external(), FLA_Apply_pivots(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Inv_scal_external(), 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_Trsv_external().

Referenced by FLA_LU_piv_internal().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj AL, AR, A0, a1, A2;
22 
23  FLA_Obj pT, p0,
24  pB, pi1,
25  p2;
26 
27  FLA_Obj AB0, aB1;
28 
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_TL );
32 
33  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
34 
35  FLA_Part_2x1( p, &pT,
36  &pB, 0, FLA_TOP );
37 
38  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
39  FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
42  /* ************* */ /* ************************** */
43  &a10t, /**/ &alpha11, &a12t,
44  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
45  1, 1, FLA_BR );
46 
47  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
48  1, FLA_RIGHT );
49 
50  FLA_Repart_2x1_to_3x1( pT, &p0,
51  /* ** */ /* *** */
52  &pi1,
53  pB, &p2, 1, FLA_BOTTOM );
54 
55  /*------------------------------------------------------------*/
56 
57  // Apply previously computed pivots
58  FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p0, a1 );
59 
60  // a01 = trilu( A00 ) \ a01
61  FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
62 
63  // alpha11 = alpha11 - a10t * a01
64  FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
65 
66  // a21 = a21 - A20 * a01
67  FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
68 
69  // aB1 = / alpha11 \
70  // \ a21 /
71  FLA_Merge_2x1( alpha11,
72  a21, &aB1 );
73 
74  // Determine pivot index
75  FLA_Amax_external( aB1, pi1 );
76 
77  // Apply pivots to current column
78  FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, pi1, aB1 );
79 
80  // a21 = a21 / alpha11
81  FLA_Inv_scal_external( alpha11, a21 );
82 
83  // AB0 = / a10t \
84  // \ A20 /
85  FLA_Merge_2x1( a10t,
86  A20, &AB0 );
87 
88  // Apply pivots to previous columns
89  FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, pi1, AB0 );
90 
91  /*------------------------------------------------------------*/
92 
93  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
94  a10t, alpha11, /**/ a12t,
95  /* ************** */ /* ************************ */
96  &ABL, /**/ &ABR, A20, a21, /**/ A22,
97  FLA_TL );
98 
99  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
100  FLA_LEFT );
101 
102  FLA_Cont_with_3x1_to_2x1( &pT, p0,
103  pi1,
104  /* ** */ /* *** */
105  &pB, p2, FLA_TOP );
106 
107  }
108 
109  if ( FLA_Obj_width( ATR ) > 0 )
110  {
111  /* Apply pivots to untouched columns */
112  FLA_Apply_pivots( FLA_LEFT, FLA_NO_TRANSPOSE, p, ATR );
113 
114  /* ATR = trilu( ATL ) \ ATR */
115  FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
116  FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
117  FLA_ONE, ATL, ATR );
118  }
119 
120  return FLA_SUCCESS;
121 }
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_Dots_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj beta, FLA_Obj rho)
Definition: FLA_Dots_external.c:13
FLA_Error FLA_Inv_scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Inv_scal_external.c:13
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_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv_external.c:13
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
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_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_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_Obj FLA_ONE
Definition: FLA_Init.c:18
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 FLA_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:76
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_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:110
FLA_Error FLA_Apply_pivots(FLA_Side side, FLA_Trans trans, FLA_Obj p, FLA_Obj A)
Definition: FLA_Apply_pivots.c:15
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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
FLA_Error FLA_Amax_external(FLA_Obj x, FLA_Obj index)
Definition: FLA_Amax_external.c:13
FLA_Error FLA_Trsv_external(FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj A, FLA_Obj x)
Definition: FLA_Trsv_external.c:13
FLA_Error FLA_Merge_2x1(FLA_Obj AT, FLA_Obj AB, FLA_Obj *A)
Definition: FLA_View.c:541