libflame  revision_anchor
Functions
FLA_QR2_UT_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_QR2_UT_blk_var1 (FLA_Obj U, FLA_Obj D, FLA_Obj T, fla_qr2ut_t *cntl)
 

Function Documentation

◆ FLA_QR2_UT_blk_var1()

FLA_Error FLA_QR2_UT_blk_var1 ( FLA_Obj  U,
FLA_Obj  D,
FLA_Obj  T,
fla_qr2ut_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Copy_internal(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_min_dim(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_QR2_UT_internal(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_QR2_UT_internal().

15 {
16  FLA_Obj UTL, UTR, U00, U01, U02,
17  UBL, UBR, U10, U11, U12,
18  U20, U21, U22;
19 
20  FLA_Obj DL, DR, D0, D1, D2;
21 
22  FLA_Obj TL, TR, T0, T1, W12;
23 
24  FLA_Obj W12T, W12B;
25 
26  FLA_Obj T1T, T2B;
27 
28  dim_t b_alg, b;
29 
30  // Query the algorithmic blocksize by inspecting the length of T.
31  b_alg = FLA_Obj_length( T );
32 
33  FLA_Part_2x2( U, &UTL, &UTR,
34  &UBL, &UBR, 0, 0, FLA_TL );
35 
36  FLA_Part_1x2( D, &DL, &DR, 0, FLA_LEFT );
37 
38  FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
39 
40  while ( FLA_Obj_min_dim( UBR ) > 0 ){
41 
42  b = min( b_alg, FLA_Obj_min_dim( UBR ) );
43 
44  FLA_Repart_2x2_to_3x3( UTL, /**/ UTR, &U00, /**/ &U01, &U02,
45  /* ************* */ /* ******************** */
46  &U10, /**/ &U11, &U12,
47  UBL, /**/ UBR, &U20, /**/ &U21, &U22,
48  b, b, FLA_BR );
49 
50  FLA_Repart_1x2_to_1x3( DL, /**/ DR, &D0, /**/ &D1, &D2,
51  b, FLA_RIGHT );
52 
53  FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &W12,
54  b, FLA_RIGHT );
55 
56  /*------------------------------------------------------------*/
57 
58  // T1T = FLA_Top_part( T1, b );
59 
60  FLA_Part_2x1( T1, &T1T,
61  &T2B, b, FLA_TOP );
62 
63  // [ U11, ...
64  // D1, T1 ] = FLA_QR2_UT( U11
65  // D1, T1T );
66 
68  D1, T1T,
69  FLA_Cntl_sub_qr2ut( cntl ) );
70 
71 
72  if ( FLA_Obj_width( U12 ) > 0 )
73  {
74  // W12T = FLA_Top_part( W12, b );
75 
76  FLA_Part_2x1( W12, &W12T,
77  &W12B, b, FLA_TOP );
78 
79  // W12T = inv( triu( T1T ) )' * ( U12 + D1' * D2 );
80 
81  FLA_Copy_internal( U12, W12T,
82  FLA_Cntl_sub_copy( cntl ) );
83 
84  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
85  FLA_ONE, D1, D2, FLA_ONE, W12T,
86  FLA_Cntl_sub_gemm1( cntl ) );
87 
88  FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
89  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
90  FLA_ONE, T1T, W12T,
91  FLA_Cntl_sub_trsm( cntl ) );
92 
93  // U12 = U12 - W12T;
94  // D2 = D2 - D1 * W12T;
95 
96  FLA_Axpy_internal( FLA_MINUS_ONE, W12T, U12,
97  FLA_Cntl_sub_axpy( cntl ) );
98 
99  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
100  FLA_MINUS_ONE, D1, W12T, FLA_ONE, D2,
101  FLA_Cntl_sub_gemm2( cntl ) );
102  }
103 
104  /*------------------------------------------------------------*/
105 
106  FLA_Cont_with_3x3_to_2x2( &UTL, /**/ &UTR, U00, U01, /**/ U02,
107  U10, U11, /**/ U12,
108  /* ************** */ /* ****************** */
109  &UBL, /**/ &UBR, U20, U21, /**/ U22,
110  FLA_TL );
111 
112  FLA_Cont_with_1x3_to_1x2( &DL, /**/ &DR, D0, D1, /**/ D2,
113  FLA_LEFT );
114 
115  FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ W12,
116  FLA_LEFT );
117 
118  }
119 
120  return FLA_SUCCESS;
121 }
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_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_Copy_internal(FLA_Obj A, FLA_Obj B, fla_copy_t *cntl)
Definition: FLA_Copy_internal.c:16
unsigned long dim_t
Definition: FLA_type_defs.h:71
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_QR2_UT_internal(FLA_Obj U, FLA_Obj D, FLA_Obj T, fla_qr2ut_t *cntl)
Definition: FLA_QR2_UT_internal.c:16
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_Axpy_internal(FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
Definition: FLA_Axpy_internal.c:16
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
dim_t FLA_Obj_min_dim(FLA_Obj obj)
Definition: FLA_Query.c:153
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