libflame  revision_anchor
Functions
FLA_Apply_Q_UT_rnfc_blk_var3.c File Reference

(r)

Functions

FLA_Error FLA_Apply_Q_UT_rnfc_blk_var3 (FLA_Obj A, FLA_Obj TW, FLA_Obj W, FLA_Obj B, fla_apqut_t *cntl)
 

Function Documentation

◆ FLA_Apply_Q_UT_rnfc_blk_var3()

FLA_Error FLA_Apply_Q_UT_rnfc_blk_var3 ( FLA_Obj  A,
FLA_Obj  TW,
FLA_Obj  W,
FLA_Obj  B,
fla_apqut_t cntl 
)
14 {
15  FLA_Obj ATL, ATR, A00, A01, A02,
16  ABL, ABR, A10, A11, A12,
17  A20, A21, A22;
18 
19  FLA_Obj TWTL, TWTR, TW00, TW01, TW02,
20  TWBL, TWBR, TW10, T11, W12,
21  TW20, TW21, TW22;
22 
23  FLA_Obj WTL, WTR,
24  WBL, WBR;
25 
26  FLA_Obj BL, BR, B0, B1, B2;
27 
28  dim_t b;
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_TL );
32 
33  FLA_Part_2x2( TW, &TWTL, &TWTR,
34  &TWBL, &TWBR, 0, 0, FLA_TL );
35 
36  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
37 
38  while ( FLA_Obj_min_dim( ABR ) > 0 ){
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_2x2_to_3x3( TWTL, /**/ TWTR, &TW00, /**/ &TW01, &TW02,
49  /* *************** */ /* *********************** */
50  &TW10, /**/ &T11, &W12,
51  TWBL, /**/ TWBR, &TW20, /**/ &TW21, &TW22,
52  b, b, FLA_BR );
53 
54  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
55  b, FLA_RIGHT );
56 
57  /*------------------------------------------------------------*/
58 
59  FLA_Part_2x2( W, &WTL, &WTR,
60  &WBL, &WBR, b, FLA_Obj_length( B1 ), FLA_TL );
61 
62  // WTL = B1^T;
63 
64  FLA_Copyt_internal( FLA_TRANSPOSE, B1, WTL,
65  FLA_Cntl_sub_copyt( cntl ) );
66 
67  // U11 = trilu( A11 );
68  // U21 = A21;
69  // Let WTL^T be conformal to B1.
70  //
71  // WTL^T = ( B1 * U11 + B2 * U21 ) * inv( triu(T11) );
72  // WTL = inv( triu(T11)^T ) * ( U11^T * B1^T + U21^T * B2^T );
73 
74  FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
75  FLA_TRANSPOSE, FLA_UNIT_DIAG,
76  FLA_ONE, A11, WTL,
77  FLA_Cntl_sub_trmm1( cntl ) );
78 
79  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_TRANSPOSE,
80  FLA_ONE, A21, B2, FLA_ONE, WTL,
81  FLA_Cntl_sub_gemm1( cntl ) );
82 
83  FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
84  FLA_TRANSPOSE, FLA_NONUNIT_DIAG,
85  FLA_ONE, T11, WTL,
86  FLA_Cntl_sub_trsm( cntl ) );
87 
88  // B2 = B2 - WTL^T * U21';
89  // B1 = B1 - WTL^T * U11';
90  // = B1 - ( conj(U11) * WTL )^T;
91 
92  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE,
93  FLA_MINUS_ONE, WTL, A21, FLA_ONE, B2,
94  FLA_Cntl_sub_gemm2( cntl ) );
95 
96  FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
97  FLA_CONJ_NO_TRANSPOSE, FLA_UNIT_DIAG,
98  FLA_MINUS_ONE, A11, WTL,
99  FLA_Cntl_sub_trmm2( cntl ) );
100 
101  FLA_Axpyt_internal( FLA_TRANSPOSE, FLA_ONE, WTL, B1,
102  FLA_Cntl_sub_axpyt( cntl ) );
103 
104  /*------------------------------------------------------------*/
105 
106  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
107  A10, A11, /**/ A12,
108  /* ************** */ /* ****************** */
109  &ABL, /**/ &ABR, A20, A21, /**/ A22,
110  FLA_TL );
111 
112  FLA_Cont_with_3x3_to_2x2( &TWTL, /**/ &TWTR, TW00, TW01, /**/ TW02,
113  TW10, T11, /**/ W12,
114  /* **************** */ /* ********************* */
115  &TWBL, /**/ &TWBR, TW20, TW21, /**/ TW22,
116  FLA_TL );
117 
118  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
119  FLA_LEFT );
120  }
121 
122  return FLA_SUCCESS;
123 }
FLA_Error FLA_Axpyt_internal(FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpyt_t *cntl)
Definition: FLA_Axpyt_internal.c:16
FLA_Error FLA_Copyt_internal(FLA_Trans trans, FLA_Obj A, FLA_Obj B, fla_copyt_t *cntl)
Definition: FLA_Copyt_internal.c:16
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_Trmm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl)
Definition: FLA_Trmm_internal.c:16
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_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_1x2(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:110
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
dim_t FLA_Obj_min_dim(FLA_Obj obj)
Definition: FLA_Query.c:153
unsigned long dim_t
Definition: FLA_type_defs.h:71
Definition: FLA_type_defs.h:159

References FLA_Axpyt_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Copyt_internal(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_min_dim(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Trmm_internal(), and FLA_Trsm_internal().

Referenced by FLA_Apply_Q_UT_rnfc().