libflame  revision_anchor
Functions
FLA_Apply_QUD_UT_lhfc_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_Apply_QUD_UT_lhfc_blk_var1 (FLA_Obj T, FLA_Obj W, FLA_Obj R, FLA_Obj U, FLA_Obj C, FLA_Obj V, FLA_Obj D, fla_apqudut_t *cntl)
 

Function Documentation

◆ FLA_Apply_QUD_UT_lhfc_blk_var1()

FLA_Error FLA_Apply_QUD_UT_lhfc_blk_var1 ( FLA_Obj  T,
FLA_Obj  W,
FLA_Obj  R,
FLA_Obj  U,
FLA_Obj  C,
FLA_Obj  V,
FLA_Obj  D,
fla_apqudut_t cntl 
)
17 {
18  FLA_Obj TL, TR, T0, T1, T2;
19 
20  FLA_Obj UL, UR, U0, U1, U2;
21 
22  FLA_Obj VL, VR, V0, V1, V2;
23 
24  FLA_Obj RT, R0,
25  RB, R1,
26  R2;
27 
28  FLA_Obj T1T,
29  T1B;
30 
31  FLA_Obj W1TL, W1TR,
32  W1BL, W1BR;
33 
34  dim_t b_alg, b;
35 
36  // Query the algorithmic blocksize by inspecting the length of T.
37  b_alg = FLA_Obj_length( T );
38 
39  FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
40 
41  FLA_Part_1x2( U, &UL, &UR, 0, FLA_LEFT );
42 
43  FLA_Part_1x2( V, &VL, &VR, 0, FLA_LEFT );
44 
45  FLA_Part_2x1( R, &RT,
46  &RB, 0, FLA_TOP );
47 
48  while ( FLA_Obj_width( UL ) < FLA_Obj_width( U ) ){
49 
50  b = min( b_alg, FLA_Obj_width( UR ) );
51 
52  FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &T2,
53  b, FLA_RIGHT );
54 
55  FLA_Repart_1x2_to_1x3( UL, /**/ UR, &U0, /**/ &U1, &U2,
56  b, FLA_RIGHT );
57 
58  FLA_Repart_1x2_to_1x3( VL, /**/ VR, &V0, /**/ &V1, &V2,
59  b, FLA_RIGHT );
60 
61  FLA_Repart_2x1_to_3x1( RT, &R0,
62  /* ** */ /* ** */
63  &R1,
64  RB, &R2, b, FLA_BOTTOM );
65 
66  /*------------------------------------------------------------*/
67 
68  FLA_Part_2x1( T1, &T1T,
69  &T1B, b, FLA_TOP );
70 
71  FLA_Part_2x2( W, &W1TL, &W1TR,
72  &W1BL, &W1BR, b, FLA_Obj_width( R1 ), FLA_TL );
73 
74  // W1TL = R1;
75 
76  FLA_Copyt_internal( FLA_NO_TRANSPOSE, R1, W1TL,
77  FLA_Cntl_sub_copyt( cntl ) );
78 
79  // W1TL = inv( triu( T1T ) )' * ( R1 + U1' * C + V1' * D );
80 
81  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
82  FLA_ONE, U1, C, FLA_ONE, W1TL,
83  FLA_Cntl_sub_gemm1( cntl ) );
84 
85  FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
86  FLA_ONE, V1, D, FLA_ONE, W1TL,
87  FLA_Cntl_sub_gemm2( cntl ) );
88 
89  FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
90  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
91  FLA_ONE, T1T, W1TL,
92  FLA_Cntl_sub_trsm( cntl ) );
93 
94  // R1 = R1 - W1TL;
95  // C = C - U1 * W1TL;
96  // D = D + V1 * W1TL;
97 
98  FLA_Axpyt_internal( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, W1TL, R1,
99  FLA_Cntl_sub_axpyt( cntl ) );
100 
101  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
102  FLA_MINUS_ONE, U1, W1TL, FLA_ONE, C,
103  FLA_Cntl_sub_gemm3( cntl ) );
104 
105  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
106  FLA_ONE, V1, W1TL, FLA_ONE, D,
107  FLA_Cntl_sub_gemm4( cntl ) );
108 
109  /*------------------------------------------------------------*/
110 
111  FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ T2,
112  FLA_LEFT );
113 
114  FLA_Cont_with_1x3_to_1x2( &UL, /**/ &UR, U0, U1, /**/ U2,
115  FLA_LEFT );
116 
117  FLA_Cont_with_1x3_to_1x2( &VL, /**/ &VR, V0, V1, /**/ V2,
118  FLA_LEFT );
119 
120  FLA_Cont_with_3x1_to_2x1( &RT, R0,
121  R1,
122  /* ** */ /* ** */
123  &RB, R2, FLA_TOP );
124  }
125 
126  return FLA_SUCCESS;
127 }
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_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_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_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
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_3x1_to_2x1(), FLA_Copyt_internal(), FLA_Gemm_internal(), 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(), and FLA_Trsm_internal().

Referenced by FLA_Apply_QUD_UT_lhfc().