libflame  revision_anchor
Functions
FLA_Trsm_lut_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_Trsm_lut_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl)
 

Function Documentation

◆ FLA_Trsm_lut_blk_var1()

FLA_Error FLA_Trsm_lut_blk_var1 ( FLA_Diag  diagA,
FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_trsm_t cntl 
)
16 {
17  FLA_Obj ATL, ATR, A00, A01, A02,
18  ABL, ABR, A10, A11, A12,
19  A20, A21, A22;
20 
21  FLA_Obj BT, B0,
22  BB, B1,
23  B2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( alpha, B,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_TL );
32 
33  FLA_Part_2x1( B, &BT,
34  &BB, 0, FLA_TOP );
35 
36  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
37 
38  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
39 
40  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
41  /* ************* */ /* ******************** */
42  &A10, /**/ &A11, &A12,
43  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
44  b, b, FLA_BR );
45 
46  FLA_Repart_2x1_to_3x1( BT, &B0,
47  /* ** */ /* ** */
48  &B1,
49  BB, &B2, b, FLA_BOTTOM );
50 
51  /*------------------------------------------------------------*/
52 
53  /* B1 = B1 - A01' * B0 */
54  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
55  FLA_MINUS_ONE, A01, B0, FLA_ONE, B1,
56  FLA_Cntl_sub_gemm( cntl ) );
57 
58  /* B1 = triu( A11' ) \ B1 */
59  FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, diagA,
60  FLA_ONE, A11, B1,
61  FLA_Cntl_sub_trsm( cntl ) );
62 
63  /*------------------------------------------------------------*/
64 
65  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
66  A10, A11, /**/ A12,
67  /* ************** */ /* ****************** */
68  &ABL, /**/ &ABR, A20, A21, /**/ A22,
69  FLA_TL );
70 
71  FLA_Cont_with_3x1_to_2x1( &BT, B0,
72  B1,
73  /* ** */ /* ** */
74  &BB, B2, FLA_TOP );
75 
76  }
77 
78  return FLA_SUCCESS;
79 }
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_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition: FLA_Scal_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_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_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
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
unsigned long dim_t
Definition: FLA_type_defs.h:71
Definition: FLA_type_defs.h:159

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and FLA_Trsm_internal().

Referenced by FLA_Trsm_lut().