libflame  revision_anchor
Functions
FLA_Trsm_rln_blk_var1.c File Reference

(r)

Functions

FLA_Error FLA_Trsm_rln_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl)
 

Function Documentation

◆ FLA_Trsm_rln_blk_var1()

FLA_Error FLA_Trsm_rln_blk_var1 ( FLA_Diag  diagA,
FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_trsm_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and FLA_Trsm_internal().

Referenced by FLA_Trsm_rln().

16 {
17  FLA_Obj ATL, ATR, A00, A01, A02,
18  ABL, ABR, A10, A11, A12,
19  A20, A21, A22;
20 
21  FLA_Obj BL, BR, B0, B1, B2;
22 
23  dim_t b;
24 
25  FLA_Scal_internal( alpha, B,
26  FLA_Cntl_sub_scal( cntl ) );
27 
28  FLA_Part_2x2( A, &ATL, &ATR,
29  &ABL, &ABR, 0, 0, FLA_BR );
30 
31  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
32 
33  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
34 
35  b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );
36 
37  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
38  &A10, &A11, /**/ &A12,
39  /* ************* */ /* ******************** */
40  ABL, /**/ ABR, &A20, &A21, /**/ &A22,
41  b, b, FLA_TL );
42 
43  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
44  b, FLA_LEFT );
45 
46  /*------------------------------------------------------------*/
47 
48  /* B1 = B1 - B2 * A21; */
49  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
50  FLA_MINUS_ONE, B2, A21, FLA_ONE, B1,
51  FLA_Cntl_sub_gemm( cntl ) );
52 
53  /* B1 = B1 / tril( A11 ); */
54  FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
55  FLA_ONE, A11, B1,
56  FLA_Cntl_sub_trsm( cntl ) );
57 
58  /*------------------------------------------------------------*/
59 
60  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
61  /* ************** */ /* ****************** */
62  A10, /**/ A11, A12,
63  &ABL, /**/ &ABR, A20, /**/ A21, A22,
64  FLA_BR );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
67  FLA_RIGHT );
68 
69  }
70 
71  return FLA_SUCCESS;
72 }
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
unsigned long dim_t
Definition: FLA_type_defs.h:71
FLA_Obj FLA_MINUS_ONE
Definition: FLA_Init.c:22
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
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_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
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_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_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition: FLA_Scal_internal.c:16
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116
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