libflame  revision_anchor
Functions
FLA_Chol_l_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Chol_l_blk_var2 (FLA_Obj A, fla_chol_t *cntl)
 

Function Documentation

◆ FLA_Chol_l_blk_var2()

FLA_Error FLA_Chol_l_blk_var2 ( FLA_Obj  A,
fla_chol_t cntl 
)

References FLA_Chol_internal(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_Chol_l().

16 {
17  FLA_Obj ATL, ATR, A00, A01, A02,
18  ABL, ABR, A10, A11, A12,
19  A20, A21, A22;
20 
21  dim_t b;
22 
23  int r_val = FLA_SUCCESS;
24 
25  FLA_Part_2x2( A, &ATL, &ATR,
26  &ABL, &ABR, 0, 0, FLA_TL );
27 
28  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
29 
30  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
31 
32  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
33  /* ************* */ /* ******************** */
34  &A10, /**/ &A11, &A12,
35  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
36  b, b, FLA_BR );
37 
38  /*------------------------------------------------------------*/
39 
40  // A11 = A11 - A10 * A10'
41  FLA_Herk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE,
42  FLA_MINUS_ONE, A10, FLA_ONE, A11,
43  FLA_Cntl_sub_herk( cntl ) );
44 
45  // A21 = A21 - A20 * A10'
46  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
47  FLA_MINUS_ONE, A20, A10, FLA_ONE, A21,
48  FLA_Cntl_sub_gemm( cntl ) );
49 
50  // A11 = chol( A11 )
51  r_val = FLA_Chol_internal( FLA_LOWER_TRIANGULAR, A11,
52  FLA_Cntl_sub_chol( cntl ) );
53 
54  if ( r_val != FLA_SUCCESS )
55  return ( FLA_Obj_length( A00 ) + r_val );
56 
57  // A21 = A21 * inv( tril( A11 )' )
58  FLA_Trsm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR,
59  FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
60  FLA_ONE, A11, A21,
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 
72  return r_val;
73 }
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_Herk_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl)
Definition: FLA_Herk_internal.c:16
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_Chol_internal(FLA_Uplo uplo, FLA_Obj A, fla_chol_t *cntl)
Definition: FLA_Chol_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