libflame  revision_anchor
Functions
FLA_LU_nopiv_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_LU_nopiv_blk_var2 (FLA_Obj A, fla_lu_t *cntl)
 

Function Documentation

◆ FLA_LU_nopiv_blk_var2()

FLA_Error FLA_LU_nopiv_blk_var2 ( FLA_Obj  A,
fla_lu_t cntl 
)

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

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 
24  FLA_Part_2x2( A, &ATL, &ATR,
25  &ABL, &ABR, 0, 0, FLA_TL );
26 
27  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
28  FLA_Obj_width( ATL ) < FLA_Obj_width( 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  // A10 = A10 / triu( A00 )
41  FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
42  FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
43  FLA_ONE, A00, A10,
44  FLA_Cntl_sub_trsm1( cntl ) );
45 
46  // A11 = LU( A11 - A10 * A01 )
47  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
48  FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
49  FLA_Cntl_sub_gemm1( cntl ) );
50 
51  // A11 = LU_nopiv( A11 )
52  FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
53 
54  // A12 = A12 - A10 * A02
55  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
56  FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
57  FLA_Cntl_sub_gemm2( cntl ) );
58 
59  // A12 = trilu( A11 ) \ A12
60  FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
61  FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
62  FLA_ONE, A11, A12,
63  FLA_Cntl_sub_trsm2( cntl ) );
64 
65  /*------------------------------------------------------------*/
66 
67  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
68  A10, A11, /**/ A12,
69  /* ************** */ /* ****************** */
70  &ABL, /**/ &ABR, A20, A21, /**/ A22,
71  FLA_TL );
72  }
73 
74  if ( FLA_Obj_length( ABL ) > 0 )
75  // ABL = ABL / triu( ATL )
76  FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
77  FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
78  FLA_ONE, ATL, ABL );
79 
80  return FLA_SUCCESS;
81 }
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
FLA_Error FLA_LU_nopiv_internal(FLA_Obj A, fla_lu_t *cntl)
Definition: FLA_LU_nopiv_internal.c:16
Definition: FLA_type_defs.h:158
dim_t FLA_Obj_width(FLA_Obj obj)
Definition: FLA_Query.c:123
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_Trsm_external(FLA_Side side, FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition: FLA_Trsm_external.c:13
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