libflame  revision_anchor
Functions
FLA_Trinv_ln_blk_var4.c File Reference

(r)

Functions

FLA_Error FLA_Trinv_ln_blk_var4 (FLA_Obj A, fla_trinv_t *cntl)
 

Function Documentation

◆ FLA_Trinv_ln_blk_var4()

FLA_Error FLA_Trinv_ln_blk_var4 ( FLA_Obj  A,
fla_trinv_t cntl 
)
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  FLA_Part_2x2( A, &ATL, &ATR,
24  &ABL, &ABR, 0, 0, FLA_TL );
25 
26  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
27 
28  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
29 
30  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
31  /* ************* */ /* ******************** */
32  &A10, /**/ &A11, &A12,
33  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
34  b, b, FLA_BR );
35 
36  /*------------------------------------------------------------*/
37 
38  // A21 = -tril( A22 ) \ A21;
39  FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
40  FLA_MINUS_ONE, A22, A21,
41  FLA_Cntl_sub_trsm1( cntl ) );
42 
43  // A20 = -A21 * A10 + A20;
44  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
45  FLA_MINUS_ONE, A21, A10, FLA_ONE, A20,
46  FLA_Cntl_sub_gemm( cntl ) );
47 
48  // A10 = A10 * A00;
49  FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
50  FLA_ONE, A00, A10,
51  FLA_Cntl_sub_trmm( cntl ) );
52 
53  // A11 = inv( A11 );
54  FLA_Trinv_internal( FLA_LOWER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
55  FLA_Cntl_sub_trinv( cntl ) );
56 
57  /*------------------------------------------------------------*/
58 
59  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
60  A10, A11, /**/ A12,
61  /* ************** */ /* ****************** */
62  &ABL, /**/ &ABR, A20, A21, /**/ A22,
63  FLA_TL );
64 
65  }
66 
67  return FLA_SUCCESS;
68 }
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_Trinv_internal(FLA_Uplo uplo, FLA_Diag diag, FLA_Obj A, fla_trinv_t *cntl)
Definition: FLA_Trinv_internal.c:16
FLA_Error FLA_Trmm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl)
Definition: FLA_Trmm_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
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_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trinv_internal(), FLA_Trmm_internal(), and FLA_Trsm_internal().

Referenced by FLA_Trinv_ln().