libflame  revision_anchor
Functions
FLA_Hemm_ru.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Hemm_ru_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_blk_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
 
FLA_Error FLA_Hemm_ru_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_ru_unb_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Hemm_ru_blk_var1()

FLA_Error FLA_Hemm_ru_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
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_1x2( B, &BL, &BR, 0, FLA_LEFT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
36 
37  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
42  /* ************* */ /* ******************** */
43  &A10, /**/ &A11, &A12,
44  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
45  b, b, FLA_BR );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
48  b, FLA_RIGHT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
51  b, FLA_RIGHT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C0 = C0 + B1 * A01' */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
57  alpha, B1, A01, FLA_ONE, C0,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B1 * A11 */
61  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
62  alpha, A11, B1, FLA_ONE, C1,
63  FLA_Cntl_sub_hemm( cntl ) );
64 
65  /* C1 = C1 + B0 * A01 */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
67  alpha, B0, A01, FLA_ONE, C1,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
73  A10, A11, /**/ A12,
74  /* ************** */ /* ****************** */
75  &ABL, /**/ &ABR, A20, A21, /**/ A22,
76  FLA_TL );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
79  FLA_LEFT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
82  FLA_LEFT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var10()

FLA_Error FLA_Hemm_ru_blk_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj BT, B0,
18  BB, B1,
19  B2;
20 
21  FLA_Obj CT, C0,
22  CB, C1,
23  C2;
24 
25  dim_t b;
26 
27  FLA_Part_2x1( B, &BT,
28  &BB, 0, FLA_BOTTOM );
29 
30  FLA_Part_2x1( C, &CT,
31  &CB, 0, FLA_BOTTOM );
32 
33  while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){
34 
35  b = FLA_Determine_blocksize( BT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
36 
37  FLA_Repart_2x1_to_3x1( BT, &B0,
38  &B1,
39  /* ** */ /* ** */
40  BB, &B2, b, FLA_TOP );
41 
42  FLA_Repart_2x1_to_3x1( CT, &C0,
43  &C1,
44  /* ** */ /* ** */
45  CB, &C2, b, FLA_TOP );
46 
47  /*------------------------------------------------------------*/
48 
49  /* C1 = C1 + B1 * A */
50  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
51  alpha, A, B1, beta, C1,
52  FLA_Cntl_sub_hemm( cntl ) );
53 
54  /*------------------------------------------------------------*/
55 
56  FLA_Cont_with_3x1_to_2x1( &BT, B0,
57  /* ** */ /* ** */
58  B1,
59  &BB, B2, FLA_BOTTOM );
60 
61  FLA_Cont_with_3x1_to_2x1( &CT, C0,
62  /* ** */ /* ** */
63  C1,
64  &CB, C2, FLA_BOTTOM );
65 
66  }
67 
68  return FLA_SUCCESS;
69 }
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
unsigned long dim_t
Definition: FLA_type_defs.h:71
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_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
Definition: FLA_type_defs.h:158
FLA_Error FLA_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var2()

FLA_Error FLA_Hemm_ru_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
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_1x2( B, &BL, &BR, 0, FLA_LEFT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
36 
37  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
42  /* ************* */ /* ******************** */
43  &A10, /**/ &A11, &A12,
44  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
45  b, b, FLA_BR );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
48  b, FLA_RIGHT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
51  b, FLA_RIGHT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C1 = C1 + B0 * A01 */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, B0, A01, FLA_ONE, C1,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B1 * A11 */
61  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
62  alpha, A11, B1, FLA_ONE, C1,
63  FLA_Cntl_sub_hemm( cntl ) );
64 
65  /* C1 = C1 + B2 * A12' */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
67  alpha, B2, A12, FLA_ONE, C1,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
73  A10, A11, /**/ A12,
74  /* ************** */ /* ****************** */
75  &ABL, /**/ &ABR, A20, A21, /**/ A22,
76  FLA_TL );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
79  FLA_LEFT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
82  FLA_LEFT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var3()

FLA_Error FLA_Hemm_ru_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
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_1x2( B, &BL, &BR, 0, FLA_LEFT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
36 
37  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
42  /* ************* */ /* ******************** */
43  &A10, /**/ &A11, &A12,
44  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
45  b, b, FLA_BR );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
48  b, FLA_RIGHT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
51  b, FLA_RIGHT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C1 = C1 + B1 * A11 */
56  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
57  alpha, A11, B1, FLA_ONE, C1,
58  FLA_Cntl_sub_hemm( cntl ) );
59 
60  /* C1 = C1 + B2 * A12' */
61  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
62  alpha, B2, A12, FLA_ONE, C1,
63  FLA_Cntl_sub_gemm1( cntl ) );
64 
65  /* C2 = C2 + B1 * A12 */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
67  alpha, B1, A12, FLA_ONE, C2,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
73  A10, A11, /**/ A12,
74  /* ************** */ /* ****************** */
75  &ABL, /**/ &ABR, A20, A21, /**/ A22,
76  FLA_TL );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
79  FLA_LEFT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
82  FLA_LEFT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var4()

FLA_Error FLA_Hemm_ru_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

14 {
15  FLA_Obj ATL, ATR, A00, A01, A02,
16  ABL, ABR, A10, A11, A12,
17  A20, A21, A22;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CL, CR, C0, C1, C2;
22 
23  dim_t b;
24 
25  FLA_Scal_internal( beta, C,
26  FLA_Cntl_sub_scal( cntl ) );
27 
28  FLA_Part_2x2( A, &ATL, &ATR,
29  &ABL, &ABR, 0, 0, FLA_TL );
30 
31  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
32 
33  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
34 
35  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
36 
37  b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
38 
39  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
40  /* ************* */ /* ******************** */
41  &A10, /**/ &A11, &A12,
42  ABL, /**/ ABR, &A20, /**/ &A21, &A22,
43  b, b, FLA_BR );
44 
45  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
46  b, FLA_RIGHT );
47 
48  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
49  b, FLA_RIGHT );
50 
51  /*------------------------------------------------------------*/
52 
53  /* C0 = C0 + B1 * A01' */
54  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
55  alpha, B1, A01, FLA_ONE, C0,
56  FLA_Cntl_sub_gemm1( cntl ) );
57 
58  /* C1 = C1 + B1 * A11 */
59  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
60  alpha, A11, B1, FLA_ONE, C1,
61  FLA_Cntl_sub_hemm( cntl ) );
62 
63  /* C2 = C2 + B1 * A12 */
64  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
65  alpha, B1, A12, FLA_ONE, C2,
66  FLA_Cntl_sub_gemm2( cntl ) );
67 
68  /*------------------------------------------------------------*/
69 
70  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
71  A10, A11, /**/ A12,
72  /* ************** */ /* ****************** */
73  &ABL, /**/ &ABR, A20, A21, /**/ A22,
74  FLA_TL );
75 
76  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
77  FLA_LEFT );
78 
79  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
80  FLA_LEFT );
81 
82  }
83 
84  return FLA_SUCCESS;
85 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var5()

FLA_Error FLA_Hemm_ru_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_BR );
32 
33  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
36 
37  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
42  &A10, &A11, /**/ &A12,
43  /* ************* */ /* ******************** */
44  ABL, /**/ ABR, &A20, &A21, /**/ &A22,
45  b, b, FLA_TL );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
48  b, FLA_LEFT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
51  b, FLA_LEFT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C1 = C1 + B1 * A11 */
56  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
57  alpha, A11, B1, FLA_ONE, C1,
58  FLA_Cntl_sub_hemm( cntl ) );
59 
60  /* C1 = C1 + B2 * A12' */
61  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
62  alpha, B2, A12, FLA_ONE, C1,
63  FLA_Cntl_sub_gemm1( cntl ) );
64 
65  /* C2 = C2 + B1 * A12 */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
67  alpha, B1, A12, FLA_ONE, C2,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
73  /* ************** */ /* ****************** */
74  A10, /**/ A11, A12,
75  &ABL, /**/ &ABR, A20, /**/ A21, A22,
76  FLA_BR );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
79  FLA_RIGHT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
82  FLA_RIGHT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var6()

FLA_Error FLA_Hemm_ru_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_BR );
32 
33  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
36 
37  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
42  &A10, &A11, /**/ &A12,
43  /* ************* */ /* ******************** */
44  ABL, /**/ ABR, &A20, &A21, /**/ &A22,
45  b, b, FLA_TL );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
48  b, FLA_LEFT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
51  b, FLA_LEFT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C1 = C1 + B0 * A01 */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, B0, A01, FLA_ONE, C1,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B1 * A11 */
61  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
62  alpha, A11, B1, FLA_ONE, C1,
63  FLA_Cntl_sub_hemm( cntl ) );
64 
65  /* C1 = C1 + B2 * A12' */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
67  alpha, B2, A12, FLA_ONE, C1,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
73  /* ************** */ /* ****************** */
74  A10, /**/ A11, A12,
75  &ABL, /**/ &ABR, A20, /**/ A21, A22,
76  FLA_BR );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
79  FLA_RIGHT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
82  FLA_RIGHT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var7()

FLA_Error FLA_Hemm_ru_blk_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_BR );
32 
33  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
36 
37  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
42  &A10, &A11, /**/ &A12,
43  /* ************* */ /* ******************** */
44  ABL, /**/ ABR, &A20, &A21, /**/ &A22,
45  b, b, FLA_TL );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
48  b, FLA_LEFT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
51  b, FLA_LEFT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C0 = C0 + B1 * A01' */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
57  alpha, B1, A01, FLA_ONE, C0,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B0 * A01 */
61  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B0, A01, FLA_ONE, C1,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C1 = C1 + B1 * A11 */
66  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
67  alpha, A11, B1, FLA_ONE, C1,
68  FLA_Cntl_sub_hemm( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
73  /* ************** */ /* ****************** */
74  A10, /**/ A11, A12,
75  &ABL, /**/ &ABR, A20, /**/ A21, A22,
76  FLA_BR );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
79  FLA_RIGHT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
82  FLA_RIGHT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var8()

FLA_Error FLA_Hemm_ru_blk_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

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

Referenced by FLA_Hemm_ru().

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  FLA_Obj CL, CR, C0, C1, C2;
24 
25  dim_t b;
26 
27  FLA_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( cntl ) );
29 
30  FLA_Part_2x2( A, &ATL, &ATR,
31  &ABL, &ABR, 0, 0, FLA_BR );
32 
33  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
34 
35  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
36 
37  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
38 
39  b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
42  &A10, &A11, /**/ &A12,
43  /* ************* */ /* ******************** */
44  ABL, /**/ ABR, &A20, &A21, /**/ &A22,
45  b, b, FLA_TL );
46 
47  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
48  b, FLA_LEFT );
49 
50  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
51  b, FLA_LEFT );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C0 = C0 + B1 * A01' */
56  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE,
57  alpha, B1, A01, FLA_ONE, C0,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C1 = C1 + B1 * A11 */
61  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
62  alpha, A11, B1, FLA_ONE, C1,
63  FLA_Cntl_sub_hemm( cntl ) );
64 
65  /* C2 = C2 + B1 * A12 */
66  FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
67  alpha, B1, A12, FLA_ONE, C2,
68  FLA_Cntl_sub_gemm2( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
73  /* ************** */ /* ****************** */
74  A10, /**/ A11, A12,
75  &ABL, /**/ &ABR, A20, /**/ A21, A22,
76  FLA_BR );
77 
78  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
79  FLA_RIGHT );
80 
81  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
82  FLA_RIGHT );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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
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_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_blk_var9()

FLA_Error FLA_Hemm_ru_blk_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Hemm_ru().

14 {
15  FLA_Obj BT, B0,
16  BB, B1,
17  B2;
18 
19  FLA_Obj CT, C0,
20  CB, C1,
21  C2;
22 
23  dim_t b;
24 
25  FLA_Part_2x1( B, &BT,
26  &BB, 0, FLA_TOP );
27 
28  FLA_Part_2x1( C, &CT,
29  &CB, 0, FLA_TOP );
30 
31  while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){
32 
33  b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
34 
35  FLA_Repart_2x1_to_3x1( BT, &B0,
36  /* ** */ /* ** */
37  &B1,
38  BB, &B2, b, FLA_BOTTOM );
39 
40  FLA_Repart_2x1_to_3x1( CT, &C0,
41  /* ** */ /* ** */
42  &C1,
43  CB, &C2, b, FLA_BOTTOM );
44 
45  /*------------------------------------------------------------*/
46 
47  /* C1 = C1 + B1 * A */
48  FLA_Hemm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
49  alpha, A, B1, beta, C1,
50  FLA_Cntl_sub_hemm( cntl ) );
51 
52  /*------------------------------------------------------------*/
53 
54  FLA_Cont_with_3x1_to_2x1( &BT, B0,
55  B1,
56  /* ** */ /* ** */
57  &BB, B2, FLA_TOP );
58 
59  FLA_Cont_with_3x1_to_2x1( &CT, C0,
60  C1,
61  /* ** */ /* ** */
62  &CB, C2, FLA_TOP );
63 
64  }
65 
66  return FLA_SUCCESS;
67 }
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
unsigned long dim_t
Definition: FLA_type_defs.h:71
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_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
Definition: FLA_type_defs.h:158
FLA_Error FLA_Hemm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_hemm_t *cntl)
Definition: FLA_Hemm_internal.c:16
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_Hemm_ru_unb_var1()

FLA_Error FLA_Hemm_ru_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_TL );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
33 
34  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
37  /* ************* */ /* ************************** */
38  &a10t, /**/ &alpha11, &a12t,
39  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
40  1, 1, FLA_BR );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
43  1, FLA_RIGHT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
46  1, FLA_RIGHT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C0 = C0 + b1t * a01' */
51  FLA_Gerc_external( FLA_NO_CONJUGATE, FLA_CONJUGATE, alpha, b1t, a01, C0 );
52 
53  /* c1t = c1t + B0 * a01 */
54  FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B0, a01, FLA_ONE, c1t );
55 
56  /* c1t = c1t + b1t * alpha11 */
57  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
62  a10t, alpha11, /**/ a12t,
63  /* ************** */ /* ************************ */
64  &ABL, /**/ &ABR, A20, a21, /**/ A22,
65  FLA_TL );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
68  FLA_LEFT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
71  FLA_LEFT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv_external.c:13
FLA_Error FLA_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Gerc_external(FLA_Conj conjx, FLA_Conj conjy, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Gerc_external.c:13
Definition: FLA_type_defs.h:158
FLA_Error FLA_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var10()

FLA_Error FLA_Hemm_ru_unb_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Hemvc_external(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj BT, B0,
18  BB, b1t,
19  B2;
20 
21  FLA_Obj CT, C0,
22  CB, c1t,
23  C2;
24 
25  FLA_Part_2x1( B, &BT,
26  &BB, 0, FLA_BOTTOM );
27 
28  FLA_Part_2x1( C, &CT,
29  &CB, 0, FLA_BOTTOM );
30 
31  while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){
32 
33  FLA_Repart_2x1_to_3x1( BT, &B0,
34  &b1t,
35  /* ** */ /* ** */
36  BB, &B2, 1, FLA_TOP );
37 
38  FLA_Repart_2x1_to_3x1( CT, &C0,
39  &c1t,
40  /* ** */ /* ** */
41  CB, &C2, 1, FLA_TOP );
42 
43  /*------------------------------------------------------------*/
44 
45  /* c1t = c1t + b1t * A */
46  /* c1t' = c1t' + A' * b1t' */
47  FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE, alpha, A, b1t, beta, c1t );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_3x1_to_2x1( &BT, B0,
52  /* ** */ /* ** */
53  b1t,
54  &BB, B2, FLA_BOTTOM );
55 
56  FLA_Cont_with_3x1_to_2x1( &CT, C0,
57  /* ** */ /* ** */
58  c1t,
59  &CB, C2, FLA_BOTTOM );
60 
61  }
62 
63  return FLA_SUCCESS;
64 }
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_Hemvc_external(FLA_Uplo uplo, FLA_Conj conja, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Hemvc_external.c:13
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
Definition: FLA_type_defs.h:158
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_Hemm_ru_unb_var2()

FLA_Error FLA_Hemm_ru_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_TL );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
33 
34  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
37  /* ************* */ /* ************************** */
38  &a10t, /**/ &alpha11, &a12t,
39  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
40  1, 1, FLA_BR );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
43  1, FLA_RIGHT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
46  1, FLA_RIGHT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* c1t = c1t + B0 * a01 */
51  FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B0, a01, FLA_ONE, c1t );
52 
53  /* c1t = c1t + b1t * alpha11 */
54  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
55 
56  /* c1t = c1t + B2 * a12t' */
57  FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a12t, FLA_ONE, c1t );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
62  a10t, alpha11, /**/ a12t,
63  /* ************** */ /* ************************ */
64  &ABL, /**/ &ABR, A20, a21, /**/ A22,
65  FLA_TL );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
68  FLA_LEFT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
71  FLA_LEFT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv_external.c:13
FLA_Error FLA_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Gemvc_external(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemvc_external.c:13
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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var3()

FLA_Error FLA_Hemm_ru_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_TL );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
33 
34  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
37  /* ************* */ /* ************************** */
38  &a10t, /**/ &alpha11, &a12t,
39  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
40  1, 1, FLA_BR );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
43  1, FLA_RIGHT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
46  1, FLA_RIGHT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* c1t = c1t + b1t * alpha11 */
51  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
52 
53  /* c1t = c1t + B2 * a12t' */
54  FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a12t, FLA_ONE, c1t );
55 
56  /* C2 = C2 + b1t * a12t */
57  FLA_Ger_external( alpha, b1t, a12t, C2 );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
62  a10t, alpha11, /**/ a12t,
63  /* ************** */ /* ************************ */
64  &ABL, /**/ &ABR, A20, a21, /**/ A22,
65  FLA_TL );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
68  FLA_LEFT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
71  FLA_LEFT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Gemvc_external(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemvc_external.c:13
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_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Ger_external.c:13
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var4()

FLA_Error FLA_Hemm_ru_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_TL );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
33 
34  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
37  /* ************* */ /* ************************** */
38  &a10t, /**/ &alpha11, &a12t,
39  ABL, /**/ ABR, &A20, /**/ &a21, &A22,
40  1, 1, FLA_BR );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
43  1, FLA_RIGHT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
46  1, FLA_RIGHT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C0 = C0 + b1t * a01' */
51  FLA_Gerc_external( FLA_NO_CONJUGATE, FLA_CONJUGATE, alpha, b1t, a01, C0 );
52 
53  /* c1t = c1t + b1t * alpha11 */
54  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
55 
56  /* C2 = C2 + b1t * a12t */
57  FLA_Ger_external( alpha, b1t, a12t, C2 );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
62  a10t, alpha11, /**/ a12t,
63  /* ************** */ /* ************************ */
64  &ABL, /**/ &ABR, A20, a21, /**/ A22,
65  FLA_TL );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
68  FLA_LEFT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
71  FLA_LEFT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Gerc_external(FLA_Conj conjx, FLA_Conj conjy, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Gerc_external.c:13
Definition: FLA_type_defs.h:158
FLA_Error FLA_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Ger_external.c:13
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var5()

FLA_Error FLA_Hemm_ru_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_BR );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
33 
34  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
37  &a10t, &alpha11, /**/ &a12t,
38  /* ************* */ /* ************************** */
39  ABL, /**/ ABR, &A20, &a21, /**/ &A22,
40  1, 1, FLA_TL );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
43  1, FLA_LEFT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
46  1, FLA_LEFT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* c1t = c1t + b1t * alpha11 */
51  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
52 
53  /* c1t = c1t + B2 * a12t' */
54  FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a12t, FLA_ONE, c1t );
55 
56  /* C2 = C2 + b1t * a12t */
57  FLA_Ger_external( alpha, b1t, a12t, C2 );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
62  /* ************** */ /* ************************ */
63  a10t, /**/ alpha11, a12t,
64  &ABL, /**/ &ABR, A20, /**/ a21, A22,
65  FLA_BR );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
68  FLA_RIGHT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
71  FLA_RIGHT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Gemvc_external(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemvc_external.c:13
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_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Ger_external.c:13
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var6()

FLA_Error FLA_Hemm_ru_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_BR );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
33 
34  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
37  &a10t, &alpha11, /**/ &a12t,
38  /* ************* */ /* ************************** */
39  ABL, /**/ ABR, &A20, &a21, /**/ &A22,
40  1, 1, FLA_TL );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
43  1, FLA_LEFT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
46  1, FLA_LEFT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* c1t = c1t + B0 * a01 */
51  FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B0, a01, FLA_ONE, c1t );
52 
53  /* c1t = c1t + b1t * alpha11 */
54  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
55 
56  /* c1t = c1t + B2 * a12t' */
57  FLA_Gemvc_external( FLA_NO_TRANSPOSE, FLA_CONJUGATE, alpha, B2, a12t, FLA_ONE, c1t );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
62  /* ************** */ /* ************************ */
63  a10t, /**/ alpha11, a12t,
64  &ABL, /**/ &ABR, A20, /**/ a21, A22,
65  FLA_BR );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
68  FLA_RIGHT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
71  FLA_RIGHT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv_external.c:13
FLA_Error FLA_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Gemvc_external(FLA_Trans transa, FLA_Conj conjx, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemvc_external.c:13
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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var7()

FLA_Error FLA_Hemm_ru_unb_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_BR );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
33 
34  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
37  &a10t, &alpha11, /**/ &a12t,
38  /* ************* */ /* ************************** */
39  ABL, /**/ ABR, &A20, &a21, /**/ &A22,
40  1, 1, FLA_TL );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
43  1, FLA_LEFT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
46  1, FLA_LEFT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C0 = C0 + b1t * a01' */
51  FLA_Gerc_external( FLA_NO_CONJUGATE, FLA_CONJUGATE, alpha, b1t, a01, C0 );
52 
53  /* c1t = c1t + B0 * a01 */
54  FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B0, a01, FLA_ONE, c1t );
55 
56  /* c1t = c1t + b1t * alpha11 */
57  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
62  /* ************** */ /* ************************ */
63  a10t, /**/ alpha11, a12t,
64  &ABL, /**/ &ABR, A20, /**/ a21, A22,
65  FLA_BR );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
68  FLA_RIGHT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
71  FLA_RIGHT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Gemv_external(FLA_Trans transa, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Gemv_external.c:13
FLA_Error FLA_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Gerc_external(FLA_Conj conjx, FLA_Conj conjy, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Gerc_external.c:13
Definition: FLA_type_defs.h:158
FLA_Error FLA_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var8()

FLA_Error FLA_Hemm_ru_unb_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Axpys_external(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scal_external().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj ATL, ATR, A00, a01, A02,
18  ABL, ABR, a10t, alpha11, a12t,
19  A20, a21, A22;
20 
21  FLA_Obj BL, BR, B0, b1t, B2;
22 
23  FLA_Obj CL, CR, C0, c1t, C2;
24 
25  FLA_Scal_external( beta, C );
26 
27  FLA_Part_2x2( A, &ATL, &ATR,
28  &ABL, &ABR, 0, 0, FLA_BR );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
33 
34  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
35 
36  FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
37  &a10t, &alpha11, /**/ &a12t,
38  /* ************* */ /* ************************** */
39  ABL, /**/ ABR, &A20, &a21, /**/ &A22,
40  1, 1, FLA_TL );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
43  1, FLA_LEFT );
44 
45  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
46  1, FLA_LEFT );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C0 = C0 + b1t * a01' */
51  FLA_Gerc_external( FLA_NO_CONJUGATE, FLA_CONJUGATE, alpha, b1t, a01, C0 );
52 
53  /* c1t = c1t + b1t * alpha11 */
54  FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );
55 
56  /* C2 = C2 + b1t * a12t */
57  FLA_Ger_external( alpha, b1t, a12t, C2 );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
62  /* ************** */ /* ************************ */
63  a10t, /**/ alpha11, a12t,
64  &ABL, /**/ &ABR, A20, /**/ a21, A22,
65  FLA_BR );
66 
67  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
68  FLA_RIGHT );
69 
70  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
71  FLA_RIGHT );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }
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_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_external.c:13
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_Gerc_external(FLA_Conj conjx, FLA_Conj conjy, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Gerc_external.c:13
Definition: FLA_type_defs.h:158
FLA_Error FLA_Axpys_external(FLA_Obj alpha0, FLA_Obj alpha1, FLA_Obj A, FLA_Obj beta, FLA_Obj B)
Definition: FLA_Axpys_external.c:13
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_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Ger_external.c:13
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

◆ FLA_Hemm_ru_unb_var9()

FLA_Error FLA_Hemm_ru_unb_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Hemvc_external(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Hemm_ru().

16 {
17  FLA_Obj BT, B0,
18  BB, b1t,
19  B2;
20 
21  FLA_Obj CT, C0,
22  CB, c1t,
23  C2;
24 
25  FLA_Part_2x1( B, &BT,
26  &BB, 0, FLA_TOP );
27 
28  FLA_Part_2x1( C, &CT,
29  &CB, 0, FLA_TOP );
30 
31  while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){
32 
33  FLA_Repart_2x1_to_3x1( BT, &B0,
34  /* ** */ /* ** */
35  &b1t,
36  BB, &B2, 1, FLA_BOTTOM );
37 
38  FLA_Repart_2x1_to_3x1( CT, &C0,
39  /* ** */ /* ** */
40  &c1t,
41  CB, &C2, 1, FLA_BOTTOM );
42 
43  /*------------------------------------------------------------*/
44 
45  /* c1t = c1t + b1t * A */
46  /* c1t' = c1t' + A' * b1t' */
47  FLA_Hemvc_external( FLA_UPPER_TRIANGULAR, FLA_CONJUGATE, alpha, A, b1t, beta, c1t );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_3x1_to_2x1( &BT, B0,
52  b1t,
53  /* ** */ /* ** */
54  &BB, B2, FLA_TOP );
55 
56  FLA_Cont_with_3x1_to_2x1( &CT, C0,
57  c1t,
58  /* ** */ /* ** */
59  &CB, C2, FLA_TOP );
60 
61  }
62 
63  return FLA_SUCCESS;
64 }
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_Hemvc_external(FLA_Uplo uplo, FLA_Conj conja, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition: FLA_Hemvc_external.c:13
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
Definition: FLA_type_defs.h:158
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