libflame revision_anchor
Functions
FLA_Hemm_lu.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Hemm_lu_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_lu_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_lu_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_lu_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_lu_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_lu_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_lu_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_lu_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_lu_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_lu_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_lu_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Hemm_lu_unb_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Hemm_lu_blk_var1()

FLA_Error FLA_Hemm_lu_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_TL );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_TOP );
42
43 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
48 /* ************* */ /* ******************** */
49 &A10, /**/ &A11, &A12,
50 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
51 b, b, FLA_BR );
52
54 /* ** */ /* ** */
55 &B1,
56 BB, &B2, b, FLA_BOTTOM );
57
59 /* ** */ /* ** */
60 &C1,
61 CB, &C2, b, FLA_BOTTOM );
62
63 /*------------------------------------------------------------*/
64
65 /* C0 = C0 + A01 * B1 */
67 alpha, A01, B1, FLA_ONE, C0,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C1 = C1 + A11 * B1 */
72 alpha, A11, B1, FLA_ONE, C1,
73 FLA_Cntl_sub_hemm( cntl ) );
74
75 /* C1 = C1 + A01' * B0 */
77 alpha, A01, B0, FLA_ONE, C1,
78 FLA_Cntl_sub_gemm2( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
83 A10, A11, /**/ A12,
84 /* ************** */ /* ****************** */
85 &ABL, /**/ &ABR, A20, A21, /**/ A22,
86 FLA_TL );
87
89 B1,
90 /* ** */ /* ** */
91 &BB, B2, FLA_TOP );
92
94 C1,
95 /* ** */ /* ** */
96 &CB, C2, FLA_TOP );
97
98 }
99
100 return FLA_SUCCESS;
101}
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_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_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition FLA_Scal_internal.c:16
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
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
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_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_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
int i
Definition bl1_axmyv2.c:145
Definition FLA_type_defs.h:159

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var10()

FLA_Error FLA_Hemm_lu_blk_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj BL, BR, B0, B1, B2;
18
19 FLA_Obj CL, CR, C0, C1, C2;
20
21 dim_t b;
22
23 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
24
25 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
26
27 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
28
30
31 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
32 b, FLA_LEFT );
33
34 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
35 b, FLA_LEFT );
36
37 /*------------------------------------------------------------*/
38
39 /* C1 = C1 + A * B1 */
41 alpha, A, B1, beta, C1,
42 FLA_Cntl_sub_hemm( cntl ) );
43
44 /*------------------------------------------------------------*/
45
46 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
47 FLA_RIGHT );
48
49 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
50 FLA_RIGHT );
51
52 }
53
54 return FLA_SUCCESS;
55}
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
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
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_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

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var2()

FLA_Error FLA_Hemm_lu_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_TL );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_TOP );
42
43 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
48 /* ************* */ /* ******************** */
49 &A10, /**/ &A11, &A12,
50 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
51 b, b, FLA_BR );
52
54 /* ** */ /* ** */
55 &B1,
56 BB, &B2, b, FLA_BOTTOM );
57
59 /* ** */ /* ** */
60 &C1,
61 CB, &C2, b, FLA_BOTTOM );
62
63 /*------------------------------------------------------------*/
64
65 /* C1 = C1 + A01' * B0 */
67 alpha, A01, B0, FLA_ONE, C1,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C1 = C1 + A12 * B2 */
72 alpha, A12, B2, FLA_ONE, C1,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C1 = C1 + A11 * B1 */
77 alpha, A11, B1, FLA_ONE, C1,
78 FLA_Cntl_sub_hemm( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
83 A10, A11, /**/ A12,
84 /* ************** */ /* ****************** */
85 &ABL, /**/ &ABR, A20, A21, /**/ A22,
86 FLA_TL );
87
89 B1,
90 /* ** */ /* ** */
91 &BB, B2, FLA_TOP );
92
94 C1,
95 /* ** */ /* ** */
96 &CB, C2, FLA_TOP );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var3()

FLA_Error FLA_Hemm_lu_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_TL );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_TOP );
42
43 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
48 /* ************* */ /* ******************** */
49 &A10, /**/ &A11, &A12,
50 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
51 b, b, FLA_BR );
52
54 /* ** */ /* ** */
55 &B1,
56 BB, &B2, b, FLA_BOTTOM );
57
59 /* ** */ /* ** */
60 &C1,
61 CB, &C2, b, FLA_BOTTOM );
62
63 /*------------------------------------------------------------*/
64
65 /* C1 = C1 + A11 * B1 */
67 alpha, A11, B1, FLA_ONE, C1,
68 FLA_Cntl_sub_hemm( cntl ) );
69
70 /* C1 = C1 + A12 * B2 */
72 alpha, A12, B2, FLA_ONE, C1,
73 FLA_Cntl_sub_gemm1( cntl ) );
74
75 /* C2 = C2 + A12' * B1 */
77 alpha, A12, B1, FLA_ONE, C2,
78 FLA_Cntl_sub_gemm2( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
83 A10, A11, /**/ A12,
84 /* ************** */ /* ****************** */
85 &ABL, /**/ &ABR, A20, A21, /**/ A22,
86 FLA_TL );
87
89 B1,
90 /* ** */ /* ** */
91 &BB, B2, FLA_TOP );
92
94 C1,
95 /* ** */ /* ** */
96 &CB, C2, FLA_TOP );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var4()

FLA_Error FLA_Hemm_lu_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
14{
15 FLA_Obj ATL, ATR, A00, A01, A02,
16 ABL, ABR, A10, A11, A12,
17 A20, A21, A22;
18
19 FLA_Obj BT, B0,
20 BB, B1,
21 B2;
22
23 FLA_Obj CT, C0,
24 CB, C1,
25 C2;
26
27 dim_t b;
28
30 FLA_Cntl_sub_scal( cntl ) );
31
32 FLA_Part_2x2( A, &ATL, &ATR,
33 &ABL, &ABR, 0, 0, FLA_TL );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_TOP );
37
38 FLA_Part_2x1( C, &CT,
39 &CB, 0, FLA_TOP );
40
41 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
42
44
45 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
46 /* ************* */ /* ******************** */
47 &A10, /**/ &A11, &A12,
48 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
49 b, b, FLA_BR );
50
52 /* ** */ /* ** */
53 &B1,
54 BB, &B2, b, FLA_BOTTOM );
55
57 /* ** */ /* ** */
58 &C1,
59 CB, &C2, b, FLA_BOTTOM );
60
61 /*------------------------------------------------------------*/
62
63 /* C0 = C0 + A01 * B1 */
65 alpha, A01, B1, FLA_ONE, C0,
66 FLA_Cntl_sub_gemm1( cntl ) );
67
68 /* C1 = C1 + A11 * B1 */
70 alpha, A11, B1, FLA_ONE, C1,
71 FLA_Cntl_sub_hemm( cntl ) );
72
73 /* C2 = C2 + A12' * B1 */
75 alpha, A12, B1, FLA_ONE, C2,
76 FLA_Cntl_sub_gemm2( cntl ) );
77
78 /*------------------------------------------------------------*/
79
80 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
81 A10, A11, /**/ A12,
82 /* ************** */ /* ****************** */
83 &ABL, /**/ &ABR, A20, A21, /**/ A22,
84 FLA_TL );
85
87 B1,
88 /* ** */ /* ** */
89 &BB, B2, FLA_TOP );
90
92 C1,
93 /* ** */ /* ** */
94 &CB, C2, FLA_TOP );
95
96 }
97
98 return FLA_SUCCESS;
99}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var5()

FLA_Error FLA_Hemm_lu_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_BR );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_BOTTOM );
42
43 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
48 &A10, &A11, /**/ &A12,
49 /* ************* */ /* ******************** */
50 ABL, /**/ ABR, &A20, &A21, /**/ &A22,
51 b, b, FLA_TL );
52
54 &B1,
55 /* ** */ /* ** */
56 BB, &B2, b, FLA_TOP );
57
59 &C1,
60 /* ** */ /* ** */
61 CB, &C2, b, FLA_TOP );
62
63 /*------------------------------------------------------------*/
64
65 /* C1 = C1 + A11 * B1 */
67 alpha, A11, B1, FLA_ONE, C1,
68 FLA_Cntl_sub_hemm( cntl ) );
69
70 /* C1 = C1 + A12 * B2 */
72 alpha, A12, B2, FLA_ONE, C1,
73 FLA_Cntl_sub_gemm1( cntl ) );
74
75 /* C2 = C2 + A12' * B1 */
77 alpha, A12, B1, FLA_ONE, C2,
78 FLA_Cntl_sub_gemm2( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
83 /* ************** */ /* ****************** */
84 A10, /**/ A11, A12,
85 &ABL, /**/ &ABR, A20, /**/ A21, A22,
86 FLA_BR );
87
89 /* ** */ /* ** */
90 B1,
91 &BB, B2, FLA_BOTTOM );
92
94 /* ** */ /* ** */
95 C1,
96 &CB, C2, FLA_BOTTOM );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var6()

FLA_Error FLA_Hemm_lu_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_BR );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_BOTTOM );
42
43 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
48 &A10, &A11, /**/ &A12,
49 /* ************* */ /* ******************** */
50 ABL, /**/ ABR, &A20, &A21, /**/ &A22,
51 b, b, FLA_TL );
52
54 &B1,
55 /* ** */ /* ** */
56 BB, &B2, b, FLA_TOP );
57
59 &C1,
60 /* ** */ /* ** */
61 CB, &C2, b, FLA_TOP );
62
63 /*------------------------------------------------------------*/
64
65 /* C1 = C1 + A01' * B0 */
67 alpha, A01, B0, FLA_ONE, C1,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C1 = C1 + A11 * B1 */
72 alpha, A11, B1, FLA_ONE, C1,
73 FLA_Cntl_sub_hemm( cntl ) );
74
75 /* C1 = C1 + A12 * B2 */
77 alpha, A12, B2, FLA_ONE, C1,
78 FLA_Cntl_sub_gemm2( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
83 /* ************** */ /* ****************** */
84 A10, /**/ A11, A12,
85 &ABL, /**/ &ABR, A20, /**/ A21, A22,
86 FLA_BR );
87
89 /* ** */ /* ** */
90 B1,
91 &BB, B2, FLA_BOTTOM );
92
94 /* ** */ /* ** */
95 C1,
96 &CB, C2, FLA_BOTTOM );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var7()

FLA_Error FLA_Hemm_lu_blk_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_BR );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_BOTTOM );
42
43 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
48 &A10, &A11, /**/ &A12,
49 /* ************* */ /* ******************** */
50 ABL, /**/ ABR, &A20, &A21, /**/ &A22,
51 b, b, FLA_TL );
52
54 &B1,
55 /* ** */ /* ** */
56 BB, &B2, b, FLA_TOP );
57
59 &C1,
60 /* ** */ /* ** */
61 CB, &C2, b, FLA_TOP );
62
63 /*------------------------------------------------------------*/
64
65 /* C0 = C0 + A01 * B1 */
67 alpha, A01, B1, FLA_ONE, C0,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C1 = C1 + A01' * B0 */
72 alpha, A01, B0, FLA_ONE, C1,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C1 = C1 + A11 * B1 */
77 alpha, A11, B1, FLA_ONE, C1,
78 FLA_Cntl_sub_hemm( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
83 /* ************** */ /* ****************** */
84 A10, /**/ A11, A12,
85 &ABL, /**/ &ABR, A20, /**/ A21, A22,
86 FLA_BR );
87
89 /* ** */ /* ** */
90 B1,
91 &BB, B2, FLA_BOTTOM );
92
94 /* ** */ /* ** */
95 C1,
96 &CB, C2, FLA_BOTTOM );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var8()

FLA_Error FLA_Hemm_lu_blk_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, C1,
27 C2;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scal( cntl ) );
33
34 FLA_Part_2x2( A, &ATL, &ATR,
35 &ABL, &ABR, 0, 0, FLA_BR );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x1( C, &CT,
41 &CB, 0, FLA_BOTTOM );
42
43 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
44
46
47 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02,
48 &A10, &A11, /**/ &A12,
49 /* ************* */ /* ******************** */
50 ABL, /**/ ABR, &A20, &A21, /**/ &A22,
51 b, b, FLA_TL );
52
54 &B1,
55 /* ** */ /* ** */
56 BB, &B2, b, FLA_TOP );
57
59 &C1,
60 /* ** */ /* ** */
61 CB, &C2, b, FLA_TOP );
62
63 /*------------------------------------------------------------*/
64
65 /* C0 = C0 + A01 * B1 */
67 alpha, A01, B1, FLA_ONE, C0,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C1 = C1 + A11 * B1 */
72 alpha, A11, B1, FLA_ONE, C1,
73 FLA_Cntl_sub_hemm( cntl ) );
74
75 /* C2 = C2 + A12' * B1 */
77 alpha, A12, B1, FLA_ONE, C2,
78 FLA_Cntl_sub_gemm2( cntl ) );
79
80 /*------------------------------------------------------------*/
81
82 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02,
83 /* ************** */ /* ****************** */
84 A10, /**/ A11, A12,
85 &ABL, /**/ &ABR, A20, /**/ A21, A22,
86 FLA_BR );
87
89 /* ** */ /* ** */
90 B1,
91 &BB, B2, FLA_BOTTOM );
92
94 /* ** */ /* ** */
95 C1,
96 &CB, C2, FLA_BOTTOM );
97
98 }
99
100 return FLA_SUCCESS;
101}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_internal(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_blk_var9()

FLA_Error FLA_Hemm_lu_blk_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_hemm_t cntl 
)
14{
15 FLA_Obj BL, BR, B0, B1, B2;
16
17 FLA_Obj CL, CR, C0, C1, C2;
18
19 dim_t b;
20
21 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
22
23 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
24
25 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
26
28
29 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
30 b, FLA_RIGHT );
31
32 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
33 b, FLA_RIGHT );
34
35 /*------------------------------------------------------------*/
36
37 /* C1 = C1 + A * B1 */
39 alpha, A, B1, beta, C1,
40 FLA_Cntl_sub_hemm( cntl ) );
41
42 /*------------------------------------------------------------*/
43
44 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
45 FLA_LEFT );
46
47 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
48 FLA_LEFT );
49
50 }
51
52 return FLA_SUCCESS;
53}

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var1()

FLA_Error FLA_Hemm_lu_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_TL );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_TOP );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_TOP );
39
40 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
43 /* ************* */ /* ************************** */
44 &a10t, /**/ &alpha11, &a12t,
45 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
46 1, 1, FLA_BR );
47
49 /* ** */ /* ** */
50 &b1t,
51 BB, &B2, 1, FLA_BOTTOM );
52
54 /* ** */ /* ** */
55 &c1t,
56 CB, &C2, 1, FLA_BOTTOM );
57
58 /*------------------------------------------------------------*/
59
60 /* C0 = C0 + a01 * b1t */
62
63 /* c1t = c1t + a01' * B0 */
64 /* c1t' = c1t' + B0' * a01 */
66
67 /* c1t = c1t + alpha11 * b1t */
69
70 /*------------------------------------------------------------*/
71
72 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
73 a10t, alpha11, /**/ a12t,
74 /* ************** */ /* ************************ */
75 &ABL, /**/ &ABR, A20, a21, /**/ A22,
76 FLA_TL );
77
79 b1t,
80 /* ** */ /* ** */
81 &BB, B2, FLA_TOP );
82
84 c1t,
85 /* ** */ /* ** */
86 &CB, C2, FLA_TOP );
87
88 }
89
90 return FLA_SUCCESS;
91}
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_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition FLA_Scal_external.c:13
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_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition FLA_Ger_external.c:13

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var10()

FLA_Error FLA_Hemm_lu_unb_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj BL, BR, B0, b1t, B2;
18
19 FLA_Obj CL, CR, C0, c1t, C2;
20
21 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
22
23 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
24
25 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
26
27 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
28 1, FLA_LEFT );
29
30 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2,
31 1, FLA_LEFT );
32
33 /*------------------------------------------------------------*/
34
35 /* c1t = c1t + A * b1t */
37
38 /*------------------------------------------------------------*/
39
40 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
41 FLA_RIGHT );
42
43 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2,
44 FLA_RIGHT );
45
46 }
47
48 return FLA_SUCCESS;
49}
FLA_Error FLA_Hemv_external(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y)
Definition FLA_Hemv_external.c:13

References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var2()

FLA_Error FLA_Hemm_lu_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_TL );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_TOP );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_TOP );
39
40 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
43 /* ************* */ /* ************************** */
44 &a10t, /**/ &alpha11, &a12t,
45 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
46 1, 1, FLA_BR );
47
49 /* ** */ /* ** */
50 &b1t,
51 BB, &B2, 1, FLA_BOTTOM );
52
54 /* ** */ /* ** */
55 &c1t,
56 CB, &C2, 1, FLA_BOTTOM );
57
58 /*------------------------------------------------------------*/
59
60 /* c1t = c1t + a01' * B0 */
61 /* c1t' = c1t' + B0' * a01 */
63
64 /* c1t = c1t + a12t * B2 */
65 /* c1t' = c1t' + B2' * a12t' */
67
68 /* c1t = c1t + alpha11 * b1t */
70
71 /*------------------------------------------------------------*/
72
73 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
74 a10t, alpha11, /**/ a12t,
75 /* ************** */ /* ************************ */
76 &ABL, /**/ &ABR, A20, a21, /**/ A22,
77 FLA_TL );
78
80 b1t,
81 /* ** */ /* ** */
82 &BB, B2, FLA_TOP );
83
85 c1t,
86 /* ** */ /* ** */
87 &CB, C2, FLA_TOP );
88
89 }
90
91 return FLA_SUCCESS;
92}
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

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var3()

FLA_Error FLA_Hemm_lu_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_TL );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_TOP );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_TOP );
39
40 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
43 /* ************* */ /* ************************** */
44 &a10t, /**/ &alpha11, &a12t,
45 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
46 1, 1, FLA_BR );
47
49 /* ** */ /* ** */
50 &b1t,
51 BB, &B2, 1, FLA_BOTTOM );
52
54 /* ** */ /* ** */
55 &c1t,
56 CB, &C2, 1, FLA_BOTTOM );
57
58 /*------------------------------------------------------------*/
59
60 /* c1t = c1t + alpha11 * b1t */
62
63 /* c1t = c1t + a12t * B2 */
64 /* c1t' = c1t' + B2' * a12t' */
66
67 /* C2 = C2 + a12t' * b1t */
69
70 /*------------------------------------------------------------*/
71
72 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
73 a10t, alpha11, /**/ a12t,
74 /* ************** */ /* ************************ */
75 &ABL, /**/ &ABR, A20, a21, /**/ A22,
76 FLA_TL );
77
79 b1t,
80 /* ** */ /* ** */
81 &BB, B2, FLA_TOP );
82
84 c1t,
85 /* ** */ /* ** */
86 &CB, C2, FLA_TOP );
87
88 }
89
90 return FLA_SUCCESS;
91}
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

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var4()

FLA_Error FLA_Hemm_lu_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_TL );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_TOP );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_TOP );
39
40 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
43 /* ************* */ /* ************************** */
44 &a10t, /**/ &alpha11, &a12t,
45 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
46 1, 1, FLA_BR );
47
49 /* ** */ /* ** */
50 &b1t,
51 BB, &B2, 1, FLA_BOTTOM );
52
54 /* ** */ /* ** */
55 &c1t,
56 CB, &C2, 1, FLA_BOTTOM );
57
58 /*------------------------------------------------------------*/
59
60 /* C0 = C0 + a01 * b1t */
62
63 /* c1t = c1t + alpha11 * b1t */
65
66 /* C2 = C2 + a12t' * b1t */
68
69 /*------------------------------------------------------------*/
70
71 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
72 a10t, alpha11, /**/ a12t,
73 /* ************** */ /* ************************ */
74 &ABL, /**/ &ABR, A20, a21, /**/ A22,
75 FLA_TL );
76
78 b1t,
79 /* ** */ /* ** */
80 &BB, B2, FLA_TOP );
81
83 c1t,
84 /* ** */ /* ** */
85 &CB, C2, FLA_TOP );
86
87 }
88
89 return FLA_SUCCESS;
90}

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var5()

FLA_Error FLA_Hemm_lu_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_BR );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_BOTTOM );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
43 &a10t, &alpha11, /**/ &a12t,
44 /* ************* */ /* ************************** */
45 ABL, /**/ ABR, &A20, &a21, /**/ &A22,
46 1, 1, FLA_TL );
47
49 &b1t,
50 /* ** */ /* ** */
51 BB, &B2, 1, FLA_TOP );
52
54 &c1t,
55 /* ** */ /* ** */
56 CB, &C2, 1, FLA_TOP );
57
58 /*------------------------------------------------------------*/
59
60 /* c1t = c1t + alpha11 * b1t */
62
63 /* c1t = c1t + a12t * B2 */
64 /* c1t' = c1t' + B2' * a12t' */
66
67 /* C2 = C2 + a12t' * b1t */
69
70 /*------------------------------------------------------------*/
71
72 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
73 /* ************** */ /* ************************ */
74 a10t, /**/ alpha11, a12t,
75 &ABL, /**/ &ABR, A20, /**/ a21, A22,
76 FLA_BR );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
84 /* ** */ /* ** */
85 c1t,
86 &CB, C2, FLA_BOTTOM );
87
88 }
89
90 return FLA_SUCCESS;
91}

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var6()

FLA_Error FLA_Hemm_lu_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_BR );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_BOTTOM );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
43 &a10t, &alpha11, /**/ &a12t,
44 /* ************* */ /* ************************** */
45 ABL, /**/ ABR, &A20, &a21, /**/ &A22,
46 1, 1, FLA_TL );
47
49 &b1t,
50 /* ** */ /* ** */
51 BB, &B2, 1, FLA_TOP );
52
54 &c1t,
55 /* ** */ /* ** */
56 CB, &C2, 1, FLA_TOP );
57
58 /*------------------------------------------------------------*/
59
60 /* c1t = c1t + a01' * B0 */
61 /* c1t' = c1t' + B0' * a01 */
63
64 /* c1t = c1t + alpha11 * b1t */
66
67 /* c1t = c1t + a12t * B2 */
68 /* c1t' = c1t' + B2' * a12t' */
70
71 /*------------------------------------------------------------*/
72
73 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
74 /* ************** */ /* ************************ */
75 a10t, /**/ alpha11, a12t,
76 &ABL, /**/ &ABR, A20, /**/ a21, A22,
77 FLA_BR );
78
80 /* ** */ /* ** */
81 b1t,
82 &BB, B2, FLA_BOTTOM );
83
85 /* ** */ /* ** */
86 c1t,
87 &CB, C2, FLA_BOTTOM );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var7()

FLA_Error FLA_Hemm_lu_unb_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_BR );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_BOTTOM );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
43 &a10t, &alpha11, /**/ &a12t,
44 /* ************* */ /* ************************** */
45 ABL, /**/ ABR, &A20, &a21, /**/ &A22,
46 1, 1, FLA_TL );
47
49 &b1t,
50 /* ** */ /* ** */
51 BB, &B2, 1, FLA_TOP );
52
54 &c1t,
55 /* ** */ /* ** */
56 CB, &C2, 1, FLA_TOP );
57
58 /*------------------------------------------------------------*/
59
60 /* C0 = C0 + a01 * b1t */
62
63 /* c1t = c1t + a01' * B0 */
64 /* c1t' = c1t' + B0' * a01 */
66
67 /* c1t = c1t + alpha11 * b1t */
69
70 /*------------------------------------------------------------*/
71
72 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
73 /* ************** */ /* ************************ */
74 a10t, /**/ alpha11, a12t,
75 &ABL, /**/ &ABR, A20, /**/ a21, A22,
76 FLA_BR );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
84 /* ** */ /* ** */
85 c1t,
86 &CB, C2, FLA_BOTTOM );
87
88 }
89
90 return FLA_SUCCESS;
91}

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var8()

FLA_Error FLA_Hemm_lu_unb_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
19 A20, a21, A22;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CT, C0,
26 CB, c1t,
27 C2;
28
30
31 FLA_Part_2x2( A, &ATL, &ATR,
32 &ABL, &ABR, 0, 0, FLA_BR );
33
34 FLA_Part_2x1( B, &BT,
35 &BB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( C, &CT,
38 &CB, 0, FLA_BOTTOM );
39
40 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){
41
42 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02,
43 &a10t, &alpha11, /**/ &a12t,
44 /* ************* */ /* ************************** */
45 ABL, /**/ ABR, &A20, &a21, /**/ &A22,
46 1, 1, FLA_TL );
47
49 &b1t,
50 /* ** */ /* ** */
51 BB, &B2, 1, FLA_TOP );
52
54 &c1t,
55 /* ** */ /* ** */
56 CB, &C2, 1, FLA_TOP );
57
58 /*------------------------------------------------------------*/
59
60 /* C0 = C0 + a01 * b1t */
62
63 /* c1t = c1t + alpha11 * b1t */
65
66 /* C2 = C2 + a12t' * b1t */
68
69 /*------------------------------------------------------------*/
70
71 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02,
72 /* ************** */ /* ************************ */
73 a10t, /**/ alpha11, a12t,
74 &ABL, /**/ &ABR, A20, /**/ a21, A22,
75 FLA_BR );
76
78 /* ** */ /* ** */
79 b1t,
80 &BB, B2, FLA_BOTTOM );
81
83 /* ** */ /* ** */
84 c1t,
85 &CB, C2, FLA_BOTTOM );
86
87 }
88
89 return FLA_SUCCESS;
90}

References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and i.

Referenced by FLA_Hemm_lu().

◆ FLA_Hemm_lu_unb_var9()

FLA_Error FLA_Hemm_lu_unb_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj BL, BR, B0, b1t, B2;
18
19 FLA_Obj CL, CR, C0, c1t, C2;
20
21 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
22
23 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
24
25 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
26
27 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
28 1, FLA_RIGHT );
29
30 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2,
31 1, FLA_RIGHT );
32
33 /*------------------------------------------------------------*/
34
35 /* c1t = c1t + A * b1t */
37
38 /*------------------------------------------------------------*/
39
40 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
41 FLA_LEFT );
42
43 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2,
44 FLA_LEFT );
45
46 }
47
48 return FLA_SUCCESS;
49}

References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and i.

Referenced by FLA_Hemm_lu().