libflame revision_anchor
Functions
FLA_Her2k_ln.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Her2k_ln_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_blk_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
 
FLA_Error FLA_Her2k_ln_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Her2k_ln_unb_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Her2k_ln_blk_var1()

FLA_Error FLA_Her2k_ln_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_TOP );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_TL );
42
43 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
44
46
48 /* ** */ /* ** */
49 &A1,
50 AB, &A2, b, FLA_BOTTOM );
51
53 /* ** */ /* ** */
54 &B1,
55 BB, &B2, b, FLA_BOTTOM );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
58 /* ************* */ /* ******************** */
59 &C10, /**/ &C11, &C12,
60 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
61 b, b, FLA_BR );
62
63 /*------------------------------------------------------------*/
64
65 /* C10 = C10 + A1 * B0' */
67 alpha, A1, B0, FLA_ONE, C10,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C10 = C10 + B1 * A0' */
72 alpha, B1, A0, FLA_ONE, C10,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 A1,
84 /* ** */ /* ** */
85 &AB, A2, FLA_TOP );
86
88 B1,
89 /* ** */ /* ** */
90 &BB, B2, FLA_TOP );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
93 C10, C11, /**/ C12,
94 /* ************** */ /* ****************** */
95 &CBL, /**/ &CBR, C20, C21, /**/ C22,
96 FLA_TL );
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_Her2k_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)
Definition FLA_Her2k_internal.c:16
FLA_Error FLA_Scalr_internal(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, fla_scalr_t *cntl)
Definition FLA_Scalr_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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var10()

FLA_Error FLA_Her2k_ln_blk_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AL, AR, A0, A1, A2;
18
19 FLA_Obj BL, BR, B0, B1, B2;
20
21 dim_t b;
22
24 FLA_Cntl_sub_scalr( cntl ) );
25
26 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
27
28 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
29
30 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31
33
34 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
35 b, FLA_LEFT );
36
37 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
38 b, FLA_LEFT );
39
40 /*------------------------------------------------------------*/
41
42 /* C = C + A1 * B1' + B1 * A1' */
44 alpha, A1, B1, FLA_ONE, C,
45 FLA_Cntl_sub_her2k( cntl ) );
46
47 /*------------------------------------------------------------*/
48
49 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
50 FLA_RIGHT );
51
52 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
53 FLA_RIGHT );
54
55 }
56
57 return FLA_SUCCESS;
58}
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_Her2k_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var2()

FLA_Error FLA_Her2k_ln_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_TOP );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_TL );
42
43 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
44
46
48 /* ** */ /* ** */
49 &A1,
50 AB, &A2, b, FLA_BOTTOM );
51
53 /* ** */ /* ** */
54 &B1,
55 BB, &B2, b, FLA_BOTTOM );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
58 /* ************* */ /* ******************** */
59 &C10, /**/ &C11, &C12,
60 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
61 b, b, FLA_BR );
62
63 /*------------------------------------------------------------*/
64
65 /* C10 = C10 + B1 * A0' */
67 alpha, B1, A0, FLA_ONE, C10,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C21 = C21 + A2 * B1' */
72 alpha, A2, B1, FLA_ONE, C21,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 A1,
84 /* ** */ /* ** */
85 &AB, A2, FLA_TOP );
86
88 B1,
89 /* ** */ /* ** */
90 &BB, B2, FLA_TOP );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
93 C10, C11, /**/ C12,
94 /* ************** */ /* ****************** */
95 &CBL, /**/ &CBR, C20, C21, /**/ C22,
96 FLA_TL );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var3()

FLA_Error FLA_Her2k_ln_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
14{
15 FLA_Obj AT, A0,
16 AB, A1,
17 A2;
18
19 FLA_Obj BT, B0,
20 BB, B1,
21 B2;
22
23 FLA_Obj CTL, CTR, C00, C01, C02,
24 CBL, CBR, C10, C11, C12,
25 C20, C21, C22;
26
27 dim_t b;
28
30 FLA_Cntl_sub_scalr( cntl ) );
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_TOP );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_TOP );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_TL );
40
41 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
42
44
46 /* ** */ /* ** */
47 &A1,
48 AB, &A2, b, FLA_BOTTOM );
49
51 /* ** */ /* ** */
52 &B1,
53 BB, &B2, b, FLA_BOTTOM );
54
55 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
56 /* ************* */ /* ******************** */
57 &C10, /**/ &C11, &C12,
58 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
59 b, b, FLA_BR );
60
61 /*------------------------------------------------------------*/
62
63 /* C10 = C10 + A1 * B0' */
65 alpha, A1, B0, FLA_ONE, C10,
66 FLA_Cntl_sub_gemm1( cntl ) );
67
68 /* C21 = C21 + B2 * A1' */
70 alpha, B2, A1, FLA_ONE, C21,
71 FLA_Cntl_sub_gemm2( cntl ) );
72
73 /* C11 = C11 + A1 * B1' + B1 * A1' */
75 alpha, A1, B1, FLA_ONE, C11,
76 FLA_Cntl_sub_her2k( cntl ) );
77
78 /*------------------------------------------------------------*/
79
81 A1,
82 /* ** */ /* ** */
83 &AB, A2, FLA_TOP );
84
86 B1,
87 /* ** */ /* ** */
88 &BB, B2, FLA_TOP );
89
90 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
91 C10, C11, /**/ C12,
92 /* ************** */ /* ****************** */
93 &CBL, /**/ &CBR, C20, C21, /**/ C22,
94 FLA_TL );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var4()

FLA_Error FLA_Her2k_ln_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_TOP );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_TOP );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_TL );
42
43 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
44
46
48 /* ** */ /* ** */
49 &A1,
50 AB, &A2, b, FLA_BOTTOM );
51
53 /* ** */ /* ** */
54 &B1,
55 BB, &B2, b, FLA_BOTTOM );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
58 /* ************* */ /* ******************** */
59 &C10, /**/ &C11, &C12,
60 CBL, /**/ CBR, &C20, /**/ &C21, &C22,
61 b, b, FLA_BR );
62
63 /*------------------------------------------------------------*/
64
65 /* C21 = C21 + A2 * B1' */
67 alpha, A2, B1, FLA_ONE, C21,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C21 = C21 + B2 * A1' */
72 alpha, B2, A1, FLA_ONE, C21,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 A1,
84 /* ** */ /* ** */
85 &AB, A2, FLA_TOP );
86
88 B1,
89 /* ** */ /* ** */
90 &BB, B2, FLA_TOP );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
93 C10, C11, /**/ C12,
94 /* ************** */ /* ****************** */
95 &CBL, /**/ &CBR, C20, C21, /**/ C22,
96 FLA_TL );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var5()

FLA_Error FLA_Her2k_ln_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_BR );
42
43 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
44
46
48 &A1,
49 /* ** */ /* ** */
50 AB, &A2, b, FLA_TOP );
51
53 &B1,
54 /* ** */ /* ** */
55 BB, &B2, b, FLA_TOP );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
58 &C10, &C11, /**/ &C12,
59 /* ************* */ /* ******************** */
60 CBL, /**/ CBR, &C20, &C21, /**/ &C22,
61 b, b, FLA_TL );
62
63 /*------------------------------------------------------------*/
64
65 /* C21 = C21 + A2 * B1' */
67 alpha, A2, B1, FLA_ONE, C21,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C21 = C21 + B2 * A1' */
72 alpha, B2, A1, FLA_ONE, C21,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 /* ** */ /* ** */
84 A1,
85 &AB, A2, FLA_BOTTOM );
86
88 /* ** */ /* ** */
89 B1,
90 &BB, B2, FLA_BOTTOM );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
93 /* ************** */ /* ****************** */
94 C10, /**/ C11, C12,
95 &CBL, /**/ &CBR, C20, /**/ C21, C22,
96 FLA_BR );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var6()

FLA_Error FLA_Her2k_ln_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_BR );
42
43 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
44
46
48 &A1,
49 /* ** */ /* ** */
50 AB, &A2, b, FLA_TOP );
51
53 &B1,
54 /* ** */ /* ** */
55 BB, &B2, b, FLA_TOP );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
58 &C10, &C11, /**/ &C12,
59 /* ************* */ /* ******************** */
60 CBL, /**/ CBR, &C20, &C21, /**/ &C22,
61 b, b, FLA_TL );
62
63 /*------------------------------------------------------------*/
64
65 /* C10 = C10 + A1 * B0' */
67 alpha, A1, B0, FLA_ONE, C10,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C21 = C21 + B2 * A1' */
72 alpha, B2, A1, FLA_ONE, C21,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 /* ** */ /* ** */
84 A1,
85 &AB, A2, FLA_BOTTOM );
86
88 /* ** */ /* ** */
89 B1,
90 &BB, B2, FLA_BOTTOM );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
93 /* ************** */ /* ****************** */
94 C10, /**/ C11, C12,
95 &CBL, /**/ &CBR, C20, /**/ C21, C22,
96 FLA_BR );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var7()

FLA_Error FLA_Her2k_ln_blk_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_BR );
42
43 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
44
46
48 &A1,
49 /* ** */ /* ** */
50 AB, &A2, b, FLA_TOP );
51
53 &B1,
54 /* ** */ /* ** */
55 BB, &B2, b, FLA_TOP );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
58 &C10, &C11, /**/ &C12,
59 /* ************* */ /* ******************** */
60 CBL, /**/ CBR, &C20, &C21, /**/ &C22,
61 b, b, FLA_TL );
62
63 /*------------------------------------------------------------*/
64
65 /* C10 = C10 + B1 * A0' */
67 alpha, B1, A0, FLA_ONE, C10,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C21 = C21 + A2 * B1' */
72 alpha, A2, B1, FLA_ONE, C21,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 /* ** */ /* ** */
84 A1,
85 &AB, A2, FLA_BOTTOM );
86
88 /* ** */ /* ** */
89 B1,
90 &BB, B2, FLA_BOTTOM );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
93 /* ************** */ /* ****************** */
94 C10, /**/ C11, C12,
95 &CBL, /**/ &CBR, C20, /**/ C21, C22,
96 FLA_BR );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var8()

FLA_Error FLA_Her2k_ln_blk_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
16{
17 FLA_Obj AT, A0,
18 AB, A1,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, B1,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, C01, C02,
26 CBL, CBR, C10, C11, C12,
27 C20, C21, C22;
28
29 dim_t b;
30
32 FLA_Cntl_sub_scalr( cntl ) );
33
34 FLA_Part_2x1( A, &AT,
35 &AB, 0, FLA_BOTTOM );
36
37 FLA_Part_2x1( B, &BT,
38 &BB, 0, FLA_BOTTOM );
39
40 FLA_Part_2x2( C, &CTL, &CTR,
41 &CBL, &CBR, 0, 0, FLA_BR );
42
43 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
44
46
48 &A1,
49 /* ** */ /* ** */
50 AB, &A2, b, FLA_TOP );
51
53 &B1,
54 /* ** */ /* ** */
55 BB, &B2, b, FLA_TOP );
56
57 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
58 &C10, &C11, /**/ &C12,
59 /* ************* */ /* ******************** */
60 CBL, /**/ CBR, &C20, &C21, /**/ &C22,
61 b, b, FLA_TL );
62
63 /*------------------------------------------------------------*/
64
65 /* C10 = C10 + A1 * B0' */
67 alpha, A1, B0, FLA_ONE, C10,
68 FLA_Cntl_sub_gemm1( cntl ) );
69
70 /* C10 = C10 + B1 * A0' */
72 alpha, B1, A0, FLA_ONE, C10,
73 FLA_Cntl_sub_gemm2( cntl ) );
74
75 /* C11 = C11 + A1 * B1' + B1 * A1' */
77 alpha, A1, B1, FLA_ONE, C11,
78 FLA_Cntl_sub_her2k( cntl ) );
79
80 /*------------------------------------------------------------*/
81
83 /* ** */ /* ** */
84 A1,
85 &AB, A2, FLA_BOTTOM );
86
88 /* ** */ /* ** */
89 B1,
90 &BB, B2, FLA_BOTTOM );
91
92 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
93 /* ************** */ /* ****************** */
94 C10, /**/ C11, C12,
95 &CBL, /**/ &CBR, C20, /**/ C21, C22,
96 FLA_BR );
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_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_blk_var9()

FLA_Error FLA_Her2k_ln_blk_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)
14{
15 FLA_Obj AL, AR, A0, A1, A2;
16
17 FLA_Obj BL, BR, B0, B1, B2;
18
19 dim_t b;
20
22 FLA_Cntl_sub_scalr( cntl ) );
23
24 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
25
26 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
27
28 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
29
31
32 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
33 b, FLA_RIGHT );
34
35 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
36 b, FLA_RIGHT );
37
38 /*------------------------------------------------------------*/
39
40 /* C = C + A1 * B1' + B1 * A1' */
42 alpha, A1, B1, FLA_ONE, C,
43 FLA_Cntl_sub_her2k( cntl ) );
44
45 /*------------------------------------------------------------*/
46
47 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
48 FLA_LEFT );
49
50 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
51 FLA_LEFT );
52
53 }
54
55 return FLA_SUCCESS;
56}

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Her2k_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_internal(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var1()

FLA_Error FLA_Her2k_ln_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_TOP );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_TOP );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_TL );
40
41 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
42
43
45 /* ** */ /* ** */
46 &a1t,
47 AB, &A2, 1, FLA_BOTTOM );
48
50 /* ** */ /* ** */
51 &b1t,
52 BB, &B2, 1, FLA_BOTTOM );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
55 /* ************* */ /* ************************** */
56 &c10t, /**/ &gamma11, &c12t,
57 CBL, /**/ CBR, &C20, /**/ &c21, &C22,
58 1, 1, FLA_BR );
59
60 /*------------------------------------------------------------*/
61
62 /* c10t = c10t + a1t * B0' */
64
65 /* c10t = c10t + b1t * A0' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 a1t,
75 /* ** */ /* ** */
76 &AB, A2, FLA_TOP );
77
79 b1t,
80 /* ** */ /* ** */
81 &BB, B2, FLA_TOP );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
84 c10t, gamma11, /**/ c12t,
85 /* ************** */ /* ************************ */
86 &CBL, /**/ &CBR, C20, c21, /**/ C22,
87 FLA_TL );
88
89 }
90
91 return FLA_SUCCESS;
92}
FLA_Error FLA_Dot2cs_external(FLA_Conj conj, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj beta, FLA_Obj rho)
Definition FLA_Dot2cs_external.c:13
FLA_Error FLA_Scalr_external(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A)
Definition FLA_Scalr_external.c:13
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_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var10()

FLA_Error FLA_Her2k_ln_unb_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AL, AR, A0, a1t, A2;
18
19 FLA_Obj BL, BR, B0, b1t, B2;
20
22
23 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
24
25 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
26
27 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
28
29
30 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1t, /**/ &A2,
31 1, FLA_LEFT );
32
33 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2,
34 1, FLA_LEFT );
35
36 /*------------------------------------------------------------*/
37
38 /* C = C + a1t * b1t' + b1t * a1t' */
40
41 /*------------------------------------------------------------*/
42
43 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1t, A2,
44 FLA_RIGHT );
45
46 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2,
47 FLA_RIGHT );
48
49 }
50
51 return FLA_SUCCESS;
52}
FLA_Error FLA_Her2_external(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition FLA_Her2_external.c:13

References FLA_Cont_with_1x3_to_1x2(), FLA_Her2_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var2()

FLA_Error FLA_Her2k_ln_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_TOP );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_TOP );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_TL );
40
41 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
42
43
45 /* ** */ /* ** */
46 &a1t,
47 AB, &A2, 1, FLA_BOTTOM );
48
50 /* ** */ /* ** */
51 &b1t,
52 BB, &B2, 1, FLA_BOTTOM );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
55 /* ************* */ /* ************************** */
56 &c10t, /**/ &gamma11, &c12t,
57 CBL, /**/ CBR, &C20, /**/ &c21, &C22,
58 1, 1, FLA_BR );
59
60 /*------------------------------------------------------------*/
61
62 /* c21 = c21 + A2 * b1t' */
64
65 /* c10t = c10t + b1t * A0' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 a1t,
75 /* ** */ /* ** */
76 &AB, A2, FLA_TOP );
77
79 b1t,
80 /* ** */ /* ** */
81 &BB, B2, FLA_TOP );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
84 c10t, gamma11, /**/ c12t,
85 /* ************** */ /* ************************ */
86 &CBL, /**/ &CBR, C20, c21, /**/ C22,
87 FLA_TL );
88
89 }
90
91 return FLA_SUCCESS;
92}
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

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), 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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var3()

FLA_Error FLA_Her2k_ln_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
14{
15 FLA_Obj AT, A0,
16 AB, a1t,
17 A2;
18
19 FLA_Obj BT, B0,
20 BB, b1t,
21 B2;
22
23 FLA_Obj CTL, CTR, C00, c01, C02,
25 C20, c21, C22;
26
27
29
30 FLA_Part_2x1( A, &AT,
31 &AB, 0, FLA_TOP );
32
33 FLA_Part_2x1( B, &BT,
34 &BB, 0, FLA_TOP );
35
36 FLA_Part_2x2( C, &CTL, &CTR,
37 &CBL, &CBR, 0, 0, FLA_TL );
38
39 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
40
41
43 /* ** */ /* ** */
44 &a1t,
45 AB, &A2, 1, FLA_BOTTOM );
46
48 /* ** */ /* ** */
49 &b1t,
50 BB, &B2, 1, FLA_BOTTOM );
51
52 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
53 /* ************* */ /* ************************** */
54 &c10t, /**/ &gamma11, &c12t,
55 CBL, /**/ CBR, &C20, /**/ &c21, &C22,
56 1, 1, FLA_BR );
57
58 /*------------------------------------------------------------*/
59
60 /* c10t = c10t + a1t * B0' */
62
63 /* c21 = c21 + B2 * a1t' */
65
66 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
68
69 /*------------------------------------------------------------*/
70
72 a1t,
73 /* ** */ /* ** */
74 &AB, A2, FLA_TOP );
75
77 b1t,
78 /* ** */ /* ** */
79 &BB, B2, FLA_TOP );
80
81 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
82 c10t, gamma11, /**/ c12t,
83 /* ************** */ /* ************************ */
84 &CBL, /**/ &CBR, C20, c21, /**/ C22,
85 FLA_TL );
86
87 }
88
89 return FLA_SUCCESS;
90}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), 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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var4()

FLA_Error FLA_Her2k_ln_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_TOP );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_TOP );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_TL );
40
41 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
42
43
45 /* ** */ /* ** */
46 &a1t,
47 AB, &A2, 1, FLA_BOTTOM );
48
50 /* ** */ /* ** */
51 &b1t,
52 BB, &B2, 1, FLA_BOTTOM );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
55 /* ************* */ /* ************************** */
56 &c10t, /**/ &gamma11, &c12t,
57 CBL, /**/ CBR, &C20, /**/ &c21, &C22,
58 1, 1, FLA_BR );
59
60 /*------------------------------------------------------------*/
61
62 /* c21 = c21 + A2 * b1t' */
64
65 /* c21 = c21 + B2 * a1t' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 a1t,
75 /* ** */ /* ** */
76 &AB, A2, FLA_TOP );
77
79 b1t,
80 /* ** */ /* ** */
81 &BB, B2, FLA_TOP );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
84 c10t, gamma11, /**/ c12t,
85 /* ************** */ /* ************************ */
86 &CBL, /**/ &CBR, C20, c21, /**/ C22,
87 FLA_TL );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var5()

FLA_Error FLA_Her2k_ln_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_BOTTOM );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_BOTTOM );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_BR );
40
41 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
42
43
45 &a1t,
46 /* ** */ /* ** */
47 AB, &A2, 1, FLA_TOP );
48
50 &b1t,
51 /* ** */ /* ** */
52 BB, &B2, 1, FLA_TOP );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
55 &c10t, &gamma11, /**/ &c12t,
56 /* ************* */ /* ************************** */
57 CBL, /**/ CBR, &C20, &c21, /**/ &C22,
58 1, 1, FLA_TL );
59
60 /*------------------------------------------------------------*/
61
62 /* c21 = c21 + A2 * b1t' */
64
65 /* c21 = c21 + B2 * a1t' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 /* ** */ /* ** */
75 a1t,
76 &AB, A2, FLA_BOTTOM );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
84 /* ************** */ /* ************************ */
85 c10t, /**/ gamma11, c12t,
86 &CBL, /**/ &CBR, C20, /**/ c21, C22,
87 FLA_BR );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var6()

FLA_Error FLA_Her2k_ln_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_BOTTOM );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_BOTTOM );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_BR );
40
41 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
42
43
45 &a1t,
46 /* ** */ /* ** */
47 AB, &A2, 1, FLA_TOP );
48
50 &b1t,
51 /* ** */ /* ** */
52 BB, &B2, 1, FLA_TOP );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
55 &c10t, &gamma11, /**/ &c12t,
56 /* ************* */ /* ************************** */
57 CBL, /**/ CBR, &C20, &c21, /**/ &C22,
58 1, 1, FLA_TL );
59
60 /*------------------------------------------------------------*/
61
62 /* c10t = c10t + a1t * B0' */
64
65 /* c21 = c21 + B2 * a1t' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 /* ** */ /* ** */
75 a1t,
76 &AB, A2, FLA_BOTTOM );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
84 /* ************** */ /* ************************ */
85 c10t, /**/ gamma11, c12t,
86 &CBL, /**/ &CBR, C20, /**/ c21, C22,
87 FLA_BR );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), 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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var7()

FLA_Error FLA_Her2k_ln_unb_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_BOTTOM );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_BOTTOM );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_BR );
40
41 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
42
43
45 &a1t,
46 /* ** */ /* ** */
47 AB, &A2, 1, FLA_TOP );
48
50 &b1t,
51 /* ** */ /* ** */
52 BB, &B2, 1, FLA_TOP );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
55 &c10t, &gamma11, /**/ &c12t,
56 /* ************* */ /* ************************** */
57 CBL, /**/ CBR, &C20, &c21, /**/ &C22,
58 1, 1, FLA_TL );
59
60 /*------------------------------------------------------------*/
61
62 /* c10t = c10t + b1t * A0' */
64
65 /* c21 = c21 + A2 * b1t' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 /* ** */ /* ** */
75 a1t,
76 &AB, A2, FLA_BOTTOM );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
84 /* ************** */ /* ************************ */
85 c10t, /**/ gamma11, c12t,
86 &CBL, /**/ &CBR, C20, /**/ c21, C22,
87 FLA_BR );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), 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_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var8()

FLA_Error FLA_Her2k_ln_unb_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AT, A0,
18 AB, a1t,
19 A2;
20
21 FLA_Obj BT, B0,
22 BB, b1t,
23 B2;
24
25 FLA_Obj CTL, CTR, C00, c01, C02,
27 C20, c21, C22;
28
29
31
32 FLA_Part_2x1( A, &AT,
33 &AB, 0, FLA_BOTTOM );
34
35 FLA_Part_2x1( B, &BT,
36 &BB, 0, FLA_BOTTOM );
37
38 FLA_Part_2x2( C, &CTL, &CTR,
39 &CBL, &CBR, 0, 0, FLA_BR );
40
41 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
42
43
45 &a1t,
46 /* ** */ /* ** */
47 AB, &A2, 1, FLA_TOP );
48
50 &b1t,
51 /* ** */ /* ** */
52 BB, &B2, 1, FLA_TOP );
53
54 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
55 &c10t, &gamma11, /**/ &c12t,
56 /* ************* */ /* ************************** */
57 CBL, /**/ CBR, &C20, &c21, /**/ &C22,
58 1, 1, FLA_TL );
59
60 /*------------------------------------------------------------*/
61
62 /* c10t = c10t + a1t * B0' */
64
65 /* c10t = c10t + b1t * A0' */
67
68 /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
70
71 /*------------------------------------------------------------*/
72
74 /* ** */ /* ** */
75 a1t,
76 &AB, A2, FLA_BOTTOM );
77
79 /* ** */ /* ** */
80 b1t,
81 &BB, B2, FLA_BOTTOM );
82
83 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
84 /* ************** */ /* ************************ */
85 c10t, /**/ gamma11, c12t,
86 &CBL, /**/ &CBR, C20, /**/ c21, C22,
87 FLA_BR );
88
89 }
90
91 return FLA_SUCCESS;
92}

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2cs_external(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().

◆ FLA_Her2k_ln_unb_var9()

FLA_Error FLA_Her2k_ln_unb_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16{
17 FLA_Obj AL, AR, A0, a1t, A2;
18
19 FLA_Obj BL, BR, B0, b1t, B2;
20
22
23 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
24
25 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
26
27 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
28
29
30 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2,
31 1, FLA_RIGHT );
32
33 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2,
34 1, FLA_RIGHT );
35
36 /*------------------------------------------------------------*/
37
38 /* C = C + a1t * b1t' + b1t * a1t' */
40
41 /*------------------------------------------------------------*/
42
43 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2,
44 FLA_LEFT );
45
46 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2,
47 FLA_LEFT );
48
49 }
50
51 return FLA_SUCCESS;
52}

References FLA_Cont_with_1x3_to_1x2(), FLA_Her2_external(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scalr_external(), and i.

Referenced by FLA_Her2k_ln().