libflame  revision_anchor
Functions
FLA_Syr2k_lt.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Syr2k_lt_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_blk_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
 
FLA_Error FLA_Syr2k_lt_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var7 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var8 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var9 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syr2k_lt_unb_var10 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Syr2k_lt_blk_var1()

FLA_Error FLA_Syr2k_lt_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_TL );
36 
37  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
42  b, FLA_RIGHT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
45  b, FLA_RIGHT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
48  /* ************* */ /* ******************** */
49  &C10, /**/ &C11, &C12,
50  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
51  b, b, FLA_BR );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C10 = C10 + A1' * B0 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, A1, B0, FLA_ONE, C10,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C10 = C10 + B1' * A0 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B1, A0, FLA_ONE, C10,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
73  FLA_LEFT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
76  FLA_LEFT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
79  C10, C11, /**/ C12,
80  /* ************** */ /* ****************** */
81  &CBL, /**/ &CBR, C20, C21, /**/ C22,
82  FLA_TL );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }
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_Scalr_internal(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj A, fla_scalr_t *cntl)
Definition: FLA_Scalr_internal.c:16
FLA_Error FLA_Syr2k_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl)
Definition: FLA_Syr2k_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_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_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_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
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition: FLA_Blocksize.c:234
unsigned long dim_t
Definition: FLA_type_defs.h:71
Definition: FLA_type_defs.h:159

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var10()

FLA_Error FLA_Syr2k_lt_blk_var10 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_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  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_2x1( A, &AT,
31  &AB, 0, FLA_BOTTOM );
32 
33  FLA_Part_2x1( B, &BT,
34  &BB, 0, FLA_BOTTOM );
35 
36  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
37 
38  b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
39 
40  FLA_Repart_2x1_to_3x1( AT, &A0,
41  &A1,
42  /* ** */ /* ** */
43  AB, &A2, b, FLA_TOP );
44 
45  FLA_Repart_2x1_to_3x1( BT, &B0,
46  &B1,
47  /* ** */ /* ** */
48  BB, &B2, b, FLA_TOP );
49 
50  /*------------------------------------------------------------*/
51 
52  /* C = C + A1' * B1 + B1' * A1 */
53  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
54  alpha, A1, B1, FLA_ONE, C,
55  FLA_Cntl_sub_syr2k( cntl ) );
56 
57  /*------------------------------------------------------------*/
58 
59  FLA_Cont_with_3x1_to_2x1( &AT, A0,
60  /* ** */ /* ** */
61  A1,
62  &AB, A2, FLA_BOTTOM );
63 
64  FLA_Cont_with_3x1_to_2x1( &BT, B0,
65  /* ** */ /* ** */
66  B1,
67  &BB, B2, FLA_BOTTOM );
68 
69  }
70 
71  return FLA_SUCCESS;
72 }
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

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var2()

FLA_Error FLA_Syr2k_lt_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_TL );
36 
37  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
42  b, FLA_RIGHT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
45  b, FLA_RIGHT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
48  /* ************* */ /* ******************** */
49  &C10, /**/ &C11, &C12,
50  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
51  b, b, FLA_BR );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C10 = C10 + B1' * A0 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, B1, A0, FLA_ONE, C10,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C21 = C21 + A2' * B1 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, A2, B1, FLA_ONE, C21,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
73  FLA_LEFT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
76  FLA_LEFT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
79  C10, C11, /**/ C12,
80  /* ************** */ /* ****************** */
81  &CBL, /**/ &CBR, C20, C21, /**/ C22,
82  FLA_TL );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var3()

FLA_Error FLA_Syr2k_lt_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
14 {
15  FLA_Obj AL, AR, A0, A1, A2;
16 
17  FLA_Obj BL, BR, B0, B1, B2;
18 
19  FLA_Obj CTL, CTR, C00, C01, C02,
20  CBL, CBR, C10, C11, C12,
21  C20, C21, C22;
22 
23  dim_t b;
24 
25  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
26  FLA_Cntl_sub_scalr( cntl ) );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_TL );
34 
35  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
36 
37  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
38 
39  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
40  b, FLA_RIGHT );
41 
42  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
43  b, FLA_RIGHT );
44 
45  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
46  /* ************* */ /* ******************** */
47  &C10, /**/ &C11, &C12,
48  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
49  b, b, FLA_BR );
50 
51  /*------------------------------------------------------------*/
52 
53  /* C10 = C10 + A1' * B0 */
54  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
55  alpha, A1, B0, FLA_ONE, C10,
56  FLA_Cntl_sub_gemm1( cntl ) );
57 
58  /* C21 = C21 + B2' * A1 */
59  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
60  alpha, B2, A1, FLA_ONE, C21,
61  FLA_Cntl_sub_gemm2( cntl ) );
62 
63  /* C11 = C11 + A1' * B1 + B1' * A1 */
64  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
65  alpha, A1, B1, FLA_ONE, C11,
66  FLA_Cntl_sub_syr2k( cntl ) );
67 
68  /*------------------------------------------------------------*/
69 
70  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
71  FLA_LEFT );
72 
73  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
74  FLA_LEFT );
75 
76  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
77  C10, C11, /**/ C12,
78  /* ************** */ /* ****************** */
79  &CBL, /**/ &CBR, C20, C21, /**/ C22,
80  FLA_TL );
81 
82  }
83 
84  return FLA_SUCCESS;
85 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var4()

FLA_Error FLA_Syr2k_lt_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_TL );
36 
37  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
42  b, FLA_RIGHT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
45  b, FLA_RIGHT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
48  /* ************* */ /* ******************** */
49  &C10, /**/ &C11, &C12,
50  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
51  b, b, FLA_BR );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C21 = C21 + A2' * B1 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, A2, B1, FLA_ONE, C21,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C21 = C21 + B2' * A1 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B2, A1, FLA_ONE, C21,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
73  FLA_LEFT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
76  FLA_LEFT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
79  C10, C11, /**/ C12,
80  /* ************** */ /* ****************** */
81  &CBL, /**/ &CBR, C20, C21, /**/ C22,
82  FLA_TL );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var5()

FLA_Error FLA_Syr2k_lt_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_BR );
36 
37  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
42  b, FLA_LEFT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
45  b, FLA_LEFT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
48  &C10, &C11, /**/ &C12,
49  /* ************* */ /* ******************** */
50  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
51  b, b, FLA_TL );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C21 = C21 + A2' * B1 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, A2, B1, FLA_ONE, C21,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C21 = C21 + B2' * A1 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B2, A1, FLA_ONE, C21,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
73  FLA_RIGHT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
76  FLA_RIGHT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
79  /* ************** */ /* ****************** */
80  C10, /**/ C11, C12,
81  &CBL, /**/ &CBR, C20, /**/ C21, C22,
82  FLA_BR );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var6()

FLA_Error FLA_Syr2k_lt_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_BR );
36 
37  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
42  b, FLA_LEFT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
45  b, FLA_LEFT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
48  &C10, &C11, /**/ &C12,
49  /* ************* */ /* ******************** */
50  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
51  b, b, FLA_TL );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C10 = C10 + A1' * B0 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, A1, B0, FLA_ONE, C10,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C21 = C21 + B2' * A1 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B2, A1, FLA_ONE, C21,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
73  FLA_RIGHT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
76  FLA_RIGHT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
79  /* ************** */ /* ****************** */
80  C10, /**/ C11, C12,
81  &CBL, /**/ &CBR, C20, /**/ C21, C22,
82  FLA_BR );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var7()

FLA_Error FLA_Syr2k_lt_blk_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_BR );
36 
37  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
42  b, FLA_LEFT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
45  b, FLA_LEFT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
48  &C10, &C11, /**/ &C12,
49  /* ************* */ /* ******************** */
50  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
51  b, b, FLA_TL );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C10 = C10 + B1' * A0 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, B1, A0, FLA_ONE, C10,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C21 = C21 + A2' * B1 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, A2, B1, FLA_ONE, C21,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
73  FLA_RIGHT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
76  FLA_RIGHT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
79  /* ************** */ /* ****************** */
80  C10, /**/ C11, C12,
81  &CBL, /**/ &CBR, C20, /**/ C21, C22,
82  FLA_BR );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var8()

FLA_Error FLA_Syr2k_lt_blk_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj BL, BR, B0, B1, B2;
20 
21  FLA_Obj CTL, CTR, C00, C01, C02,
22  CBL, CBR, C10, C11, C12,
23  C20, C21, C22;
24 
25  dim_t b;
26 
27  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
28  FLA_Cntl_sub_scalr( cntl ) );
29 
30  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
31 
32  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
33 
34  FLA_Part_2x2( C, &CTL, &CTR,
35  &CBL, &CBR, 0, 0, FLA_BR );
36 
37  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
38 
39  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
40 
41  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
42  b, FLA_LEFT );
43 
44  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
45  b, FLA_LEFT );
46 
47  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
48  &C10, &C11, /**/ &C12,
49  /* ************* */ /* ******************** */
50  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
51  b, b, FLA_TL );
52 
53  /*------------------------------------------------------------*/
54 
55  /* C10 = C10 + A1' * B0 */
56  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
57  alpha, A1, B0, FLA_ONE, C10,
58  FLA_Cntl_sub_gemm1( cntl ) );
59 
60  /* C10 = C10 + B1' * A0 */
61  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
62  alpha, B1, A0, FLA_ONE, C10,
63  FLA_Cntl_sub_gemm2( cntl ) );
64 
65  /* C11 = C11 + A1' * B1 + B1' * A1 */
66  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
67  alpha, A1, B1, FLA_ONE, C11,
68  FLA_Cntl_sub_syr2k( cntl ) );
69 
70  /*------------------------------------------------------------*/
71 
72  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
73  FLA_RIGHT );
74 
75  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
76  FLA_RIGHT );
77 
78  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
79  /* ************** */ /* ****************** */
80  C10, /**/ C11, C12,
81  &CBL, /**/ &CBR, C20, /**/ C21, C22,
82  FLA_BR );
83 
84  }
85 
86  return FLA_SUCCESS;
87 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_blk_var9()

FLA_Error FLA_Syr2k_lt_blk_var9 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_syr2k_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  dim_t b;
24 
25  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
26  FLA_Cntl_sub_scalr( cntl ) );
27 
28  FLA_Part_2x1( A, &AT,
29  &AB, 0, FLA_TOP );
30 
31  FLA_Part_2x1( B, &BT,
32  &BB, 0, FLA_TOP );
33 
34  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
35 
36  b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
37 
38  FLA_Repart_2x1_to_3x1( AT, &A0,
39  /* ** */ /* ** */
40  &A1,
41  AB, &A2, b, FLA_BOTTOM );
42 
43  FLA_Repart_2x1_to_3x1( BT, &B0,
44  /* ** */ /* ** */
45  &B1,
46  BB, &B2, b, FLA_BOTTOM );
47 
48  /*------------------------------------------------------------*/
49 
50  /* C = C + A1' * B1 + B1' * A1 */
51  FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
52  alpha, A1, B1, FLA_ONE, C,
53  FLA_Cntl_sub_syr2k( cntl ) );
54 
55  /*------------------------------------------------------------*/
56 
57  FLA_Cont_with_3x1_to_2x1( &AT, A0,
58  A1,
59  /* ** */ /* ** */
60  &AB, A2, FLA_TOP );
61 
62  FLA_Cont_with_3x1_to_2x1( &BT, B0,
63  B1,
64  /* ** */ /* ** */
65  &BB, B2, FLA_TOP );
66 
67  }
68 
69  return FLA_SUCCESS;
70 }

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_internal(), and FLA_Syr2k_internal().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var1()

FLA_Error FLA_Syr2k_lt_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_TL );
34 
35  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
39  1, FLA_RIGHT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
42  1, FLA_RIGHT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
45  /* ************* */ /* ************************** */
46  &c10t, /**/ &gamma11, &c12t,
47  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
48  1, 1, FLA_BR );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c10t = c10t + a1' * B0 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1, FLA_ONE, c10t );
54 
55  /* c10t = c10t + b1' * A0 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1, FLA_ONE, c10t );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
64  FLA_LEFT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
67  FLA_LEFT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
70  c10t, gamma11, /**/ c12t,
71  /* ************** */ /* ************************ */
72  &CBL, /**/ &CBR, C20, c21, /**/ C22,
73  FLA_TL );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }
FLA_Error FLA_Dot2s_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj beta, FLA_Obj rho)
Definition: FLA_Dot2s_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_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var10()

FLA_Error FLA_Syr2k_lt_unb_var10 ( 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_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
26 
27  FLA_Part_2x1( A, &AT,
28  &AB, 0, FLA_BOTTOM );
29 
30  FLA_Part_2x1( B, &BT,
31  &BB, 0, FLA_BOTTOM );
32 
33  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
34 
35 
36  FLA_Repart_2x1_to_3x1( AT, &A0,
37  &a1t,
38  /* ** */ /* ** */
39  AB, &A2, 1, FLA_TOP );
40 
41  FLA_Repart_2x1_to_3x1( BT, &B0,
42  &b1t,
43  /* ** */ /* ** */
44  BB, &B2, 1, FLA_TOP );
45 
46  /*------------------------------------------------------------*/
47 
48  /* C = C + a1t' * b1t + b1t' * a1t */
49  FLA_Syr2_external( FLA_LOWER_TRIANGULAR, alpha, a1t, b1t, C );
50 
51  /*------------------------------------------------------------*/
52 
53  FLA_Cont_with_3x1_to_2x1( &AT, A0,
54  /* ** */ /* ** */
55  a1t,
56  &AB, A2, FLA_BOTTOM );
57 
58  FLA_Cont_with_3x1_to_2x1( &BT, B0,
59  /* ** */ /* ** */
60  b1t,
61  &BB, B2, FLA_BOTTOM );
62 
63  }
64 
65  return FLA_SUCCESS;
66 }
FLA_Error FLA_Syr2_external(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Syr2_external.c:13

References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr2_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var2()

FLA_Error FLA_Syr2k_lt_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_TL );
34 
35  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
39  1, FLA_RIGHT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
42  1, FLA_RIGHT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
45  /* ************* */ /* ************************** */
46  &c10t, /**/ &gamma11, &c12t,
47  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
48  1, 1, FLA_BR );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c10t = c10t + b1' * A0 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1, FLA_ONE, c10t );
54 
55  /* c21 = c21 + A2' * b1 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
64  FLA_LEFT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
67  FLA_LEFT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
70  c10t, gamma11, /**/ c12t,
71  /* ************** */ /* ************************ */
72  &CBL, /**/ &CBR, C20, c21, /**/ C22,
73  FLA_TL );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var3()

FLA_Error FLA_Syr2k_lt_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
14 {
15  FLA_Obj AL, AR, A0, a1, A2;
16 
17  FLA_Obj BL, BR, B0, b1, B2;
18 
19  FLA_Obj CTL, CTR, C00, c01, C02,
20  CBL, CBR, c10t, gamma11, c12t,
21  C20, c21, C22;
22 
23 
24  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
25 
26  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
27 
28  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
29 
30  FLA_Part_2x2( C, &CTL, &CTR,
31  &CBL, &CBR, 0, 0, FLA_TL );
32 
33  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
34 
35 
36  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
37  1, FLA_RIGHT );
38 
39  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
40  1, FLA_RIGHT );
41 
42  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
43  /* ************* */ /* ************************** */
44  &c10t, /**/ &gamma11, &c12t,
45  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
46  1, 1, FLA_BR );
47 
48  /*------------------------------------------------------------*/
49 
50  /* c10t = c10t + a1' * B0 */
51  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1, FLA_ONE, c10t );
52 
53  /* c21 = c21 + B2' * a1 */
54  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1, FLA_ONE, c21 );
55 
56  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
57  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
58 
59  /*------------------------------------------------------------*/
60 
61  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
62  FLA_LEFT );
63 
64  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
65  FLA_LEFT );
66 
67  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
68  c10t, gamma11, /**/ c12t,
69  /* ************** */ /* ************************ */
70  &CBL, /**/ &CBR, C20, c21, /**/ C22,
71  FLA_TL );
72 
73  }
74 
75  return FLA_SUCCESS;
76 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var4()

FLA_Error FLA_Syr2k_lt_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_TL );
34 
35  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
39  1, FLA_RIGHT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
42  1, FLA_RIGHT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
45  /* ************* */ /* ************************** */
46  &c10t, /**/ &gamma11, &c12t,
47  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
48  1, 1, FLA_BR );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c21 = c21 + A2' * b1 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );
54 
55  /* c21 = c21 + B2' * a1 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1, FLA_ONE, c21 );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
64  FLA_LEFT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
67  FLA_LEFT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
70  c10t, gamma11, /**/ c12t,
71  /* ************** */ /* ************************ */
72  &CBL, /**/ &CBR, C20, c21, /**/ C22,
73  FLA_TL );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var5()

FLA_Error FLA_Syr2k_lt_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_BR );
34 
35  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
39  1, FLA_LEFT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
42  1, FLA_LEFT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
45  &c10t, &gamma11, /**/ &c12t,
46  /* ************* */ /* ************************** */
47  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
48  1, 1, FLA_TL );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c21 = c21 + A2' * b1 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );
54 
55  /* c21 = c21 + B2' * a1 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1, FLA_ONE, c21 );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
64  FLA_RIGHT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
67  FLA_RIGHT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
70  /* ************** */ /* ************************ */
71  c10t, /**/ gamma11, c12t,
72  &CBL, /**/ &CBR, C20, /**/ c21, C22,
73  FLA_BR );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var6()

FLA_Error FLA_Syr2k_lt_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_BR );
34 
35  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
39  1, FLA_LEFT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
42  1, FLA_LEFT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
45  &c10t, &gamma11, /**/ &c12t,
46  /* ************* */ /* ************************** */
47  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
48  1, 1, FLA_TL );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c10t = c10t + a1' * B0 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1, FLA_ONE, c10t );
54 
55  /* c21 = c21 + B2' * a1 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1, FLA_ONE, c21 );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
64  FLA_RIGHT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
67  FLA_RIGHT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
70  /* ************** */ /* ************************ */
71  c10t, /**/ gamma11, c12t,
72  &CBL, /**/ &CBR, C20, /**/ c21, C22,
73  FLA_BR );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var7()

FLA_Error FLA_Syr2k_lt_unb_var7 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_BR );
34 
35  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
39  1, FLA_LEFT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
42  1, FLA_LEFT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
45  &c10t, &gamma11, /**/ &c12t,
46  /* ************* */ /* ************************** */
47  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
48  1, 1, FLA_TL );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c10t = c10t + b1' * A0 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1, FLA_ONE, c10t );
54 
55  /* c21 = c21 + A2' * b1 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
64  FLA_RIGHT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
67  FLA_RIGHT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
70  /* ************** */ /* ************************ */
71  c10t, /**/ gamma11, c12t,
72  &CBL, /**/ &CBR, C20, /**/ c21, C22,
73  FLA_BR );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var8()

FLA_Error FLA_Syr2k_lt_unb_var8 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj BL, BR, B0, b1, B2;
20 
21  FLA_Obj CTL, CTR, C00, c01, C02,
22  CBL, CBR, c10t, gamma11, c12t,
23  C20, c21, C22;
24 
25 
26  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
27 
28  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
29 
30  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
31 
32  FLA_Part_2x2( C, &CTL, &CTR,
33  &CBL, &CBR, 0, 0, FLA_BR );
34 
35  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
36 
37 
38  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
39  1, FLA_LEFT );
40 
41  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
42  1, FLA_LEFT );
43 
44  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
45  &c10t, &gamma11, /**/ &c12t,
46  /* ************* */ /* ************************** */
47  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
48  1, 1, FLA_TL );
49 
50  /*------------------------------------------------------------*/
51 
52  /* c10t = c10t + a1' * B0 */
53  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1, FLA_ONE, c10t );
54 
55  /* c10t = c10t + b1' * A0 */
56  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1, FLA_ONE, c10t );
57 
58  /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
59  FLA_Dot2s_external( alpha, a1, b1, FLA_ONE, gamma11 );
60 
61  /*------------------------------------------------------------*/
62 
63  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
64  FLA_RIGHT );
65 
66  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
67  FLA_RIGHT );
68 
69  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
70  /* ************** */ /* ************************ */
71  c10t, /**/ gamma11, c12t,
72  &CBL, /**/ &CBR, C20, /**/ c21, C22,
73  FLA_BR );
74 
75  }
76 
77  return FLA_SUCCESS;
78 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_external().

Referenced by FLA_Syr2k_lt().

◆ FLA_Syr2k_lt_unb_var9()

FLA_Error FLA_Syr2k_lt_unb_var9 ( 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_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
26 
27  FLA_Part_2x1( A, &AT,
28  &AB, 0, FLA_TOP );
29 
30  FLA_Part_2x1( B, &BT,
31  &BB, 0, FLA_TOP );
32 
33  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
34 
35 
36  FLA_Repart_2x1_to_3x1( AT, &A0,
37  /* ** */ /* ** */
38  &a1t,
39  AB, &A2, 1, FLA_BOTTOM );
40 
41  FLA_Repart_2x1_to_3x1( BT, &B0,
42  /* ** */ /* ** */
43  &b1t,
44  BB, &B2, 1, FLA_BOTTOM );
45 
46  /*------------------------------------------------------------*/
47 
48  /* C = C + a1t' * b1t + b1t' * a1t */
49  FLA_Syr2_external( FLA_LOWER_TRIANGULAR, alpha, a1t, b1t, C );
50 
51  /*------------------------------------------------------------*/
52 
53  FLA_Cont_with_3x1_to_2x1( &AT, A0,
54  a1t,
55  /* ** */ /* ** */
56  &AB, A2, FLA_TOP );
57 
58  FLA_Cont_with_3x1_to_2x1( &BT, B0,
59  b1t,
60  /* ** */ /* ** */
61  &BB, B2, FLA_TOP );
62 
63  }
64 
65  return FLA_SUCCESS;
66 }

References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr2_external().

Referenced by FLA_Syr2k_lt().