libflame  revision_anchor
Functions
FLA_Syrk_lt.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Syrk_lt_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
 
FLA_Error FLA_Syrk_lt_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syrk_lt_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syrk_lt_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syrk_lt_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syrk_lt_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Syrk_lt_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Syrk_lt_blk_var1()

FLA_Error FLA_Syrk_lt_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
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_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_TL );
29 
30  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
31 
32  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
33 
34  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
35  b, FLA_RIGHT );
36 
37  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
38  /* ************* */ /* ******************** */
39  &C10, /**/ &C11, &C12,
40  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
41  b, b, FLA_BR );
42 
43  /*------------------------------------------------------------*/
44 
45  /* C10 = C10 + A1' * A0 */
46  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
47  alpha, A1, A0, beta, C10,
48  FLA_Cntl_sub_gemm( cntl ) );
49 
50  /* C11 = C11 + A1' * A1 */
51  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
52  alpha, A1, beta, C11,
53  FLA_Cntl_sub_syrk( cntl ) );
54 
55  /*------------------------------------------------------------*/
56 
57  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
58  FLA_LEFT );
59 
60  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
61  C10, C11, /**/ C12,
62  /* ************** */ /* ****************** */
63  &CBL, /**/ &CBR, C20, C21, /**/ C22,
64  FLA_TL );
65 
66  }
67 
68  return FLA_SUCCESS;
69 }
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_Syrk_internal(FLA_Uplo uplo, FLA_Trans trans, FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t *cntl)
Definition: FLA_Syrk_internal.c:16
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_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_blk_var2()

FLA_Error FLA_Syrk_lt_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
14 {
15  FLA_Obj AL, AR, A0, A1, A2;
16 
17  FLA_Obj CTL, CTR, C00, C01, C02,
18  CBL, CBR, C10, C11, C12,
19  C20, C21, C22;
20 
21  dim_t b;
22 
23  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
24 
25  FLA_Part_2x2( C, &CTL, &CTR,
26  &CBL, &CBR, 0, 0, FLA_TL );
27 
28  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
29 
30  b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
31 
32  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2,
33  b, FLA_RIGHT );
34 
35  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02,
36  /* ************* */ /* ******************** */
37  &C10, /**/ &C11, &C12,
38  CBL, /**/ CBR, &C20, /**/ &C21, &C22,
39  b, b, FLA_BR );
40 
41  /*------------------------------------------------------------*/
42 
43  /* C21 = C21 + A2' * A1 */
44  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
45  alpha, A2, A1, beta, C21,
46  FLA_Cntl_sub_gemm( cntl ) );
47 
48  /* C11 = C11 + A1' * A1 */
49  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
50  alpha, A1, beta, C11,
51  FLA_Cntl_sub_syrk( cntl ) );
52 
53  /*------------------------------------------------------------*/
54 
55  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
56  FLA_LEFT );
57 
58  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02,
59  C10, C11, /**/ C12,
60  /* ************** */ /* ****************** */
61  &CBL, /**/ &CBR, C20, C21, /**/ C22,
62  FLA_TL );
63 
64  }
65 
66  return FLA_SUCCESS;
67 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_blk_var3()

FLA_Error FLA_Syrk_lt_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
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_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_BR );
29 
30  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31 
32  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
33 
34  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
35  b, FLA_LEFT );
36 
37  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
38  &C10, &C11, /**/ &C12,
39  /* ************* */ /* ******************** */
40  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
41  b, b, FLA_TL );
42 
43  /*------------------------------------------------------------*/
44 
45  /* C21 = C21 + A2' * A1 */
46  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
47  alpha, A2, A1, beta, C21,
48  FLA_Cntl_sub_gemm( cntl ) );
49 
50  /* C11 = C11 + A1' * A1 */
51  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
52  alpha, A1, beta, C11,
53  FLA_Cntl_sub_syrk( cntl ) );
54 
55  /*------------------------------------------------------------*/
56 
57  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
58  FLA_RIGHT );
59 
60  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
61  /* ************** */ /* ****************** */
62  C10, /**/ C11, C12,
63  &CBL, /**/ &CBR, C20, /**/ C21, C22,
64  FLA_BR );
65 
66  }
67 
68  return FLA_SUCCESS;
69 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_blk_var4()

FLA_Error FLA_Syrk_lt_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
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_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_BR );
29 
30  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31 
32  b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
33 
34  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2,
35  b, FLA_LEFT );
36 
37  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02,
38  &C10, &C11, /**/ &C12,
39  /* ************* */ /* ******************** */
40  CBL, /**/ CBR, &C20, &C21, /**/ &C22,
41  b, b, FLA_TL );
42 
43  /*------------------------------------------------------------*/
44 
45  /* C10 = C10 + A1' * A0 */
46  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
47  alpha, A1, A0, beta, C10,
48  FLA_Cntl_sub_gemm( cntl ) );
49 
50  /* C11 = C11 + A1' * A1 */
51  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
52  alpha, A1, beta, C11,
53  FLA_Cntl_sub_syrk( cntl ) );
54 
55  /*------------------------------------------------------------*/
56 
57  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
58  FLA_RIGHT );
59 
60  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02,
61  /* ************** */ /* ****************** */
62  C10, /**/ C11, C12,
63  &CBL, /**/ &CBR, C20, /**/ C21, C22,
64  FLA_BR );
65 
66  }
67 
68  return FLA_SUCCESS;
69 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_blk_var5()

FLA_Error FLA_Syrk_lt_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
14 {
15  FLA_Obj AT, A0,
16  AB, A1,
17  A2;
18 
19  dim_t b;
20 
21  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
22  FLA_Cntl_sub_scalr( cntl ) );
23 
24  FLA_Part_2x1( A, &AT,
25  &AB, 0, FLA_TOP );
26 
27  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
28 
29  b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
30 
31  FLA_Repart_2x1_to_3x1( AT, &A0,
32  /* ** */ /* ** */
33  &A1,
34  AB, &A2, b, FLA_BOTTOM );
35 
36  /*------------------------------------------------------------*/
37 
38  /* C = C + A1' * A1 */
39  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
40  alpha, A1, FLA_ONE, C,
41  FLA_Cntl_sub_syrk( cntl ) );
42 
43  /*------------------------------------------------------------*/
44 
45  FLA_Cont_with_3x1_to_2x1( &AT, A0,
46  A1,
47  /* ** */ /* ** */
48  &AB, A2, FLA_TOP );
49 
50  }
51 
52  return FLA_SUCCESS;
53 }
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_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_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_blk_var6()

FLA_Error FLA_Syrk_lt_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_syrk_t cntl 
)
16 {
17  FLA_Obj AT, A0,
18  AB, A1,
19  A2;
20 
21  dim_t b;
22 
23  FLA_Scalr_internal( FLA_LOWER_TRIANGULAR, beta, C,
24  FLA_Cntl_sub_scalr( cntl ) );
25 
26  FLA_Part_2x1( A, &AT,
27  &AB, 0, FLA_BOTTOM );
28 
29  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
30 
31  b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );
32 
33  FLA_Repart_2x1_to_3x1( AT, &A0,
34  &A1,
35  /* ** */ /* ** */
36  AB, &A2, b, FLA_TOP );
37 
38  /*------------------------------------------------------------*/
39 
40  /* C = C + A1' * A1 */
41  FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE,
42  alpha, A1, FLA_ONE, C,
43  FLA_Cntl_sub_syrk( cntl ) );
44 
45  /*------------------------------------------------------------*/
46 
47  FLA_Cont_with_3x1_to_2x1( &AT, A0,
48  /* ** */ /* ** */
49  A1,
50  &AB, A2, FLA_BOTTOM );
51 
52  }
53 
54  return FLA_SUCCESS;
55 }

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_Syrk_internal().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_unb_var1()

FLA_Error FLA_Syrk_lt_unb_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj CTL, CTR, C00, c01, C02,
20  CBL, CBR, c10t, gamma11, c12t,
21  C20, c21, C22;
22 
23  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_TL );
29 
30  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
31 
32  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
33  1, FLA_RIGHT );
34 
35  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
36  /* ************* */ /* ************************** */
37  &c10t, /**/ &gamma11, &c12t,
38  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
39  1, 1, FLA_BR );
40 
41  /*------------------------------------------------------------*/
42 
43  /* c10t = c10t + A0' * a1 */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, a1, FLA_ONE, c10t );
45 
46  /* gamma11 = gamma11 + a1' * a1 */
47  FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
52  FLA_LEFT );
53 
54  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
55  c10t, gamma11, /**/ c12t,
56  /* ************** */ /* ************************ */
57  &CBL, /**/ &CBR, C20, c21, /**/ C22,
58  FLA_TL );
59 
60  }
61 
62  return FLA_SUCCESS;
63 }
FLA_Error FLA_Dots_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj beta, FLA_Obj rho)
Definition: FLA_Dots_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_Dots_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_Syrk_lt().

◆ FLA_Syrk_lt_unb_var2()

FLA_Error FLA_Syrk_lt_unb_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj CTL, CTR, C00, c01, C02,
20  CBL, CBR, c10t, gamma11, c12t,
21  C20, c21, C22;
22 
23  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_TL );
29 
30  while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){
31 
32  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2,
33  1, FLA_RIGHT );
34 
35  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02,
36  /* ************* */ /* ************************** */
37  &c10t, /**/ &gamma11, &c12t,
38  CBL, /**/ CBR, &C20, /**/ &c21, &C22,
39  1, 1, FLA_BR );
40 
41  /*------------------------------------------------------------*/
42 
43  /* c21 = c21 + A2' * a1 */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, a1, FLA_ONE, c21 );
45 
46  /* gamma11 = gamma11 + a1' * a1 */
47  FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
52  FLA_LEFT );
53 
54  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02,
55  c10t, gamma11, /**/ c12t,
56  /* ************** */ /* ************************ */
57  &CBL, /**/ &CBR, C20, c21, /**/ C22,
58  FLA_TL );
59 
60  }
61 
62  return FLA_SUCCESS;
63 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_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_Syrk_lt().

◆ FLA_Syrk_lt_unb_var3()

FLA_Error FLA_Syrk_lt_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj CTL, CTR, C00, c01, C02,
20  CBL, CBR, c10t, gamma11, c12t,
21  C20, c21, C22;
22 
23  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_BR );
29 
30  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31 
32  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
33  1, FLA_LEFT );
34 
35  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
36  &c10t, &gamma11, /**/ &c12t,
37  /* ************* */ /* ************************** */
38  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
39  1, 1, FLA_TL );
40 
41  /*------------------------------------------------------------*/
42 
43  /* c21 = c21 + A2' * a1 */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, a1, FLA_ONE, c21 );
45 
46  /* gamma11 = gamma11 + a1' * a1 */
47  FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
52  FLA_RIGHT );
53 
54  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
55  /* ************** */ /* ************************ */
56  c10t, /**/ gamma11, c12t,
57  &CBL, /**/ &CBR, C20, /**/ c21, C22,
58  FLA_BR );
59 
60  }
61 
62  return FLA_SUCCESS;
63 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_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_Syrk_lt().

◆ FLA_Syrk_lt_unb_var4()

FLA_Error FLA_Syrk_lt_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AL, AR, A0, a1, A2;
18 
19  FLA_Obj CTL, CTR, C00, c01, C02,
20  CBL, CBR, c10t, gamma11, c12t,
21  C20, c21, C22;
22 
23  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x2( C, &CTL, &CTR,
28  &CBL, &CBR, 0, 0, FLA_BR );
29 
30  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){
31 
32  FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2,
33  1, FLA_LEFT );
34 
35  FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02,
36  &c10t, &gamma11, /**/ &c12t,
37  /* ************* */ /* ************************** */
38  CBL, /**/ CBR, &C20, &c21, /**/ &C22,
39  1, 1, FLA_TL );
40 
41  /*------------------------------------------------------------*/
42 
43  /* c10t = c10t + A0' * a1 */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, a1, FLA_ONE, c10t );
45 
46  /* gamma11 = gamma11 + a1' * a1 */
47  FLA_Dots_external( alpha, a1, a1, FLA_ONE, gamma11 );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
52  FLA_RIGHT );
53 
54  FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02,
55  /* ************** */ /* ************************ */
56  c10t, /**/ gamma11, c12t,
57  &CBL, /**/ &CBR, C20, /**/ c21, C22,
58  FLA_BR );
59 
60  }
61 
62  return FLA_SUCCESS;
63 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dots_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_Syrk_lt().

◆ FLA_Syrk_lt_unb_var5()

FLA_Error FLA_Syrk_lt_unb_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AT, A0,
18  AB, a1t,
19  A2;
20 
21  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
22 
23  FLA_Part_2x1( A, &AT,
24  &AB, 0, FLA_TOP );
25 
26  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){
27 
28  FLA_Repart_2x1_to_3x1( AT, &A0,
29  /* ** */ /* *** */
30  &a1t,
31  AB, &A2, 1, FLA_BOTTOM );
32 
33  /*------------------------------------------------------------*/
34 
35  /* C := C + a1t' * a1t */
36  FLA_Syr_external( FLA_LOWER_TRIANGULAR, alpha, a1t, C );
37 
38  /*------------------------------------------------------------*/
39 
40  FLA_Cont_with_3x1_to_2x1( &AT, A0,
41  a1t,
42  /* ** */ /* *** */
43  &AB, A2, FLA_TOP );
44 
45  }
46 
47  return FLA_SUCCESS;
48 }
FLA_Error FLA_Syr_external(FLA_Uplo uplo, FLA_Obj alpha, FLA_Obj x, FLA_Obj A)
Definition: FLA_Syr_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_Syr_external().

Referenced by FLA_Syrk_lt().

◆ FLA_Syrk_lt_unb_var6()

FLA_Error FLA_Syrk_lt_unb_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj AT, A0,
18  AB, a1t,
19  A2;
20 
21  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );
22 
23  FLA_Part_2x1( A, &AT,
24  &AB, 0, FLA_BOTTOM );
25 
26  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){
27 
28  FLA_Repart_2x1_to_3x1( AT, &A0,
29  &a1t,
30  /* ** */ /* *** */
31  AB, &A2, 1, FLA_TOP );
32 
33  /*------------------------------------------------------------*/
34 
35  /* C := C + a1t' * a1t */
36  FLA_Syr_external( FLA_LOWER_TRIANGULAR, alpha, a1t, C );
37 
38  /*------------------------------------------------------------*/
39 
40  FLA_Cont_with_3x1_to_2x1( &AT, A0,
41  /* ** */ /* *** */
42  a1t,
43  &AB, A2, FLA_BOTTOM );
44 
45  }
46 
47  return FLA_SUCCESS;
48 }

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

Referenced by FLA_Syrk_lt().