libflame  revision_anchor
Functions
FLA_Gemm_tn.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_Gemm_tn_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl)
 
FLA_Error FLA_Gemm_tn_unb_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Gemm_tn_unb_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Gemm_tn_unb_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Gemm_tn_unb_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Gemm_tn_unb_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 
FLA_Error FLA_Gemm_tn_unb_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C)
 

Function Documentation

◆ FLA_Gemm_tn_blk_var1()

FLA_Error FLA_Gemm_tn_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)
14 {
15  FLA_Obj AL, AR, A0, A1, A2;
16 
17  FLA_Obj CT, C0,
18  CB, C1,
19  C2;
20 
21  dim_t b;
22 
23  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
24 
25  FLA_Part_2x1( C, &CT,
26  &CB, 0, FLA_TOP );
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_2x1_to_3x1( CT, &C0,
36  /* ** */ /* ** */
37  &C1,
38  CB, &C2, b, FLA_BOTTOM );
39 
40  /*------------------------------------------------------------*/
41 
42  /* C1 = alpha * A1' * B + C1; */
43  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
44  alpha, A1, B, beta, C1,
45  FLA_Cntl_sub_gemm( cntl ) );
46 
47  /*------------------------------------------------------------*/
48 
49  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
50  FLA_LEFT );
51 
52  FLA_Cont_with_3x1_to_2x1( &CT, C0,
53  C1,
54  /* ** */ /* ** */
55  &CB, C2, FLA_TOP );
56 
57  }
58 
59  return FLA_SUCCESS;
60 }
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_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_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_Part_2x1(FLA_Obj A, FLA_Obj *A1, FLA_Obj *A2, dim_t mb, FLA_Side side)
Definition: FLA_View.c:76
FLA_Error FLA_Repart_1x2_to_1x3(FLA_Obj AL, FLA_Obj AR, FLA_Obj *A0, FLA_Obj *A1, FLA_Obj *A2, dim_t nb, FLA_Side side)
Definition: FLA_View.c:267
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_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_blk_var2()

FLA_Error FLA_Gemm_tn_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)
16 {
17  FLA_Obj AL, AR, A0, A1, A2;
18 
19  FLA_Obj CT, C0,
20  CB, C1,
21  C2;
22 
23  dim_t b;
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x1( C, &CT,
28  &CB, 0, FLA_BOTTOM );
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_2x1_to_3x1( CT, &C0,
38  &C1,
39  /* ** */ /* ** */
40  CB, &C2, b, FLA_TOP );
41 
42  /*------------------------------------------------------------*/
43 
44  /* C1 = alpha * A1' * B + C1; */
45  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
46  alpha, A1, B, beta, C1,
47  FLA_Cntl_sub_gemm( cntl ) );
48 
49  /*------------------------------------------------------------*/
50 
51  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
52  FLA_RIGHT );
53 
54  FLA_Cont_with_3x1_to_2x1( &CT, C0,
55  /* ** */ /* ** */
56  C1,
57  &CB, C2, FLA_BOTTOM );
58 
59  }
60 
61  return FLA_SUCCESS;
62 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_blk_var3()

FLA_Error FLA_Gemm_tn_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)
14 {
15  FLA_Obj BL, BR, B0, B1, B2;
16 
17  FLA_Obj CL, CR, C0, C1, C2;
18 
19  dim_t b;
20 
21  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
22 
23  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
24 
25  while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
26 
27  b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
28 
29  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2,
30  b, FLA_RIGHT );
31 
32  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
33  b, FLA_RIGHT );
34 
35  /*------------------------------------------------------------*/
36 
37  /* C1 = alpha * A' * B1 + C1; */
38  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
39  alpha, A, B1, beta, C1,
40  FLA_Cntl_sub_gemm( cntl ) );
41 
42  /*------------------------------------------------------------*/
43 
44  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2,
45  FLA_LEFT );
46 
47  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
48  FLA_LEFT );
49 
50  }
51 
52  return FLA_SUCCESS;
53 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_blk_var4()

FLA_Error FLA_Gemm_tn_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_t cntl 
)
16 {
17  FLA_Obj BL, BR, B0, B1, B2;
18 
19  FLA_Obj CL, CR, C0, C1, C2;
20 
21  dim_t b;
22 
23  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
24 
25  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
26 
27  while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
28 
29  b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );
30 
31  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2,
32  b, FLA_LEFT );
33 
34  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2,
35  b, FLA_LEFT );
36 
37  /*------------------------------------------------------------*/
38 
39  /* C1 = alpha * A' * B1 + C1; */
40  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
41  alpha, A, B1, beta, C1,
42  FLA_Cntl_sub_gemm( cntl ) );
43 
44  /*------------------------------------------------------------*/
45 
46  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2,
47  FLA_RIGHT );
48 
49  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2,
50  FLA_RIGHT );
51 
52  }
53 
54  return FLA_SUCCESS;
55 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_blk_var5()

FLA_Error FLA_Gemm_tn_blk_var5 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_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_Scal_internal( beta, C,
26  FLA_Cntl_sub_scal( 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 = alpha * A1' * B1 + C; */
51  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
52  alpha, A1, B1, FLA_ONE, C,
53  FLA_Cntl_sub_gemm( 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 }
FLA_Error FLA_Scal_internal(FLA_Obj alpha, FLA_Obj A, fla_scal_t *cntl)
Definition: FLA_Scal_internal.c:16
FLA_Obj FLA_ONE
Definition: FLA_Init.c:18
dim_t FLA_Obj_length(FLA_Obj obj)
Definition: FLA_Query.c:116

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_blk_var6()

FLA_Error FLA_Gemm_tn_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_gemm_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_Scal_internal( beta, C,
28  FLA_Cntl_sub_scal( 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 = alpha * A1' * B1 + C; */
53  FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
54  alpha, A1, B1, FLA_ONE, C,
55  FLA_Cntl_sub_gemm( 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 }

References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var1()

FLA_Error FLA_Gemm_tn_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 CT, C0,
20  CB, c1t,
21  C2;
22 
23  FLA_Scal_external( beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
26 
27  FLA_Part_2x1( C, &CT,
28  &CB, 0, FLA_TOP );
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_2x1_to_3x1( CT, &C0,
36  /* ** */ /* *** */
37  &c1t,
38  CB, &C2, 1, FLA_BOTTOM );
39 
40  /*------------------------------------------------------------*/
41 
42  /* c1t = a1' * B + c1t */
43  /* c1t' = B' * a1 + c1t' */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B, a1, FLA_ONE, c1t );
45 
46  /*------------------------------------------------------------*/
47 
48  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2,
49  FLA_LEFT );
50 
51  FLA_Cont_with_3x1_to_2x1( &CT, C0,
52  c1t,
53  /* ** */ /* *** */
54  &CB, C2, FLA_TOP );
55 
56  }
57 
58  return FLA_SUCCESS;
59 }
FLA_Error FLA_Scal_external(FLA_Obj alpha, FLA_Obj A)
Definition: FLA_Scal_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_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var2()

FLA_Error FLA_Gemm_tn_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 CT, C0,
20  CB, c1t,
21  C2;
22 
23  FLA_Scal_external( beta, C );
24 
25  FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
26 
27  FLA_Part_2x1( C, &CT,
28  &CB, 0, FLA_BOTTOM );
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_2x1_to_3x1( CT, &C0,
36  &c1t,
37  /* ** */ /* *** */
38  CB, &C2, 1, FLA_TOP );
39 
40  /*------------------------------------------------------------*/
41 
42  /* c1t = a1' * B + c1t */
43  /* c1t' = B' * a1 + c1t' */
44  FLA_Gemv_external( FLA_TRANSPOSE, alpha, B, a1, FLA_ONE, c1t );
45 
46  /*------------------------------------------------------------*/
47 
48  FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2,
49  FLA_RIGHT );
50 
51  FLA_Cont_with_3x1_to_2x1( &CT, C0,
52  /* ** */ /* *** */
53  c1t,
54  &CB, C2, FLA_BOTTOM );
55 
56  }
57 
58  return FLA_SUCCESS;
59 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var3()

FLA_Error FLA_Gemm_tn_unb_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj BL, BR, B0, b1, B2;
18 
19  FLA_Obj CL, CR, C0, c1, C2;
20 
21  FLA_Scal_external( beta, C );
22 
23  FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT );
24 
25  FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
26 
27  while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){
28 
29  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2,
30  1, FLA_RIGHT );
31 
32  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1, &C2,
33  1, FLA_RIGHT );
34 
35  /*------------------------------------------------------------*/
36 
37  /* c1 = A' * b1 + c1 */
38  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A, b1, FLA_ONE, c1 );
39 
40  /*------------------------------------------------------------*/
41 
42  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2,
43  FLA_LEFT );
44 
45  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1, /**/ C2,
46  FLA_LEFT );
47 
48  }
49 
50  return FLA_SUCCESS;
51 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var4()

FLA_Error FLA_Gemm_tn_unb_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C 
)
16 {
17  FLA_Obj BL, BR, B0, b1, B2;
18 
19  FLA_Obj CL, CR, C0, c1, C2;
20 
21  FLA_Scal_external( beta, C );
22 
23  FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT );
24 
25  FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT );
26 
27  while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){
28 
29  FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2,
30  1, FLA_LEFT );
31 
32  FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1, /**/ &C2,
33  1, FLA_LEFT );
34 
35  /*------------------------------------------------------------*/
36 
37  /* c1 = A' * b1 + c1 */
38  FLA_Gemv_external( FLA_TRANSPOSE, alpha, A, b1, FLA_ONE, c1 );
39 
40  /*------------------------------------------------------------*/
41 
42  FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2,
43  FLA_RIGHT );
44 
45  FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1, C2,
46  FLA_RIGHT );
47 
48  }
49 
50  return FLA_SUCCESS;
51 }

References FLA_Cont_with_1x3_to_1x2(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var5()

FLA_Error FLA_Gemm_tn_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_Scal_external( 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  FLA_Repart_2x1_to_3x1( AT, &A0,
36  /* ** */ /* *** */
37  &a1t,
38  AB, &A2, 1, FLA_BOTTOM );
39 
40  FLA_Repart_2x1_to_3x1( BT, &B0,
41  /* ** */ /* *** */
42  &b1t,
43  BB, &B2, 1, FLA_BOTTOM );
44 
45  /*------------------------------------------------------------*/
46 
47  /* C = a1t' * b1t + C */
48  FLA_Ger_external( alpha, a1t, b1t, C );
49 
50  /*------------------------------------------------------------*/
51 
52  FLA_Cont_with_3x1_to_2x1( &AT, A0,
53  a1t,
54  /* ** */ /* *** */
55  &AB, A2, FLA_TOP );
56 
57  FLA_Cont_with_3x1_to_2x1( &BT, B0,
58  b1t,
59  /* ** */ /* *** */
60  &BB, B2, FLA_TOP );
61 
62  }
63 
64  return FLA_SUCCESS;
65 }
FLA_Error FLA_Ger_external(FLA_Obj alpha, FLA_Obj x, FLA_Obj y, FLA_Obj A)
Definition: FLA_Ger_external.c:13

References FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().

◆ FLA_Gemm_tn_unb_var6()

FLA_Error FLA_Gemm_tn_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_Scal_external( 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  FLA_Repart_2x1_to_3x1( AT, &A0,
36  &a1t,
37  /* ** */ /* *** */
38  AB, &A2, 1, FLA_TOP );
39 
40  FLA_Repart_2x1_to_3x1( BT, &B0,
41  &b1t,
42  /* ** */ /* *** */
43  BB, &B2, 1, FLA_TOP );
44 
45  /*------------------------------------------------------------*/
46 
47  /* C = a1t' * b1t + C */
48  FLA_Ger_external( alpha, a1t, b1t, C );
49 
50  /*------------------------------------------------------------*/
51 
52  FLA_Cont_with_3x1_to_2x1( &AT, A0,
53  /* ** */ /* *** */
54  a1t,
55  &AB, A2, FLA_BOTTOM );
56 
57  FLA_Cont_with_3x1_to_2x1( &BT, B0,
58  /* ** */ /* *** */
59  b1t,
60  &BB, B2, FLA_BOTTOM );
61 
62  }
63 
64  return FLA_SUCCESS;
65 }

References FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().

Referenced by FLA_Gemm_tn().