libflame revision_anchor
Functions
FLA_LU_nopiv_vars.h File Reference

(r)

Go to the source code of this file.

Functions

FLA_Error FLA_LU_nopiv_blk_var1 (FLA_Obj A, fla_lu_t *cntl)
 
FLA_Error FLA_LU_nopiv_blk_var2 (FLA_Obj A, fla_lu_t *cntl)
 
FLA_Error FLA_LU_nopiv_blk_var3 (FLA_Obj A, fla_lu_t *cntl)
 
FLA_Error FLA_LU_nopiv_blk_var4 (FLA_Obj A, fla_lu_t *cntl)
 
FLA_Error FLA_LU_nopiv_blk_var5 (FLA_Obj A, fla_lu_t *cntl)
 
FLA_Error FLA_LU_nopiv_unb_var1 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_unb_var2 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_unb_var3 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_unb_var4 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_unb_var5 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_opt_var1 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_ops_var1 (int m_A, int n_A, float *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opd_var1 (int m_A, int n_A, double *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opc_var1 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opz_var1 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opt_var2 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_ops_var2 (int m_A, int n_A, float *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opd_var2 (int m_A, int n_A, double *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opc_var2 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opz_var2 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opt_var3 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_ops_var3 (int m_A, int n_A, float *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opd_var3 (int m_A, int n_A, double *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opc_var3 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opz_var3 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opt_var4 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_ops_var4 (int m_A, int n_A, float *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opd_var4 (int m_A, int n_A, double *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opc_var4 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opz_var4 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opt_var5 (FLA_Obj A)
 
FLA_Error FLA_LU_nopiv_ops_var5 (int m_A, int n_A, float *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opd_var5 (int m_A, int n_A, double *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opc_var5 (int m_A, int n_A, scomplex *A, int rs_A, int cs_A)
 
FLA_Error FLA_LU_nopiv_opz_var5 (int m_A, int n_A, dcomplex *A, int rs_A, int cs_A)
 

Function Documentation

◆ FLA_LU_nopiv_blk_var1()

FLA_Error FLA_LU_nopiv_blk_var1 ( FLA_Obj A,
fla_lu_t * cntl )
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 dim_t b;
22
23
24 FLA_Part_2x2( A, &ATL, &ATR,
25 &ABL, &ABR, 0, 0, FLA_TL );
26
27 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
28 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
29
30 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
31
32 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
33 /* ************* */ /* ******************** */
34 &A10, /**/ &A11, &A12,
35 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
36 b, b, FLA_BR );
37
38 /*------------------------------------------------------------*/
39
40 // A01 = trilu( A00 ) \ A01
41 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
42 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
43 FLA_ONE, A00, A01,
44 FLA_Cntl_sub_trsm1( cntl ) );
45
46 // A10 = A10 / triu( A00 )
47 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
48 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
49 FLA_ONE, A00, A10,
50 FLA_Cntl_sub_trsm2( cntl ) );
51
52 // A11 = LU( A11 - A10 * A01 )
53 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
54 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
55 FLA_Cntl_sub_gemm1( cntl ) );
56
57 // A11 = LU_nopiv( A11 )
58 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
59
60 /*------------------------------------------------------------*/
61
62 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
63 A10, A11, /**/ A12,
64 /* ************** */ /* ****************** */
65 &ABL, /**/ &ABR, A20, A21, /**/ A22,
66 FLA_TL );
67 }
68
69 if ( FLA_Obj_length( ABL ) > 0 )
70 // ABL = ABL / triu( ATL )
71 FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
72 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
73 FLA_ONE, ATL, ABL );
74 else if ( FLA_Obj_width( ATR ) > 0 )
75 // ATR = trilu( ATL ) \ ATR
76 FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
77 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
78 FLA_ONE, ATL, ATR );
79
80 return FLA_SUCCESS;
81}
FLA_Gemm_internal
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_LU_nopiv_internal
FLA_Error FLA_LU_nopiv_internal(FLA_Obj A, fla_lu_t *cntl)
Definition FLA_LU_nopiv_internal.c:16
FLA_Trsm_internal
FLA_Error FLA_Trsm_internal(FLA_Side side, FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trsm_t *cntl)
Definition FLA_Trsm_internal.c:16
FLA_Trsm_external
FLA_Error FLA_Trsm_external(FLA_Side side, FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, FLA_Obj B)
Definition FLA_Trsm_external.c:13
FLA_MINUS_ONE
FLA_Obj FLA_MINUS_ONE
Definition FLA_Init.c:22
FLA_ONE
FLA_Obj FLA_ONE
Definition FLA_Init.c:18
FLA_Cont_with_3x3_to_2x2
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_Part_2x2
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_Obj_width
dim_t FLA_Obj_width(FLA_Obj obj)
Definition FLA_Query.c:123
FLA_Obj_length
dim_t FLA_Obj_length(FLA_Obj obj)
Definition FLA_Query.c:116
FLA_Repart_2x2_to_3x3
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_Determine_blocksize
dim_t FLA_Determine_blocksize(FLA_Obj A_unproc, FLA_Quadrant to_dir, fla_blocksize_t *cntl_blocksizes)
Definition FLA_Blocksize.c:234
dim_t
unsigned long dim_t
Definition FLA_type_defs.h:71
FLA_Obj_view
Definition FLA_type_defs.h:159

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_blk_var2()

FLA_Error FLA_LU_nopiv_blk_var2 ( FLA_Obj A,
fla_lu_t * cntl )
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 dim_t b;
22
23
24 FLA_Part_2x2( A, &ATL, &ATR,
25 &ABL, &ABR, 0, 0, FLA_TL );
26
27 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
28 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
29
30 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
31
32 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
33 /* ************* */ /* ******************** */
34 &A10, /**/ &A11, &A12,
35 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
36 b, b, FLA_BR );
37
38 /*------------------------------------------------------------*/
39
40 // A10 = A10 / triu( A00 )
41 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
42 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
43 FLA_ONE, A00, A10,
44 FLA_Cntl_sub_trsm1( cntl ) );
45
46 // A11 = LU( A11 - A10 * A01 )
47 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
48 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
49 FLA_Cntl_sub_gemm1( cntl ) );
50
51 // A11 = LU_nopiv( A11 )
52 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
53
54 // A12 = A12 - A10 * A02
55 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
56 FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
57 FLA_Cntl_sub_gemm2( cntl ) );
58
59 // A12 = trilu( A11 ) \ A12
60 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
61 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
62 FLA_ONE, A11, A12,
63 FLA_Cntl_sub_trsm2( cntl ) );
64
65 /*------------------------------------------------------------*/
66
67 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
68 A10, A11, /**/ A12,
69 /* ************** */ /* ****************** */
70 &ABL, /**/ &ABR, A20, A21, /**/ A22,
71 FLA_TL );
72 }
73
74 if ( FLA_Obj_length( ABL ) > 0 )
75 // ABL = ABL / triu( ATL )
76 FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
77 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
78 FLA_ONE, ATL, ABL );
79
80 return FLA_SUCCESS;
81}

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_blk_var3()

FLA_Error FLA_LU_nopiv_blk_var3 ( FLA_Obj A,
fla_lu_t * cntl )
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 dim_t b;
22
23
24 FLA_Part_2x2( A, &ATL, &ATR,
25 &ABL, &ABR, 0, 0, FLA_TL );
26
27 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
28 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
29
30 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
31
32 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
33 /* ************* */ /* ******************** */
34 &A10, /**/ &A11, &A12,
35 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
36 b, b, FLA_BR );
37
38 /*------------------------------------------------------------*/
39
40 // A01 = trilu( A00 ) \ A10
41 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
42 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
43 FLA_ONE, A00, A01,
44 FLA_Cntl_sub_trsm1( cntl ) );
45
46 // A11 = LU( A11 - A10 * A01 )
47 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
48 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
49 FLA_Cntl_sub_gemm1( cntl ) );
50
51 // A11 = LU_nopiv( A11 )
52 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
53
54 // A21 = A21 - A20 * A01
55 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
56 FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
57 FLA_Cntl_sub_gemm2( cntl ) );
58
59 // A21 = A21 / triu( A11 )
60 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
61 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
62 FLA_ONE, A11, A21,
63 FLA_Cntl_sub_trsm2( cntl ) );
64
65 /*------------------------------------------------------------*/
66
67 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
68 A10, A11, /**/ A12,
69 /* ************** */ /* ****************** */
70 &ABL, /**/ &ABR, A20, A21, /**/ A22,
71 FLA_TL );
72 }
73
74 if ( FLA_Obj_width( ATR ) > 0 )
75 /* ATR = trilu( ATL ) \ ATR */
76 FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
77 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
78 FLA_ONE, ATL, ATR );
79
80 return FLA_SUCCESS;
81}

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_blk_var4()

FLA_Error FLA_LU_nopiv_blk_var4 ( FLA_Obj A,
fla_lu_t * cntl )
16{
17 FLA_Obj ATL, ATR, A00, A01, A02,
18 ABL, ABR, A10, A11, A12,
19 A20, A21, A22;
20
21 dim_t b;
22
23
24 FLA_Part_2x2( A, &ATL, &ATR,
25 &ABL, &ABR, 0, 0, FLA_TL );
26
27 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
28 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
29
30 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
31
32 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
33 /* ************* */ /* ******************** */
34 &A10, /**/ &A11, &A12,
35 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
36 b, b, FLA_BR );
37
38 /*------------------------------------------------------------*/
39
40 // A11 = LU( A11 - A10 * A01 )
41 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
42 FLA_MINUS_ONE, A10, A01, FLA_ONE, A11,
43 FLA_Cntl_sub_gemm1( cntl ) );
44
45 // A11 = LU_nopiv( A11 )
46 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
47
48 // A12 = A12 - A10 * A02
49 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
50 FLA_MINUS_ONE, A10, A02, FLA_ONE, A12,
51 FLA_Cntl_sub_gemm2( cntl ) );
52
53 // A12 = trilu( A11 ) \ A12
54 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
55 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
56 FLA_ONE, A11, A12,
57 FLA_Cntl_sub_trsm1( cntl ) );
58
59 // A21 = A21 - A20 * A01
60 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
61 FLA_MINUS_ONE, A20, A01, FLA_ONE, A21,
62 FLA_Cntl_sub_gemm3( cntl ) );
63
64 // A21 = A21 / triu( A11 )
65 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
66 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
67 FLA_ONE, A11, A21,
68 FLA_Cntl_sub_trsm2( cntl ) );
69
70 /*------------------------------------------------------------*/
71
72 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
73 A10, A11, /**/ A12,
74 /* ************** */ /* ****************** */
75 &ABL, /**/ &ABR, A20, A21, /**/ A22,
76 FLA_TL );
77 }
78
79 return FLA_SUCCESS;
80}

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_blk_var5()

FLA_Error FLA_LU_nopiv_blk_var5 ( FLA_Obj A,
fla_lu_t * cntl )
14{
15 FLA_Obj ATL, ATR, A00, A01, A02,
16 ABL, ABR, A10, A11, A12,
17 A20, A21, A22;
18
19 dim_t b;
20
21
22 FLA_Part_2x2( A, &ATL, &ATR,
23 &ABL, &ABR, 0, 0, FLA_TL );
24
25 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
26 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
27
28 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) );
29
30 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02,
31 /* ************* */ /* ******************** */
32 &A10, /**/ &A11, &A12,
33 ABL, /**/ ABR, &A20, /**/ &A21, &A22,
34 b, b, FLA_BR );
35
36 /*------------------------------------------------------------*/
37
38 // A11 = LU_nopiv( A11 )
39 FLA_LU_nopiv_internal( A11, FLA_Cntl_sub_lu( cntl ) );
40
41 // A12 = trilu( A11 ) \ A12
42 FLA_Trsm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR,
43 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
44 FLA_ONE, A11, A12,
45 FLA_Cntl_sub_trsm1( cntl ) );
46
47 // A21 = A21 / triu( A11 )
48 FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
49 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
50 FLA_ONE, A11, A21,
51 FLA_Cntl_sub_trsm2( cntl ) );
52
53 // A22 = A22 - A21 * A12
54 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
55 FLA_MINUS_ONE, A21, A12, FLA_ONE, A22,
56 FLA_Cntl_sub_gemm1( cntl ) );
57
58 /*------------------------------------------------------------*/
59
60 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02,
61 A10, A11, /**/ A12,
62 /* ************** */ /* ****************** */
63 &ABL, /**/ &ABR, A20, A21, /**/ A22,
64 FLA_TL );
65 }
66
67 return FLA_SUCCESS;
68}

References FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_LU_nopiv_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), and FLA_Trsm_internal().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opc_var1()

FLA_Error FLA_LU_nopiv_opc_var1 ( int m_A,
int n_A,
scomplex * A,
int rs_A,
int cs_A )
266{
267 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
268 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
269 int min_m_n = min( m_A, n_A );
270 int i;
271
272 for ( i = 0; i < min_m_n; ++i )
273 {
274 scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
275 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
276 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
277 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
278
279 int mn_behind = i;
280
281 /*------------------------------------------------------------*/
282
283 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
284 bl1_ctrsv( BLIS1_LOWER_TRIANGULAR,
285 BLIS1_NO_TRANSPOSE,
286 BLIS1_UNIT_DIAG,
287 mn_behind,
288 A00, rs_A, cs_A,
289 a01, rs_A );
290
291 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
292 bl1_ctrsv( BLIS1_UPPER_TRIANGULAR,
293 BLIS1_TRANSPOSE,
294 BLIS1_NONUNIT_DIAG,
295 mn_behind,
296 A00, rs_A, cs_A,
297 a10t, cs_A );
298
299 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
300 bl1_cdots( BLIS1_NO_CONJUGATE,
301 mn_behind,
302 buff_m1,
303 a10t, cs_A,
304 a01, rs_A,
305 buff_1,
306 alpha11 );
307
308 /*------------------------------------------------------------*/
309
310 }
311
312 if ( m_A > n_A )
313 {
314 scomplex* ATL = buff_A;
315 scomplex* ABL = buff_A + n_A*rs_A;
316
317 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
318 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
319 // FLA_ONE, ATL, ABL );
320 bl1_ctrsm( BLIS1_RIGHT,
321 BLIS1_UPPER_TRIANGULAR,
322 BLIS1_NO_TRANSPOSE,
323 BLIS1_NONUNIT_DIAG,
324 m_A - n_A,
325 n_A,
326 buff_1,
327 ATL, rs_A, cs_A,
328 ABL, rs_A, cs_A );
329 }
330 else if ( m_A < n_A )
331 {
332 scomplex* ATL = buff_A;
333 scomplex* ATR = buff_A + m_A*cs_A;
334
335 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
336 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
337 // FLA_ONE, ATL, ATR );
338 bl1_ctrsm( BLIS1_LEFT,
339 BLIS1_LOWER_TRIANGULAR,
340 BLIS1_NO_TRANSPOSE,
341 BLIS1_UNIT_DIAG,
342 m_A,
343 n_A - m_A,
344 buff_1,
345 ATL, rs_A, cs_A,
346 ATR, rs_A, cs_A );
347 }
348
349 return FLA_SUCCESS;
350}
i
int i
Definition bl1_axmyv2.c:145
bl1_cdots
void bl1_cdots(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy, scomplex *beta, scomplex *rho)
Definition bl1_dots.c:39
bl1_ctrsm
void bl1_ctrsm(side1_t side, uplo1_t uplo, trans1_t trans, diag1_t diag, int m, int n, scomplex *alpha, scomplex *a, int a_rs, int a_cs, scomplex *b, int b_rs, int b_cs)
Definition bl1_trsm.c:219
bl1_ctrsv
void bl1_ctrsv(uplo1_t uplo, trans1_t trans, diag1_t diag, int m, scomplex *a, int a_rs, int a_cs, scomplex *x, int incx)
Definition bl1_trsv.c:99
BLIS1_LOWER_TRIANGULAR
@ BLIS1_LOWER_TRIANGULAR
Definition blis_type_defs.h:62
BLIS1_UPPER_TRIANGULAR
@ BLIS1_UPPER_TRIANGULAR
Definition blis_type_defs.h:63
BLIS1_UNIT_DIAG
@ BLIS1_UNIT_DIAG
Definition blis_type_defs.h:75
BLIS1_NONUNIT_DIAG
@ BLIS1_NONUNIT_DIAG
Definition blis_type_defs.h:74
BLIS1_NO_TRANSPOSE
@ BLIS1_NO_TRANSPOSE
Definition blis_type_defs.h:54
BLIS1_TRANSPOSE
@ BLIS1_TRANSPOSE
Definition blis_type_defs.h:55
BLIS1_NO_CONJUGATE
@ BLIS1_NO_CONJUGATE
Definition blis_type_defs.h:81
BLIS1_RIGHT
@ BLIS1_RIGHT
Definition blis_type_defs.h:69
BLIS1_LEFT
@ BLIS1_LEFT
Definition blis_type_defs.h:68
scomplex
Definition blis_type_defs.h:133

References bl1_cdots(), bl1_ctrsm(), bl1_ctrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var1().

◆ FLA_LU_nopiv_opc_var2()

FLA_Error FLA_LU_nopiv_opc_var2 ( int m_A,
int n_A,
scomplex * A,
int rs_A,
int cs_A )
242{
243 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
244 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
245 int min_m_n = min( m_A, n_A );
246 int i;
247
248 for ( i = 0; i < min_m_n; ++i )
249 {
250 scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
251 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
252 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
253 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
254 scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
255 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
256
257 int n_ahead = n_A - i - 1;
258 int mn_behind = i;
259
260 /*------------------------------------------------------------*/
261
262 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
263 bl1_ctrsv( BLIS1_UPPER_TRIANGULAR,
264 BLIS1_TRANSPOSE,
265 BLIS1_NONUNIT_DIAG,
266 mn_behind,
267 A00, rs_A, cs_A,
268 a10t, cs_A );
269
270 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
271 bl1_cdots( BLIS1_NO_CONJUGATE,
272 mn_behind,
273 buff_m1,
274 a10t, cs_A,
275 a01, rs_A,
276 buff_1,
277 alpha11 );
278
279 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
280 bl1_cgemv( BLIS1_TRANSPOSE,
281 BLIS1_NO_CONJUGATE,
282 mn_behind,
283 n_ahead,
284 buff_m1,
285 A02, rs_A, cs_A,
286 a10t, cs_A,
287 buff_1,
288 a12t, cs_A );
289
290 /*------------------------------------------------------------*/
291
292 }
293
294 if ( m_A > n_A )
295 {
296 scomplex* ATL = buff_A;
297 scomplex* ABL = buff_A + n_A*rs_A;
298
299 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
300 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
301 // FLA_ONE, ATL, ABL );
302 bl1_ctrsm( BLIS1_RIGHT,
303 BLIS1_UPPER_TRIANGULAR,
304 BLIS1_NO_TRANSPOSE,
305 BLIS1_NONUNIT_DIAG,
306 m_A - n_A,
307 n_A,
308 buff_1,
309 ATL, rs_A, cs_A,
310 ABL, rs_A, cs_A );
311 }
312
313 return FLA_SUCCESS;
314}
bl1_cgemv
void bl1_cgemv(trans1_t transa, conj1_t conjx, int m, int n, scomplex *alpha, scomplex *a, int a_rs, int a_cs, scomplex *x, int incx, scomplex *beta, scomplex *y, int incy)
Definition bl1_gemv.c:125

References bl1_cdots(), bl1_cgemv(), bl1_ctrsm(), bl1_ctrsv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var2().

◆ FLA_LU_nopiv_opc_var3()

FLA_Error FLA_LU_nopiv_opc_var3 ( int m_A,
int n_A,
scomplex * A,
int rs_A,
int cs_A )
254{
255 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
256 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
257 int min_m_n = min( m_A, n_A );
258 int i;
259
260 for ( i = 0; i < min_m_n; ++i )
261 {
262 scomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
263 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
264 scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
265 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
266 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
267 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
268
269 int m_ahead = m_A - i - 1;
270 int mn_behind = i;
271
272 /*------------------------------------------------------------*/
273
274 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
275 bl1_ctrsv( BLIS1_LOWER_TRIANGULAR,
276 BLIS1_NO_TRANSPOSE,
277 BLIS1_UNIT_DIAG,
278 mn_behind,
279 A00, rs_A, cs_A,
280 a01, rs_A );
281
282 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
283 bl1_cdots( BLIS1_NO_CONJUGATE,
284 mn_behind,
285 buff_m1,
286 a10t, cs_A,
287 a01, rs_A,
288 buff_1,
289 alpha11 );
290
291 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
292 bl1_cgemv( BLIS1_NO_TRANSPOSE,
293 BLIS1_NO_CONJUGATE,
294 m_ahead,
295 mn_behind,
296 buff_m1,
297 A20, rs_A, cs_A,
298 a01, rs_A,
299 buff_1,
300 a21, rs_A );
301
302 // FLA_Inv_scal_external( alpha11, a21 );
303 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
304 m_ahead,
305 alpha11,
306 a21, rs_A );
307
308 /*------------------------------------------------------------*/
309
310 }
311
312 if ( m_A < n_A )
313 {
314 scomplex* ATL = buff_A;
315 scomplex* ATR = buff_A + m_A*cs_A;
316
317 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
318 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
319 // FLA_ONE, ATL, ATR );
320 bl1_ctrsm( BLIS1_LEFT,
321 BLIS1_LOWER_TRIANGULAR,
322 BLIS1_NO_TRANSPOSE,
323 BLIS1_UNIT_DIAG,
324 m_A,
325 n_A - m_A,
326 buff_1,
327 ATL, rs_A, cs_A,
328 ATR, rs_A, cs_A );
329 }
330
331 return FLA_SUCCESS;
332}
bl1_cinvscalv
void bl1_cinvscalv(conj1_t conj, int n, scomplex *alpha, scomplex *x, int incx)
Definition bl1_invscalv.c:52

References bl1_cdots(), bl1_cgemv(), bl1_cinvscalv(), bl1_ctrsm(), bl1_ctrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_UNIT_DIAG, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var3().

◆ FLA_LU_nopiv_opc_var4()

FLA_Error FLA_LU_nopiv_opc_var4 ( int m_A,
int n_A,
scomplex * A,
int rs_A,
int cs_A )
226{
227 scomplex* buff_1 = FLA_COMPLEX_PTR( FLA_ONE );
228 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
229 int min_m_n = min( m_A, n_A );
230 int i;
231
232 for ( i = 0; i < min_m_n; ++i )
233 {
234 scomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
235 scomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
236 scomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
237 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
238 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
239 scomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
240 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
241
242 int m_ahead = m_A - i - 1;
243 int n_ahead = n_A - i - 1;
244 int mn_behind = i;
245
246 /*------------------------------------------------------------*/
247
248 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
249 bl1_cdots( BLIS1_NO_CONJUGATE,
250 mn_behind,
251 buff_m1,
252 a10t, cs_A,
253 a01, rs_A,
254 buff_1,
255 alpha11 );
256
257 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
258 bl1_cgemv( BLIS1_TRANSPOSE,
259 BLIS1_NO_CONJUGATE,
260 mn_behind,
261 n_ahead,
262 buff_m1,
263 A02, rs_A, cs_A,
264 a10t, cs_A,
265 buff_1,
266 a12t, cs_A );
267
268 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
269 bl1_cgemv( BLIS1_NO_TRANSPOSE,
270 BLIS1_NO_CONJUGATE,
271 m_ahead,
272 mn_behind,
273 buff_m1,
274 A20, rs_A, cs_A,
275 a01, rs_A,
276 buff_1,
277 a21, rs_A );
278
279 // FLA_Inv_scal_external( alpha11, a21 );
280 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
281 m_ahead,
282 alpha11,
283 a21, rs_A );
284
285 /*------------------------------------------------------------*/
286
287 }
288
289 return FLA_SUCCESS;
290}

References bl1_cdots(), bl1_cgemv(), bl1_cinvscalv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var4().

◆ FLA_LU_nopiv_opc_var5()

FLA_Error FLA_LU_nopiv_opc_var5 ( int m_A,
int n_A,
scomplex * A,
int rs_A,
int cs_A )
172{
173 scomplex* buff_m1 = FLA_COMPLEX_PTR( FLA_MINUS_ONE );
174 int min_m_n = min( m_A, n_A );
175 int i;
176
177 for ( i = 0; i < min_m_n; ++i )
178 {
179 scomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
180 scomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
181 scomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
182 scomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
183
184 int m_ahead = m_A - i - 1;
185 int n_ahead = n_A - i - 1;
186
187 /*------------------------------------------------------------*/
188
189 // FLA_Inv_scal_external( alpha11, a21 );
190 bl1_cinvscalv( BLIS1_NO_CONJUGATE,
191 m_ahead,
192 alpha11,
193 a21, rs_A );
194
195 // FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
196 bl1_cger( BLIS1_NO_CONJUGATE,
197 BLIS1_NO_CONJUGATE,
198 m_ahead,
199 n_ahead,
200 buff_m1,
201 a21, rs_A,
202 a12t, cs_A,
203 A22, rs_A, cs_A );
204
205 /*------------------------------------------------------------*/
206
207 }
208
209 return FLA_SUCCESS;
210}
bl1_cger
void bl1_cger(conj1_t conjx, conj1_t conjy, int m, int n, scomplex *alpha, scomplex *x, int incx, scomplex *y, int incy, scomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:111

References bl1_cger(), bl1_cinvscalv(), BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var5().

◆ FLA_LU_nopiv_opd_var1()

FLA_Error FLA_LU_nopiv_opd_var1 ( int m_A,
int n_A,
double * A,
int rs_A,
int cs_A )
175{
176 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
177 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
178 int min_m_n = min( m_A, n_A );
179 int i;
180
181 for ( i = 0; i < min_m_n; ++i )
182 {
183 double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
184 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
185 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
186 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
187
188 int mn_behind = i;
189
190 /*------------------------------------------------------------*/
191
192 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
193 bl1_dtrsv( BLIS1_LOWER_TRIANGULAR,
194 BLIS1_NO_TRANSPOSE,
195 BLIS1_UNIT_DIAG,
196 mn_behind,
197 A00, rs_A, cs_A,
198 a01, rs_A );
199
200 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
201 bl1_dtrsv( BLIS1_UPPER_TRIANGULAR,
202 BLIS1_TRANSPOSE,
203 BLIS1_NONUNIT_DIAG,
204 mn_behind,
205 A00, rs_A, cs_A,
206 a10t, cs_A );
207
208 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
209 bl1_ddots( BLIS1_NO_CONJUGATE,
210 mn_behind,
211 buff_m1,
212 a10t, cs_A,
213 a01, rs_A,
214 buff_1,
215 alpha11 );
216
217 /*------------------------------------------------------------*/
218
219 }
220
221 if ( m_A > n_A )
222 {
223 double* ATL = buff_A;
224 double* ABL = buff_A + n_A*rs_A;
225
226 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
227 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
228 // FLA_ONE, ATL, ABL );
229 bl1_dtrsm( BLIS1_RIGHT,
230 BLIS1_UPPER_TRIANGULAR,
231 BLIS1_NO_TRANSPOSE,
232 BLIS1_NONUNIT_DIAG,
233 m_A - n_A,
234 n_A,
235 buff_1,
236 ATL, rs_A, cs_A,
237 ABL, rs_A, cs_A );
238 }
239 else if ( m_A < n_A )
240 {
241 double* ATL = buff_A;
242 double* ATR = buff_A + m_A*cs_A;
243
244 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
245 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
246 // FLA_ONE, ATL, ATR );
247 bl1_dtrsm( BLIS1_LEFT,
248 BLIS1_LOWER_TRIANGULAR,
249 BLIS1_NO_TRANSPOSE,
250 BLIS1_UNIT_DIAG,
251 m_A,
252 n_A - m_A,
253 buff_1,
254 ATL, rs_A, cs_A,
255 ATR, rs_A, cs_A );
256 }
257
258 return FLA_SUCCESS;
259}
bl1_ddots
void bl1_ddots(conj1_t conj, int n, double *alpha, double *x, int incx, double *y, int incy, double *beta, double *rho)
Definition bl1_dots.c:26
bl1_dtrsm
void bl1_dtrsm(side1_t side, uplo1_t uplo, trans1_t trans, diag1_t diag, int m, int n, double *alpha, double *a, int a_rs, int a_cs, double *b, int b_rs, int b_cs)
Definition bl1_trsm.c:116
bl1_dtrsv
void bl1_dtrsv(uplo1_t uplo, trans1_t trans, diag1_t diag, int m, double *a, int a_rs, int a_cs, double *x, int incx)
Definition bl1_trsv.c:56

References bl1_ddots(), bl1_dtrsm(), bl1_dtrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var1().

◆ FLA_LU_nopiv_opd_var2()

FLA_Error FLA_LU_nopiv_opd_var2 ( int m_A,
int n_A,
double * A,
int rs_A,
int cs_A )
163{
164 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
165 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
166 int min_m_n = min( m_A, n_A );
167 int i;
168
169 for ( i = 0; i < min_m_n; ++i )
170 {
171 double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
172 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
173 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
174 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
175 double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
176 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
177
178 int n_ahead = n_A - i - 1;
179 int mn_behind = i;
180
181 /*------------------------------------------------------------*/
182
183 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
184 bl1_dtrsv( BLIS1_UPPER_TRIANGULAR,
185 BLIS1_TRANSPOSE,
186 BLIS1_NONUNIT_DIAG,
187 mn_behind,
188 A00, rs_A, cs_A,
189 a10t, cs_A );
190
191 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
192 bl1_ddots( BLIS1_NO_CONJUGATE,
193 mn_behind,
194 buff_m1,
195 a10t, cs_A,
196 a01, rs_A,
197 buff_1,
198 alpha11 );
199
200 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
201 bl1_dgemv( BLIS1_TRANSPOSE,
202 BLIS1_NO_CONJUGATE,
203 mn_behind,
204 n_ahead,
205 buff_m1,
206 A02, rs_A, cs_A,
207 a10t, cs_A,
208 buff_1,
209 a12t, cs_A );
210
211 /*------------------------------------------------------------*/
212
213 }
214
215 if ( m_A > n_A )
216 {
217 double* ATL = buff_A;
218 double* ABL = buff_A + n_A*rs_A;
219
220 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
221 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
222 // FLA_ONE, ATL, ABL );
223 bl1_dtrsm( BLIS1_RIGHT,
224 BLIS1_UPPER_TRIANGULAR,
225 BLIS1_NO_TRANSPOSE,
226 BLIS1_NONUNIT_DIAG,
227 m_A - n_A,
228 n_A,
229 buff_1,
230 ATL, rs_A, cs_A,
231 ABL, rs_A, cs_A );
232 }
233
234 return FLA_SUCCESS;
235}
bl1_dgemv
void bl1_dgemv(trans1_t transa, conj1_t conjx, int m, int n, double *alpha, double *a, int a_rs, int a_cs, double *x, int incx, double *beta, double *y, int incy)
Definition bl1_gemv.c:69

References bl1_ddots(), bl1_dgemv(), bl1_dtrsm(), bl1_dtrsv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var2().

◆ FLA_LU_nopiv_opd_var3()

FLA_Error FLA_LU_nopiv_opd_var3 ( int m_A,
int n_A,
double * A,
int rs_A,
int cs_A )
169{
170 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
171 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
172 int min_m_n = min( m_A, n_A );
173 int i;
174
175 for ( i = 0; i < min_m_n; ++i )
176 {
177 double* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
178 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
179 double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
180 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
181 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
182 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
183
184 int m_ahead = m_A - i - 1;
185 int mn_behind = i;
186
187 /*------------------------------------------------------------*/
188
189 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
190 bl1_dtrsv( BLIS1_LOWER_TRIANGULAR,
191 BLIS1_NO_TRANSPOSE,
192 BLIS1_UNIT_DIAG,
193 mn_behind,
194 A00, rs_A, cs_A,
195 a01, rs_A );
196
197 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
198 bl1_ddots( BLIS1_NO_CONJUGATE,
199 mn_behind,
200 buff_m1,
201 a10t, cs_A,
202 a01, rs_A,
203 buff_1,
204 alpha11 );
205
206 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
207 bl1_dgemv( BLIS1_NO_TRANSPOSE,
208 BLIS1_NO_CONJUGATE,
209 m_ahead,
210 mn_behind,
211 buff_m1,
212 A20, rs_A, cs_A,
213 a01, rs_A,
214 buff_1,
215 a21, rs_A );
216
217 // FLA_Inv_scal_external( alpha11, a21 );
218 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
219 m_ahead,
220 alpha11,
221 a21, rs_A );
222
223 /*------------------------------------------------------------*/
224
225 }
226
227 if ( m_A < n_A )
228 {
229 double* ATL = buff_A;
230 double* ATR = buff_A + m_A*cs_A;
231
232 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
233 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
234 // FLA_ONE, ATL, ATR );
235 bl1_dtrsm( BLIS1_LEFT,
236 BLIS1_LOWER_TRIANGULAR,
237 BLIS1_NO_TRANSPOSE,
238 BLIS1_UNIT_DIAG,
239 m_A,
240 n_A - m_A,
241 buff_1,
242 ATL, rs_A, cs_A,
243 ATR, rs_A, cs_A );
244 }
245
246 return FLA_SUCCESS;
247}
bl1_dinvscalv
void bl1_dinvscalv(conj1_t conj, int n, double *alpha, double *x, int incx)
Definition bl1_invscalv.c:26

References bl1_ddots(), bl1_dgemv(), bl1_dinvscalv(), bl1_dtrsm(), bl1_dtrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_UNIT_DIAG, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var3().

◆ FLA_LU_nopiv_opd_var4()

FLA_Error FLA_LU_nopiv_opd_var4 ( int m_A,
int n_A,
double * A,
int rs_A,
int cs_A )
155{
156 double* buff_1 = FLA_DOUBLE_PTR( FLA_ONE );
157 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
158 int min_m_n = min( m_A, n_A );
159 int i;
160
161 for ( i = 0; i < min_m_n; ++i )
162 {
163 double* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
164 double* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
165 double* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
166 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
167 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
168 double* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
169 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
170
171 int m_ahead = m_A - i - 1;
172 int n_ahead = n_A - i - 1;
173 int mn_behind = i;
174
175 /*------------------------------------------------------------*/
176
177 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
178 bl1_ddots( BLIS1_NO_CONJUGATE,
179 mn_behind,
180 buff_m1,
181 a10t, cs_A,
182 a01, rs_A,
183 buff_1,
184 alpha11 );
185
186 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
187 bl1_dgemv( BLIS1_TRANSPOSE,
188 BLIS1_NO_CONJUGATE,
189 mn_behind,
190 n_ahead,
191 buff_m1,
192 A02, rs_A, cs_A,
193 a10t, cs_A,
194 buff_1,
195 a12t, cs_A );
196
197 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
198 bl1_dgemv( BLIS1_NO_TRANSPOSE,
199 BLIS1_NO_CONJUGATE,
200 m_ahead,
201 mn_behind,
202 buff_m1,
203 A20, rs_A, cs_A,
204 a01, rs_A,
205 buff_1,
206 a21, rs_A );
207
208 // FLA_Inv_scal_external( alpha11, a21 );
209 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
210 m_ahead,
211 alpha11,
212 a21, rs_A );
213
214 /*------------------------------------------------------------*/
215
216 }
217
218 return FLA_SUCCESS;
219}

References bl1_ddots(), bl1_dgemv(), bl1_dinvscalv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var4().

◆ FLA_LU_nopiv_opd_var5()

FLA_Error FLA_LU_nopiv_opd_var5 ( int m_A,
int n_A,
double * A,
int rs_A,
int cs_A )
127{
128 double* buff_m1 = FLA_DOUBLE_PTR( FLA_MINUS_ONE );
129 int min_m_n = min( m_A, n_A );
130 int i;
131
132 for ( i = 0; i < min_m_n; ++i )
133 {
134 double* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
135 double* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
136 double* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
137 double* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
138
139 int m_ahead = m_A - i - 1;
140 int n_ahead = n_A - i - 1;
141
142 /*------------------------------------------------------------*/
143
144 // FLA_Inv_scal_external( alpha11, a21 );
145 bl1_dinvscalv( BLIS1_NO_CONJUGATE,
146 m_ahead,
147 alpha11,
148 a21, rs_A );
149
150 // FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
151 bl1_dger( BLIS1_NO_CONJUGATE,
152 BLIS1_NO_CONJUGATE,
153 m_ahead,
154 n_ahead,
155 buff_m1,
156 a21, rs_A,
157 a12t, cs_A,
158 A22, rs_A, cs_A );
159
160 /*------------------------------------------------------------*/
161
162 }
163
164 return FLA_SUCCESS;
165}
bl1_dger
void bl1_dger(conj1_t conjx, conj1_t conjy, int m, int n, double *alpha, double *x, int incx, double *y, int incy, double *a, int a_rs, int a_cs)
Definition bl1_ger.c:62

References bl1_dger(), bl1_dinvscalv(), BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var5().

◆ FLA_LU_nopiv_ops_var1()

FLA_Error FLA_LU_nopiv_ops_var1 ( int m_A,
int n_A,
float * A,
int rs_A,
int cs_A )
84{
85 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
86 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
87 int min_m_n = min( m_A, n_A );
88 int i;
89
90 for ( i = 0; i < min_m_n; ++i )
91 {
92 float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
93 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
94 float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
95 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
96
97 int mn_behind = i;
98
99 /*------------------------------------------------------------*/
100
101 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
102 bl1_strsv( BLIS1_LOWER_TRIANGULAR,
103 BLIS1_NO_TRANSPOSE,
104 BLIS1_UNIT_DIAG,
105 mn_behind,
106 A00, rs_A, cs_A,
107 a01, rs_A );
108
109 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
110 bl1_strsv( BLIS1_UPPER_TRIANGULAR,
111 BLIS1_TRANSPOSE,
112 BLIS1_NONUNIT_DIAG,
113 mn_behind,
114 A00, rs_A, cs_A,
115 a10t, cs_A );
116
117 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
118 bl1_sdots( BLIS1_NO_CONJUGATE,
119 mn_behind,
120 buff_m1,
121 a10t, cs_A,
122 a01, rs_A,
123 buff_1,
124 alpha11 );
125
126 /*------------------------------------------------------------*/
127
128 }
129
130 if ( m_A > n_A )
131 {
132 float* ATL = buff_A;
133 float* ABL = buff_A + n_A*rs_A;
134
135 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
136 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
137 // FLA_ONE, ATL, ABL );
138 bl1_strsm( BLIS1_RIGHT,
139 BLIS1_UPPER_TRIANGULAR,
140 BLIS1_NO_TRANSPOSE,
141 BLIS1_NONUNIT_DIAG,
142 m_A - n_A,
143 n_A,
144 buff_1,
145 ATL, rs_A, cs_A,
146 ABL, rs_A, cs_A );
147 }
148 else if ( m_A < n_A )
149 {
150 float* ATL = buff_A;
151 float* ATR = buff_A + m_A*cs_A;
152
153 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
154 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
155 // FLA_ONE, ATL, ATR );
156 bl1_strsm( BLIS1_LEFT,
157 BLIS1_LOWER_TRIANGULAR,
158 BLIS1_NO_TRANSPOSE,
159 BLIS1_UNIT_DIAG,
160 m_A,
161 n_A - m_A,
162 buff_1,
163 ATL, rs_A, cs_A,
164 ATR, rs_A, cs_A );
165 }
166
167 return FLA_SUCCESS;
168}
bl1_sdots
void bl1_sdots(conj1_t conj, int n, float *alpha, float *x, int incx, float *y, int incy, float *beta, float *rho)
Definition bl1_dots.c:13
bl1_strsm
void bl1_strsm(side1_t side, uplo1_t uplo, trans1_t trans, diag1_t diag, int m, int n, float *alpha, float *a, int a_rs, int a_cs, float *b, int b_rs, int b_cs)
Definition bl1_trsm.c:13
bl1_strsv
void bl1_strsv(uplo1_t uplo, trans1_t trans, diag1_t diag, int m, float *a, int a_rs, int a_cs, float *x, int incx)
Definition bl1_trsv.c:13

References bl1_sdots(), bl1_strsm(), bl1_strsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var1().

◆ FLA_LU_nopiv_ops_var2()

FLA_Error FLA_LU_nopiv_ops_var2 ( int m_A,
int n_A,
float * A,
int rs_A,
int cs_A )
84{
85 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
86 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
87 int min_m_n = min( m_A, n_A );
88 int i;
89
90 for ( i = 0; i < min_m_n; ++i )
91 {
92 float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
93 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
94 float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
95 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
96 float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
97 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
98
99 int n_ahead = n_A - i - 1;
100 int mn_behind = i;
101
102 /*------------------------------------------------------------*/
103
104 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
105 bl1_strsv( BLIS1_UPPER_TRIANGULAR,
106 BLIS1_TRANSPOSE,
107 BLIS1_NONUNIT_DIAG,
108 mn_behind,
109 A00, rs_A, cs_A,
110 a10t, cs_A );
111
112 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
113 bl1_sdots( BLIS1_NO_CONJUGATE,
114 mn_behind,
115 buff_m1,
116 a10t, cs_A,
117 a01, rs_A,
118 buff_1,
119 alpha11 );
120
121 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
122 bl1_sgemv( BLIS1_TRANSPOSE,
123 BLIS1_NO_CONJUGATE,
124 mn_behind,
125 n_ahead,
126 buff_m1,
127 A02, rs_A, cs_A,
128 a10t, cs_A,
129 buff_1,
130 a12t, cs_A );
131
132 /*------------------------------------------------------------*/
133
134 }
135
136 if ( m_A > n_A )
137 {
138 float* ATL = buff_A;
139 float* ABL = buff_A + n_A*rs_A;
140
141 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
142 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
143 // FLA_ONE, ATL, ABL );
144 bl1_strsm( BLIS1_RIGHT,
145 BLIS1_UPPER_TRIANGULAR,
146 BLIS1_NO_TRANSPOSE,
147 BLIS1_NONUNIT_DIAG,
148 m_A - n_A,
149 n_A,
150 buff_1,
151 ATL, rs_A, cs_A,
152 ABL, rs_A, cs_A );
153 }
154
155 return FLA_SUCCESS;
156}
bl1_sgemv
void bl1_sgemv(trans1_t transa, conj1_t conjx, int m, int n, float *alpha, float *a, int a_rs, int a_cs, float *x, int incx, float *beta, float *y, int incy)
Definition bl1_gemv.c:13

References bl1_sdots(), bl1_sgemv(), bl1_strsm(), bl1_strsv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var2().

◆ FLA_LU_nopiv_ops_var3()

FLA_Error FLA_LU_nopiv_ops_var3 ( int m_A,
int n_A,
float * A,
int rs_A,
int cs_A )
84{
85 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
86 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
87 int min_m_n = min( m_A, n_A );
88 int i;
89
90 for ( i = 0; i < min_m_n; ++i )
91 {
92 float* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
93 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
94 float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
95 float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
96 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
97 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
98
99 int m_ahead = m_A - i - 1;
100 int mn_behind = i;
101
102 /*------------------------------------------------------------*/
103
104 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
105 bl1_strsv( BLIS1_LOWER_TRIANGULAR,
106 BLIS1_NO_TRANSPOSE,
107 BLIS1_UNIT_DIAG,
108 mn_behind,
109 A00, rs_A, cs_A,
110 a01, rs_A );
111
112 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
113 bl1_sdots( BLIS1_NO_CONJUGATE,
114 mn_behind,
115 buff_m1,
116 a10t, cs_A,
117 a01, rs_A,
118 buff_1,
119 alpha11 );
120
121 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
122 bl1_sgemv( BLIS1_NO_TRANSPOSE,
123 BLIS1_NO_CONJUGATE,
124 m_ahead,
125 mn_behind,
126 buff_m1,
127 A20, rs_A, cs_A,
128 a01, rs_A,
129 buff_1,
130 a21, rs_A );
131
132 // FLA_Inv_scal_external( alpha11, a21 );
133 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
134 m_ahead,
135 alpha11,
136 a21, rs_A );
137
138 /*------------------------------------------------------------*/
139
140 }
141
142 if ( m_A < n_A )
143 {
144 float* ATL = buff_A;
145 float* ATR = buff_A + m_A*cs_A;
146
147 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
148 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
149 // FLA_ONE, ATL, ATR );
150 bl1_strsm( BLIS1_LEFT,
151 BLIS1_LOWER_TRIANGULAR,
152 BLIS1_NO_TRANSPOSE,
153 BLIS1_UNIT_DIAG,
154 m_A,
155 n_A - m_A,
156 buff_1,
157 ATL, rs_A, cs_A,
158 ATR, rs_A, cs_A );
159 }
160
161 return FLA_SUCCESS;
162}
bl1_sinvscalv
void bl1_sinvscalv(conj1_t conj, int n, float *alpha, float *x, int incx)
Definition bl1_invscalv.c:13

References bl1_sdots(), bl1_sgemv(), bl1_sinvscalv(), bl1_strsm(), bl1_strsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_UNIT_DIAG, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var3().

◆ FLA_LU_nopiv_ops_var4()

FLA_Error FLA_LU_nopiv_ops_var4 ( int m_A,
int n_A,
float * A,
int rs_A,
int cs_A )
84{
85 float* buff_1 = FLA_FLOAT_PTR( FLA_ONE );
86 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
87 int min_m_n = min( m_A, n_A );
88 int i;
89
90 for ( i = 0; i < min_m_n; ++i )
91 {
92 float* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
93 float* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
94 float* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
95 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
96 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
97 float* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
98 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
99
100 int m_ahead = m_A - i - 1;
101 int n_ahead = n_A - i - 1;
102 int mn_behind = i;
103
104 /*------------------------------------------------------------*/
105
106 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
107 bl1_sdots( BLIS1_NO_CONJUGATE,
108 mn_behind,
109 buff_m1,
110 a10t, cs_A,
111 a01, rs_A,
112 buff_1,
113 alpha11 );
114
115 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
116 bl1_sgemv( BLIS1_TRANSPOSE,
117 BLIS1_NO_CONJUGATE,
118 mn_behind,
119 n_ahead,
120 buff_m1,
121 A02, rs_A, cs_A,
122 a10t, cs_A,
123 buff_1,
124 a12t, cs_A );
125
126 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
127 bl1_sgemv( BLIS1_NO_TRANSPOSE,
128 BLIS1_NO_CONJUGATE,
129 m_ahead,
130 mn_behind,
131 buff_m1,
132 A20, rs_A, cs_A,
133 a01, rs_A,
134 buff_1,
135 a21, rs_A );
136
137 // FLA_Inv_scal_external( alpha11, a21 );
138 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
139 m_ahead,
140 alpha11,
141 a21, rs_A );
142
143 /*------------------------------------------------------------*/
144
145 }
146
147 return FLA_SUCCESS;
148}

References bl1_sdots(), bl1_sgemv(), bl1_sinvscalv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var4().

◆ FLA_LU_nopiv_ops_var5()

FLA_Error FLA_LU_nopiv_ops_var5 ( int m_A,
int n_A,
float * A,
int rs_A,
int cs_A )
82{
83 float* buff_m1 = FLA_FLOAT_PTR( FLA_MINUS_ONE );
84 int min_m_n = min( m_A, n_A );
85 int i;
86
87 for ( i = 0; i < min_m_n; ++i )
88 {
89 float* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
90 float* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
91 float* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
92 float* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
93
94 int m_ahead = m_A - i - 1;
95 int n_ahead = n_A - i - 1;
96
97 /*------------------------------------------------------------*/
98
99 // FLA_Inv_scal_external( alpha11, a21 );
100 bl1_sinvscalv( BLIS1_NO_CONJUGATE,
101 m_ahead,
102 alpha11,
103 a21, rs_A );
104
105 // FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
106 bl1_sger( BLIS1_NO_CONJUGATE,
107 BLIS1_NO_CONJUGATE,
108 m_ahead,
109 n_ahead,
110 buff_m1,
111 a21, rs_A,
112 a12t, cs_A,
113 A22, rs_A, cs_A );
114
115 /*------------------------------------------------------------*/
116
117 }
118
119 return FLA_SUCCESS;
120}
bl1_sger
void bl1_sger(conj1_t conjx, conj1_t conjy, int m, int n, float *alpha, float *x, int incx, float *y, int incy, float *a, int a_rs, int a_cs)
Definition bl1_ger.c:13

References bl1_sger(), bl1_sinvscalv(), BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var5().

◆ FLA_LU_nopiv_opt_var1()

FLA_Error FLA_LU_nopiv_opt_var1 ( FLA_Obj A)
16{
17 FLA_Datatype datatype;
18 int m_A, n_A;
19 int rs_A, cs_A;
20
21 datatype = FLA_Obj_datatype( A );
22
23 m_A = FLA_Obj_length( A );
24 n_A = FLA_Obj_width( A );
25 rs_A = FLA_Obj_row_stride( A );
26 cs_A = FLA_Obj_col_stride( A );
27
28
29 switch ( datatype )
30 {
31 case FLA_FLOAT:
32 {
33 float* buff_A = FLA_FLOAT_PTR( A );
34
35 FLA_LU_nopiv_ops_var1( m_A,
36 n_A,
37 buff_A, rs_A, cs_A );
38
39 break;
40 }
41
42 case FLA_DOUBLE:
43 {
44 double* buff_A = FLA_DOUBLE_PTR( A );
45
46 FLA_LU_nopiv_opd_var1( m_A,
47 n_A,
48 buff_A, rs_A, cs_A );
49
50 break;
51 }
52
53 case FLA_COMPLEX:
54 {
55 scomplex* buff_A = FLA_COMPLEX_PTR( A );
56
57 FLA_LU_nopiv_opc_var1( m_A,
58 n_A,
59 buff_A, rs_A, cs_A );
60
61 break;
62 }
63
64 case FLA_DOUBLE_COMPLEX:
65 {
66 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
67
68 FLA_LU_nopiv_opz_var1( m_A,
69 n_A,
70 buff_A, rs_A, cs_A );
71
72 break;
73 }
74 }
75
76 return FLA_SUCCESS;
77}
FLA_LU_nopiv_opc_var1
FLA_Error FLA_LU_nopiv_opc_var1(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var1.c:263
FLA_LU_nopiv_opz_var1
FLA_Error FLA_LU_nopiv_opz_var1(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var1.c:354
FLA_LU_nopiv_opd_var1
FLA_Error FLA_LU_nopiv_opd_var1(int m_A, int n_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var1.c:172
FLA_LU_nopiv_ops_var1
FLA_Error FLA_LU_nopiv_ops_var1(int m_A, int n_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var1.c:81
FLA_Obj_row_stride
dim_t FLA_Obj_row_stride(FLA_Obj obj)
Definition FLA_Query.c:167
FLA_Obj_col_stride
dim_t FLA_Obj_col_stride(FLA_Obj obj)
Definition FLA_Query.c:174
FLA_Obj_datatype
FLA_Datatype FLA_Obj_datatype(FLA_Obj obj)
Definition FLA_Query.c:13
FLA_Datatype
int FLA_Datatype
Definition FLA_type_defs.h:49
dcomplex
Definition blis_type_defs.h:138

References FLA_LU_nopiv_opc_var1(), FLA_LU_nopiv_opd_var1(), FLA_LU_nopiv_ops_var1(), FLA_LU_nopiv_opz_var1(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opt_var2()

FLA_Error FLA_LU_nopiv_opt_var2 ( FLA_Obj A)
16{
17 FLA_Datatype datatype;
18 int m_A, n_A;
19 int rs_A, cs_A;
20
21 datatype = FLA_Obj_datatype( A );
22
23 m_A = FLA_Obj_length( A );
24 n_A = FLA_Obj_width( A );
25 rs_A = FLA_Obj_row_stride( A );
26 cs_A = FLA_Obj_col_stride( A );
27
28
29 switch ( datatype )
30 {
31 case FLA_FLOAT:
32 {
33 float* buff_A = FLA_FLOAT_PTR( A );
34
35 FLA_LU_nopiv_ops_var2( m_A,
36 n_A,
37 buff_A, rs_A, cs_A );
38
39 break;
40 }
41
42 case FLA_DOUBLE:
43 {
44 double* buff_A = FLA_DOUBLE_PTR( A );
45
46 FLA_LU_nopiv_opd_var2( m_A,
47 n_A,
48 buff_A, rs_A, cs_A );
49
50 break;
51 }
52
53 case FLA_COMPLEX:
54 {
55 scomplex* buff_A = FLA_COMPLEX_PTR( A );
56
57 FLA_LU_nopiv_opc_var2( m_A,
58 n_A,
59 buff_A, rs_A, cs_A );
60
61 break;
62 }
63
64 case FLA_DOUBLE_COMPLEX:
65 {
66 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
67
68 FLA_LU_nopiv_opz_var2( m_A,
69 n_A,
70 buff_A, rs_A, cs_A );
71
72 break;
73 }
74 }
75
76 return FLA_SUCCESS;
77}
FLA_LU_nopiv_opz_var2
FLA_Error FLA_LU_nopiv_opz_var2(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var2.c:318
FLA_LU_nopiv_opd_var2
FLA_Error FLA_LU_nopiv_opd_var2(int m_A, int n_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var2.c:160
FLA_LU_nopiv_opc_var2
FLA_Error FLA_LU_nopiv_opc_var2(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var2.c:239
FLA_LU_nopiv_ops_var2
FLA_Error FLA_LU_nopiv_ops_var2(int m_A, int n_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var2.c:81

References FLA_LU_nopiv_opc_var2(), FLA_LU_nopiv_opd_var2(), FLA_LU_nopiv_ops_var2(), FLA_LU_nopiv_opz_var2(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opt_var3()

FLA_Error FLA_LU_nopiv_opt_var3 ( FLA_Obj A)
16{
17 FLA_Datatype datatype;
18 int m_A, n_A;
19 int rs_A, cs_A;
20
21 datatype = FLA_Obj_datatype( A );
22
23 m_A = FLA_Obj_length( A );
24 n_A = FLA_Obj_width( A );
25 rs_A = FLA_Obj_row_stride( A );
26 cs_A = FLA_Obj_col_stride( A );
27
28
29 switch ( datatype )
30 {
31 case FLA_FLOAT:
32 {
33 float* buff_A = FLA_FLOAT_PTR( A );
34
35 FLA_LU_nopiv_ops_var3( m_A,
36 n_A,
37 buff_A, rs_A, cs_A );
38
39 break;
40 }
41
42 case FLA_DOUBLE:
43 {
44 double* buff_A = FLA_DOUBLE_PTR( A );
45
46 FLA_LU_nopiv_opd_var3( m_A,
47 n_A,
48 buff_A, rs_A, cs_A );
49
50 break;
51 }
52
53 case FLA_COMPLEX:
54 {
55 scomplex* buff_A = FLA_COMPLEX_PTR( A );
56
57 FLA_LU_nopiv_opc_var3( m_A,
58 n_A,
59 buff_A, rs_A, cs_A );
60
61 break;
62 }
63
64 case FLA_DOUBLE_COMPLEX:
65 {
66 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
67
68 FLA_LU_nopiv_opz_var3( m_A,
69 n_A,
70 buff_A, rs_A, cs_A );
71
72 break;
73 }
74 }
75
76 return FLA_SUCCESS;
77}
FLA_LU_nopiv_opz_var3
FLA_Error FLA_LU_nopiv_opz_var3(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var3.c:336
FLA_LU_nopiv_ops_var3
FLA_Error FLA_LU_nopiv_ops_var3(int m_A, int n_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var3.c:81
FLA_LU_nopiv_opc_var3
FLA_Error FLA_LU_nopiv_opc_var3(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var3.c:251
FLA_LU_nopiv_opd_var3
FLA_Error FLA_LU_nopiv_opd_var3(int m_A, int n_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var3.c:166

References FLA_LU_nopiv_opc_var3(), FLA_LU_nopiv_opd_var3(), FLA_LU_nopiv_ops_var3(), FLA_LU_nopiv_opz_var3(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opt_var4()

FLA_Error FLA_LU_nopiv_opt_var4 ( FLA_Obj A)
16{
17 FLA_Datatype datatype;
18 int m_A, n_A;
19 int rs_A, cs_A;
20
21 datatype = FLA_Obj_datatype( A );
22
23 m_A = FLA_Obj_length( A );
24 n_A = FLA_Obj_width( A );
25 rs_A = FLA_Obj_row_stride( A );
26 cs_A = FLA_Obj_col_stride( A );
27
28
29 switch ( datatype )
30 {
31 case FLA_FLOAT:
32 {
33 float* buff_A = FLA_FLOAT_PTR( A );
34
35 FLA_LU_nopiv_ops_var4( m_A,
36 n_A,
37 buff_A, rs_A, cs_A );
38
39 break;
40 }
41
42 case FLA_DOUBLE:
43 {
44 double* buff_A = FLA_DOUBLE_PTR( A );
45
46 FLA_LU_nopiv_opd_var4( m_A,
47 n_A,
48 buff_A, rs_A, cs_A );
49
50 break;
51 }
52
53 case FLA_COMPLEX:
54 {
55 scomplex* buff_A = FLA_COMPLEX_PTR( A );
56
57 FLA_LU_nopiv_opc_var4( m_A,
58 n_A,
59 buff_A, rs_A, cs_A );
60
61 break;
62 }
63
64 case FLA_DOUBLE_COMPLEX:
65 {
66 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
67
68 FLA_LU_nopiv_opz_var4( m_A,
69 n_A,
70 buff_A, rs_A, cs_A );
71
72 break;
73 }
74 }
75
76 return FLA_SUCCESS;
77}
FLA_LU_nopiv_opc_var4
FLA_Error FLA_LU_nopiv_opc_var4(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var4.c:223
FLA_LU_nopiv_ops_var4
FLA_Error FLA_LU_nopiv_ops_var4(int m_A, int n_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var4.c:81
FLA_LU_nopiv_opd_var4
FLA_Error FLA_LU_nopiv_opd_var4(int m_A, int n_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var4.c:152
FLA_LU_nopiv_opz_var4
FLA_Error FLA_LU_nopiv_opz_var4(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var4.c:294

References FLA_LU_nopiv_opc_var4(), FLA_LU_nopiv_opd_var4(), FLA_LU_nopiv_ops_var4(), FLA_LU_nopiv_opz_var4(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opt_var5()

FLA_Error FLA_LU_nopiv_opt_var5 ( FLA_Obj A)
14{
15 FLA_Datatype datatype;
16 int m_A, n_A;
17 int rs_A, cs_A;
18
19 datatype = FLA_Obj_datatype( A );
20
21 m_A = FLA_Obj_length( A );
22 n_A = FLA_Obj_width( A );
23 rs_A = FLA_Obj_row_stride( A );
24 cs_A = FLA_Obj_col_stride( A );
25
26
27 switch ( datatype )
28 {
29 case FLA_FLOAT:
30 {
31 float* buff_A = FLA_FLOAT_PTR( A );
32
33 FLA_LU_nopiv_ops_var5( m_A,
34 n_A,
35 buff_A, rs_A, cs_A );
36
37 break;
38 }
39
40 case FLA_DOUBLE:
41 {
42 double* buff_A = FLA_DOUBLE_PTR( A );
43
44 FLA_LU_nopiv_opd_var5( m_A,
45 n_A,
46 buff_A, rs_A, cs_A );
47
48 break;
49 }
50
51 case FLA_COMPLEX:
52 {
53 scomplex* buff_A = FLA_COMPLEX_PTR( A );
54
55 FLA_LU_nopiv_opc_var5( m_A,
56 n_A,
57 buff_A, rs_A, cs_A );
58
59 break;
60 }
61
62 case FLA_DOUBLE_COMPLEX:
63 {
64 dcomplex* buff_A = FLA_DOUBLE_COMPLEX_PTR( A );
65
66 FLA_LU_nopiv_opz_var5( m_A,
67 n_A,
68 buff_A, rs_A, cs_A );
69
70 break;
71 }
72 }
73
74 return FLA_SUCCESS;
75}
FLA_LU_nopiv_opd_var5
FLA_Error FLA_LU_nopiv_opd_var5(int m_A, int n_A, double *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var5.c:124
FLA_LU_nopiv_opz_var5
FLA_Error FLA_LU_nopiv_opz_var5(int m_A, int n_A, dcomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var5.c:214
FLA_LU_nopiv_ops_var5
FLA_Error FLA_LU_nopiv_ops_var5(int m_A, int n_A, float *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var5.c:79
FLA_LU_nopiv_opc_var5
FLA_Error FLA_LU_nopiv_opc_var5(int m_A, int n_A, scomplex *buff_A, int rs_A, int cs_A)
Definition FLA_LU_nopiv_opt_var5.c:169

References FLA_LU_nopiv_opc_var5(), FLA_LU_nopiv_opd_var5(), FLA_LU_nopiv_ops_var5(), FLA_LU_nopiv_opz_var5(), FLA_Obj_col_stride(), FLA_Obj_datatype(), FLA_Obj_length(), FLA_Obj_row_stride(), and FLA_Obj_width().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_opz_var1()

FLA_Error FLA_LU_nopiv_opz_var1 ( int m_A,
int n_A,
dcomplex * A,
int rs_A,
int cs_A )
357{
358 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
359 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
360 int min_m_n = min( m_A, n_A );
361 int i;
362
363 for ( i = 0; i < min_m_n; ++i )
364 {
365 dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
366 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
367 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
368 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
369
370 int mn_behind = i;
371
372 /*------------------------------------------------------------*/
373
374 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
375 bl1_ztrsv( BLIS1_LOWER_TRIANGULAR,
376 BLIS1_NO_TRANSPOSE,
377 BLIS1_UNIT_DIAG,
378 mn_behind,
379 A00, rs_A, cs_A,
380 a01, rs_A );
381
382 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
383 bl1_ztrsv( BLIS1_UPPER_TRIANGULAR,
384 BLIS1_TRANSPOSE,
385 BLIS1_NONUNIT_DIAG,
386 mn_behind,
387 A00, rs_A, cs_A,
388 a10t, cs_A );
389
390 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
391 bl1_zdots( BLIS1_NO_CONJUGATE,
392 mn_behind,
393 buff_m1,
394 a10t, cs_A,
395 a01, rs_A,
396 buff_1,
397 alpha11 );
398
399 /*------------------------------------------------------------*/
400
401 }
402
403 if ( m_A > n_A )
404 {
405 dcomplex* ATL = buff_A;
406 dcomplex* ABL = buff_A + n_A*rs_A;
407
408 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
409 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
410 // FLA_ONE, ATL, ABL );
411 bl1_ztrsm( BLIS1_RIGHT,
412 BLIS1_UPPER_TRIANGULAR,
413 BLIS1_NO_TRANSPOSE,
414 BLIS1_NONUNIT_DIAG,
415 m_A - n_A,
416 n_A,
417 buff_1,
418 ATL, rs_A, cs_A,
419 ABL, rs_A, cs_A );
420 }
421 else if ( m_A < n_A )
422 {
423 dcomplex* ATL = buff_A;
424 dcomplex* ATR = buff_A + m_A*cs_A;
425
426 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
427 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
428 // FLA_ONE, ATL, ATR );
429 bl1_ztrsm( BLIS1_LEFT,
430 BLIS1_LOWER_TRIANGULAR,
431 BLIS1_NO_TRANSPOSE,
432 BLIS1_UNIT_DIAG,
433 m_A,
434 n_A - m_A,
435 buff_1,
436 ATL, rs_A, cs_A,
437 ATR, rs_A, cs_A );
438 }
439
440 return FLA_SUCCESS;
441}
bl1_zdots
void bl1_zdots(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *beta, dcomplex *rho)
Definition bl1_dots.c:56
bl1_ztrsm
void bl1_ztrsm(side1_t side, uplo1_t uplo, trans1_t trans, diag1_t diag, int m, int n, dcomplex *alpha, dcomplex *a, int a_rs, int a_cs, dcomplex *b, int b_rs, int b_cs)
Definition bl1_trsm.c:369
bl1_ztrsv
void bl1_ztrsv(uplo1_t uplo, trans1_t trans, diag1_t diag, int m, dcomplex *a, int a_rs, int a_cs, dcomplex *x, int incx)
Definition bl1_trsv.c:177

References bl1_zdots(), bl1_ztrsm(), bl1_ztrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UNIT_DIAG, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var1().

◆ FLA_LU_nopiv_opz_var2()

FLA_Error FLA_LU_nopiv_opz_var2 ( int m_A,
int n_A,
dcomplex * A,
int rs_A,
int cs_A )
321{
322 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
323 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
324 int min_m_n = min( m_A, n_A );
325 int i;
326
327 for ( i = 0; i < min_m_n; ++i )
328 {
329 dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
330 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
331 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
332 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
333 dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
334 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
335
336 int n_ahead = n_A - i - 1;
337 int mn_behind = i;
338
339 /*------------------------------------------------------------*/
340
341 // FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
342 bl1_ztrsv( BLIS1_UPPER_TRIANGULAR,
343 BLIS1_TRANSPOSE,
344 BLIS1_NONUNIT_DIAG,
345 mn_behind,
346 A00, rs_A, cs_A,
347 a10t, cs_A );
348
349 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
350 bl1_zdots( BLIS1_NO_CONJUGATE,
351 mn_behind,
352 buff_m1,
353 a10t, cs_A,
354 a01, rs_A,
355 buff_1,
356 alpha11 );
357
358 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
359 bl1_zgemv( BLIS1_TRANSPOSE,
360 BLIS1_NO_CONJUGATE,
361 mn_behind,
362 n_ahead,
363 buff_m1,
364 A02, rs_A, cs_A,
365 a10t, cs_A,
366 buff_1,
367 a12t, cs_A );
368
369 /*------------------------------------------------------------*/
370
371 }
372
373 if ( m_A > n_A )
374 {
375 dcomplex* ATL = buff_A;
376 dcomplex* ABL = buff_A + n_A*rs_A;
377
378 // FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
379 // FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
380 // FLA_ONE, ATL, ABL );
381 bl1_ztrsm( BLIS1_RIGHT,
382 BLIS1_UPPER_TRIANGULAR,
383 BLIS1_NO_TRANSPOSE,
384 BLIS1_NONUNIT_DIAG,
385 m_A - n_A,
386 n_A,
387 buff_1,
388 ATL, rs_A, cs_A,
389 ABL, rs_A, cs_A );
390 }
391
392 return FLA_SUCCESS;
393}
bl1_zgemv
void bl1_zgemv(trans1_t transa, conj1_t conjx, int m, int n, dcomplex *alpha, dcomplex *a, int a_rs, int a_cs, dcomplex *x, int incx, dcomplex *beta, dcomplex *y, int incy)
Definition bl1_gemv.c:255

References bl1_zdots(), bl1_zgemv(), bl1_ztrsm(), bl1_ztrsv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_NONUNIT_DIAG, BLIS1_RIGHT, BLIS1_TRANSPOSE, BLIS1_UPPER_TRIANGULAR, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var2().

◆ FLA_LU_nopiv_opz_var3()

FLA_Error FLA_LU_nopiv_opz_var3 ( int m_A,
int n_A,
dcomplex * A,
int rs_A,
int cs_A )
339{
340 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
341 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
342 int min_m_n = min( m_A, n_A );
343 int i;
344
345 for ( i = 0; i < min_m_n; ++i )
346 {
347 dcomplex* A00 = buff_A + (0 )*cs_A + (0 )*rs_A;
348 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
349 dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
350 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
351 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
352 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
353
354 int m_ahead = m_A - i - 1;
355 int mn_behind = i;
356
357 /*------------------------------------------------------------*/
358
359 // FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
360 bl1_ztrsv( BLIS1_LOWER_TRIANGULAR,
361 BLIS1_NO_TRANSPOSE,
362 BLIS1_UNIT_DIAG,
363 mn_behind,
364 A00, rs_A, cs_A,
365 a01, rs_A );
366
367 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
368 bl1_zdots( BLIS1_NO_CONJUGATE,
369 mn_behind,
370 buff_m1,
371 a10t, cs_A,
372 a01, rs_A,
373 buff_1,
374 alpha11 );
375
376 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
377 bl1_zgemv( BLIS1_NO_TRANSPOSE,
378 BLIS1_NO_CONJUGATE,
379 m_ahead,
380 mn_behind,
381 buff_m1,
382 A20, rs_A, cs_A,
383 a01, rs_A,
384 buff_1,
385 a21, rs_A );
386
387 // FLA_Inv_scal_external( alpha11, a21 );
388 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
389 m_ahead,
390 alpha11,
391 a21, rs_A );
392
393 /*------------------------------------------------------------*/
394
395 }
396
397 if ( m_A < n_A )
398 {
399 dcomplex* ATL = buff_A;
400 dcomplex* ATR = buff_A + m_A*cs_A;
401
402 // FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
403 // FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
404 // FLA_ONE, ATL, ATR );
405 bl1_ztrsm( BLIS1_LEFT,
406 BLIS1_LOWER_TRIANGULAR,
407 BLIS1_NO_TRANSPOSE,
408 BLIS1_UNIT_DIAG,
409 m_A,
410 n_A - m_A,
411 buff_1,
412 ATL, rs_A, cs_A,
413 ATR, rs_A, cs_A );
414 }
415
416 return FLA_SUCCESS;
417}
bl1_zinvscalv
void bl1_zinvscalv(conj1_t conj, int n, dcomplex *alpha, dcomplex *x, int incx)
Definition bl1_invscalv.c:78

References bl1_zdots(), bl1_zgemv(), bl1_zinvscalv(), bl1_ztrsm(), bl1_ztrsv(), BLIS1_LEFT, BLIS1_LOWER_TRIANGULAR, BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_UNIT_DIAG, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var3().

◆ FLA_LU_nopiv_opz_var4()

FLA_Error FLA_LU_nopiv_opz_var4 ( int m_A,
int n_A,
dcomplex * A,
int rs_A,
int cs_A )
297{
298 dcomplex* buff_1 = FLA_DOUBLE_COMPLEX_PTR( FLA_ONE );
299 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
300 int min_m_n = min( m_A, n_A );
301 int i;
302
303 for ( i = 0; i < min_m_n; ++i )
304 {
305 dcomplex* a10t = buff_A + (0 )*cs_A + (i )*rs_A;
306 dcomplex* A20 = buff_A + (0 )*cs_A + (i+1)*rs_A;
307 dcomplex* a01 = buff_A + (i )*cs_A + (0 )*rs_A;
308 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
309 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
310 dcomplex* A02 = buff_A + (i+1)*cs_A + (0 )*rs_A;
311 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
312
313 int m_ahead = m_A - i - 1;
314 int n_ahead = n_A - i - 1;
315 int mn_behind = i;
316
317 /*------------------------------------------------------------*/
318
319 // FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
320 bl1_zdots( BLIS1_NO_CONJUGATE,
321 mn_behind,
322 buff_m1,
323 a10t, cs_A,
324 a01, rs_A,
325 buff_1,
326 alpha11 );
327
328 // FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
329 bl1_zgemv( BLIS1_TRANSPOSE,
330 BLIS1_NO_CONJUGATE,
331 mn_behind,
332 n_ahead,
333 buff_m1,
334 A02, rs_A, cs_A,
335 a10t, cs_A,
336 buff_1,
337 a12t, cs_A );
338
339 // FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
340 bl1_zgemv( BLIS1_NO_TRANSPOSE,
341 BLIS1_NO_CONJUGATE,
342 m_ahead,
343 mn_behind,
344 buff_m1,
345 A20, rs_A, cs_A,
346 a01, rs_A,
347 buff_1,
348 a21, rs_A );
349
350 // FLA_Inv_scal_external( alpha11, a21 );
351 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
352 m_ahead,
353 alpha11,
354 a21, rs_A );
355
356 /*------------------------------------------------------------*/
357
358 }
359
360 return FLA_SUCCESS;
361}

References bl1_zdots(), bl1_zgemv(), bl1_zinvscalv(), BLIS1_NO_CONJUGATE, BLIS1_NO_TRANSPOSE, BLIS1_TRANSPOSE, FLA_MINUS_ONE, FLA_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var4().

◆ FLA_LU_nopiv_opz_var5()

FLA_Error FLA_LU_nopiv_opz_var5 ( int m_A,
int n_A,
dcomplex * A,
int rs_A,
int cs_A )
217{
218 dcomplex* buff_m1 = FLA_DOUBLE_COMPLEX_PTR( FLA_MINUS_ONE );
219 int min_m_n = min( m_A, n_A );
220 int i;
221
222 for ( i = 0; i < min_m_n; ++i )
223 {
224 dcomplex* alpha11 = buff_A + (i )*cs_A + (i )*rs_A;
225 dcomplex* a21 = buff_A + (i )*cs_A + (i+1)*rs_A;
226 dcomplex* a12t = buff_A + (i+1)*cs_A + (i )*rs_A;
227 dcomplex* A22 = buff_A + (i+1)*cs_A + (i+1)*rs_A;
228
229 int m_ahead = m_A - i - 1;
230 int n_ahead = n_A - i - 1;
231
232 /*------------------------------------------------------------*/
233
234 // FLA_Inv_scal_external( alpha11, a21 );
235 bl1_zinvscalv( BLIS1_NO_CONJUGATE,
236 m_ahead,
237 alpha11,
238 a21, rs_A );
239
240 // FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
241 bl1_zger( BLIS1_NO_CONJUGATE,
242 BLIS1_NO_CONJUGATE,
243 m_ahead,
244 n_ahead,
245 buff_m1,
246 a21, rs_A,
247 a12t, cs_A,
248 A22, rs_A, cs_A );
249
250 /*------------------------------------------------------------*/
251
252 }
253
254 return FLA_SUCCESS;
255}
bl1_zger
void bl1_zger(conj1_t conjx, conj1_t conjy, int m, int n, dcomplex *alpha, dcomplex *x, int incx, dcomplex *y, int incy, dcomplex *a, int a_rs, int a_cs)
Definition bl1_ger.c:194

References bl1_zger(), bl1_zinvscalv(), BLIS1_NO_CONJUGATE, FLA_MINUS_ONE, and i.

Referenced by FLA_LU_nopiv_opt_var5().

◆ FLA_LU_nopiv_unb_var1()

FLA_Error FLA_LU_nopiv_unb_var1 ( FLA_Obj A)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
25 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
26
27 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
28 /* ************* */ /* ************************** */
29 &a10t, /**/ &alpha11, &a12t,
30 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
31 1, 1, FLA_BR );
32
33 /*------------------------------------------------------------*/
34
35 // a01 = trilu( A00 ) \ a01
36 FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
37
38 // a10t = a10t / triu( A00 )
39 FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
40
41 // alpha11 = alpha11 - a10t * a01
42 FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
43
44 /*------------------------------------------------------------*/
45
46 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
47 a10t, alpha11, /**/ a12t,
48 /* ************** */ /* ************************ */
49 &ABL, /**/ &ABR, A20, a21, /**/ A22,
50 FLA_TL );
51
52 }
53
54 if ( FLA_Obj_length( ABL ) > 0 )
55 // ABL = ABL / triu( ATL )
56 FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
57 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
58 FLA_ONE, ATL, ABL );
59 else if ( FLA_Obj_width( ATR ) > 0 )
60 // ATR = trilu( ATL ) \ ATR
61 FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
62 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
63 FLA_ONE, ATL, ATR );
64
65 return FLA_SUCCESS;
66}
FLA_Dots_external
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_Trsv_external
FLA_Error FLA_Trsv_external(FLA_Uplo uplo, FLA_Trans transa, FLA_Diag diag, FLA_Obj A, FLA_Obj x)
Definition FLA_Trsv_external.c:13

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_unb_var2()

FLA_Error FLA_LU_nopiv_unb_var2 ( FLA_Obj A)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
25 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
26
27 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
28 /* ************* */ /* ************************** */
29 &a10t, /**/ &alpha11, &a12t,
30 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
31 1, 1, FLA_BR );
32
33 /*------------------------------------------------------------*/
34
35 // a10t = a10t / triu( A00 )
36 FLA_Trsv_external( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, FLA_NONUNIT_DIAG, A00, a10t );
37
38 // alpha11 = alpha11 - a10t * a01
39 FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
40
41 // a12t = a12t - a10t * A02
42 FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
43
44 /*------------------------------------------------------------*/
45
46 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
47 a10t, alpha11, /**/ a12t,
48 /* ************** */ /* ************************ */
49 &ABL, /**/ &ABR, A20, a21, /**/ A22,
50 FLA_TL );
51
52 }
53
54 if ( FLA_Obj_length( ABL ) > 0 )
55 // ABL = ABL / triu( ATL )
56 FLA_Trsm_external( FLA_RIGHT, FLA_UPPER_TRIANGULAR,
57 FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
58 FLA_ONE, ATL, ABL );
59
60 return FLA_SUCCESS;
61}
FLA_Gemv_external
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_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_unb_var3()

FLA_Error FLA_LU_nopiv_unb_var3 ( FLA_Obj A)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
25 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
26
27 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
28 /* ************* */ /* ************************** */
29 &a10t, /**/ &alpha11, &a12t,
30 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
31 1, 1, FLA_BR );
32
33 /*------------------------------------------------------------*/
34
35 // a01 = trilu( A00 ) \ a10
36 FLA_Trsv_external( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_UNIT_DIAG, A00, a01 );
37
38 // alpha11 = alpha11 - a10t * a01
39 FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
40
41 // a21 = a21 - A20 * a01
42 FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
43
44 // a21 = a21 / alpha11
45 FLA_Inv_scal_external( alpha11, a21 );
46
47 /*------------------------------------------------------------*/
48
49 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
50 a10t, alpha11, /**/ a12t,
51 /* ************** */ /* ************************ */
52 &ABL, /**/ &ABR, A20, a21, /**/ A22,
53 FLA_TL );
54
55 }
56
57 if ( FLA_Obj_width( ATR ) > 0 )
58 // ATR = trilu( ATL ) \ ATR
59 FLA_Trsm_external( FLA_LEFT, FLA_LOWER_TRIANGULAR,
60 FLA_NO_TRANSPOSE, FLA_UNIT_DIAG,
61 FLA_ONE, ATL, ATR );
62
63 return FLA_SUCCESS;
64}
FLA_Inv_scal_external
FLA_Error FLA_Inv_scal_external(FLA_Obj alpha, FLA_Obj A)
Definition FLA_Inv_scal_external.c:13

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), FLA_Repart_2x2_to_3x3(), FLA_Trsm_external(), and FLA_Trsv_external().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_unb_var4()

FLA_Error FLA_LU_nopiv_unb_var4 ( FLA_Obj A)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
25 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
26
27 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
28 /* ************* */ /* ************************** */
29 &a10t, /**/ &alpha11, &a12t,
30 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
31 1, 1, FLA_BR );
32
33 /*------------------------------------------------------------*/
34
35 // alpha11 = alpha11 - a10t * a01
36 FLA_Dots_external( FLA_MINUS_ONE, a10t, a01, FLA_ONE, alpha11 );
37
38 // a12t = a12t - a10t * A02
39 FLA_Gemv_external( FLA_TRANSPOSE, FLA_MINUS_ONE, A02, a10t, FLA_ONE, a12t );
40
41 // a21 = a21 - A20 * a01
42 FLA_Gemv_external( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, A20, a01, FLA_ONE, a21 );
43
44 // a21 = a21 / alpha11
45 FLA_Inv_scal_external( alpha11, a21 );
46
47 /*------------------------------------------------------------*/
48
49 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
50 a10t, alpha11, /**/ a12t,
51 /* ************** */ /* ************************ */
52 &ABL, /**/ &ABR, A20, a21, /**/ A22,
53 FLA_TL );
54
55 }
56
57 return FLA_SUCCESS;
58}

References FLA_Cont_with_3x3_to_2x2(), FLA_Dots_external(), FLA_Gemv_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_2x2(), and FLA_Repart_2x2_to_3x3().

Referenced by FLA_LU_nopiv_internal().

◆ FLA_LU_nopiv_unb_var5()

FLA_Error FLA_LU_nopiv_unb_var5 ( FLA_Obj A)
16{
17 FLA_Obj ATL, ATR, A00, a01, A02,
18 ABL, ABR, a10t, alpha11, a12t,
19 A20, a21, A22;
20
21 FLA_Part_2x2( A, &ATL, &ATR,
22 &ABL, &ABR, 0, 0, FLA_TL );
23
24 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) &&
25 FLA_Obj_width( ATL ) < FLA_Obj_width( A )){
26
27 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02,
28 /* ************* */ /* ************************** */
29 &a10t, /**/ &alpha11, &a12t,
30 ABL, /**/ ABR, &A20, /**/ &a21, &A22,
31 1, 1, FLA_BR );
32
33 /*------------------------------------------------------------*/
34
35 // a21 = a21 / alpha11
36 FLA_Inv_scal_external( alpha11, a21 );
37
38 // A22 = A22 - a21 * a12t
39 FLA_Ger_external( FLA_MINUS_ONE, a21, a12t, A22 );
40
41 /*------------------------------------------------------------*/
42
43 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02,
44 a10t, alpha11, /**/ a12t,
45 /* ************** */ /* ************************ */
46 &ABL, /**/ &ABR, A20, a21, /**/ A22,
47 FLA_TL );
48
49 }
50
51 return FLA_SUCCESS;
52}
FLA_Ger_external
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_3x3_to_2x2(), FLA_Ger_external(), FLA_Inv_scal_external(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_Part_2x2(), and FLA_Repart_2x2_to_3x3().

Referenced by FLA_LU_nopiv_internal().

Generated on Thu Sep 23 2021 17:46:17 for libflame by doxygen 1.12.0