Go to the source code of this file.
Functions | |
FLA_Error | FLA_Trmm_luh_blk_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error | FLA_Trmm_luh_blk_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error | FLA_Trmm_luh_blk_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error | FLA_Trmm_luh_blk_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t *cntl) |
FLA_Error | FLA_Trmm_luh_unb_var1 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trmm_luh_unb_var2 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trmm_luh_unb_var3 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error | FLA_Trmm_luh_unb_var4 (FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B) |
FLA_Error FLA_Trmm_luh_blk_var1 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trmm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_luh().
00038 { 00039 FLA_Obj ATL, ATR, A00, A01, A02, 00040 ABL, ABR, A10, A11, A12, 00041 A20, A21, A22; 00042 00043 FLA_Obj BT, B0, 00044 BB, B1, 00045 B2; 00046 00047 dim_t b; 00048 00049 FLA_Part_2x2( A, &ATL, &ATR, 00050 &ABL, &ABR, 0, 0, FLA_BR ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_BOTTOM ); 00054 00055 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00056 00057 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00058 00059 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00060 &A10, &A11, /**/ &A12, 00061 /* ************* */ /* ******************** */ 00062 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00063 b, b, FLA_TL ); 00064 00065 FLA_Repart_2x1_to_3x1( BT, &B0, 00066 &B1, 00067 /* ** */ /* ** */ 00068 BB, &B2, b, FLA_TOP ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* B1 = triu( A11' ) * B1; */ 00073 FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, 00074 alpha, A11, B1, 00075 FLA_Cntl_sub_trmm( cntl ) ); 00076 00077 /* B1 = B1 + A01' * B0; */ 00078 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00079 FLA_ONE, A01, B0, FLA_ONE, B1, 00080 FLA_Cntl_sub_gemm( cntl ) ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00085 /* ************** */ /* ****************** */ 00086 A10, /**/ A11, A12, 00087 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00088 FLA_BR ); 00089 00090 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00091 /* ** */ /* ** */ 00092 B1, 00093 &BB, B2, FLA_BOTTOM ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
FLA_Error FLA_Trmm_luh_blk_var2 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trmm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_luh().
00036 { 00037 FLA_Obj ATL, ATR, A00, A01, A02, 00038 ABL, ABR, A10, A11, A12, 00039 A20, A21, A22; 00040 00041 FLA_Obj BT, B0, 00042 BB, B1, 00043 B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x2( A, &ATL, &ATR, 00048 &ABL, &ABR, 0, 0, FLA_BR ); 00049 00050 FLA_Part_2x1( B, &BT, 00051 &BB, 0, FLA_BOTTOM ); 00052 00053 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00054 00055 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00056 00057 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00058 &A10, &A11, /**/ &A12, 00059 /* ************* */ /* ******************** */ 00060 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00061 b, b, FLA_TL ); 00062 00063 FLA_Repart_2x1_to_3x1( BT, &B0, 00064 &B1, 00065 /* ** */ /* ** */ 00066 BB, &B2, b, FLA_TOP ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* B2 = B2 + A12' * B1; */ 00071 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00072 FLA_ONE, A12, B1, FLA_ONE, B2, 00073 FLA_Cntl_sub_gemm( cntl ) ); 00074 00075 /* B1 = triu( A11' ) * B1; */ 00076 FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, 00077 alpha, A11, B1, 00078 FLA_Cntl_sub_trmm( cntl ) ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00083 /* ************** */ /* ****************** */ 00084 A10, /**/ A11, A12, 00085 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00086 FLA_BR ); 00087 00088 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00089 /* ** */ /* ** */ 00090 B1, 00091 &BB, B2, FLA_BOTTOM ); 00092 00093 } 00094 00095 return FLA_SUCCESS; 00096 }
FLA_Error FLA_Trmm_luh_blk_var3 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trmm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_luh().
00036 { 00037 FLA_Obj BL, BR, B0, B1, B2; 00038 00039 dim_t b; 00040 00041 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00042 00043 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ 00044 00045 b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00046 00047 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00048 b, FLA_RIGHT ); 00049 00050 /*------------------------------------------------------------*/ 00051 00052 /* B1 = triu( A )' * B1; */ 00053 FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, 00054 alpha, A, B1, 00055 FLA_Cntl_sub_trmm( cntl ) ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00060 FLA_LEFT ); 00061 00062 } 00063 00064 return FLA_SUCCESS; 00065 }
FLA_Error FLA_Trmm_luh_blk_var4 | ( | FLA_Diag | diagA, | |
FLA_Obj | alpha, | |||
FLA_Obj | A, | |||
FLA_Obj | B, | |||
fla_trmm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), and FLA_Trmm_internal().
Referenced by FLA_Trmm_luh().
00038 { 00039 FLA_Obj BL, BR, B0, B1, B2; 00040 00041 dim_t b; 00042 00043 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00044 00045 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){ 00046 00047 b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00048 00049 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00050 b, FLA_LEFT ); 00051 00052 /*------------------------------------------------------------*/ 00053 00054 /* B1 = triu( A )' * B1; */ 00055 FLA_Trmm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, 00056 alpha, A, B1, 00057 FLA_Cntl_sub_trmm( cntl ) ); 00058 00059 /*------------------------------------------------------------*/ 00060 00061 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00062 FLA_RIGHT ); 00063 00064 } 00065 00066 return FLA_SUCCESS; 00067 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Scalc_external().
00038 { 00039 FLA_Obj ATL, ATR, A00, a01, A02, 00040 ABL, ABR, a10t, alpha11, a12t, 00041 A20, a21, A22; 00042 00043 FLA_Obj BT, B0, 00044 BB, b1t, 00045 B2; 00046 00047 FLA_Scal_external( alpha, B ); 00048 00049 FLA_Part_2x2( A, &ATL, &ATR, 00050 &ABL, &ABR, 0, 0, FLA_BR ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_BOTTOM ); 00054 00055 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00056 00057 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00058 &a10t, &alpha11, /**/ &a12t, 00059 /* ************* */ /* ************************** */ 00060 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00061 1, 1, FLA_TL ); 00062 00063 FLA_Repart_2x1_to_3x1( BT, &B0, 00064 &b1t, 00065 /* ** */ /* *** */ 00066 BB, &B2, 1, FLA_TOP ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* b1t = alpha11 * b1t */ 00071 if ( diagA != FLA_UNIT_DIAG ) 00072 FLA_Scalc_external( FLA_CONJUGATE, alpha11, b1t ); 00073 00074 /* b1t = b1t + a01' * B0; */ 00075 FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, FLA_ONE, B0, a01, FLA_ONE, b1t ); 00076 00077 /*------------------------------------------------------------*/ 00078 00079 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00080 /* ************** */ /* ************************ */ 00081 a10t, /**/ alpha11, a12t, 00082 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00083 FLA_BR ); 00084 00085 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00086 /* ** */ /* *** */ 00087 b1t, 00088 &BB, B2, FLA_BOTTOM ); 00089 00090 } 00091 00092 return FLA_SUCCESS; 00093 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), FLA_Scal_external(), and FLA_Scalc_external().
00038 { 00039 FLA_Obj ATL, ATR, A00, a01, A02, 00040 ABL, ABR, a10t, alpha11, a12t, 00041 A20, a21, A22; 00042 00043 FLA_Obj BT, B0, 00044 BB, b1t, 00045 B2; 00046 00047 FLA_Scal_external( alpha, B ); 00048 00049 FLA_Part_2x2( A, &ATL, &ATR, 00050 &ABL, &ABR, 0, 0, FLA_BR ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_BOTTOM ); 00054 00055 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00056 00057 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00058 &a10t, &alpha11, /**/ &a12t, 00059 /* ************* */ /* ************************** */ 00060 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00061 1, 1, FLA_TL ); 00062 00063 FLA_Repart_2x1_to_3x1( BT, &B0, 00064 &b1t, 00065 /* ** */ /* *** */ 00066 BB, &B2, 1, FLA_TOP ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* B2 = B2 + a12t' * b1t; */ 00071 FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, FLA_ONE, a12t, b1t, B2 ); 00072 00073 /* b1t = alpha11 * b1t; */ 00074 if ( diagA != FLA_UNIT_DIAG ) 00075 FLA_Scalc_external( FLA_CONJUGATE, alpha11, b1t ); 00076 00077 /*------------------------------------------------------------*/ 00078 00079 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00080 /* ************** */ /* ************************ */ 00081 a10t, /**/ alpha11, a12t, 00082 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00083 FLA_BR ); 00084 00085 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00086 /* ** */ /* *** */ 00087 b1t, 00088 &BB, B2, FLA_BOTTOM ); 00089 00090 } 00091 00092 return FLA_SUCCESS; 00093 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scal_external(), and FLA_Trmv_external().
00038 { 00039 FLA_Obj BL, BR, B0, b1, B2; 00040 00041 FLA_Scal_external( alpha, B ); 00042 00043 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00044 00045 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ 00046 00047 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2, 00048 1, FLA_RIGHT ); 00049 00050 /*------------------------------------------------------------*/ 00051 00052 /* b1 = triu( A )' * b1; */ 00053 FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, A, b1 ); 00054 00055 /*------------------------------------------------------------*/ 00056 00057 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2, 00058 FLA_LEFT ); 00059 00060 } 00061 00062 return FLA_SUCCESS; 00063 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Repart_1x2_to_1x3(), FLA_Scal_external(), and FLA_Trmv_external().
00038 { 00039 FLA_Obj BL, BR, B0, b1, B2; 00040 00041 FLA_Scal_external( alpha, B ); 00042 00043 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00044 00045 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){ 00046 00047 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2, 00048 1, FLA_LEFT ); 00049 00050 /*------------------------------------------------------------*/ 00051 00052 /* b1 = triu( A )' * b1; */ 00053 FLA_Trmv_external( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, diagA, A, b1 ); 00054 00055 /*------------------------------------------------------------*/ 00056 00057 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2, 00058 FLA_RIGHT ); 00059 00060 } 00061 00062 return FLA_SUCCESS; 00063 }