Go to the source code of this file.
FLA_Error FLA_Hemm_lu_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_TL ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_TOP ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_TOP ); 00061 00062 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00067 /* ************* */ /* ******************** */ 00068 &A10, /**/ &A11, &A12, 00069 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00070 b, b, FLA_BR ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 /* ** */ /* ** */ 00074 &B1, 00075 BB, &B2, b, FLA_BOTTOM ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 /* ** */ /* ** */ 00079 &C1, 00080 CB, &C2, b, FLA_BOTTOM ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C0 = C0 + A01 * B1 */ 00085 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A01, B1, FLA_ONE, C0, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A01' * B0 */ 00090 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00091 alpha, A01, B0, FLA_ONE, C1, 00092 FLA_Cntl_sub_gemm2( cntl ) ); 00093 00094 /* C1 = C1 + A11 * B1 */ 00095 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00096 alpha, A11, B1, beta, C1, 00097 FLA_Cntl_sub_hemm( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00102 A10, A11, /**/ A12, 00103 /* ************** */ /* ****************** */ 00104 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00105 FLA_TL ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 B1, 00109 /* ** */ /* ** */ 00110 &BB, B2, FLA_TOP ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 C1, 00114 /* ** */ /* ** */ 00115 &CB, C2, FLA_TOP ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var10 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_lu().
00038 { 00039 FLA_Obj BL, BR, B0, B1, B2; 00040 00041 FLA_Obj CL, CR, C0, C1, C2; 00042 00043 dim_t b; 00044 00045 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00046 00047 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); 00048 00049 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){ 00050 00051 b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00052 00053 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00054 b, FLA_LEFT ); 00055 00056 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2, 00057 b, FLA_LEFT ); 00058 00059 /*------------------------------------------------------------*/ 00060 00061 /* C1 = C1 + A * B1 */ 00062 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00063 alpha, A, B1, beta, C1, 00064 FLA_Cntl_sub_hemm( cntl ) ); 00065 00066 /*------------------------------------------------------------*/ 00067 00068 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00069 FLA_RIGHT ); 00070 00071 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2, 00072 FLA_RIGHT ); 00073 00074 } 00075 00076 return FLA_SUCCESS; 00077 }
FLA_Error FLA_Hemm_lu_blk_var2 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_TL ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_TOP ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_TOP ); 00061 00062 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00067 /* ************* */ /* ******************** */ 00068 &A10, /**/ &A11, &A12, 00069 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00070 b, b, FLA_BR ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 /* ** */ /* ** */ 00074 &B1, 00075 BB, &B2, b, FLA_BOTTOM ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 /* ** */ /* ** */ 00079 &C1, 00080 CB, &C2, b, FLA_BOTTOM ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C1 = C1 + A01' * B0 */ 00085 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A01, B0, FLA_ONE, C1, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A12 * B2 */ 00090 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00091 alpha, A12, B2, FLA_ONE, C1, 00092 FLA_Cntl_sub_gemm2( cntl ) ); 00093 00094 /* C1 = C1 + A11 * B1 */ 00095 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00096 alpha, A11, B1, beta, C1, 00097 FLA_Cntl_sub_hemm( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00102 A10, A11, /**/ A12, 00103 /* ************** */ /* ****************** */ 00104 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00105 FLA_TL ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 B1, 00109 /* ** */ /* ** */ 00110 &BB, B2, FLA_TOP ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 C1, 00114 /* ** */ /* ** */ 00115 &CB, C2, FLA_TOP ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var3 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_TL ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_TOP ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_TOP ); 00061 00062 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00067 /* ************* */ /* ******************** */ 00068 &A10, /**/ &A11, &A12, 00069 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00070 b, b, FLA_BR ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 /* ** */ /* ** */ 00074 &B1, 00075 BB, &B2, b, FLA_BOTTOM ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 /* ** */ /* ** */ 00079 &C1, 00080 CB, &C2, b, FLA_BOTTOM ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C1 = C1 + A11 * B1 */ 00085 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00086 alpha, A11, B1, beta, C1, 00087 FLA_Cntl_sub_hemm( cntl ) ); 00088 00089 /* C1 = C1 + A12 * B2 */ 00090 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00091 alpha, A12, B2, FLA_ONE, C1, 00092 FLA_Cntl_sub_gemm1( cntl ) ); 00093 00094 /* C2 = C2 + A12' * B1 */ 00095 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 alpha, A12, B1, FLA_ONE, C2, 00097 FLA_Cntl_sub_gemm2( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00102 A10, A11, /**/ A12, 00103 /* ************** */ /* ****************** */ 00104 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00105 FLA_TL ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 B1, 00109 /* ** */ /* ** */ 00110 &BB, B2, FLA_TOP ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 C1, 00114 /* ** */ /* ** */ 00115 &CB, C2, FLA_TOP ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var4 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00046 CB, C1, 00047 C2; 00048 00049 dim_t b; 00050 00051 FLA_Part_2x2( A, &ATL, &ATR, 00052 &ABL, &ABR, 0, 0, FLA_TL ); 00053 00054 FLA_Part_2x1( B, &BT, 00055 &BB, 0, FLA_TOP ); 00056 00057 FLA_Part_2x1( C, &CT, 00058 &CB, 0, FLA_TOP ); 00059 00060 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00061 00062 b = FLA_Determine_blocksize( ABR, FLA_BR, FLA_Cntl_blocksize( cntl ) ); 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &A01, &A02, 00065 /* ************* */ /* ******************** */ 00066 &A10, /**/ &A11, &A12, 00067 ABL, /**/ ABR, &A20, /**/ &A21, &A22, 00068 b, b, FLA_BR ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 /* ** */ /* ** */ 00072 &B1, 00073 BB, &B2, b, FLA_BOTTOM ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 /* ** */ /* ** */ 00077 &C1, 00078 CB, &C2, b, FLA_BOTTOM ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C0 = C0 + A01 * B1 */ 00083 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00084 alpha, A01, B1, FLA_ONE, C0, 00085 FLA_Cntl_sub_gemm1( cntl ) ); 00086 00087 /* C1 = C1 + A11 * B1 */ 00088 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00089 alpha, A11, B1, beta, C1, 00090 FLA_Cntl_sub_hemm( cntl ) ); 00091 00092 /* C2 = C2 + A12' * B1 */ 00093 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00094 alpha, A12, B1, FLA_ONE, C2, 00095 FLA_Cntl_sub_gemm2( cntl ) ); 00096 00097 /*------------------------------------------------------------*/ 00098 00099 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, A01, /**/ A02, 00100 A10, A11, /**/ A12, 00101 /* ************** */ /* ****************** */ 00102 &ABL, /**/ &ABR, A20, A21, /**/ A22, 00103 FLA_TL ); 00104 00105 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00106 B1, 00107 /* ** */ /* ** */ 00108 &BB, B2, FLA_TOP ); 00109 00110 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00111 C1, 00112 /* ** */ /* ** */ 00113 &CB, C2, FLA_TOP ); 00114 00115 } 00116 00117 return FLA_SUCCESS; 00118 }
FLA_Error FLA_Hemm_lu_blk_var5 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_BOTTOM ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_BOTTOM ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00067 &A10, &A11, /**/ &A12, 00068 /* ************* */ /* ******************** */ 00069 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00070 b, b, FLA_TL ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 &B1, 00074 /* ** */ /* ** */ 00075 BB, &B2, b, FLA_TOP ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 &C1, 00079 /* ** */ /* ** */ 00080 CB, &C2, b, FLA_TOP ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C1 = C1 + A11 * B1 */ 00085 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00086 alpha, A11, B1, beta, C1, 00087 FLA_Cntl_sub_hemm( cntl ) ); 00088 00089 /* C1 = C1 + A12 * B2 */ 00090 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00091 alpha, A12, B2, FLA_ONE, C1, 00092 FLA_Cntl_sub_gemm1( cntl ) ); 00093 00094 /* C2 = C2 + A12' * B1 */ 00095 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 alpha, A12, B1, FLA_ONE, C2, 00097 FLA_Cntl_sub_gemm2( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00102 /* ************** */ /* ****************** */ 00103 A10, /**/ A11, A12, 00104 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00105 FLA_BR ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 /* ** */ /* ** */ 00109 B1, 00110 &BB, B2, FLA_BOTTOM ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 /* ** */ /* ** */ 00114 C1, 00115 &CB, C2, FLA_BOTTOM ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var6 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_BOTTOM ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_BOTTOM ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00067 &A10, &A11, /**/ &A12, 00068 /* ************* */ /* ******************** */ 00069 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00070 b, b, FLA_TL ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 &B1, 00074 /* ** */ /* ** */ 00075 BB, &B2, b, FLA_TOP ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 &C1, 00079 /* ** */ /* ** */ 00080 CB, &C2, b, FLA_TOP ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C1 = C1 + A01' * B0 */ 00085 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A01, B0, FLA_ONE, C1, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A11 * B1 */ 00090 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00091 alpha, A11, B1, beta, C1, 00092 FLA_Cntl_sub_hemm( cntl ) ); 00093 00094 /* C1 = C1 + A12 * B2 */ 00095 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 alpha, A12, B2, FLA_ONE, C1, 00097 FLA_Cntl_sub_gemm2( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00102 /* ************** */ /* ****************** */ 00103 A10, /**/ A11, A12, 00104 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00105 FLA_BR ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 /* ** */ /* ** */ 00109 B1, 00110 &BB, B2, FLA_BOTTOM ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 /* ** */ /* ** */ 00114 C1, 00115 &CB, C2, FLA_BOTTOM ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var7 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_BOTTOM ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_BOTTOM ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00067 &A10, &A11, /**/ &A12, 00068 /* ************* */ /* ******************** */ 00069 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00070 b, b, FLA_TL ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 &B1, 00074 /* ** */ /* ** */ 00075 BB, &B2, b, FLA_TOP ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 &C1, 00079 /* ** */ /* ** */ 00080 CB, &C2, b, FLA_TOP ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C0 = C0 + A01 * B1 */ 00085 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A01, B1, FLA_ONE, C0, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A01' * B0 */ 00090 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00091 alpha, A01, B0, FLA_ONE, C1, 00092 FLA_Cntl_sub_gemm2( cntl ) ); 00093 00094 /* C1 = C1 + A11 * B1 */ 00095 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00096 alpha, A11, B1, beta, C1, 00097 FLA_Cntl_sub_hemm( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00102 /* ************** */ /* ****************** */ 00103 A10, /**/ A11, A12, 00104 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00105 FLA_BR ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 /* ** */ /* ** */ 00109 B1, 00110 &BB, B2, FLA_BOTTOM ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 /* ** */ /* ** */ 00114 C1, 00115 &CB, C2, FLA_BOTTOM ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var8 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Hemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
Referenced by FLA_Hemm_lu().
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 FLA_Obj CT, C0, 00048 CB, C1, 00049 C2; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x2( A, &ATL, &ATR, 00054 &ABL, &ABR, 0, 0, FLA_BR ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_BOTTOM ); 00058 00059 FLA_Part_2x1( C, &CT, 00060 &CB, 0, FLA_BOTTOM ); 00061 00062 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &A01, /**/ &A02, 00067 &A10, &A11, /**/ &A12, 00068 /* ************* */ /* ******************** */ 00069 ABL, /**/ ABR, &A20, &A21, /**/ &A22, 00070 b, b, FLA_TL ); 00071 00072 FLA_Repart_2x1_to_3x1( BT, &B0, 00073 &B1, 00074 /* ** */ /* ** */ 00075 BB, &B2, b, FLA_TOP ); 00076 00077 FLA_Repart_2x1_to_3x1( CT, &C0, 00078 &C1, 00079 /* ** */ /* ** */ 00080 CB, &C2, b, FLA_TOP ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C0 = C0 + A01 * B1 */ 00085 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 00086 alpha, A01, B1, FLA_ONE, C0, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C1 = C1 + A11 * B1 */ 00090 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00091 alpha, A11, B1, beta, C1, 00092 FLA_Cntl_sub_hemm( cntl ) ); 00093 00094 /* C2 = C2 + A12' * B1 */ 00095 FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 00096 alpha, A12, B1, FLA_ONE, C2, 00097 FLA_Cntl_sub_gemm2( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ A01, A02, 00102 /* ************** */ /* ****************** */ 00103 A10, /**/ A11, A12, 00104 &ABL, /**/ &ABR, A20, /**/ A21, A22, 00105 FLA_BR ); 00106 00107 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00108 /* ** */ /* ** */ 00109 B1, 00110 &BB, B2, FLA_BOTTOM ); 00111 00112 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00113 /* ** */ /* ** */ 00114 C1, 00115 &CB, C2, FLA_BOTTOM ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }
FLA_Error FLA_Hemm_lu_blk_var9 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_hemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Hemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
Referenced by FLA_Hemm_lu().
00036 { 00037 FLA_Obj BL, BR, B0, B1, B2; 00038 00039 FLA_Obj CL, CR, C0, C1, C2; 00040 00041 dim_t b; 00042 00043 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00044 00045 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00046 00047 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ 00048 00049 b = FLA_Determine_blocksize( BR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00050 00051 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00052 b, FLA_RIGHT ); 00053 00054 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, 00055 b, FLA_RIGHT ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 /* C1 = C1 + A * B1 */ 00060 FLA_Hemm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 00061 alpha, A, B1, beta, C1, 00062 FLA_Cntl_sub_hemm( cntl ) ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00067 FLA_LEFT ); 00068 00069 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, 00070 FLA_LEFT ); 00071 00072 } 00073 00074 return FLA_SUCCESS; 00075 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_TL ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_TOP ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_TOP ); 00060 00061 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00065 /* ************* */ /* ************************** */ 00066 &a10t, /**/ &alpha11, &a12t, 00067 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00068 1, 1, FLA_BR ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 /* ** */ /* ** */ 00072 &b1t, 00073 BB, &B2, 1, FLA_BOTTOM ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 /* ** */ /* ** */ 00077 &c1t, 00078 CB, &C2, 1, FLA_BOTTOM ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C0 = C0 + a01 * b1t */ 00083 FLA_Ger_external( alpha, a01, b1t, C0 ); 00084 00085 /* c1t = c1t + a01' * B0 */ 00086 /* c1t' = c1t' + B0' * a01 */ 00087 FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B0, a01, FLA_ONE, c1t ); 00088 00089 /* c1t = c1t + alpha11 * b1t */ 00090 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00091 00092 /*------------------------------------------------------------*/ 00093 00094 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00095 a10t, alpha11, /**/ a12t, 00096 /* ************** */ /* ************************ */ 00097 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00098 FLA_TL ); 00099 00100 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00101 b1t, 00102 /* ** */ /* ** */ 00103 &BB, B2, FLA_TOP ); 00104 00105 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00106 c1t, 00107 /* ** */ /* ** */ 00108 &CB, C2, FLA_TOP ); 00109 00110 } 00111 00112 return FLA_SUCCESS; 00113 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
00038 { 00039 FLA_Obj BL, BR, B0, b1t, B2; 00040 00041 FLA_Obj CL, CR, C0, c1t, C2; 00042 00043 00044 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00045 00046 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); 00047 00048 while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){ 00049 00050 00051 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2, 00052 1, FLA_LEFT ); 00053 00054 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1t, /**/ &C2, 00055 1, FLA_LEFT ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 /* c1t = c1t + A * b1t */ 00060 FLA_Hemv_external( FLA_UPPER_TRIANGULAR, alpha, A, b1t, beta, c1t ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2, 00065 FLA_RIGHT ); 00066 00067 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1t, C2, 00068 FLA_RIGHT ); 00069 00070 } 00071 00072 return FLA_SUCCESS; 00073 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_TL ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_TOP ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_TOP ); 00060 00061 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00065 /* ************* */ /* ************************** */ 00066 &a10t, /**/ &alpha11, &a12t, 00067 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00068 1, 1, FLA_BR ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 /* ** */ /* ** */ 00072 &b1t, 00073 BB, &B2, 1, FLA_BOTTOM ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 /* ** */ /* ** */ 00077 &c1t, 00078 CB, &C2, 1, FLA_BOTTOM ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* c1t = c1t + a01' * B0 */ 00083 /* c1t' = c1t' + B0' * a01 */ 00084 FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B0, a01, FLA_ONE, c1t ); 00085 00086 /* c1t = c1t + a12t * B2 */ 00087 /* c1t' = c1t' + B2' * a12t' */ 00088 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a12t, FLA_ONE, c1t ); 00089 00090 /* c1t = c1t + alpha11 * b1t */ 00091 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00092 00093 /*------------------------------------------------------------*/ 00094 00095 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00096 a10t, alpha11, /**/ a12t, 00097 /* ************** */ /* ************************ */ 00098 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00099 FLA_TL ); 00100 00101 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00102 b1t, 00103 /* ** */ /* ** */ 00104 &BB, B2, FLA_TOP ); 00105 00106 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00107 c1t, 00108 /* ** */ /* ** */ 00109 &CB, C2, FLA_TOP ); 00110 00111 } 00112 00113 return FLA_SUCCESS; 00114 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_TL ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_TOP ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_TOP ); 00060 00061 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00065 /* ************* */ /* ************************** */ 00066 &a10t, /**/ &alpha11, &a12t, 00067 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00068 1, 1, FLA_BR ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 /* ** */ /* ** */ 00072 &b1t, 00073 BB, &B2, 1, FLA_BOTTOM ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 /* ** */ /* ** */ 00077 &c1t, 00078 CB, &C2, 1, FLA_BOTTOM ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* c1t = c1t + alpha11 * b1t */ 00083 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00084 00085 /* c1t = c1t + a12t * B2 */ 00086 /* c1t' = c1t' + B2' * a12t' */ 00087 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a12t, FLA_ONE, c1t ); 00088 00089 /* C2 = C2 + a12t' * b1t */ 00090 FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a12t, b1t, C2 ); 00091 00092 /*------------------------------------------------------------*/ 00093 00094 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00095 a10t, alpha11, /**/ a12t, 00096 /* ************** */ /* ************************ */ 00097 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00098 FLA_TL ); 00099 00100 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00101 b1t, 00102 /* ** */ /* ** */ 00103 &BB, B2, FLA_TOP ); 00104 00105 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00106 c1t, 00107 /* ** */ /* ** */ 00108 &CB, C2, FLA_TOP ); 00109 00110 } 00111 00112 return FLA_SUCCESS; 00113 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_TL ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_TOP ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_TOP ); 00060 00061 while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, /**/ &a01, &A02, 00065 /* ************* */ /* ************************** */ 00066 &a10t, /**/ &alpha11, &a12t, 00067 ABL, /**/ ABR, &A20, /**/ &a21, &A22, 00068 1, 1, FLA_BR ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 /* ** */ /* ** */ 00072 &b1t, 00073 BB, &B2, 1, FLA_BOTTOM ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 /* ** */ /* ** */ 00077 &c1t, 00078 CB, &C2, 1, FLA_BOTTOM ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C0 = C0 + a01 * b1t */ 00083 FLA_Ger_external( alpha, a01, b1t, C0 ); 00084 00085 /* c1t = c1t + alpha11 * b1t */ 00086 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00087 00088 /* C2 = C2 + a12t' * b1t */ 00089 FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a12t, b1t, C2 ); 00090 00091 /*------------------------------------------------------------*/ 00092 00093 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, a01, /**/ A02, 00094 a10t, alpha11, /**/ a12t, 00095 /* ************** */ /* ************************ */ 00096 &ABL, /**/ &ABR, A20, a21, /**/ A22, 00097 FLA_TL ); 00098 00099 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00100 b1t, 00101 /* ** */ /* ** */ 00102 &BB, B2, FLA_TOP ); 00103 00104 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00105 c1t, 00106 /* ** */ /* ** */ 00107 &CB, C2, FLA_TOP ); 00108 00109 } 00110 00111 return FLA_SUCCESS; 00112 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_BR ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_BOTTOM ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_BOTTOM ); 00060 00061 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00065 &a10t, &alpha11, /**/ &a12t, 00066 /* ************* */ /* ************************** */ 00067 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00068 1, 1, FLA_TL ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 &b1t, 00072 /* ** */ /* ** */ 00073 BB, &B2, 1, FLA_TOP ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 &c1t, 00077 /* ** */ /* ** */ 00078 CB, &C2, 1, FLA_TOP ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* c1t = c1t + alpha11 * b1t */ 00083 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00084 00085 /* c1t = c1t + a12t * B2 */ 00086 /* c1t' = c1t' + B2' * a12t' */ 00087 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a12t, FLA_ONE, c1t ); 00088 00089 /* C2 = C2 + a12t' * b1t */ 00090 FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a12t, b1t, C2 ); 00091 00092 /*------------------------------------------------------------*/ 00093 00094 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00095 /* ************** */ /* ************************ */ 00096 a10t, /**/ alpha11, a12t, 00097 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00098 FLA_BR ); 00099 00100 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00101 /* ** */ /* ** */ 00102 b1t, 00103 &BB, B2, FLA_BOTTOM ); 00104 00105 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00106 /* ** */ /* ** */ 00107 c1t, 00108 &CB, C2, FLA_BOTTOM ); 00109 00110 } 00111 00112 return FLA_SUCCESS; 00113 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemv_external(), FLA_Gemvc_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_BR ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_BOTTOM ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_BOTTOM ); 00060 00061 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00065 &a10t, &alpha11, /**/ &a12t, 00066 /* ************* */ /* ************************** */ 00067 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00068 1, 1, FLA_TL ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 &b1t, 00072 /* ** */ /* ** */ 00073 BB, &B2, 1, FLA_TOP ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 &c1t, 00077 /* ** */ /* ** */ 00078 CB, &C2, 1, FLA_TOP ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* c1t = c1t + a01' * B0 */ 00083 /* c1t' = c1t' + B0' * a01 */ 00084 FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B0, a01, FLA_ONE, c1t ); 00085 00086 /* c1t = c1t + alpha11 * b1t */ 00087 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00088 00089 /* c1t = c1t + a12t * B2 */ 00090 /* c1t' = c1t' + B2' * a12t' */ 00091 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a12t, FLA_ONE, c1t ); 00092 00093 /*------------------------------------------------------------*/ 00094 00095 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00096 /* ************** */ /* ************************ */ 00097 a10t, /**/ alpha11, a12t, 00098 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00099 FLA_BR ); 00100 00101 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00102 /* ** */ /* ** */ 00103 b1t, 00104 &BB, B2, FLA_BOTTOM ); 00105 00106 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00107 /* ** */ /* ** */ 00108 c1t, 00109 &CB, C2, FLA_BOTTOM ); 00110 00111 } 00112 00113 return FLA_SUCCESS; 00114 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Gemvc_external(), FLA_Ger_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_BR ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_BOTTOM ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_BOTTOM ); 00060 00061 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00065 &a10t, &alpha11, /**/ &a12t, 00066 /* ************* */ /* ************************** */ 00067 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00068 1, 1, FLA_TL ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 &b1t, 00072 /* ** */ /* ** */ 00073 BB, &B2, 1, FLA_TOP ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 &c1t, 00077 /* ** */ /* ** */ 00078 CB, &C2, 1, FLA_TOP ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C0 = C0 + a01 * b1t */ 00083 FLA_Ger_external( alpha, a01, b1t, C0 ); 00084 00085 /* c1t = c1t + a01' * B0 */ 00086 /* c1t' = c1t' + B0' * a01 */ 00087 FLA_Gemvc_external( FLA_TRANSPOSE, FLA_CONJUGATE, alpha, B0, a01, FLA_ONE, c1t ); 00088 00089 /* c1t = c1t + alpha11 * b1t */ 00090 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00091 00092 /*------------------------------------------------------------*/ 00093 00094 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00095 /* ************** */ /* ************************ */ 00096 a10t, /**/ alpha11, a12t, 00097 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00098 FLA_BR ); 00099 00100 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00101 /* ** */ /* ** */ 00102 b1t, 00103 &BB, B2, FLA_BOTTOM ); 00104 00105 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00106 /* ** */ /* ** */ 00107 c1t, 00108 &CB, C2, FLA_BOTTOM ); 00109 00110 } 00111 00112 return FLA_SUCCESS; 00113 }
References FLA_Axpys_external(), FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Ger_external(), FLA_Gerc_external(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), and FLA_Repart_2x2_to_3x3().
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_Obj CT, C0, 00048 CB, c1t, 00049 C2; 00050 00051 00052 FLA_Part_2x2( A, &ATL, &ATR, 00053 &ABL, &ABR, 0, 0, FLA_BR ); 00054 00055 FLA_Part_2x1( B, &BT, 00056 &BB, 0, FLA_BOTTOM ); 00057 00058 FLA_Part_2x1( C, &CT, 00059 &CB, 0, FLA_BOTTOM ); 00060 00061 while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){ 00062 00063 00064 FLA_Repart_2x2_to_3x3( ATL, /**/ ATR, &A00, &a01, /**/ &A02, 00065 &a10t, &alpha11, /**/ &a12t, 00066 /* ************* */ /* ************************** */ 00067 ABL, /**/ ABR, &A20, &a21, /**/ &A22, 00068 1, 1, FLA_TL ); 00069 00070 FLA_Repart_2x1_to_3x1( BT, &B0, 00071 &b1t, 00072 /* ** */ /* ** */ 00073 BB, &B2, 1, FLA_TOP ); 00074 00075 FLA_Repart_2x1_to_3x1( CT, &C0, 00076 &c1t, 00077 /* ** */ /* ** */ 00078 CB, &C2, 1, FLA_TOP ); 00079 00080 /*------------------------------------------------------------*/ 00081 00082 /* C0 = C0 + a01 * b1t */ 00083 FLA_Ger_external( alpha, a01, b1t, C0 ); 00084 00085 /* c1t = c1t + alpha11 * b1t */ 00086 FLA_Axpys_external( alpha, alpha11, b1t, beta, c1t ); 00087 00088 /* C2 = C2 + a12t' * b1t */ 00089 FLA_Gerc_external( FLA_CONJUGATE, FLA_NO_CONJUGATE, alpha, a12t, b1t, C2 ); 00090 00091 /*------------------------------------------------------------*/ 00092 00093 FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR, A00, /**/ a01, A02, 00094 /* ************** */ /* ************************ */ 00095 a10t, /**/ alpha11, a12t, 00096 &ABL, /**/ &ABR, A20, /**/ a21, A22, 00097 FLA_BR ); 00098 00099 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00100 /* ** */ /* ** */ 00101 b1t, 00102 &BB, B2, FLA_BOTTOM ); 00103 00104 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00105 /* ** */ /* ** */ 00106 c1t, 00107 &CB, C2, FLA_BOTTOM ); 00108 00109 } 00110 00111 return FLA_SUCCESS; 00112 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Hemv_external(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().
00038 { 00039 FLA_Obj BL, BR, B0, b1t, B2; 00040 00041 FLA_Obj CL, CR, C0, c1t, C2; 00042 00043 00044 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00045 00046 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00047 00048 while ( FLA_Obj_width( BL ) < FLA_Obj_width( B ) ){ 00049 00050 00051 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2, 00052 1, FLA_RIGHT ); 00053 00054 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1t, &C2, 00055 1, FLA_RIGHT ); 00056 00057 /*------------------------------------------------------------*/ 00058 00059 /* c1t = c1t + A * b1t */ 00060 FLA_Hemv_external( FLA_UPPER_TRIANGULAR, alpha, A, b1t, beta, c1t ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2, 00065 FLA_LEFT ); 00066 00067 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1t, /**/ C2, 00068 FLA_LEFT ); 00069 00070 } 00071 00072 return FLA_SUCCESS; 00073 }