Go to the source code of this file.
FLA_Error FLA_Syr2k_ut_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_TL ); 00055 00056 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00061 b, FLA_RIGHT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00064 b, FLA_RIGHT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00067 /* ************* */ /* ******************** */ 00068 &C10, /**/ &C11, &C12, 00069 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00070 b, b, FLA_BR ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C12 = C12 + A1' * B2 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A1, B2, beta, C12, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C12 = C12 + B1' * A2 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B1, A2, FLA_ONE, C12, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00092 FLA_LEFT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00095 FLA_LEFT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00098 C10, C11, /**/ C12, 00099 /* ************** */ /* ****************** */ 00100 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00101 FLA_TL ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var10 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, A1, 00041 A2; 00042 00043 FLA_Obj BT, B0, 00044 BB, B1, 00045 B2; 00046 00047 dim_t b; 00048 00049 FLA_Part_2x1( A, &AT, 00050 &AB, 0, FLA_BOTTOM ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_BOTTOM ); 00054 00055 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00056 00057 b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00058 00059 FLA_Repart_2x1_to_3x1( AT, &A0, 00060 &A1, 00061 /* ** */ /* ** */ 00062 AB, &A2, b, FLA_TOP ); 00063 00064 FLA_Repart_2x1_to_3x1( BT, &B0, 00065 &B1, 00066 /* ** */ /* ** */ 00067 BB, &B2, b, FLA_TOP ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 /* C = C + A1' * B1 + B1' * A1 */ 00072 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00073 alpha, A1, B1, beta, C, 00074 FLA_Cntl_sub_syr2k( cntl ) ); 00075 00076 /*------------------------------------------------------------*/ 00077 00078 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00079 /* ** */ /* ** */ 00080 A1, 00081 &AB, A2, FLA_BOTTOM ); 00082 00083 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00084 /* ** */ /* ** */ 00085 B1, 00086 &BB, B2, FLA_BOTTOM ); 00087 00088 } 00089 00090 return FLA_SUCCESS; 00091 }
FLA_Error FLA_Syr2k_ut_blk_var2 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_TL ); 00055 00056 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00061 b, FLA_RIGHT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00064 b, FLA_RIGHT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00067 /* ************* */ /* ******************** */ 00068 &C10, /**/ &C11, &C12, 00069 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00070 b, b, FLA_BR ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C01 = C01 + B0' * A1 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, B0, A1, beta, C01, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C12 = C12 + A1' * B2 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, A1, B2, FLA_ONE, C12, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00092 FLA_LEFT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00095 FLA_LEFT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00098 C10, C11, /**/ C12, 00099 /* ************** */ /* ****************** */ 00100 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00101 FLA_TL ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var3 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00036 { 00037 FLA_Obj AL, AR, A0, A1, A2; 00038 00039 FLA_Obj BL, BR, B0, B1, B2; 00040 00041 FLA_Obj CTL, CTR, C00, C01, C02, 00042 CBL, CBR, C10, C11, C12, 00043 C20, C21, C22; 00044 00045 dim_t b; 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00048 00049 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00050 00051 FLA_Part_2x2( C, &CTL, &CTR, 00052 &CBL, &CBR, 0, 0, FLA_TL ); 00053 00054 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00055 00056 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00059 b, FLA_RIGHT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00062 b, FLA_RIGHT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00065 /* ************* */ /* ******************** */ 00066 &C10, /**/ &C11, &C12, 00067 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00068 b, b, FLA_BR ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* C01 = C01 + A0' * B1 */ 00073 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00074 alpha, A0, B1, beta, C01, 00075 FLA_Cntl_sub_gemm1( cntl ) ); 00076 00077 /* C12 = C12 + B1' * A2 */ 00078 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00079 alpha, B1, A2, FLA_ONE, C12, 00080 FLA_Cntl_sub_gemm2( cntl ) ); 00081 00082 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00083 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00084 alpha, A1, B1, beta, C11, 00085 FLA_Cntl_sub_syr2k( cntl ) ); 00086 00087 /*------------------------------------------------------------*/ 00088 00089 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00090 FLA_LEFT ); 00091 00092 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00093 FLA_LEFT ); 00094 00095 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00096 C10, C11, /**/ C12, 00097 /* ************** */ /* ****************** */ 00098 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00099 FLA_TL ); 00100 00101 } 00102 00103 return FLA_SUCCESS; 00104 }
FLA_Error FLA_Syr2k_ut_blk_var4 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_TL ); 00055 00056 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00061 b, FLA_RIGHT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00064 b, FLA_RIGHT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00067 /* ************* */ /* ******************** */ 00068 &C10, /**/ &C11, &C12, 00069 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00070 b, b, FLA_BR ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C01 = C01 + A0' * B1 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A0, B1, beta, C01, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C01 = C01 + B0' * A1 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B0, A1, FLA_ONE, C01, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00092 FLA_LEFT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00095 FLA_LEFT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00098 C10, C11, /**/ C12, 00099 /* ************** */ /* ****************** */ 00100 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00101 FLA_TL ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var5 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_BR ); 00055 00056 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00061 b, FLA_LEFT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00064 b, FLA_LEFT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02, 00067 &C10, &C11, /**/ &C12, 00068 /* ************* */ /* ******************** */ 00069 CBL, /**/ CBR, &C20, &C21, /**/ &C22, 00070 b, b, FLA_TL ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C01 = C01 + A0' * B1 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A0, B1, beta, C01, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C01 = C01 + B0' * A1 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B0, A1, FLA_ONE, C01, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00092 FLA_RIGHT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00095 FLA_RIGHT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02, 00098 /* ************** */ /* ****************** */ 00099 C10, /**/ C11, C12, 00100 &CBL, /**/ &CBR, C20, /**/ C21, C22, 00101 FLA_BR ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var6 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_BR ); 00055 00056 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00061 b, FLA_LEFT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00064 b, FLA_LEFT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02, 00067 &C10, &C11, /**/ &C12, 00068 /* ************* */ /* ******************** */ 00069 CBL, /**/ CBR, &C20, &C21, /**/ &C22, 00070 b, b, FLA_TL ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C01 = C01 + A0' * B1 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A0, B1, beta, C01, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C12 = C12 + B1' * A2 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B1, A2, FLA_ONE, C12, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00092 FLA_RIGHT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00095 FLA_RIGHT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02, 00098 /* ************** */ /* ****************** */ 00099 C10, /**/ C11, C12, 00100 &CBL, /**/ &CBR, C20, /**/ C21, C22, 00101 FLA_BR ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var7 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_BR ); 00055 00056 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00061 b, FLA_LEFT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00064 b, FLA_LEFT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02, 00067 &C10, &C11, /**/ &C12, 00068 /* ************* */ /* ******************** */ 00069 CBL, /**/ CBR, &C20, &C21, /**/ &C22, 00070 b, b, FLA_TL ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C01 = C01 + B0' * A1 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, B0, A1, beta, C01, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C12 = C12 + A1' * B2 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, A1, B2, FLA_ONE, C12, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00092 FLA_RIGHT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00095 FLA_RIGHT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02, 00098 /* ************** */ /* ****************** */ 00099 C10, /**/ C11, C12, 00100 &CBL, /**/ &CBR, C20, /**/ C21, C22, 00101 FLA_BR ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var8 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 FLA_Obj CTL, CTR, C00, C01, C02, 00044 CBL, CBR, C10, C11, C12, 00045 C20, C21, C22; 00046 00047 dim_t b; 00048 00049 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00050 00051 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00052 00053 FLA_Part_2x2( C, &CTL, &CTR, 00054 &CBL, &CBR, 0, 0, FLA_BR ); 00055 00056 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00057 00058 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00059 00060 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00061 b, FLA_LEFT ); 00062 00063 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00064 b, FLA_LEFT ); 00065 00066 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &C01, /**/ &C02, 00067 &C10, &C11, /**/ &C12, 00068 /* ************* */ /* ******************** */ 00069 CBL, /**/ CBR, &C20, &C21, /**/ &C22, 00070 b, b, FLA_TL ); 00071 00072 /*------------------------------------------------------------*/ 00073 00074 /* C12 = C12 + A1' * B2 */ 00075 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00076 alpha, A1, B2, beta, C12, 00077 FLA_Cntl_sub_gemm1( cntl ) ); 00078 00079 /* C12 = C12 + B1' * A2 */ 00080 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 00081 alpha, B1, A2, FLA_ONE, C12, 00082 FLA_Cntl_sub_gemm2( cntl ) ); 00083 00084 /* C11 = C11 + A1' * B1 + B1' * A1 */ 00085 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00086 alpha, A1, B1, beta, C11, 00087 FLA_Cntl_sub_syr2k( cntl ) ); 00088 00089 /*------------------------------------------------------------*/ 00090 00091 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00092 FLA_RIGHT ); 00093 00094 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00095 FLA_RIGHT ); 00096 00097 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ C01, C02, 00098 /* ************** */ /* ****************** */ 00099 C10, /**/ C11, C12, 00100 &CBL, /**/ &CBR, C20, /**/ C21, C22, 00101 FLA_BR ); 00102 00103 } 00104 00105 return FLA_SUCCESS; 00106 }
FLA_Error FLA_Syr2k_ut_blk_var9 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_syr2k_t * | cntl | |||
) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ut().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj BT, B0, 00042 BB, B1, 00043 B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_TOP ); 00049 00050 FLA_Part_2x1( B, &BT, 00051 &BB, 0, FLA_TOP ); 00052 00053 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00054 00055 b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00056 00057 FLA_Repart_2x1_to_3x1( AT, &A0, 00058 /* ** */ /* ** */ 00059 &A1, 00060 AB, &A2, b, FLA_BOTTOM ); 00061 00062 FLA_Repart_2x1_to_3x1( BT, &B0, 00063 /* ** */ /* ** */ 00064 &B1, 00065 BB, &B2, b, FLA_BOTTOM ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 /* C = C + A1' * B1 + B1' * A1 */ 00070 FLA_Syr2k_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 00071 alpha, A1, B1, beta, C, 00072 FLA_Cntl_sub_syr2k( cntl ) ); 00073 00074 /*------------------------------------------------------------*/ 00075 00076 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00077 A1, 00078 /* ** */ /* ** */ 00079 &AB, A2, FLA_TOP ); 00080 00081 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00082 B1, 00083 /* ** */ /* ** */ 00084 &BB, B2, FLA_TOP ); 00085 00086 } 00087 00088 return FLA_SUCCESS; 00089 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_TL ); 00054 00055 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2, 00059 1, FLA_RIGHT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2, 00062 1, FLA_RIGHT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02, 00065 /* ************* */ /* ************************** */ 00066 &c10t, /**/ &gamma11, &c12t, 00067 CBL, /**/ CBR, &C20, /**/ &c21, &C22, 00068 1, 1, FLA_BR ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c12t = c12t + a1t' * B2 */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1t, beta, c12t ); 00074 00075 /* c12t = c12t + b1t' * A2 */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1t, FLA_ONE, c12t ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2, 00084 FLA_LEFT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2, 00087 FLA_LEFT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02, 00090 c10t, gamma11, /**/ c12t, 00091 /* ************** */ /* ************************ */ 00092 &CBL, /**/ &CBR, C20, c21, /**/ C22, 00093 FLA_TL ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr2_external().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, a1t, 00041 A2; 00042 00043 FLA_Obj BT, B0, 00044 BB, b1t, 00045 B2; 00046 00047 FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C ); 00048 00049 FLA_Part_2x1( A, &AT, 00050 &AB, 0, FLA_BOTTOM ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_BOTTOM ); 00054 00055 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00056 00057 00058 FLA_Repart_2x1_to_3x1( AT, &A0, 00059 &a1t, 00060 /* ** */ /* ** */ 00061 AB, &A2, 1, FLA_TOP ); 00062 00063 FLA_Repart_2x1_to_3x1( BT, &B0, 00064 &b1t, 00065 /* ** */ /* ** */ 00066 BB, &B2, 1, FLA_TOP ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* C = C + a1t' * b1t + b1t' * a1t */ 00071 FLA_Syr2_external( FLA_UPPER_TRIANGULAR, alpha, a1t, b1t, C ); 00072 00073 /*------------------------------------------------------------*/ 00074 00075 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00076 /* ** */ /* ** */ 00077 a1t, 00078 &AB, A2, FLA_BOTTOM ); 00079 00080 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00081 /* ** */ /* ** */ 00082 b1t, 00083 &BB, B2, FLA_BOTTOM ); 00084 00085 } 00086 00087 return FLA_SUCCESS; 00088 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_TL ); 00054 00055 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2, 00059 1, FLA_RIGHT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2, 00062 1, FLA_RIGHT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02, 00065 /* ************* */ /* ************************** */ 00066 &c10t, /**/ &gamma11, &c12t, 00067 CBL, /**/ CBR, &C20, /**/ &c21, &C22, 00068 1, 1, FLA_BR ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c01 = c01 + B0' * a1t */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1t, beta, c01 ); 00074 00075 /* c12t = c12t + a1t' * B2 */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1t, FLA_ONE, c12t ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2, 00084 FLA_LEFT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2, 00087 FLA_LEFT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02, 00090 c10t, gamma11, /**/ c12t, 00091 /* ************** */ /* ************************ */ 00092 &CBL, /**/ &CBR, C20, c21, /**/ C22, 00093 FLA_TL ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00036 { 00037 FLA_Obj AL, AR, A0, a1t, A2; 00038 00039 FLA_Obj BL, BR, B0, b1t, B2; 00040 00041 FLA_Obj CTL, CTR, C00, c01, C02, 00042 CBL, CBR, c10t, gamma11, c12t, 00043 C20, c21, C22; 00044 00045 00046 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00047 00048 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00049 00050 FLA_Part_2x2( C, &CTL, &CTR, 00051 &CBL, &CBR, 0, 0, FLA_TL ); 00052 00053 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00054 00055 00056 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2, 00057 1, FLA_RIGHT ); 00058 00059 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2, 00060 1, FLA_RIGHT ); 00061 00062 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02, 00063 /* ************* */ /* ************************** */ 00064 &c10t, /**/ &gamma11, &c12t, 00065 CBL, /**/ CBR, &C20, /**/ &c21, &C22, 00066 1, 1, FLA_BR ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* c01 = c01 + A0' * b1t */ 00071 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1t, beta, c01 ); 00072 00073 /* c12t = c12t + b1t' * A2 */ 00074 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1t, FLA_ONE, c12t ); 00075 00076 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00077 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00078 00079 /*------------------------------------------------------------*/ 00080 00081 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2, 00082 FLA_LEFT ); 00083 00084 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2, 00085 FLA_LEFT ); 00086 00087 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02, 00088 c10t, gamma11, /**/ c12t, 00089 /* ************** */ /* ************************ */ 00090 &CBL, /**/ &CBR, C20, c21, /**/ C22, 00091 FLA_TL ); 00092 00093 } 00094 00095 return FLA_SUCCESS; 00096 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_TL ); 00054 00055 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1t, &A2, 00059 1, FLA_RIGHT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1t, &B2, 00062 1, FLA_RIGHT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &c01, &C02, 00065 /* ************* */ /* ************************** */ 00066 &c10t, /**/ &gamma11, &c12t, 00067 CBL, /**/ CBR, &C20, /**/ &c21, &C22, 00068 1, 1, FLA_BR ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c01 = c01 + A0' * b1t */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1t, beta, c01 ); 00074 00075 /* c01 = c01 + B0' * a1t */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1t, FLA_ONE, c01 ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1t, /**/ A2, 00084 FLA_LEFT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1t, /**/ B2, 00087 FLA_LEFT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, c01, /**/ C02, 00090 c10t, gamma11, /**/ c12t, 00091 /* ************** */ /* ************************ */ 00092 &CBL, /**/ &CBR, C20, c21, /**/ C22, 00093 FLA_TL ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_BR ); 00054 00055 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1t, /**/ &A2, 00059 1, FLA_LEFT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2, 00062 1, FLA_LEFT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02, 00065 &c10t, &gamma11, /**/ &c12t, 00066 /* ************* */ /* ************************** */ 00067 CBL, /**/ CBR, &C20, &c21, /**/ &C22, 00068 1, 1, FLA_TL ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c01 = c01 + A0' * b1t */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1t, beta, c01 ); 00074 00075 /* c01 = c01 + B0' * a1t */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1t, FLA_ONE, c01 ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1t, A2, 00084 FLA_RIGHT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2, 00087 FLA_RIGHT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02, 00090 /* ************** */ /* ************************ */ 00091 c10t, /**/ gamma11, c12t, 00092 &CBL, /**/ &CBR, C20, /**/ c21, C22, 00093 FLA_BR ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_BR ); 00054 00055 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1t, /**/ &A2, 00059 1, FLA_LEFT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2, 00062 1, FLA_LEFT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02, 00065 &c10t, &gamma11, /**/ &c12t, 00066 /* ************* */ /* ************************** */ 00067 CBL, /**/ CBR, &C20, &c21, /**/ &C22, 00068 1, 1, FLA_TL ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c01 = c01 + A0' * b1t */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A0, b1t, beta, c01 ); 00074 00075 /* c12t = c12t + b1t' * A2 */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1t, FLA_ONE, c12t ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1t, A2, 00084 FLA_RIGHT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2, 00087 FLA_RIGHT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02, 00090 /* ************** */ /* ************************ */ 00091 c10t, /**/ gamma11, c12t, 00092 &CBL, /**/ &CBR, C20, /**/ c21, C22, 00093 FLA_BR ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_BR ); 00054 00055 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1t, /**/ &A2, 00059 1, FLA_LEFT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2, 00062 1, FLA_LEFT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02, 00065 &c10t, &gamma11, /**/ &c12t, 00066 /* ************* */ /* ************************** */ 00067 CBL, /**/ CBR, &C20, &c21, /**/ &C22, 00068 1, 1, FLA_TL ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c01 = c01 + B0' * a1t */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B0, a1t, beta, c01 ); 00074 00075 /* c12t = c12t + a1t' * B2 */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1t, FLA_ONE, c12t ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1t, A2, 00084 FLA_RIGHT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2, 00087 FLA_RIGHT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02, 00090 /* ************** */ /* ************************ */ 00091 c10t, /**/ gamma11, c12t, 00092 &CBL, /**/ &CBR, C20, /**/ c21, C22, 00093 FLA_BR ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Dot2s_external(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().
00038 { 00039 FLA_Obj AL, AR, A0, a1t, A2; 00040 00041 FLA_Obj BL, BR, B0, b1t, B2; 00042 00043 FLA_Obj CTL, CTR, C00, c01, C02, 00044 CBL, CBR, c10t, gamma11, c12t, 00045 C20, c21, C22; 00046 00047 00048 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 FLA_Part_2x2( C, &CTL, &CTR, 00053 &CBL, &CBR, 0, 0, FLA_BR ); 00054 00055 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00056 00057 00058 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1t, /**/ &A2, 00059 1, FLA_LEFT ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1t, /**/ &B2, 00062 1, FLA_LEFT ); 00063 00064 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, &c01, /**/ &C02, 00065 &c10t, &gamma11, /**/ &c12t, 00066 /* ************* */ /* ************************** */ 00067 CBL, /**/ CBR, &C20, &c21, /**/ &C22, 00068 1, 1, FLA_TL ); 00069 00070 /*------------------------------------------------------------*/ 00071 00072 /* c12t = c12t + a1t' * B2 */ 00073 FLA_Gemv_external( FLA_TRANSPOSE, alpha, B2, a1t, beta, c12t ); 00074 00075 /* c12t = c12t + b1t' * A2 */ 00076 FLA_Gemv_external( FLA_TRANSPOSE, alpha, A2, b1t, FLA_ONE, c12t ); 00077 00078 /* gamma11 = gamma11 + a1t' * b1t + b1t' * a1t */ 00079 FLA_Dot2s_external( alpha, a1t, b1t, beta, gamma11 ); 00080 00081 /*------------------------------------------------------------*/ 00082 00083 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1t, A2, 00084 FLA_RIGHT ); 00085 00086 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1t, B2, 00087 FLA_RIGHT ); 00088 00089 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, /**/ c01, C02, 00090 /* ************** */ /* ************************ */ 00091 c10t, /**/ gamma11, c12t, 00092 &CBL, /**/ &CBR, C20, /**/ c21, C22, 00093 FLA_BR ); 00094 00095 } 00096 00097 return FLA_SUCCESS; 00098 }
References FLA_Cont_with_3x1_to_2x1(), FLA_Obj_length(), FLA_Part_2x1(), FLA_Repart_2x1_to_3x1(), FLA_Scalr_external(), and FLA_Syr2_external().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, a1t, 00041 A2; 00042 00043 FLA_Obj BT, B0, 00044 BB, b1t, 00045 B2; 00046 00047 FLA_Scalr_external( FLA_UPPER_TRIANGULAR, beta, C ); 00048 00049 FLA_Part_2x1( A, &AT, 00050 &AB, 0, FLA_TOP ); 00051 00052 FLA_Part_2x1( B, &BT, 00053 &BB, 0, FLA_TOP ); 00054 00055 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00056 00057 00058 FLA_Repart_2x1_to_3x1( AT, &A0, 00059 /* ** */ /* ** */ 00060 &a1t, 00061 AB, &A2, 1, FLA_BOTTOM ); 00062 00063 FLA_Repart_2x1_to_3x1( BT, &B0, 00064 /* ** */ /* ** */ 00065 &b1t, 00066 BB, &B2, 1, FLA_BOTTOM ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 /* C = C + a1t' * b1t + b1t' * a1t */ 00071 FLA_Syr2_external( FLA_UPPER_TRIANGULAR, alpha, a1t, b1t, C ); 00072 00073 /*------------------------------------------------------------*/ 00074 00075 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00076 a1t, 00077 /* ** */ /* ** */ 00078 &AB, A2, FLA_TOP ); 00079 00080 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00081 b1t, 00082 /* ** */ /* ** */ 00083 &BB, B2, FLA_TOP ); 00084 00085 } 00086 00087 return FLA_SUCCESS; 00088 }