Go to the source code of this file.
FLA_Error FLA_Gemm_th_blk_var1 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00036 { 00037 FLA_Obj AL, AR, A0, A1, A2; 00038 00039 FLA_Obj CT, C0, 00040 CB, C1, 00041 C2; 00042 00043 dim_t b; 00044 00045 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00046 00047 FLA_Part_2x1( C, &CT, 00048 &CB, 0, FLA_TOP ); 00049 00050 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00051 00052 b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, 00055 b, FLA_RIGHT ); 00056 00057 FLA_Repart_2x1_to_3x1( CT, &C0, 00058 /* ** */ /* ** */ 00059 &C1, 00060 CB, &C2, b, FLA_BOTTOM ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C1 = alpha * A1' * B' + C1; */ 00065 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00066 alpha, A1, B, beta, C1, 00067 FLA_Cntl_sub_gemm( cntl ) ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, 00072 FLA_LEFT ); 00073 00074 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00075 C1, 00076 /* ** */ /* ** */ 00077 &CB, C2, FLA_TOP ); 00078 00079 } 00080 00081 return FLA_SUCCESS; 00082 }
FLA_Error FLA_Gemm_th_blk_var2 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00038 { 00039 FLA_Obj AL, AR, A0, A1, A2; 00040 00041 FLA_Obj CT, C0, 00042 CB, C1, 00043 C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00048 00049 FLA_Part_2x1( C, &CT, 00050 &CB, 0, FLA_BOTTOM ); 00051 00052 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00053 00054 b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &A1, /**/ &A2, 00057 b, FLA_LEFT ); 00058 00059 FLA_Repart_2x1_to_3x1( CT, &C0, 00060 &C1, 00061 /* ** */ /* ** */ 00062 CB, &C2, b, FLA_TOP ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C1 = alpha * A1' * B' + C1; */ 00067 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00068 alpha, A1, B, beta, C1, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2, 00074 FLA_RIGHT ); 00075 00076 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00077 /* ** */ /* ** */ 00078 C1, 00079 &CB, C2, FLA_BOTTOM ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }
FLA_Error FLA_Gemm_th_blk_var3 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00036 { 00037 FLA_Obj BT, B0, 00038 BB, B1, 00039 B2; 00040 00041 FLA_Obj CL, CR, C0, C1, C2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x1( B, &BT, 00046 &BB, 0, FLA_TOP ); 00047 00048 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){ 00051 00052 b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x1_to_3x1( BT, &B0, 00055 /* ** */ /* ** */ 00056 &B1, 00057 BB, &B2, b, FLA_BOTTOM ); 00058 00059 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, 00060 b, FLA_RIGHT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C1 = alpha * A' * B1' + C1; */ 00065 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00066 alpha, A, B1, beta, C1, 00067 FLA_Cntl_sub_gemm( cntl ) ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00072 B1, 00073 /* ** */ /* ** */ 00074 &BB, B2, FLA_TOP ); 00075 00076 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, 00077 FLA_LEFT ); 00078 00079 } 00080 00081 return FLA_SUCCESS; 00082 }
FLA_Error FLA_Gemm_th_blk_var4 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, B1, 00041 B2; 00042 00043 FLA_Obj CL, CR, C0, C1, C2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( B, &BT, 00048 &BB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){ 00053 00054 b = FLA_Determine_blocksize( BT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x1_to_3x1( BT, &B0, 00057 &B1, 00058 /* ** */ /* ** */ 00059 BB, &B2, b, FLA_TOP ); 00060 00061 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &C1, /**/ &C2, 00062 b, FLA_LEFT ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C1 = alpha * A' * B1' + C1; */ 00067 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00068 alpha, A, B1, beta, C1, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00074 /* ** */ /* ** */ 00075 B1, 00076 &BB, B2, FLA_BOTTOM ); 00077 00078 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ C1, C2, 00079 FLA_RIGHT ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }
FLA_Error FLA_Gemm_th_blk_var5 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00036 { 00037 FLA_Obj AT, A0, 00038 AB, A1, 00039 A2; 00040 00041 FLA_Obj BL, BR, B0, B1, B2; 00042 00043 dim_t b; 00044 00045 FLA_Part_2x1( A, &AT, 00046 &AB, 0, FLA_TOP ); 00047 00048 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00049 00050 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00051 00052 b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00053 00054 FLA_Repart_2x1_to_3x1( AT, &A0, 00055 /* ** */ /* ** */ 00056 &A1, 00057 AB, &A2, b, FLA_BOTTOM ); 00058 00059 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, 00060 b, FLA_RIGHT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C = alpha * A1' * B1' + C; */ 00065 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00066 alpha, A1, B1, beta, C, 00067 FLA_Cntl_sub_gemm( cntl ) ); 00068 00069 /*------------------------------------------------------------*/ 00070 00071 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00072 A1, 00073 /* ** */ /* ** */ 00074 &AB, A2, FLA_TOP ); 00075 00076 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, 00077 FLA_LEFT ); 00078 00079 } 00080 00081 return FLA_SUCCESS; 00082 }
FLA_Error FLA_Gemm_th_blk_var6 | ( | FLA_Obj | alpha, | |
FLA_Obj | A, | |||
FLA_Obj | B, | |||
FLA_Obj | beta, | |||
FLA_Obj | C, | |||
fla_gemm_t * | cntl | |||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemm_th().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, A1, 00041 A2; 00042 00043 FLA_Obj BL, BR, B0, B1, B2; 00044 00045 dim_t b; 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00053 00054 b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) ); 00055 00056 FLA_Repart_2x1_to_3x1( AT, &A0, 00057 &A1, 00058 /* ** */ /* ** */ 00059 AB, &A2, b, FLA_TOP ); 00060 00061 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &B1, /**/ &B2, 00062 b, FLA_LEFT ); 00063 00064 /*------------------------------------------------------------*/ 00065 00066 /* C = alpha * A1' * B1' + C; */ 00067 FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_TRANSPOSE, 00068 alpha, A1, B1, beta, C, 00069 FLA_Cntl_sub_gemm( cntl ) ); 00070 00071 /*------------------------------------------------------------*/ 00072 00073 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00074 /* ** */ /* ** */ 00075 A1, 00076 &AB, A2, FLA_BOTTOM ); 00077 00078 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ B1, B2, 00079 FLA_RIGHT ); 00080 00081 } 00082 00083 return FLA_SUCCESS; 00084 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj AL, AR, A0, a1, A2; 00040 00041 FLA_Obj CT, C0, 00042 CB, c1t, 00043 C2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); 00048 00049 FLA_Part_2x1( C, &CT, 00050 &CB, 0, FLA_TOP ); 00051 00052 while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ 00053 00054 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &a1, &A2, 00055 1, FLA_RIGHT ); 00056 00057 FLA_Repart_2x1_to_3x1( CT, &C0, 00058 /* ** */ /* *** */ 00059 &c1t, 00060 CB, &C2, 1, FLA_BOTTOM ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* c1t = a1' * B' + c1t */ 00065 /* c1t' = B * a1 + c1t' */ 00066 FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B, a1, FLA_ONE, c1t ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, a1, /**/ A2, 00071 FLA_LEFT ); 00072 00073 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00074 c1t, 00075 /* ** */ /* *** */ 00076 &CB, C2, FLA_TOP ); 00077 00078 } 00079 00080 return FLA_SUCCESS; 00081 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj AL, AR, A0, a1, A2; 00040 00041 FLA_Obj CT, C0, 00042 CB, c1t, 00043 C2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT ); 00048 00049 FLA_Part_2x1( C, &CT, 00050 &CB, 0, FLA_BOTTOM ); 00051 00052 while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){ 00053 00054 FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, &a1, /**/ &A2, 00055 1, FLA_LEFT ); 00056 00057 FLA_Repart_2x1_to_3x1( CT, &C0, 00058 &c1t, 00059 /* ** */ /* *** */ 00060 CB, &C2, 1, FLA_TOP ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* c1t = a1' * B' + c1t */ 00065 /* c1t' = B * a1 + c1t' */ 00066 FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B, a1, FLA_ONE, c1t ); 00067 00068 /*------------------------------------------------------------*/ 00069 00070 FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ a1, A2, 00071 FLA_RIGHT ); 00072 00073 FLA_Cont_with_3x1_to_2x1( &CT, C0, 00074 /* ** */ /* *** */ 00075 c1t, 00076 &CB, C2, FLA_BOTTOM ); 00077 00078 } 00079 00080 return FLA_SUCCESS; 00081 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, b1t, 00041 B2; 00042 00043 FLA_Obj CL, CR, C0, c1, C2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_2x1( B, &BT, 00048 &BB, 0, FLA_TOP ); 00049 00050 FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); 00051 00052 while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){ 00053 00054 FLA_Repart_2x1_to_3x1( BT, &B0, 00055 /* ** */ /* *** */ 00056 &b1t, 00057 BB, &B2, 1, FLA_BOTTOM ); 00058 00059 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &c1, &C2, 00060 1, FLA_RIGHT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* c1 = A' * b1t' + c1 */ 00065 FLA_Gemv_external( FLA_CONJ_TRANSPOSE, alpha, A, b1t, FLA_ONE, c1 ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00070 b1t, 00071 /* ** */ /* *** */ 00072 &BB, B2, FLA_TOP ); 00073 00074 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, c1, /**/ C2, 00075 FLA_LEFT ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Gemv_external(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj BT, B0, 00040 BB, b1t, 00041 B2; 00042 00043 FLA_Obj CL, CR, C0, c1, C2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_2x1( B, &BT, 00048 &BB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( C, &CL, &CR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( BB ) < FLA_Obj_length( B ) ){ 00053 00054 FLA_Repart_2x1_to_3x1( BT, &B0, 00055 &b1t, 00056 /* ** */ /* *** */ 00057 BB, &B2, 1, FLA_TOP ); 00058 00059 FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, &c1, /**/ &C2, 00060 1, FLA_LEFT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* c1 = A' * b1t' + c1 */ 00065 FLA_Gemv_external( FLA_CONJ_TRANSPOSE, alpha, A, b1t, FLA_ONE, c1 ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00070 /* ** */ /* *** */ 00071 b1t, 00072 &BB, B2, FLA_BOTTOM ); 00073 00074 FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, /**/ c1, C2, 00075 FLA_RIGHT ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, a1t, 00041 A2; 00042 00043 FLA_Obj BL, BR, B0, b1, B2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_TOP ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); 00051 00052 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00053 00054 FLA_Repart_2x1_to_3x1( AT, &A0, 00055 /* ** */ /* *** */ 00056 &a1t, 00057 AB, &A2, 1, FLA_BOTTOM ); 00058 00059 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &b1, &B2, 00060 1, FLA_RIGHT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C = a1t' * b1' + C */ 00065 FLA_Ger_external( alpha, a1t, b1, C ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00070 a1t, 00071 /* ** */ /* *** */ 00072 &AB, A2, FLA_TOP ); 00073 00074 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, b1, /**/ B2, 00075 FLA_LEFT ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Ger_external(), FLA_Obj_length(), FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_external().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, a1t, 00041 A2; 00042 00043 FLA_Obj BL, BR, B0, b1, B2; 00044 00045 FLA_Scal_external( beta, C ); 00046 00047 FLA_Part_2x1( A, &AT, 00048 &AB, 0, FLA_BOTTOM ); 00049 00050 FLA_Part_1x2( B, &BL, &BR, 0, FLA_RIGHT ); 00051 00052 while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){ 00053 00054 FLA_Repart_2x1_to_3x1( AT, &A0, 00055 &a1t, 00056 /* ** */ /* *** */ 00057 AB, &A2, 1, FLA_TOP ); 00058 00059 FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, &b1, /**/ &B2, 00060 1, FLA_LEFT ); 00061 00062 /*------------------------------------------------------------*/ 00063 00064 /* C = a1t' * b1' + C */ 00065 FLA_Ger_external( alpha, a1t, b1, C ); 00066 00067 /*------------------------------------------------------------*/ 00068 00069 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00070 /* ** */ /* *** */ 00071 a1t, 00072 &AB, A2, FLA_BOTTOM ); 00073 00074 FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, /**/ b1, B2, 00075 FLA_RIGHT ); 00076 00077 } 00078 00079 return FLA_SUCCESS; 00080 }