Functions | |
FLA_Error | FLA_Syr2k_ln_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_syr2k_t *cntl) |
FLA_Error FLA_Syr2k_ln_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_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Syr2k_internal().
Referenced by FLA_Syr2k_ln().
00038 { 00039 FLA_Obj AT, A0, 00040 AB, A1, 00041 A2; 00042 00043 FLA_Obj BT, B0, 00044 BB, B1, 00045 B2; 00046 00047 FLA_Obj CTL, CTR, C00, C01, C02, 00048 CBL, CBR, C10, C11, C12, 00049 C20, C21, C22; 00050 00051 dim_t b; 00052 00053 FLA_Part_2x1( A, &AT, 00054 &AB, 0, FLA_TOP ); 00055 00056 FLA_Part_2x1( B, &BT, 00057 &BB, 0, FLA_TOP ); 00058 00059 FLA_Part_2x2( C, &CTL, &CTR, 00060 &CBL, &CBR, 0, 0, FLA_TL ); 00061 00062 while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ 00063 00064 b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); 00065 00066 FLA_Repart_2x1_to_3x1( AT, &A0, 00067 /* ** */ /* ** */ 00068 &A1, 00069 AB, &A2, b, FLA_BOTTOM ); 00070 00071 FLA_Repart_2x1_to_3x1( BT, &B0, 00072 /* ** */ /* ** */ 00073 &B1, 00074 BB, &B2, b, FLA_BOTTOM ); 00075 00076 FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, 00077 /* ************* */ /* ******************** */ 00078 &C10, /**/ &C11, &C12, 00079 CBL, /**/ CBR, &C20, /**/ &C21, &C22, 00080 b, b, FLA_BR ); 00081 00082 /*------------------------------------------------------------*/ 00083 00084 /* C10 = C10 + B1 * A0' */ 00085 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00086 alpha, B1, A0, beta, C10, 00087 FLA_Cntl_sub_gemm1( cntl ) ); 00088 00089 /* C21 = C21 + A2 * B1' */ 00090 FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 00091 alpha, A2, B1, FLA_ONE, C21, 00092 FLA_Cntl_sub_gemm2( cntl ) ); 00093 00094 /* C11 = C11 + A1 * B1' + B1 * A1' */ 00095 FLA_Syr2k_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 00096 alpha, A1, B1, beta, C11, 00097 FLA_Cntl_sub_syr2k( cntl ) ); 00098 00099 /*------------------------------------------------------------*/ 00100 00101 FLA_Cont_with_3x1_to_2x1( &AT, A0, 00102 A1, 00103 /* ** */ /* ** */ 00104 &AB, A2, FLA_TOP ); 00105 00106 FLA_Cont_with_3x1_to_2x1( &BT, B0, 00107 B1, 00108 /* ** */ /* ** */ 00109 &BB, B2, FLA_TOP ); 00110 00111 FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, 00112 C10, C11, /**/ C12, 00113 /* ************** */ /* ****************** */ 00114 &CBL, /**/ &CBR, C20, C21, /**/ C22, 00115 FLA_TL ); 00116 00117 } 00118 00119 return FLA_SUCCESS; 00120 }