Functions | |
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) |
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 }