libflame revision_anchor
|
Functions | |
FLA_Error | FLA_Her2k_uh_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl) |
FLA_Error FLA_Her2k_uh_blk_var3 | ( | FLA_Obj | alpha, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | beta, | ||
FLA_Obj | C, | ||
fla_her2k_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Her2k_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_internal().
Referenced by FLA_Her2k_uh().
{ FLA_Obj AL, AR, A0, A1, A2; FLA_Obj BL, BR, B0, B1, B2; FLA_Obj CTL, CTR, C00, C01, C02, CBL, CBR, C10, C11, C12, C20, C21, C22; dim_t b; FLA_Scalr_internal( FLA_UPPER_TRIANGULAR, beta, C, FLA_Cntl_sub_scalr( cntl ) ); FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT ); FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); FLA_Part_2x2( C, &CTL, &CTR, &CBL, &CBR, 0, 0, FLA_TL ); while ( FLA_Obj_width( AL ) < FLA_Obj_width( A ) ){ b = FLA_Determine_blocksize( AR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_1x2_to_1x3( AL, /**/ AR, &A0, /**/ &A1, &A2, b, FLA_RIGHT ); FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, b, FLA_RIGHT ); FLA_Repart_2x2_to_3x3( CTL, /**/ CTR, &C00, /**/ &C01, &C02, /* ************* */ /* ******************** */ &C10, /**/ &C11, &C12, CBL, /**/ CBR, &C20, /**/ &C21, &C22, b, b, FLA_BR ); /*------------------------------------------------------------*/ /* C01 = C01 + A0' * B1 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, A0, B1, FLA_ONE, C01, FLA_Cntl_sub_gemm1( cntl ) ); /* C12 = C12 + B1' * A2 */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, alpha, B1, A2, FLA_ONE, C12, FLA_Cntl_sub_gemm2( cntl ) ); /* C11 = C11 + A1' * B1 + B1' * A1 */ FLA_Her2k_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, alpha, A1, B1, FLA_ONE, C11, FLA_Cntl_sub_her2k( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2, FLA_LEFT ); FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, FLA_LEFT ); FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR, C00, C01, /**/ C02, C10, C11, /**/ C12, /* ************** */ /* ****************** */ &CBL, /**/ &CBR, C20, C21, /**/ C22, FLA_TL ); } return FLA_SUCCESS; }