libflame revision_anchor
|
Functions | |
FLA_Error | FLA_Sylv_hn_blk_var18 (FLA_Obj isgn, FLA_Obj A, FLA_Obj B, FLA_Obj C, FLA_Obj scale, fla_sylv_t *cntl) |
FLA_Error FLA_Sylv_hn_blk_var18 | ( | FLA_Obj | isgn, |
FLA_Obj | A, | ||
FLA_Obj | B, | ||
FLA_Obj | C, | ||
FLA_Obj | scale, | ||
fla_sylv_t * | cntl | ||
) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x2_to_3x3(), and FLA_Sylv_internal().
Referenced by FLA_Sylv_hn().
{ FLA_Obj BTL, BTR, B00, B01, B02, BBL, BBR, B10, B11, B12, B20, B21, B22; FLA_Obj CL, CR, C0, C1, C2; dim_t b; FLA_Part_2x2( B, &BTL, &BTR, &BBL, &BBR, 0, 0, FLA_TL ); FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT ); while ( FLA_Obj_length( BTL ) < FLA_Obj_length( B ) ){ b = FLA_Determine_blocksize( CR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x2_to_3x3( BTL, /**/ BTR, &B00, /**/ &B01, &B02, /* ************* */ /* ******************** */ &B10, /**/ &B11, &B12, BBL, /**/ BBR, &B20, /**/ &B21, &B22, b, b, FLA_BR ); FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2, b, FLA_RIGHT ); // Loop Invariant: // CL = // CR = /*------------------------------------------------------------*/ // C1 = sylv( A', B11, C1 ); FLA_Sylv_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, isgn, A, B11, C1, scale, FLA_Cntl_sub_sylv1( cntl ) ); // C2 = C2 -/+ C1 * B12; FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, FLA_NEGATE( isgn ), C1, B12, FLA_ONE, C2, FLA_Cntl_sub_gemm1( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x3_to_2x2( &BTL, /**/ &BTR, B00, B01, /**/ B02, B10, B11, /**/ B12, /* ************** */ /* ****************** */ &BBL, /**/ &BBR, B20, B21, /**/ B22, FLA_TL ); FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2, FLA_LEFT ); } return FLA_SUCCESS; }