libflame
12600
|
Functions | |
FLA_Error | FLA_Gemm_ht_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t *cntl) |
FLA_Error FLA_Gemm_ht_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_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().
Referenced by FLA_Gemm_ht().
{ FLA_Obj AT, A0, AB, A1, A2; FLA_Obj BL, BR, B0, B1, B2; dim_t b; FLA_Scal_internal( beta, C, FLA_Cntl_sub_scal( cntl ) ); FLA_Part_2x1( A, &AT, &AB, 0, FLA_TOP ); FLA_Part_1x2( B, &BL, &BR, 0, FLA_LEFT ); while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){ b = FLA_Determine_blocksize( AB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) ); FLA_Repart_2x1_to_3x1( AT, &A0, /* ** */ /* ** */ &A1, AB, &A2, b, FLA_BOTTOM ); FLA_Repart_1x2_to_1x3( BL, /**/ BR, &B0, /**/ &B1, &B2, b, FLA_RIGHT ); /*------------------------------------------------------------*/ /* C = alpha * A1' * B1' + C; */ FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_TRANSPOSE, alpha, A1, B1, FLA_ONE, C, FLA_Cntl_sub_gemm( cntl ) ); /*------------------------------------------------------------*/ FLA_Cont_with_3x1_to_2x1( &AT, A0, A1, /* ** */ /* ** */ &AB, A2, FLA_TOP ); FLA_Cont_with_1x3_to_1x2( &BL, /**/ &BR, B0, B1, /**/ B2, FLA_LEFT ); } return FLA_SUCCESS; }