libflame  revision_anchor
Functions
FLA_Her2k_un_blk_var6.c File Reference

(r)

Functions

FLA_Error FLA_Her2k_un_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_her2k_t *cntl)

Function Documentation

FLA_Error FLA_Her2k_un_blk_var6 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
FLA_Obj  beta,
FLA_Obj  C,
fla_her2k_t cntl 
)

References FLA_Cont_with_3x1_to_2x1(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Her2k_internal(), FLA_Obj_length(), FLA_ONE, FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_2x1_to_3x1(), FLA_Repart_2x2_to_3x3(), and FLA_Scalr_internal().

Referenced by FLA_Her2k_un().

{
  FLA_Obj AT,              A0,
          AB,              A1,
                           A2;

  FLA_Obj BT,              B0,
          BB,              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_2x1( A,    &AT, 
                      &AB,            0, FLA_BOTTOM );

  FLA_Part_2x1( B,    &BT, 
                      &BB,            0, FLA_BOTTOM );

  FLA_Part_2x2( C,    &CTL, &CTR,
                      &CBL, &CBR,     0, 0, FLA_BR );

  while ( FLA_Obj_length( AB ) < FLA_Obj_length( A ) ){

    b = FLA_Determine_blocksize( AT, FLA_TOP, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_2x1_to_3x1( AT,                &A0, 
                                              &A1, 
                        /* ** */            /* ** */
                           AB,                &A2,        b, FLA_TOP );

    FLA_Repart_2x1_to_3x1( BT,                &B0, 
                                              &B1, 
                        /* ** */            /* ** */
                           BB,                &B2,        b, FLA_TOP );

    FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00, &C01, /**/ &C02,
                                                &C10, &C11, /**/ &C12,
                        /* ************* */   /* ******************** */
                           CBL, /**/ CBR,       &C20, &C21, /**/ &C22,
                           b, b, FLA_TL );

    /*------------------------------------------------------------*/

    /* C01 = C01 + A0 * B1' */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_TRANSPOSE, 
                       alpha, A0, B1, FLA_ONE, C01,
                       FLA_Cntl_sub_gemm1( cntl ) );

    /* C12 = C12 + B1 * A2' */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_CONJ_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_NO_TRANSPOSE, 
                        alpha, A1, B1, FLA_ONE, C11,
                        FLA_Cntl_sub_her2k( cntl ) );

    /*------------------------------------------------------------*/

    FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
                            /* ** */           /* ** */
                                                  A1, 
                              &AB,                A2,     FLA_BOTTOM );

    FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
                            /* ** */           /* ** */
                                                  B1, 
                              &BB,                B2,     FLA_BOTTOM );

    FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00, /**/ C01, C02,
                            /* ************** */  /* ****************** */
                                                     C10, /**/ C11, C12,
                              &CBL, /**/ &CBR,       C20, /**/ C21, C22,
                              FLA_BR );

  }

  return FLA_SUCCESS;
}