libflame  revision_anchor
Functions
FLA_Herk_uh_blk_var2.c File Reference

(r)

Functions

FLA_Error FLA_Herk_uh_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_herk_t *cntl)

Function Documentation

FLA_Error FLA_Herk_uh_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  beta,
FLA_Obj  C,
fla_herk_t cntl 
)

References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x3_to_2x2(), FLA_Determine_blocksize(), FLA_Gemm_internal(), FLA_Herk_internal(), FLA_Obj_width(), FLA_Part_1x2(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), and FLA_Repart_2x2_to_3x3().

Referenced by FLA_Herk_uh().

{
  FLA_Obj AL,    AR,       A0,  A1,  A2;

  FLA_Obj CTL,   CTR,      C00, C01, C02, 
          CBL,   CBR,      C10, C11, C12,
                           C20, C21, C22;

  dim_t b;

  FLA_Part_1x2( A,    &AL,  &AR,      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_2x2_to_3x3( CTL, /**/ CTR,       &C00, /**/ &C01, &C02,
                        /* ************* */   /* ******************** */
                                                &C10, /**/ &C11, &C12,
                           CBL, /**/ CBR,       &C20, /**/ &C21, &C22,
                           b, b, FLA_BR );

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

    /* C01 = C01 + A0' * A1 */
    FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A0, A1, beta, C01,
                       FLA_Cntl_sub_gemm( cntl ) );

    /* C11 = C11 + A1' * A1 */
    FLA_Herk_internal( FLA_UPPER_TRIANGULAR, FLA_CONJ_TRANSPOSE, 
                       alpha, A1, beta, C11,
                       FLA_Cntl_sub_herk( cntl ) );

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

    FLA_Cont_with_1x3_to_1x2( &AL,  /**/ &AR,        A0, A1, /**/ A2,
                              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;
}