Пример #1
0
FLA_Error FLA_Her2k_lh_unb_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
  FLA_Obj AL,    AR,       A0,  a1,  A2;

  FLA_Obj BL,    BR,       B0,  b1,  B2;

  FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
          CBL,   CBR,      c10t, gamma11, c12t,
                           C20,  c21,     C22;


  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );

  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 ) ){


    FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, /**/ &a1, &A2,
                           1, FLA_RIGHT );

    FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &b1, &B2,
                           1, FLA_RIGHT );

    FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
                        /* ************* */   /* ************************** */
                                                &c10t, /**/ &gamma11, &c12t,
                           CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
                           1, 1, FLA_BR );

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

    /* c21 = c21 + A2' * b1 */
    FLA_Gemv_external( FLA_CONJ_TRANSPOSE, alpha, A2, b1, FLA_ONE, c21 );    

    /* c21 = c21 + B2' * a1 */
    FLA_Gemv_external( FLA_CONJ_TRANSPOSE, alpha, B2, a1, FLA_ONE, c21 );    

    /* gamma11 = gamma11 + a1' * b1 + b1' * a1 */
    FLA_Dot2cs_external( FLA_CONJUGATE, alpha, a1, b1, FLA_ONE, gamma11 );

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

    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,
                                                     c10t, gamma11, /**/ c12t,
                            /* ************** */  /* ************************ */
                              &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
                              FLA_TL );

  }

  return FLA_SUCCESS;
}
Пример #2
0
FLA_Error FLA_Trsv_un_blk_var2( FLA_Diag diagA, FLA_Obj A, FLA_Obj x, fla_trsv_t* cntl )
{
  FLA_Obj ATL,   ATR,      A00, A01, A02, 
          ABL,   ABR,      A10, A11, A12,
                           A20, A21, A22;

  FLA_Obj xT,              x0,
          xB,              x1,
                           x2;

  dim_t b;

  FLA_Part_2x2( A,    &ATL, &ATR,
                      &ABL, &ABR,     0, 0, FLA_BR );

  FLA_Part_2x1( x,    &xT, 
                      &xB,            0, FLA_BOTTOM );

  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){

    b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, &A01, /**/ &A02,
                                                &A10, &A11, /**/ &A12,
                        /* ************* */   /* ******************** */
                           ABL, /**/ ABR,       &A20, &A21, /**/ &A22,
                           b, b, FLA_TL );

    FLA_Repart_2x1_to_3x1( xT,                &x0, 
                                              &x1, 
                        /* ** */            /* ** */
                           xB,                &x2,        b, FLA_TOP );

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

    /* x1 = triu( A11 ) \ x1; */
    FLA_Trsv_internal( FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, diagA,
                       A11, x1,
                       FLA_Cntl_sub_trsv( cntl ) );

    /* x0 = x0 - A01 * x1; */
    FLA_Gemv_internal( FLA_NO_TRANSPOSE,
                       FLA_MINUS_ONE, A01, x1, FLA_ONE, x0,
                       FLA_Cntl_sub_gemv( cntl ) );

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

    FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, /**/ A01, A02,
                            /* ************** */  /* ****************** */
                                                     A10, /**/ A11, A12,
                              &ABL, /**/ &ABR,       A20, /**/ A21, A22,
                              FLA_BR );

    FLA_Cont_with_3x1_to_2x1( &xT,                x0, 
                            /* ** */           /* ** */
                                                  x1, 
                              &xB,                x2,     FLA_BOTTOM );

  }

  return FLA_SUCCESS;
}
Пример #3
0
FLA_Error FLA_Symm_lu_blk_var7( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_symm_t* cntl )
{
  FLA_Obj ATL,   ATR,      A00, A01, A02, 
          ABL,   ABR,      A10, A11, A12,
                           A20, A21, A22;

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  FLA_Obj CT,              C0,
          CB,              C1,
                           C2;

  dim_t b;

  FLA_Scal_internal( beta, C,
                     FLA_Cntl_sub_scal( cntl ) );

  FLA_Part_2x2( A,    &ATL, &ATR,
                      &ABL, &ABR,     0, 0, FLA_BR );

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

  FLA_Part_2x1( C,    &CT, 
                      &CB,            0, FLA_BOTTOM );

  while ( FLA_Obj_length( ABR ) < FLA_Obj_length( A ) ){

    b = FLA_Determine_blocksize( ATL, FLA_TL, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00, &A01, /**/ &A02,
                                                &A10, &A11, /**/ &A12,
                        /* ************* */   /* ******************** */
                           ABL, /**/ ABR,       &A20, &A21, /**/ &A22,
                           b, b, FLA_TL );

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

    FLA_Repart_2x1_to_3x1( CT,                &C0, 
                                              &C1, 
                        /* ** */            /* ** */
                           CB,                &C2,        b, FLA_TOP );

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

    /* C0 = C0 + A01 * B1 */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A01, B1, FLA_ONE, C0,
                       FLA_Cntl_sub_gemm1( cntl ) );

    /* C1 = C1 + A01' * B0 */
    FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A01, B0, FLA_ONE, C1,
                       FLA_Cntl_sub_gemm2( cntl ) );

    /* C1 = C1 + A11 * B1 */
    FLA_Symm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, 
                       alpha, A11, B1, FLA_ONE, C1,
                       FLA_Cntl_sub_symm( cntl ) );

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

    FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00, /**/ A01, A02,
                            /* ************** */  /* ****************** */
                                                     A10, /**/ A11, A12,
                              &ABL, /**/ &ABR,       A20, /**/ A21, A22,
                              FLA_BR );

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

    FLA_Cont_with_3x1_to_2x1( &CT,                C0, 
                            /* ** */           /* ** */
                                                  C1, 
                              &CB,                C2,     FLA_BOTTOM );

  }

  return FLA_SUCCESS;
}
Пример #4
0
FLA_Error FLA_Symm_rl_unb_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
  FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
          ABL,   ABR,      a10t, alpha11, a12t,
                           A20,  a21,     A22;

  FLA_Obj BL,    BR,       B0,  b1t,  B2;

  FLA_Obj CL,    CR,       C0,  c1t,  C2;

  FLA_Scal_external( beta, C );

  FLA_Part_2x2( A,    &ATL, &ATR,
                      &ABL, &ABR,     0, 0, FLA_TL );

  FLA_Part_1x2( B,    &BL,  &BR,      0, FLA_LEFT );

  FLA_Part_1x2( C,    &CL,  &CR,      0, FLA_LEFT );

  while ( FLA_Obj_length( ATL ) < FLA_Obj_length( A ) ){

    FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
                        /* ************* */   /* ************************** */
                                                &a10t, /**/ &alpha11, &a12t,
                           ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
                           1, 1, FLA_BR );

    FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, /**/ &b1t, &B2,
                           1, FLA_RIGHT );

    FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &c1t, &C2,
                           1, FLA_RIGHT );

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

    /* C0 = C0 + b1t * a10t */
    FLA_Ger_external( alpha, b1t, a10t, C0 );
    
    /* c1t = c1t + b1t * alpha11 */
    FLA_Axpys_external( alpha, alpha11, b1t, FLA_ONE, c1t );

    /* C2 = C2 + b1t * a21' */
    FLA_Ger_external( alpha, b1t, a21, C2 );

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

    FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
                                                     a10t, alpha11, /**/ a12t,
                            /* ************** */  /* ************************ */
                              &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
                              FLA_TL );

    FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, b1t, /**/ B2,
                              FLA_LEFT );

    FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, c1t, /**/ C2,
                              FLA_LEFT );

  }

  return FLA_SUCCESS;
}
Пример #5
0
FLA_Error FLA_QR_UT_unb_var2( FLA_Obj A, FLA_Obj T )
{
  FLA_Obj ATL,   ATR,      A00,  a01,     A02, 
          ABL,   ABR,      a10t, alpha11, a12t,
                           A20,  a21,     A22;

  FLA_Obj TTL,   TTR,      T00,  t01,   T02, 
          TBL,   TBR,      t10t, tau11, t12t,
                           T20,  t21,   T22;


  FLA_Part_2x2( A,    &ATL, &ATR,
                      &ABL, &ABR,     0, 0, FLA_TL );

  FLA_Part_2x2( T,    &TTL, &TTR,
                      &TBL, &TBR,     0, 0, FLA_TL );

  while ( FLA_Obj_min_dim( ABR ) > 0 ){

    FLA_Repart_2x2_to_3x3( ATL, /**/ ATR,       &A00,  /**/ &a01,     &A02,
                        /* ************* */   /* ************************** */
                                                &a10t, /**/ &alpha11, &a12t,
                           ABL, /**/ ABR,       &A20,  /**/ &a21,     &A22,
                           1, 1, FLA_BR );

    FLA_Repart_2x2_to_3x3( TTL, /**/ TTR,       &T00,  /**/ &t01,   &T02,
                        /* ************* */   /* ************************ */
                                                &t10t, /**/ &tau11, &t12t,
                           TBL, /**/ TBR,       &T20,  /**/ &t21,   &T22,
                           1, 1, FLA_BR );

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

    // Compute tau11 and u21 from alpha11 and a21 such that tau11 and u21
    // determine a Householder transform H such that applying H from the
    // left to the column vector consisting of alpha11 and a21 annihilates
    // the entries in a21 (and updates alpha11).
    FLA_Househ2_UT( FLA_LEFT,
                    alpha11,
                    a21, tau11 );

    // / a12t \ =  H / a12t \
    // \ A22  /      \ A22  /
    //
    // where H is formed from tau11 and u21.
    FLA_Apply_H2_UT( FLA_LEFT, tau11, a21, a12t,
                                           A22 );

    // t01 = a10t' + A20' * u21;
    FLA_Copyt_external( FLA_CONJ_TRANSPOSE, a10t, t01 );
    FLA_Gemv_external( FLA_CONJ_TRANSPOSE, FLA_ONE, A20, a21, FLA_ONE, t01 );

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

    FLA_Cont_with_3x3_to_2x2( &ATL, /**/ &ATR,       A00,  a01,     /**/ A02,
                                                     a10t, alpha11, /**/ a12t,
                            /* ************** */  /* ************************ */
                              &ABL, /**/ &ABR,       A20,  a21,     /**/ A22,
                              FLA_TL );

    FLA_Cont_with_3x3_to_2x2( &TTL, /**/ &TTR,       T00,  t01,   /**/ T02,
                                                     t10t, tau11, /**/ t12t,
                            /* ************** */  /* ********************** */
                              &TBL, /**/ &TBR,       T20,  t21,   /**/ T22,
                              FLA_TL );

  }

  return FLA_SUCCESS;
}
Пример #6
0
FLA_Error FLA_Syr2k_ln_unb_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C )
{
  FLA_Obj AT,              A0,
          AB,              a1t,
                           A2;

  FLA_Obj BT,              B0,
          BB,              b1t,
                           B2;

  FLA_Obj CTL,   CTR,      C00,  c01,     C02, 
          CBL,   CBR,      c10t, gamma11, c12t,
                           C20,  c21,     C22;


  FLA_Scalr_external( FLA_LOWER_TRIANGULAR, beta, C );

  FLA_Part_2x1( A,    &AT, 
                      &AB,            0, FLA_TOP );

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

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

  while ( FLA_Obj_length( AT ) < FLA_Obj_length( A ) ){


    FLA_Repart_2x1_to_3x1( AT,                &A0, 
                        /* ** */            /* ** */
                                              &a1t, 
                           AB,                &A2,        1, FLA_BOTTOM );

    FLA_Repart_2x1_to_3x1( BT,                &B0, 
                        /* ** */            /* ** */
                                              &b1t, 
                           BB,                &B2,        1, FLA_BOTTOM );

    FLA_Repart_2x2_to_3x3( CTL, /**/ CTR,       &C00,  /**/ &c01,     &C02,
                        /* ************* */   /* ************************** */
                                                &c10t, /**/ &gamma11, &c12t,
                           CBL, /**/ CBR,       &C20,  /**/ &c21,     &C22,
                           1, 1, FLA_BR );

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

    /* c10t = c10t + a1t * B0' */
    FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B0, a1t, FLA_ONE, c10t );

    /* c21 = c21 + B2 * a1t' */
    FLA_Gemv_external( FLA_NO_TRANSPOSE, alpha, B2, a1t, FLA_ONE, c21 );    

    /* gamma11 = gamma11 + a1t * b1t' + b1t * a1t' */
    FLA_Dot2s_external( alpha, a1t, b1t, FLA_ONE, gamma11 );

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

    FLA_Cont_with_3x1_to_2x1( &AT,                A0, 
                                                  a1t, 
                            /* ** */           /* ** */
                              &AB,                A2,     FLA_TOP );

    FLA_Cont_with_3x1_to_2x1( &BT,                B0, 
                                                  b1t, 
                            /* ** */           /* ** */
                              &BB,                B2,     FLA_TOP );

    FLA_Cont_with_3x3_to_2x2( &CTL, /**/ &CTR,       C00,  c01,     /**/ C02,
                                                     c10t, gamma11, /**/ c12t,
                            /* ************** */  /* ************************ */
                              &CBL, /**/ &CBR,       C20,  c21,     /**/ C22,
                              FLA_TL );

  }

  return FLA_SUCCESS;
}