예제 #1
0
FLA_Error FLA_Gemm_tn_blk_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj AT,              A0,
          AB,              A1,
                           A2;

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  dim_t b;

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

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

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

  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 );

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

    /* C = alpha * A1' * B1 + C; */
    FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A1, B1, FLA_ONE, C,
                       FLA_Cntl_sub_gemm( 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 );

  }

  return FLA_SUCCESS;
}
예제 #2
0
FLA_Error FLA_Trmm_rlt_blk_var2( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t* cntl )
{
  FLA_Obj ATL,   ATR,      A00, A01, A02, 
          ABL,   ABR,      A10, A11, A12,
                           A20, A21, A22;

  FLA_Obj BL,    BR,       B0,  B1,  B2;

  dim_t b;

  FLA_Scal_internal( alpha, B,
                     FLA_Cntl_sub_scal( cntl ) );

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

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

  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_1x2_to_1x3( BL,  /**/ BR,        &B0, &B1, /**/ &B2,
                           b, FLA_LEFT );

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

    /* B2 = B2 + B1 * A21'; */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
                       FLA_ONE, B1, A21, FLA_ONE, B2,
                       FLA_Cntl_sub_gemm( cntl ) );

    /* B1 = B1 * tril( A11 )'; */
    FLA_Trmm_internal( FLA_RIGHT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, diagA,
                       FLA_ONE, A11, B1,
                       FLA_Cntl_sub_trmm( 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_1x3_to_1x2( &BL,  /**/ &BR,        B0, /**/ B1, B2,
                              FLA_RIGHT );

  }

  return FLA_SUCCESS;
}
예제 #3
0
FLA_Error FLA_Syrk_ut_blk_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t* cntl )
{
  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 );

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

    /* C12 = C12 + A1' * A2 */
    FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A1, A2, beta, C12,
                       FLA_Cntl_sub_gemm( cntl ) );

    /* C11 = C11 + A1' * A1 */
    FLA_Syrk_internal( FLA_UPPER_TRIANGULAR, FLA_TRANSPOSE, 
                       alpha, A1, beta, C11,
                       FLA_Cntl_sub_syrk( 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;
}
예제 #4
0
FLA_Error FLA_Gemm_ct_blk_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj AT,              A0,
          AB,              A1,
                           A2;

  FLA_Obj CT,              C0,
          CB,              C1,
                           C2;

  dim_t b;

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

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

  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_2x1_to_3x1( CT,                &C0, 
                        /* ** */            /* ** */
                                              &C1, 
                           CB,                &C2,        b, FLA_BOTTOM );

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

    /* C1 = alpha * A1 * B' + C1; */
    FLA_Gemm_internal( FLA_CONJ_NO_TRANSPOSE, FLA_TRANSPOSE,
                       alpha, A1, B, beta, C1,
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

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

  }

  return FLA_SUCCESS;
}
예제 #5
0
FLA_Error FLA_Trinv_uu_blk_var3( FLA_Obj A, fla_trinv_t* cntl )
{
    FLA_Obj ATL,   ATR,      A00, A01, A02,
            ABL,   ABR,      A10, A11, A12,
            A20, A21, A22;

    dim_t b;

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

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

        b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR );

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

        // A12 = -triuu( A11 ) \ A12;
        FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
                           FLA_MINUS_ONE, A11, A12,
                           FLA_Cntl_sub_trsm1( cntl ) );

        // A02 = A01 * A12 + A02;
        FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
                           FLA_ONE, A01, A12, FLA_ONE, A02,
                           FLA_Cntl_sub_gemm( cntl ) );

        // A01 = A01 / triuu( A11 );
        FLA_Trsm_internal( FLA_RIGHT, FLA_UPPER_TRIANGULAR, FLA_NO_TRANSPOSE, FLA_NONUNIT_DIAG,
                           FLA_ONE, A11, A01,
                           FLA_Cntl_sub_trsm2( cntl ) );

        // A11 = inv( A11 );
        FLA_Trinv_internal( FLA_UPPER_TRIANGULAR, FLA_NONUNIT_DIAG, A11,
                            FLA_Cntl_sub_trinv( cntl ) );

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

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

    }

    return FLA_SUCCESS;
}
예제 #6
0
FLA_Error FLA_Gemm_cn_blk_var5( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj AL,    AR,       A0,  A1,  A2;

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  dim_t b;

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

  FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_LEFT );

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

  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_2x1_to_3x1( BT,                &B0, 
                        /* ** */            /* ** */
                                              &B1, 
                           BB,                &B2,        b, FLA_BOTTOM );

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

    /* C = alpha * A1 * B1 + C; */
    FLA_Gemm_internal( FLA_CONJ_NO_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A1, B1, FLA_ONE, C, 
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

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

  }

  return FLA_SUCCESS;
}
예제 #7
0
FLA_Error FLA_Gemm_tt_blk_var5( FLA_Obj alpha, FLA_Obj A , FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  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_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;
}
예제 #8
0
FLA_Error FLA_Gemm_hn_blk_var2( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj AL,    AR,       A0,  A1,  A2;

  FLA_Obj CT,              C0,
          CB,              C1,
                           C2;

  dim_t b;

  FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );

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

  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){

    b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &A1, /**/ &A2,
                           b, FLA_LEFT );

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

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

    /* C1 = alpha * A1' * B + C1; */
    FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE, 
                       alpha, A1, B, beta, C1,
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

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

  }

  return FLA_SUCCESS;
}
예제 #9
0
FLA_Error FLA_Gemm_hh_blk_var3( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  FLA_Obj CL,    CR,       C0,  C1,  C2;

  dim_t b;

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

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

  while ( FLA_Obj_length( BT ) < FLA_Obj_length( B ) ){

    b = FLA_Determine_blocksize( BB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );

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

    FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, /**/ &C1, &C2,
                           b, FLA_RIGHT );

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

    /* C1 = alpha * A' * B1' + C1; */
    FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_CONJ_TRANSPOSE,
                       alpha, A, B1, beta, C1,
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

    FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, C1, /**/ C2,
                              FLA_LEFT );

  }

  return FLA_SUCCESS;
}
예제 #10
0
FLA_Error FLA_Gemm_nt_blk_var6( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj AL,    AR,       A0,  A1,  A2;

  FLA_Obj BL,    BR,       B0,  B1,  B2;

  dim_t b;

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

  FLA_Part_1x2( A,    &AL,  &AR,      0, FLA_RIGHT );

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

  while ( FLA_Obj_width( AR ) < FLA_Obj_width( A ) ){

    b = FLA_Determine_blocksize( AL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_1x2_to_1x3( AL,  /**/ AR,        &A0, &A1, /**/ &A2,
                           b, FLA_LEFT );

    FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, &B1, /**/ &B2,
                           b, FLA_LEFT );

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

    /* C = alpha * A1 * B1' + C; */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE,
                       alpha, A1, B1, FLA_ONE, C,
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

    FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, /**/ B1, B2,
                              FLA_RIGHT );

  }

  return FLA_SUCCESS;
}
예제 #11
0
FLA_Error FLA_Gemm_tc_blk_var4( FLA_Obj alpha, FLA_Obj A, FLA_Obj B, FLA_Obj beta, FLA_Obj C, fla_gemm_t* cntl )
{
  FLA_Obj BL,    BR,       B0,  B1,  B2;

  FLA_Obj CL,    CR,       C0,  C1,  C2;

  dim_t b;

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

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

  while ( FLA_Obj_width( BR ) < FLA_Obj_width( B ) ){

    b = FLA_Determine_blocksize( BL, FLA_LEFT, FLA_Cntl_blocksize( cntl ) );

    FLA_Repart_1x2_to_1x3( BL,  /**/ BR,        &B0, &B1, /**/ &B2,
                           b, FLA_LEFT );

    FLA_Repart_1x2_to_1x3( CL,  /**/ CR,        &C0, &C1, /**/ &C2,
                           b, FLA_LEFT );

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

    /* C1 = alpha * A' * B1 + C1; */
    FLA_Gemm_internal( FLA_TRANSPOSE, FLA_CONJ_NO_TRANSPOSE, 
                       alpha, A, B1, beta, C1,
                       FLA_Cntl_sub_gemm( cntl ) );

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

    FLA_Cont_with_1x3_to_1x2( &BL,  /**/ &BR,        B0, /**/ B1, B2,
                              FLA_RIGHT );

    FLA_Cont_with_1x3_to_1x2( &CL,  /**/ &CR,        C0, /**/ C1, C2,
                              FLA_RIGHT );

  }

  return FLA_SUCCESS;
}
예제 #12
0
FLA_Error FLA_Syrk_ln_blk_var1( FLA_Obj alpha, FLA_Obj A, FLA_Obj beta, FLA_Obj C, fla_syrk_t* cntl )
{
  FLA_Obj AT,              A0,
          AB,              A1,
                           A2;

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

  dim_t b;

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

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

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

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

    /* C10 = C10 + A1 * A0' */
    FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_TRANSPOSE, 
                       alpha, A1, A0, beta, C10,
                       FLA_Cntl_sub_gemm( cntl ) );

    /* C11 = C11 + A1 * A1' */
    FLA_Syrk_internal( FLA_LOWER_TRIANGULAR, FLA_NO_TRANSPOSE, 
                       alpha, A1, beta, C11,
                       FLA_Cntl_sub_syrk( cntl ) );

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

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

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

  }

  return FLA_SUCCESS;
}
예제 #13
0
FLA_Error FLA_Trmm_llt_blk_var1( FLA_Diag diagA, FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_trmm_t* cntl )
{
  FLA_Obj ATL,   ATR,      A00, A01, A02, 
          ABL,   ABR,      A10, A11, A12,
                           A20, A21, A22;

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  dim_t b;

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

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

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

    b = FLA_Determine_blocksize( ABR, FLA_BR, 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_BR );

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

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

    /* B1 = tril( A11 )' * B1; */
    FLA_Trmm_internal( FLA_LEFT, FLA_LOWER_TRIANGULAR, FLA_TRANSPOSE, diagA,
                       alpha, A11, B1,
                       FLA_Cntl_sub_trmm( cntl ) );

    /* B1 = B1 + A21' * B2; */
    FLA_Gemm_internal( FLA_TRANSPOSE, FLA_NO_TRANSPOSE,
                       alpha, A21, B2, FLA_ONE, B1,
                       FLA_Cntl_sub_gemm( cntl ) );

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

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

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

  }

  return FLA_SUCCESS;
}