libflame  12600
Functions
FLA_Axpy_vars.h File Reference

(r12600)

Go to the source code of this file.

Functions

FLA_Error FLA_Axpy_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
FLA_Error FLA_Axpy_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
FLA_Error FLA_Axpy_blk_var3 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)
FLA_Error FLA_Axpy_blk_var4 (FLA_Obj alpha, FLA_Obj A, FLA_Obj B, fla_axpy_t *cntl)

Function Documentation

FLA_Error FLA_Axpy_blk_var1 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_axpy_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Axpy_internal().

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

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  dim_t b;

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

  FLA_Part_2x1( B,    &BT, 
                      &BB,            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( BT,                &B0, 
                        /* ** */            /* ** */
                                              &B1, 
                           BB,                &B2,        b, FLA_BOTTOM );

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

    FLA_Axpy_internal( alpha, A1, B1,
                       FLA_Cntl_sub_axpy( cntl ) );

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

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

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

  return FLA_SUCCESS;
}
FLA_Error FLA_Axpy_blk_var2 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_axpy_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().

Referenced by FLA_Axpy_internal().

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

  FLA_Obj BT,              B0,
          BB,              B1,
                           B2;

  dim_t b;

  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 );

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

    FLA_Axpy_internal( alpha, A1, B1,
                       FLA_Cntl_sub_axpy( 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;
}
FLA_Error FLA_Axpy_blk_var3 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_axpy_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().

Referenced by FLA_Axpy_internal().

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

  FLA_Obj BL,    BR,       B0,  B1,  B2;

  dim_t b;

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

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

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

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

    FLA_Axpy_internal( alpha, A1, B1,
                       FLA_Cntl_sub_axpy( cntl ) );

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

    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 );
  }

  return FLA_SUCCESS;
}
FLA_Error FLA_Axpy_blk_var4 ( FLA_Obj  alpha,
FLA_Obj  A,
FLA_Obj  B,
fla_axpy_t cntl 
)

References FLA_Axpy_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Determine_blocksize(), FLA_Obj_width(), FLA_Part_1x2(), and FLA_Repart_1x2_to_1x3().

Referenced by FLA_Axpy_internal().

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

  FLA_Obj BL,    BR,       B0,  B1,  B2;

  dim_t b;

  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 );

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

    FLA_Axpy_internal( alpha, A1, B1,
                       FLA_Cntl_sub_axpy( 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;
}