|
libflame
12600
|
Go to the source code of this file.
Functions | |
| FLA_Error | FLA_Gemv_n_blk_var1 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
| FLA_Error | FLA_Gemv_n_blk_var2 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
| FLA_Error | FLA_Gemv_n_blk_var5 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
| FLA_Error | FLA_Gemv_n_blk_var6 (FLA_Obj alpha, FLA_Obj A, FLA_Obj x, FLA_Obj beta, FLA_Obj y, fla_gemv_t *cntl) |
| FLA_Error FLA_Gemv_n_blk_var1 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | x, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | y, | ||
| fla_gemv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj yT, y0,
yB, y1,
y2;
dim_t b;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_TOP );
FLA_Part_2x1( y, &yT,
&yB, 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( yT, &y0,
/* ** */ /* ** */
&y1,
yB, &y2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* y1 = alpha * A1 * x + y1; */
FLA_Gemv_internal( FLA_NO_TRANSPOSE,
alpha, A1, x, beta, y1,
FLA_Cntl_sub_gemv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
A1,
/* ** */ /* ** */
&AB, A2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &yT, y0,
y1,
/* ** */ /* ** */
&yB, y2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Gemv_n_blk_var2 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | x, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | y, | ||
| fla_gemv_t * | cntl | ||
| ) |
References FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_length(), FLA_Part_2x1(), and FLA_Repart_2x1_to_3x1().
Referenced by FLA_Gemv_n().
{
FLA_Obj AT, A0,
AB, A1,
A2;
FLA_Obj yT, y0,
yB, y1,
y2;
dim_t b;
FLA_Part_2x1( A, &AT,
&AB, 0, FLA_BOTTOM );
FLA_Part_2x1( y, &yT,
&yB, 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( yT, &y0,
&y1,
/* ** */ /* ** */
yB, &y2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* y1 = alpha * A1 * x + y1; */
FLA_Gemv_internal( FLA_NO_TRANSPOSE,
alpha, A1, x, beta, y1,
FLA_Cntl_sub_gemv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &AT, A0,
/* ** */ /* ** */
A1,
&AB, A2, FLA_BOTTOM );
FLA_Cont_with_3x1_to_2x1( &yT, y0,
/* ** */ /* ** */
y1,
&yB, y2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Gemv_n_blk_var5 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | x, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | y, | ||
| fla_gemv_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().
Referenced by FLA_Gemv_n().
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj xT, x0,
xB, x1,
x2;
dim_t b;
FLA_Scal_internal( beta, y,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_LEFT );
FLA_Part_2x1( x, &xT,
&xB, 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( xT, &x0,
/* ** */ /* ** */
&x1,
xB, &x2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
/* y = alpha * A1 * x1 + y; */
FLA_Gemv_internal( FLA_NO_TRANSPOSE,
alpha, A1, x1, FLA_ONE, y,
FLA_Cntl_sub_gemv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, A1, /**/ A2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &xT, x0,
x1,
/* ** */ /* ** */
&xB, x2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Gemv_n_blk_var6 | ( | FLA_Obj | alpha, |
| FLA_Obj | A, | ||
| FLA_Obj | x, | ||
| FLA_Obj | beta, | ||
| FLA_Obj | y, | ||
| fla_gemv_t * | cntl | ||
| ) |
References FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Determine_blocksize(), FLA_Gemv_internal(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Scal_internal().
Referenced by FLA_Gemv_n().
{
FLA_Obj AL, AR, A0, A1, A2;
FLA_Obj xT, x0,
xB, x1,
x2;
dim_t b;
FLA_Scal_internal( beta, y,
FLA_Cntl_sub_scal( cntl ) );
FLA_Part_1x2( A, &AL, &AR, 0, FLA_RIGHT );
FLA_Part_2x1( x, &xT,
&xB, 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( xT, &x0,
&x1,
/* ** */ /* ** */
xB, &x2, b, FLA_TOP );
/*------------------------------------------------------------*/
/* y = alpha * A1 * x1 + y; */
FLA_Gemv_internal( FLA_NO_TRANSPOSE,
alpha, A1, x1, FLA_ONE, y,
FLA_Cntl_sub_gemv( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &AL, /**/ &AR, A0, /**/ A1, A2,
FLA_RIGHT );
FLA_Cont_with_3x1_to_2x1( &xT, x0,
/* ** */ /* ** */
x1,
&xB, x2, FLA_BOTTOM );
}
return FLA_SUCCESS;
}
1.7.6.1