|
libflame
revision_anchor
|
Go to the source code of this file.
Functions | |
| FLA_Error | FLA_Apply_QUD_UT_lhfc_blk_var1 (FLA_Obj T, FLA_Obj W, FLA_Obj R, FLA_Obj U, FLA_Obj C, FLA_Obj V, FLA_Obj D, fla_apqudut_t *cntl) |
| FLA_Error | FLA_Apply_QUD_UT_lhfc_blk_var2 (FLA_Obj T, FLA_Obj W, FLA_Obj R, FLA_Obj U, FLA_Obj C, FLA_Obj V, FLA_Obj D, fla_apqudut_t *cntl) |
| FLA_Error | FLA_Apply_QUD_UT_lhfc_blk_var3 (FLA_Obj T, FLA_Obj W, FLA_Obj R, FLA_Obj U, FLA_Obj C, FLA_Obj V, FLA_Obj D, fla_apqudut_t *cntl) |
| FLA_Error FLA_Apply_QUD_UT_lhfc_blk_var1 | ( | FLA_Obj | T, |
| FLA_Obj | W, | ||
| FLA_Obj | R, | ||
| FLA_Obj | U, | ||
| FLA_Obj | C, | ||
| FLA_Obj | V, | ||
| FLA_Obj | D, | ||
| fla_apqudut_t * | cntl | ||
| ) |
References FLA_Axpyt_internal(), FLA_Cont_with_1x3_to_1x2(), FLA_Cont_with_3x1_to_2x1(), FLA_Copyt_internal(), FLA_Gemm_internal(), FLA_MINUS_ONE, FLA_Obj_length(), FLA_Obj_width(), FLA_ONE, FLA_Part_1x2(), FLA_Part_2x1(), FLA_Part_2x2(), FLA_Repart_1x2_to_1x3(), FLA_Repart_2x1_to_3x1(), and FLA_Trsm_internal().
Referenced by FLA_Apply_QUD_UT_lhfc().
{
FLA_Obj TL, TR, T0, T1, T2;
FLA_Obj UL, UR, U0, U1, U2;
FLA_Obj VL, VR, V0, V1, V2;
FLA_Obj RT, R0,
RB, R1,
R2;
FLA_Obj T1T,
T1B;
FLA_Obj W1TL, W1TR,
W1BL, W1BR;
dim_t b_alg, b;
// Query the algorithmic blocksize by inspecting the length of T.
b_alg = FLA_Obj_length( T );
FLA_Part_1x2( T, &TL, &TR, 0, FLA_LEFT );
FLA_Part_1x2( U, &UL, &UR, 0, FLA_LEFT );
FLA_Part_1x2( V, &VL, &VR, 0, FLA_LEFT );
FLA_Part_2x1( R, &RT,
&RB, 0, FLA_TOP );
while ( FLA_Obj_width( UL ) < FLA_Obj_width( U ) ){
b = min( b_alg, FLA_Obj_width( UR ) );
FLA_Repart_1x2_to_1x3( TL, /**/ TR, &T0, /**/ &T1, &T2,
b, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( UL, /**/ UR, &U0, /**/ &U1, &U2,
b, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( VL, /**/ VR, &V0, /**/ &V1, &V2,
b, FLA_RIGHT );
FLA_Repart_2x1_to_3x1( RT, &R0,
/* ** */ /* ** */
&R1,
RB, &R2, b, FLA_BOTTOM );
/*------------------------------------------------------------*/
FLA_Part_2x1( T1, &T1T,
&T1B, b, FLA_TOP );
FLA_Part_2x2( W, &W1TL, &W1TR,
&W1BL, &W1BR, b, FLA_Obj_width( R1 ), FLA_TL );
// W1TL = R1;
FLA_Copyt_internal( FLA_NO_TRANSPOSE, R1, W1TL,
FLA_Cntl_sub_copyt( cntl ) );
// W1TL = inv( triu( T1T ) )' * ( R1 + U1' * C + V1' * D );
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, U1, C, FLA_ONE, W1TL,
FLA_Cntl_sub_gemm1( cntl ) );
FLA_Gemm_internal( FLA_CONJ_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, V1, D, FLA_ONE, W1TL,
FLA_Cntl_sub_gemm2( cntl ) );
FLA_Trsm_internal( FLA_LEFT, FLA_UPPER_TRIANGULAR,
FLA_CONJ_TRANSPOSE, FLA_NONUNIT_DIAG,
FLA_ONE, T1T, W1TL,
FLA_Cntl_sub_trsm( cntl ) );
// R1 = R1 - W1TL;
// C = C - U1 * W1TL;
// D = D + V1 * W1TL;
FLA_Axpyt_internal( FLA_NO_TRANSPOSE, FLA_MINUS_ONE, W1TL, R1,
FLA_Cntl_sub_axpyt( cntl ) );
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_MINUS_ONE, U1, W1TL, FLA_ONE, C,
FLA_Cntl_sub_gemm3( cntl ) );
FLA_Gemm_internal( FLA_NO_TRANSPOSE, FLA_NO_TRANSPOSE,
FLA_ONE, V1, W1TL, FLA_ONE, D,
FLA_Cntl_sub_gemm4( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &TL, /**/ &TR, T0, T1, /**/ T2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &UL, /**/ &UR, U0, U1, /**/ U2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &VL, /**/ &VR, V0, V1, /**/ V2,
FLA_LEFT );
FLA_Cont_with_3x1_to_2x1( &RT, R0,
R1,
/* ** */ /* ** */
&RB, R2, FLA_TOP );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_QUD_UT_lhfc_blk_var2 | ( | FLA_Obj | T, |
| FLA_Obj | W, | ||
| FLA_Obj | R, | ||
| FLA_Obj | U, | ||
| FLA_Obj | C, | ||
| FLA_Obj | V, | ||
| FLA_Obj | D, | ||
| fla_apqudut_t * | cntl | ||
| ) |
References FLA_Apply_QUD_UT_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_Apply_QUD_UT_lhfc().
{
FLA_Obj WL, WR, W0, W1, W2;
FLA_Obj RL, RR, R0, R1, R2;
FLA_Obj CL, CR, C0, C1, C2;
FLA_Obj DL, DR, D0, D1, D2;
dim_t b;
FLA_Part_1x2( W, &WL, &WR, 0, FLA_LEFT );
FLA_Part_1x2( R, &RL, &RR, 0, FLA_LEFT );
FLA_Part_1x2( C, &CL, &CR, 0, FLA_LEFT );
FLA_Part_1x2( D, &DL, &DR, 0, FLA_LEFT );
while ( FLA_Obj_width( RL ) < FLA_Obj_width( R ) ){
b = FLA_Determine_blocksize( RR, FLA_RIGHT, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_1x2_to_1x3( WL, /**/ WR, &W0, /**/ &W1, &W2,
b, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( RL, /**/ RR, &R0, /**/ &R1, &R2,
b, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( CL, /**/ CR, &C0, /**/ &C1, &C2,
b, FLA_RIGHT );
FLA_Repart_1x2_to_1x3( DL, /**/ DR, &D0, /**/ &D1, &D2,
b, FLA_RIGHT );
/*------------------------------------------------------------*/
// Apply Q' to R1, C1, and D1 from the left:
//
// / R1 \ / R1 \
// | C1 | = Q' | C1 |
// \ D1 / \ D1 /
//
// where Q is formed from U, V, and T.
FLA_Apply_QUD_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
T, W1,
R1,
U, C1,
V, D1, FLA_Cntl_sub_apqudut( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_1x3_to_1x2( &WL, /**/ &WR, W0, W1, /**/ W2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &RL, /**/ &RR, R0, R1, /**/ R2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &CL, /**/ &CR, C0, C1, /**/ C2,
FLA_LEFT );
FLA_Cont_with_1x3_to_1x2( &DL, /**/ &DR, D0, D1, /**/ D2,
FLA_LEFT );
}
return FLA_SUCCESS;
}
| FLA_Error FLA_Apply_QUD_UT_lhfc_blk_var3 | ( | FLA_Obj | T, |
| FLA_Obj | W, | ||
| FLA_Obj | R, | ||
| FLA_Obj | U, | ||
| FLA_Obj | C, | ||
| FLA_Obj | V, | ||
| FLA_Obj | D, | ||
| fla_apqudut_t * | cntl | ||
| ) |
References FLA_Apply_QUD_UT_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_Apply_QUD_UT_lhfc().
{
FLA_Obj TT, T0,
TB, T1,
T2;
FLA_Obj UT, U0,
UB, U1,
U2;
FLA_Obj VT, V0,
VB, V1,
V2;
FLA_Obj CT, C0,
CB, C1,
C2;
FLA_Obj DT, D0,
DB, D1,
D2;
dim_t b_T, b_UC, b_VD;
FLA_Part_2x1( T, &TT,
&TB, 0, FLA_TOP );
FLA_Part_2x1( U, &UT,
&UB, 0, FLA_TOP );
FLA_Part_2x1( V, &VT,
&VB, 0, FLA_TOP );
FLA_Part_2x1( C, &CT,
&CB, 0, FLA_TOP );
FLA_Part_2x1( D, &DT,
&DB, 0, FLA_TOP );
while ( FLA_Obj_length( TT ) < FLA_Obj_length( T ) ){
b_T = FLA_Determine_blocksize( TB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
b_UC = FLA_Determine_blocksize( UB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
b_VD = FLA_Determine_blocksize( VB, FLA_BOTTOM, FLA_Cntl_blocksize( cntl ) );
FLA_Repart_2x1_to_3x1( TT, &T0,
/* ** */ /* ** */
&T1,
TB, &T2, b_T, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( UT, &U0,
/* ** */ /* ** */
&U1,
UB, &U2, b_UC, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( VT, &V0,
/* ** */ /* ** */
&V1,
VB, &V2, b_VD, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( CT, &C0,
/* ** */ /* ** */
&C1,
CB, &C2, b_UC, FLA_BOTTOM );
FLA_Repart_2x1_to_3x1( DT, &D0,
/* ** */ /* ** */
&D1,
DB, &D2, b_VD, FLA_BOTTOM );
/*------------------------------------------------------------*/
// / R \ / R \
// | C1 | = Q' | C1 |
// \ D1 / \ D1 /
//
// where Q is formed from U1, V1, and T1.
FLA_Apply_QUD_UT_internal( FLA_LEFT, FLA_CONJ_TRANSPOSE, FLA_FORWARD, FLA_COLUMNWISE,
T1, W,
R,
U1, C1,
V1, D1, FLA_Cntl_sub_apqudut( cntl ) );
/*------------------------------------------------------------*/
FLA_Cont_with_3x1_to_2x1( &TT, T0,
T1,
/* ** */ /* ** */
&TB, T2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &UT, U0,
U1,
/* ** */ /* ** */
&UB, U2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &VT, V0,
V1,
/* ** */ /* ** */
&VB, V2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &CT, C0,
C1,
/* ** */ /* ** */
&CB, C2, FLA_TOP );
FLA_Cont_with_3x1_to_2x1( &DT, D0,
D1,
/* ** */ /* ** */
&DB, D2, FLA_TOP );
}
return FLA_SUCCESS;
}
1.7.6.1