PLASMA
2.4.5
PLASMA - Parallel Linear Algebra for Scalable Multi-core Architectures
Main Page
Modules
Namespaces
Data Structures
Files
File List
Globals
All
Data Structures
Namespaces
Files
Functions
Variables
Typedefs
Enumerations
Enumerator
Macros
Groups
pstrsmrv.c
Go to the documentation of this file.
1
17
#include "
common.h
"
18
19
#define A(m,n) BLKADDR(A, float, m, n)
20
#define W(m) BLKADDR(W, float, m, 0)
21
22
/***************************************************************************/
25
void
plasma_pstrsmrv_quark
(
PLASMA_enum
side
,
PLASMA_enum
uplo
,
PLASMA_enum
trans
,
PLASMA_enum
diag
,
26
float
alpha,
PLASMA_desc
A
,
PLASMA_desc
W
,
27
PLASMA_sequence
*sequence,
PLASMA_request
*request)
28
{
29
plasma_context_t
*
plasma
;
30
Quark_Task_Flags
task_flags =
Quark_Task_Flags_Initializer
;
31
32
int
k, m;
33
int
ldak, ldam, ldwk, ldwm;
34
int
tempm, tempkm, tempkn, tempmm;
35
36
plasma =
plasma_context_self
();
37
if
(sequence->
status
!=
PLASMA_SUCCESS
)
38
return
;
39
QUARK_Task_Flag_Set
(&task_flags,
TASK_SEQUENCE
, (intptr_t)sequence->
quark_sequence
);
40
41
for
(k = A.
mt
-1; k >= 0; k--) {
42
tempm = (k+1)*A.
mb
;
43
tempkm = k == A.
mt
-1 ? A.
m
-k*A.
mb
: A.
mb
;
44
tempkn = k == A.
nt
-1 ? A.
n
-k*A.
nb
: A.
nb
;
45
ldak =
BLKLDD
(A, k);
46
ldwk =
BLKLDD
(W, k);
47
48
QUARK_CORE_slacpy
(
49
plasma->
quark
, &task_flags,
50
PlasmaLower
, tempkm, tempkn, A.
mb
,
51
A(k, k), ldak,
W
(k), ldwk );
52
53
QUARK_CORE_slaset2
(
54
plasma->
quark
, &task_flags,
55
PlasmaLower
, tempkn, tempkn,
56
0.0, A(k, k), ldak );
57
58
for
(m=k+1; m<A.
mt
; m++) {
59
tempmm = m == A.
mt
-1 ? A.
m
-m*A.
mb
: A.
mb
;
60
ldam =
BLKLDD
(A, m);
61
ldwm =
BLKLDD
(W, m);
62
63
QUARK_CORE_slacpy
(
64
plasma->
quark
, &task_flags,
65
PlasmaUpperLower
, tempmm, tempkn, A.
mb
,
66
A(m, k), ldam,
W
(m), ldwm );
67
68
QUARK_CORE_slaset
(
69
plasma->
quark
, &task_flags,
70
PlasmaUpperLower
, tempmm, tempkn,
71
0.0, 0.0, A(m, k), ldam );
72
}
73
74
if
(k*A.
mb
+tempkn < A.
m
) {
75
plasma_psgemm_quark
(
76
PlasmaNoTrans
,
PlasmaNoTrans
, -alpha,
77
plasma_desc_submatrix
(A, 0, tempm, tempm, A.
n
-tempm),
78
plasma_desc_submatrix
(W, tempm, 0, W.
m
-tempm, tempkn ),
79
alpha,
80
plasma_desc_submatrix
(A, 0, k*A.
mb
, tempm, tempkn ),
81
sequence, request);
82
83
plasma_psgemm_quark
(
84
PlasmaNoTrans
,
PlasmaNoTrans
, -alpha,
85
plasma_desc_submatrix
(A, tempm, tempm, A.
m
-tempm, A.
n
-tempm),
86
plasma_desc_submatrix
(W, tempm, 0, W.
m
-tempm, tempkn ),
87
(
float
)0.,
88
plasma_desc_submatrix
(A, tempm, k*A.
mb
, A.
m
-tempm, tempkn ),
89
sequence, request);
90
}
91
92
plasma_pstrsm_quark
(
93
PlasmaRight
,
PlasmaLower
,
94
PlasmaNoTrans
,
PlasmaUnit
,
95
alpha,
96
plasma_desc_submatrix
(W, k*W.
mb
, 0, tempkm, tempkn),
97
plasma_desc_submatrix
(A, 0, k*A.
mb
, A.
m
, tempkn),
98
sequence, request);
99
}
100
}
plasma_2.4.5
compute
pstrsmrv.c
Generated on Mon Jul 9 2012 12:44:56 for PLASMA by
1.8.1