PLASMA  2.4.5
PLASMA - Parallel Linear Algebra for Scalable Multi-core Architectures
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups
pslauum.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 /***************************************************************************/
23 void plasma_pslauum_quark(PLASMA_enum uplo, PLASMA_desc A,
24  PLASMA_sequence *sequence, PLASMA_request *request)
25 {
26  plasma_context_t *plasma;
27  Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;
28 
29  int k, m, n;
30  int ldam;
31  int tempkm, tempmm, tempnn;
32 
33  float zone = (float)1.0;
34 
35  plasma = plasma_context_self();
36  if (sequence->status != PLASMA_SUCCESS)
37  return;
38  QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);
39  /*
40  * PlasmaLower
41  */
42  if (uplo == PlasmaLower) {
43  for (m = 0; m < A.mt; m++) {
44  tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
45  ldam = BLKLDD(A, m);
46  for(n = 0; n < m; n++) {
47  tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
48  QUARK_CORE_ssyrk(
49  plasma->quark, &task_flags,
50  uplo, PlasmaTrans,
51  tempnn, tempmm, A.mb,
52  1.0, A(m, n), ldam,
53  1.0, A(n, n), A.mb);
54 
55  for(k = n+1; k < m; k++) {
56  tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
57  QUARK_CORE_sgemm(
58  plasma->quark, &task_flags,
59  PlasmaTrans, PlasmaNoTrans,
60  tempkm, tempnn, tempmm, A.mb,
61  zone, A(m, k), ldam,
62  A(m, n), ldam,
63  zone, A(k, n), A.mb);
64  }
65  }
66  for (n = 0; n < m; n++) {
67  tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
68  QUARK_CORE_strmm(
69  plasma->quark, &task_flags,
70  PlasmaLeft, uplo, PlasmaTrans, PlasmaNonUnit,
71  tempmm, tempnn, A.mb,
72  zone, A(m, m), ldam,
73  A(m, n), ldam);
74  }
75  QUARK_CORE_slauum(
76  plasma->quark, &task_flags,
77  uplo,
78  tempmm,
79  A.mb, A(m, m), ldam);
80  }
81  }
82  /*
83  * PlasmaUpper
84  */
85  else {
86  for (m = 0; m < A.mt; m++) {
87  tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
88  ldam = BLKLDD(A, m);
89  for (n = 0; n < m; n++) {
90  tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
91  QUARK_CORE_ssyrk(
92  plasma->quark, &task_flags,
93  uplo, PlasmaNoTrans,
94  tempnn, tempmm, A.mb,
95  1.0, A(n, m), A.mb,
96  1.0, A(n, n), A.mb);
97 
98  for (k = n+1; k < m; k++){
99  tempkm = k == A.mt-1 ? A.m-k*A.mb : A.mb;
100  QUARK_CORE_sgemm(
101  plasma->quark, &task_flags,
102  PlasmaNoTrans, PlasmaTrans,
103  tempnn, tempkm, tempmm, A.mb,
104  zone, A(n, m), A.mb,
105  A(k, m), A.mb,
106  zone, A(n, k), A.mb);
107  }
108  }
109  for (n = 0; n < m; n++) {
110  tempnn = n == A.nt-1 ? A.n-n*A.nb : A.nb;
111  QUARK_CORE_strmm(
112  plasma->quark, &task_flags,
113  PlasmaRight, uplo, PlasmaTrans, PlasmaNonUnit,
114  tempnn, tempmm, A.mb,
115  zone, A(m, m), ldam,
116  A(n, m), A.mb);
117  }
118  QUARK_CORE_slauum(
119  plasma->quark, &task_flags,
120  uplo,
121  tempmm,
122  A.mb, A(m, m), ldam);
123  }
124  }
125 }