scs 0.4.0 → 0.4.1
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +4 -0
- data/LICENSE.txt +1 -1
- data/README.md +1 -1
- data/lib/scs/version.rb +1 -1
- data/vendor/scs/CITATION.cff +2 -2
- data/vendor/scs/CMakeLists.txt +284 -168
- data/vendor/scs/Makefile +43 -18
- data/vendor/scs/README.md +1 -1
- data/vendor/scs/include/glbopts.h +32 -13
- data/vendor/scs/include/linsys.h +8 -8
- data/vendor/scs/include/scs.h +6 -2
- data/vendor/scs/include/scs_types.h +3 -1
- data/vendor/scs/include/scs_work.h +9 -8
- data/vendor/scs/include/util.h +1 -1
- data/vendor/scs/linsys/cpu/direct/private.c +32 -153
- data/vendor/scs/linsys/cpu/direct/private.h +6 -6
- data/vendor/scs/linsys/cpu/indirect/private.c +9 -22
- data/vendor/scs/linsys/cpu/indirect/private.h +4 -2
- data/vendor/scs/linsys/csparse.c +140 -12
- data/vendor/scs/linsys/csparse.h +10 -17
- data/vendor/scs/linsys/gpu/gpu.c +4 -4
- data/vendor/scs/linsys/gpu/gpu.h +1 -1
- data/vendor/scs/linsys/gpu/indirect/private.c +15 -26
- data/vendor/scs/linsys/mkl/direct/private.c +182 -0
- data/vendor/scs/linsys/mkl/direct/private.h +38 -0
- data/vendor/scs/linsys/scs_matrix.c +11 -5
- data/vendor/scs/scs.mk +39 -26
- data/vendor/scs/src/cones.c +15 -159
- data/vendor/scs/src/exp_cone.c +399 -0
- data/vendor/scs/src/normalize.c +4 -2
- data/vendor/scs/src/rw.c +93 -38
- data/vendor/scs/src/scs.c +83 -52
- data/vendor/scs/src/util.c +12 -3
- data/vendor/scs/test/minunit.h +2 -1
- data/vendor/scs/test/problem_utils.h +2 -1
- data/vendor/scs/test/problems/hs21_tiny_qp.h +1 -1
- data/vendor/scs/test/problems/hs21_tiny_qp_rw.h +8 -3
- data/vendor/scs/test/problems/max_ent +0 -0
- data/vendor/scs/test/problems/max_ent.h +8 -0
- data/vendor/scs/test/problems/random_prob.h +2 -43
- data/vendor/scs/test/problems/rob_gauss_cov_est.h +7 -2
- data/vendor/scs/test/problems/test_exp_cone.h +84 -0
- data/vendor/scs/test/problems/test_prob_from_data_file.h +57 -0
- data/vendor/scs/test/run_from_file.c +7 -1
- data/vendor/scs/test/run_tests.c +22 -9
- metadata +10 -3
@@ -9,15 +9,15 @@ extern "C" {
|
|
9
9
|
#include "external/amd/amd.h"
|
10
10
|
#include "external/qdldl/qdldl.h"
|
11
11
|
#include "glbopts.h"
|
12
|
-
#include "
|
12
|
+
#include "linsys.h"
|
13
13
|
#include "scs_matrix.h"
|
14
14
|
|
15
15
|
struct SCS_LIN_SYS_WORK {
|
16
|
-
scs_int m, n;
|
17
|
-
|
18
|
-
scs_float *Dinv;
|
19
|
-
scs_int *perm;
|
20
|
-
scs_float *bp;
|
16
|
+
scs_int m, n; /* linear system dimensions */
|
17
|
+
ScsMatrix *kkt, *L; /* KKT, and factorization matrix L resp. */
|
18
|
+
scs_float *Dinv; /* inverse diagonal matrix of factorization */
|
19
|
+
scs_int *perm; /* permutation of KKT matrix for factorization */
|
20
|
+
scs_float *bp; /* workspace memory for solves */
|
21
21
|
scs_int *diag_r_idxs;
|
22
22
|
scs_int factorizations;
|
23
23
|
/* ldl factorization workspace */
|
@@ -1,22 +1,9 @@
|
|
1
1
|
#include "private.h"
|
2
|
-
#include "linsys.h"
|
3
|
-
#include "util.h"
|
4
|
-
#include <limits.h>
|
5
2
|
|
6
|
-
const char *
|
7
|
-
return "sparse-indirect";
|
3
|
+
const char *scs_get_lin_sys_method() {
|
4
|
+
return "sparse-indirect-scs";
|
8
5
|
}
|
9
6
|
|
10
|
-
/*
|
11
|
-
char *SCS(get_lin_sys_summary)(ScsLinSysWork *p, const ScsInfo *info) {
|
12
|
-
char *str = (char *)scs_malloc(sizeof(char) * 128);
|
13
|
-
sprintf(str, "lin-sys: avg cg its: %2.2f\n",
|
14
|
-
(scs_float)p->tot_cg_its / (info->iter + 1));
|
15
|
-
p->tot_cg_its = 0;
|
16
|
-
return str;
|
17
|
-
}
|
18
|
-
*/
|
19
|
-
|
20
7
|
/* Not possible to do this on the fly due to M_ii += a_i' (R_y)^-1 a_i */
|
21
8
|
/* set M = inv ( diag ( R_x + P + A' R_y^{-1} A ) ) */
|
22
9
|
static void set_preconditioner(ScsLinSysWork *p) {
|
@@ -97,7 +84,7 @@ static void transpose(const ScsMatrix *A, ScsLinSysWork *p) {
|
|
97
84
|
#endif
|
98
85
|
}
|
99
86
|
|
100
|
-
void
|
87
|
+
void scs_free_lin_sys_work(ScsLinSysWork *p) {
|
101
88
|
if (p) {
|
102
89
|
scs_free(p->p);
|
103
90
|
scs_free(p->r);
|
@@ -162,13 +149,13 @@ static void apply_pre_conditioner(scs_float *z, scs_float *r, scs_int n,
|
|
162
149
|
}
|
163
150
|
|
164
151
|
/* no need to update anything in this case */
|
165
|
-
void
|
152
|
+
void scs_update_lin_sys_diag_r(ScsLinSysWork *p, const scs_float *diag_r) {
|
166
153
|
p->diag_r = diag_r; /* this isn't needed but do it to be safe */
|
167
154
|
set_preconditioner(p);
|
168
155
|
}
|
169
156
|
|
170
|
-
ScsLinSysWork *
|
171
|
-
|
157
|
+
ScsLinSysWork *scs_init_lin_sys_work(const ScsMatrix *A, const ScsMatrix *P,
|
158
|
+
const scs_float *diag_r) {
|
172
159
|
ScsLinSysWork *p = (ScsLinSysWork *)scs_calloc(1, sizeof(ScsLinSysWork));
|
173
160
|
p->A = A;
|
174
161
|
p->P = P;
|
@@ -198,7 +185,7 @@ ScsLinSysWork *SCS(init_lin_sys_work)(const ScsMatrix *A, const ScsMatrix *P,
|
|
198
185
|
p->tot_cg_its = 0;
|
199
186
|
if (!p->p || !p->r || !p->Gp || !p->tmp || !p->At || !p->At->i || !p->At->p ||
|
200
187
|
!p->At->x) {
|
201
|
-
|
188
|
+
scs_free_lin_sys_work(p);
|
202
189
|
return SCS_NULL;
|
203
190
|
}
|
204
191
|
return p;
|
@@ -288,8 +275,8 @@ static scs_int pcg(ScsLinSysWork *pr, const scs_float *s, scs_float *b,
|
|
288
275
|
* y = R_y^{-1} (Ax - ry)
|
289
276
|
*
|
290
277
|
*/
|
291
|
-
scs_int
|
292
|
-
|
278
|
+
scs_int scs_solve_lin_sys(ScsLinSysWork *p, scs_float *b, const scs_float *s,
|
279
|
+
scs_float tol) {
|
293
280
|
scs_int cg_its, max_iters;
|
294
281
|
|
295
282
|
if (tol <= 0.) {
|
@@ -5,11 +5,13 @@
|
|
5
5
|
extern "C" {
|
6
6
|
#endif
|
7
7
|
|
8
|
+
#include "csparse.h"
|
8
9
|
#include "glbopts.h"
|
9
10
|
#include "linalg.h"
|
10
|
-
#include "
|
11
|
+
#include "linsys.h"
|
11
12
|
#include "scs_matrix.h"
|
12
|
-
#include
|
13
|
+
#include "util.h" /* timer */
|
14
|
+
#include <string.h>
|
13
15
|
|
14
16
|
struct SCS_LIN_SYS_WORK {
|
15
17
|
scs_int n, m; /* linear system dimensions */
|
data/vendor/scs/linsys/csparse.c
CHANGED
@@ -2,39 +2,37 @@
|
|
2
2
|
|
3
3
|
#include "csparse.h"
|
4
4
|
|
5
|
-
|
6
|
-
|
7
|
-
|
5
|
+
ScsMatrix *SCS(cs_spalloc)(scs_int m, scs_int n, scs_int nzmax, scs_int values,
|
6
|
+
scs_int triplet) {
|
7
|
+
ScsMatrix *A = (ScsMatrix *)scs_calloc(1, sizeof(ScsMatrix));
|
8
8
|
if (!A) {
|
9
9
|
return SCS_NULL;
|
10
10
|
} /* out of memory */
|
11
11
|
A->m = m; /* define dimensions and nzmax */
|
12
12
|
A->n = n;
|
13
|
-
A->nzmax = nzmax = MAX(nzmax, 1);
|
14
|
-
A->nz = triplet ? 0 : -1; /* allocate triplet or comp.col */
|
15
13
|
A->p = (scs_int *)scs_calloc((triplet ? nzmax : n + 1), sizeof(scs_int));
|
16
14
|
A->i = (scs_int *)scs_calloc(nzmax, sizeof(scs_int));
|
17
15
|
A->x = values ? (scs_float *)scs_calloc(nzmax, sizeof(scs_float)) : SCS_NULL;
|
18
16
|
return (!A->p || !A->i || (values && !A->x)) ? SCS(cs_spfree)(A) : A;
|
19
17
|
}
|
20
18
|
|
21
|
-
|
19
|
+
ScsMatrix *SCS(cs_done)(ScsMatrix *C, void *w, void *x, scs_int ok) {
|
22
20
|
scs_free(w); /* free workspace */
|
23
21
|
scs_free(x);
|
24
22
|
return ok ? C : SCS(cs_spfree)(C); /* return result if OK, else free it */
|
25
23
|
}
|
26
24
|
|
27
25
|
/* C = compressed-column form of a triplet matrix T */
|
28
|
-
|
29
|
-
|
26
|
+
ScsMatrix *SCS(cs_compress)(const ScsMatrix *T, scs_int nz,
|
27
|
+
scs_int *idx_mapping) {
|
28
|
+
scs_int m, n, p, k, *Cp, *Ci, *w, *Ti, *Tj;
|
30
29
|
scs_float *Cx, *Tx;
|
31
|
-
|
30
|
+
ScsMatrix *C;
|
32
31
|
m = T->m;
|
33
32
|
n = T->n;
|
34
33
|
Ti = T->i;
|
35
34
|
Tj = T->p;
|
36
35
|
Tx = T->x;
|
37
|
-
nz = T->nz;
|
38
36
|
C = SCS(cs_spalloc)(m, n, nz, Tx != SCS_NULL, 0); /* allocate result */
|
39
37
|
w = (scs_int *)scs_calloc(n, sizeof(scs_int)); /* get workspace */
|
40
38
|
if (!C || !w) {
|
@@ -75,7 +73,7 @@ scs_float SCS(cumsum)(scs_int *p, scs_int *c, scs_int n) {
|
|
75
73
|
return nz2; /* return sum (c [0..n-1]) */
|
76
74
|
}
|
77
75
|
|
78
|
-
|
76
|
+
ScsMatrix *SCS(cs_spfree)(ScsMatrix *A) {
|
79
77
|
if (!A) {
|
80
78
|
return SCS_NULL;
|
81
79
|
} /* do nothing if A already SCS_NULL */
|
@@ -83,5 +81,135 @@ csc *SCS(cs_spfree)(csc *A) {
|
|
83
81
|
scs_free(A->i);
|
84
82
|
scs_free(A->x);
|
85
83
|
scs_free(A);
|
86
|
-
|
84
|
+
/* free the ScsMatrix struct and return SCS_NULL */
|
85
|
+
return (ScsMatrix *)SCS_NULL;
|
86
|
+
}
|
87
|
+
|
88
|
+
/* Build the KKT matrix */
|
89
|
+
ScsMatrix *SCS(form_kkt)(const ScsMatrix *A, const ScsMatrix *P,
|
90
|
+
scs_float *diag_p, const scs_float *diag_r,
|
91
|
+
scs_int *diag_r_idxs, scs_int upper) {
|
92
|
+
/*
|
93
|
+
* Forms column compressed KKT matrix assumes column compressed A,P matrices.
|
94
|
+
* Only upper OR lower triangular part is stuffed, depending on `upper` flag.
|
95
|
+
*
|
96
|
+
* Forms upper/lower triangular part of [(R_x + P) A'; A -R_y]
|
97
|
+
* Shapes: P : n x n, A: m x n.
|
98
|
+
*
|
99
|
+
* Output: `diag_p` will contain values of P diagonal upon completion,
|
100
|
+
* and `diag_r_idxs` will contain the indices corresponding to the entries
|
101
|
+
* in the returned matrix corresponding to the entries of R.
|
102
|
+
*
|
103
|
+
*/
|
104
|
+
scs_int h, i, j, count;
|
105
|
+
ScsMatrix *Kcsc, *K;
|
106
|
+
scs_int n = A->n;
|
107
|
+
scs_int m = A->m;
|
108
|
+
scs_int Anz = A->p[n];
|
109
|
+
scs_int Knzmax;
|
110
|
+
scs_int *idx_mapping;
|
111
|
+
if (P) {
|
112
|
+
/* Upper bound P + I triangular component NNZs as Pnz + n */
|
113
|
+
Knzmax = n + m + Anz + P->p[n];
|
114
|
+
} else {
|
115
|
+
Knzmax = n + m + Anz;
|
116
|
+
}
|
117
|
+
K = SCS(cs_spalloc)(m + n, m + n, Knzmax, 1, 1);
|
118
|
+
|
119
|
+
#if VERBOSITY > 0
|
120
|
+
scs_printf("forming kkt\n");
|
121
|
+
#endif
|
122
|
+
/* Here we generate a triplet matrix and then compress to CSC */
|
123
|
+
if (!K) {
|
124
|
+
return SCS_NULL;
|
125
|
+
}
|
126
|
+
|
127
|
+
count = 0; /* element counter */
|
128
|
+
if (P) {
|
129
|
+
/* R_x + P in top left */
|
130
|
+
for (j = 0; j < n; j++) { /* cols */
|
131
|
+
diag_p[j] = 0.;
|
132
|
+
/* empty column, add diagonal */
|
133
|
+
if (P->p[j] == P->p[j + 1]) {
|
134
|
+
K->i[count] = j;
|
135
|
+
K->p[count] = j;
|
136
|
+
K->x[count] = diag_r[j];
|
137
|
+
diag_r_idxs[j] = count; /* store the indices where diag_r occurs */
|
138
|
+
count++;
|
139
|
+
}
|
140
|
+
for (h = P->p[j]; h < P->p[j + 1]; h++) {
|
141
|
+
i = P->i[h]; /* row */
|
142
|
+
if (i > j) { /* only upper triangular needed */
|
143
|
+
break;
|
144
|
+
}
|
145
|
+
if (upper) {
|
146
|
+
K->i[count] = i;
|
147
|
+
K->p[count] = j;
|
148
|
+
} else { /* lower triangular */
|
149
|
+
/* P is passed in upper triangular, need to flip that here */
|
150
|
+
K->i[count] = j; /* col -> row */
|
151
|
+
K->p[count] = i; /* row -> col */
|
152
|
+
}
|
153
|
+
K->x[count] = P->x[h];
|
154
|
+
if (i == j) {
|
155
|
+
/* P has diagonal element */
|
156
|
+
diag_p[j] = P->x[h];
|
157
|
+
K->x[count] += diag_r[j];
|
158
|
+
diag_r_idxs[j] = count; /* store the indices where diag_r occurs */
|
159
|
+
}
|
160
|
+
count++;
|
161
|
+
/* reached the end without adding diagonal, do it now */
|
162
|
+
if ((i < j) && (h + 1 == P->p[j + 1] || P->i[h + 1] > j)) {
|
163
|
+
K->i[count] = j;
|
164
|
+
K->p[count] = j;
|
165
|
+
K->x[count] = diag_r[j];
|
166
|
+
diag_r_idxs[j] = count; /* store the indices where diag_r occurs */
|
167
|
+
count++;
|
168
|
+
}
|
169
|
+
}
|
170
|
+
}
|
171
|
+
} else {
|
172
|
+
/* R_x in top left */
|
173
|
+
for (j = 0; j < n; j++) {
|
174
|
+
diag_p[j] = 0.;
|
175
|
+
K->i[count] = j;
|
176
|
+
K->p[count] = j;
|
177
|
+
K->x[count] = diag_r[j];
|
178
|
+
diag_r_idxs[j] = count; /* store the indices where diag_r occurs */
|
179
|
+
count++;
|
180
|
+
}
|
181
|
+
}
|
182
|
+
|
183
|
+
/* A in bottom left or A^T top right */
|
184
|
+
for (j = 0; j < n; j++) { /* column */
|
185
|
+
for (h = A->p[j]; h < A->p[j + 1]; h++) {
|
186
|
+
if (upper) {
|
187
|
+
K->p[count] = A->i[h] + n; /* column */
|
188
|
+
K->i[count] = j; /*row */
|
189
|
+
} else { /* lower triangular */
|
190
|
+
K->p[count] = j; /* column */
|
191
|
+
K->i[count] = A->i[h] + n; /* row */
|
192
|
+
}
|
193
|
+
K->x[count] = A->x[h];
|
194
|
+
count++;
|
195
|
+
}
|
196
|
+
}
|
197
|
+
|
198
|
+
/* -R_y at bottom right */
|
199
|
+
for (j = 0; j < m; j++) {
|
200
|
+
K->i[count] = j + n;
|
201
|
+
K->p[count] = j + n;
|
202
|
+
K->x[count] = -diag_r[j + n];
|
203
|
+
diag_r_idxs[j + n] = count; /* store the indices where diag_r occurs */
|
204
|
+
count++;
|
205
|
+
}
|
206
|
+
|
207
|
+
idx_mapping = (scs_int *)scs_calloc(count, sizeof(scs_int));
|
208
|
+
Kcsc = SCS(cs_compress)(K, count, idx_mapping);
|
209
|
+
for (i = 0; i < m + n; i++) {
|
210
|
+
diag_r_idxs[i] = idx_mapping[diag_r_idxs[i]];
|
211
|
+
}
|
212
|
+
SCS(cs_spfree)(K);
|
213
|
+
scs_free(idx_mapping);
|
214
|
+
return Kcsc;
|
87
215
|
}
|
data/vendor/scs/linsys/csparse.h
CHANGED
@@ -10,24 +10,17 @@ extern "C" {
|
|
10
10
|
#include "glbopts.h"
|
11
11
|
#include "scs.h"
|
12
12
|
|
13
|
-
|
14
|
-
|
15
|
-
|
16
|
-
|
17
|
-
|
18
|
-
|
19
|
-
scs_int *i; /* row indices, size nzmax */
|
20
|
-
scs_float *x; /* numerical values, size nzmax */
|
21
|
-
scs_int nz; /* # of entries in triplet matrix, -1 for compressed-col */
|
22
|
-
} csc;
|
23
|
-
|
24
|
-
csc *SCS(cs_spalloc)(scs_int m, scs_int n, scs_int nzmax, scs_int values,
|
25
|
-
scs_int triplet);
|
26
|
-
csc *SCS(cs_done)(csc *C, void *w, void *x, scs_int ok);
|
27
|
-
csc *SCS(cs_compress)(const csc *T, scs_int *idx_mapping);
|
13
|
+
ScsMatrix *SCS(cs_spalloc)(scs_int m, scs_int n, scs_int nzmax, scs_int values,
|
14
|
+
scs_int triplet);
|
15
|
+
ScsMatrix *SCS(cs_done)(ScsMatrix *C, void *w, void *x, scs_int ok);
|
16
|
+
ScsMatrix *SCS(cs_compress)(const ScsMatrix *T, scs_int nz,
|
17
|
+
scs_int *idx_mapping);
|
18
|
+
ScsMatrix *SCS(cs_spfree)(ScsMatrix *A);
|
28
19
|
scs_float SCS(cumsum)(scs_int *p, scs_int *c, scs_int n);
|
29
|
-
|
30
|
-
|
20
|
+
/* Forms KKT matrix */
|
21
|
+
ScsMatrix *SCS(form_kkt)(const ScsMatrix *A, const ScsMatrix *P,
|
22
|
+
scs_float *diag_p, const scs_float *diag_r,
|
23
|
+
scs_int *diag_r_idxs, scs_int upper);
|
31
24
|
#ifdef __cplusplus
|
32
25
|
}
|
33
26
|
#endif
|
data/vendor/scs/linsys/gpu/gpu.c
CHANGED
@@ -19,13 +19,13 @@ void SCS(accum_by_atrans_gpu)(const ScsGpuMatrix *Ag,
|
|
19
19
|
if (*buffer != SCS_NULL) {
|
20
20
|
cudaFree(*buffer);
|
21
21
|
}
|
22
|
-
cudaMalloc(buffer,
|
22
|
+
cudaMalloc(buffer, new_buffer_size);
|
23
23
|
*buffer_size = new_buffer_size;
|
24
24
|
}
|
25
25
|
|
26
26
|
CUSPARSE_GEN(SpMV)
|
27
27
|
(cusparse_handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &onef, Ag->descr, x,
|
28
|
-
&onef, y, SCS_CUDA_FLOAT, SCS_CSRMV_ALG, buffer);
|
28
|
+
&onef, y, SCS_CUDA_FLOAT, SCS_CSRMV_ALG, *buffer);
|
29
29
|
}
|
30
30
|
|
31
31
|
/* this is slow, use trans routine if possible */
|
@@ -48,13 +48,13 @@ void SCS(accum_by_a_gpu)(const ScsGpuMatrix *Ag, const cusparseDnVecDescr_t x,
|
|
48
48
|
if (*buffer != SCS_NULL) {
|
49
49
|
cudaFree(*buffer);
|
50
50
|
}
|
51
|
-
cudaMalloc(buffer,
|
51
|
+
cudaMalloc(buffer, new_buffer_size);
|
52
52
|
*buffer_size = new_buffer_size;
|
53
53
|
}
|
54
54
|
|
55
55
|
CUSPARSE_GEN(SpMV)
|
56
56
|
(cusparse_handle, CUSPARSE_OPERATION_TRANSPOSE, &onef, Ag->descr, x, &onef, y,
|
57
|
-
SCS_CUDA_FLOAT, SCS_CSRMV_ALG, buffer);
|
57
|
+
SCS_CUDA_FLOAT, SCS_CSRMV_ALG, *buffer);
|
58
58
|
}
|
59
59
|
|
60
60
|
/* This assumes that P has been made full (ie not triangular) and uses the
|
data/vendor/scs/linsys/gpu/gpu.h
CHANGED
@@ -21,20 +21,10 @@ static scs_float cg_gpu_norm(cublasHandle_t cublas_handle, scs_float *r,
|
|
21
21
|
return nrm;
|
22
22
|
}
|
23
23
|
|
24
|
-
const char *
|
24
|
+
const char *scs_get_lin_sys_method() {
|
25
25
|
return "sparse-indirect GPU";
|
26
26
|
}
|
27
27
|
|
28
|
-
/*
|
29
|
-
char *SCS(get_lin_sys_summary)(ScsLinSysWork *p, const ScsInfo *info) {
|
30
|
-
char *str = (char *)scs_malloc(sizeof(char) * 128);
|
31
|
-
sprintf(str, "lin-sys: avg cg its: %2.2f\n",
|
32
|
-
(scs_float)p->tot_cg_its / (info->iter + 1));
|
33
|
-
p->tot_cg_its = 0;
|
34
|
-
return str;
|
35
|
-
}
|
36
|
-
*/
|
37
|
-
|
38
28
|
/* Not possible to do this on the fly due to M_ii += a_i' (R_y)^-1 a_i */
|
39
29
|
/* set M = inv ( diag ( R_x + P + A' R_y^{-1} A ) ) */
|
40
30
|
static void set_preconditioner(ScsLinSysWork *p, const scs_float *diag_r) {
|
@@ -76,7 +66,7 @@ static void set_preconditioner(ScsLinSysWork *p, const scs_float *diag_r) {
|
|
76
66
|
}
|
77
67
|
|
78
68
|
/* no need to update anything in this case */
|
79
|
-
void
|
69
|
+
void scs_update_lin_sys_diag_r(ScsLinSysWork *p, const scs_float *diag_r) {
|
80
70
|
scs_int i;
|
81
71
|
|
82
72
|
/* R_x to gpu */
|
@@ -93,7 +83,7 @@ void SCS(update_lin_sys_diag_r)(ScsLinSysWork *p, const scs_float *diag_r) {
|
|
93
83
|
set_preconditioner(p, diag_r);
|
94
84
|
}
|
95
85
|
|
96
|
-
void
|
86
|
+
void scs_free_lin_sys_work(ScsLinSysWork *p) {
|
97
87
|
if (p) {
|
98
88
|
scs_free(p->M);
|
99
89
|
scs_free(p->inv_r_y);
|
@@ -182,13 +172,13 @@ static void mat_vec(ScsLinSysWork *p, const scs_float *x, scs_float *y) {
|
|
182
172
|
}
|
183
173
|
|
184
174
|
/* P comes in upper triangular, expand to full
|
185
|
-
* First compute triplet version of full matrix, then compress to
|
175
|
+
* First compute triplet version of full matrix, then compress to CSC
|
186
176
|
* */
|
187
|
-
static
|
177
|
+
static ScsMatrix *fill_p_matrix(const ScsMatrix *P) {
|
188
178
|
scs_int i, j, k, kk;
|
189
179
|
scs_int Pnzmax = 2 * P->p[P->n]; /* upper bound */
|
190
|
-
|
191
|
-
|
180
|
+
ScsMatrix *P_tmp = SCS(cs_spalloc)(P->n, P->n, Pnzmax, 1, 1);
|
181
|
+
ScsMatrix *P_full;
|
192
182
|
kk = 0;
|
193
183
|
for (j = 0; j < P->n; j++) { /* cols */
|
194
184
|
for (k = P->p[j]; k < P->p[j + 1]; k++) {
|
@@ -209,16 +199,15 @@ static csc *fill_p_matrix(const ScsMatrix *P) {
|
|
209
199
|
kk++;
|
210
200
|
}
|
211
201
|
}
|
212
|
-
|
213
|
-
P_full = SCS(cs_compress)(P_tmp, SCS_NULL);
|
202
|
+
P_full = SCS(cs_compress)(P_tmp, kk, SCS_NULL);
|
214
203
|
SCS(cs_spfree)(P_tmp);
|
215
204
|
return P_full;
|
216
205
|
}
|
217
206
|
|
218
|
-
ScsLinSysWork *
|
219
|
-
|
207
|
+
ScsLinSysWork *scs_init_lin_sys_work(const ScsMatrix *A, const ScsMatrix *P,
|
208
|
+
const scs_float *diag_r) {
|
220
209
|
cudaError_t err;
|
221
|
-
|
210
|
+
ScsMatrix *P_full;
|
222
211
|
ScsLinSysWork *p = SCS_NULL;
|
223
212
|
ScsGpuMatrix *Ag = SCS_NULL;
|
224
213
|
ScsGpuMatrix *Pg = SCS_NULL;
|
@@ -324,7 +313,7 @@ ScsLinSysWork *SCS(init_lin_sys_work)(const ScsMatrix *A, const ScsMatrix *P,
|
|
324
313
|
cusparseCreateDnVec(&p->dn_vec_m, Ag->m, p->tmp_m, SCS_CUDA_FLOAT);
|
325
314
|
|
326
315
|
/* Form preconditioner and copy R_x, 1/R_y to gpu */
|
327
|
-
|
316
|
+
scs_update_lin_sys_diag_r(p, diag_r);
|
328
317
|
|
329
318
|
#if GPU_TRANSPOSE_MAT > 0
|
330
319
|
p->Agt = (ScsGpuMatrix *)scs_malloc(sizeof(ScsGpuMatrix));
|
@@ -367,7 +356,7 @@ ScsLinSysWork *SCS(init_lin_sys_work)(const ScsMatrix *A, const ScsMatrix *P,
|
|
367
356
|
if (err != cudaSuccess) {
|
368
357
|
printf("%s:%d:%s\nERROR_CUDA (*): %s\n", __FILE__, __LINE__, __func__,
|
369
358
|
cudaGetErrorString(err));
|
370
|
-
|
359
|
+
scs_free_lin_sys_work(p);
|
371
360
|
return SCS_NULL;
|
372
361
|
}
|
373
362
|
return p;
|
@@ -466,8 +455,8 @@ static scs_int pcg(ScsLinSysWork *pr, const scs_float *s, scs_float *bg,
|
|
466
455
|
* y = R_y^{-1} (Ax - ry)
|
467
456
|
*
|
468
457
|
*/
|
469
|
-
scs_int
|
470
|
-
|
458
|
+
scs_int scs_solve_lin_sys(ScsLinSysWork *p, scs_float *b, const scs_float *s,
|
459
|
+
scs_float tol) {
|
471
460
|
scs_int cg_its, max_iters;
|
472
461
|
scs_float neg_onef = -1.0;
|
473
462
|
|
@@ -0,0 +1,182 @@
|
|
1
|
+
#include "private.h"
|
2
|
+
|
3
|
+
#define PARDISO_SYMBOLIC (11)
|
4
|
+
#define PARDISO_NUMERIC (22)
|
5
|
+
#define PARDISO_SOLVE (33)
|
6
|
+
#define PARDISO_CLEANUP (-1)
|
7
|
+
|
8
|
+
/* TODO: is it necessary to use pardiso_64 and MKL_Set_Interface_Layer ? */
|
9
|
+
/*
|
10
|
+
#define MKL_INTERFACE_LP64 0
|
11
|
+
#define MKL_INTERFACE_ILP64 1
|
12
|
+
*/
|
13
|
+
#ifdef DLONG
|
14
|
+
#define _PARDISO pardiso_64
|
15
|
+
#else
|
16
|
+
#define _PARDISO pardiso
|
17
|
+
#endif
|
18
|
+
|
19
|
+
/* Prototypes for Pardiso functions */
|
20
|
+
void _PARDISO(void **pt, const scs_int *maxfct, const scs_int *mnum,
|
21
|
+
const scs_int *mtype, const scs_int *phase, const scs_int *n,
|
22
|
+
const scs_float *a, const scs_int *ia, const scs_int *ja,
|
23
|
+
scs_int *perm, const scs_int *nrhs, scs_int *iparm,
|
24
|
+
const scs_int *msglvl, scs_float *b, scs_float *x,
|
25
|
+
scs_int *error);
|
26
|
+
/* scs_int MKL_Set_Interface_Layer(scs_int); */
|
27
|
+
|
28
|
+
const char *scs_get_lin_sys_method() {
|
29
|
+
return "sparse-direct-mkl-pardiso";
|
30
|
+
}
|
31
|
+
|
32
|
+
void scs_free_lin_sys_work(ScsLinSysWork *p) {
|
33
|
+
if (p) {
|
34
|
+
p->phase = PARDISO_CLEANUP;
|
35
|
+
_PARDISO(p->pt, &(p->maxfct), &(p->mnum), &(p->mtype), &(p->phase),
|
36
|
+
&(p->n_plus_m), SCS_NULL, p->kkt->p, p->kkt->i, SCS_NULL,
|
37
|
+
&(p->nrhs), p->iparm, &(p->msglvl), SCS_NULL, SCS_NULL,
|
38
|
+
&(p->error));
|
39
|
+
if (p->error != 0) {
|
40
|
+
scs_printf("Error during MKL Pardiso cleanup: %d", (int)p->error);
|
41
|
+
}
|
42
|
+
if (p->kkt)
|
43
|
+
SCS(cs_spfree)(p->kkt);
|
44
|
+
if (p->sol)
|
45
|
+
scs_free(p->sol);
|
46
|
+
if (p->diag_r_idxs)
|
47
|
+
scs_free(p->diag_r_idxs);
|
48
|
+
if (p->diag_p)
|
49
|
+
scs_free(p->diag_p);
|
50
|
+
scs_free(p);
|
51
|
+
}
|
52
|
+
}
|
53
|
+
|
54
|
+
ScsLinSysWork *scs_init_lin_sys_work(const ScsMatrix *A, const ScsMatrix *P,
|
55
|
+
const scs_float *diag_r) {
|
56
|
+
scs_int i;
|
57
|
+
ScsLinSysWork *p = scs_calloc(1, sizeof(ScsLinSysWork));
|
58
|
+
|
59
|
+
/* TODO: is this necessary with pardiso_64? */
|
60
|
+
/* Set MKL interface layer */
|
61
|
+
/*
|
62
|
+
#ifdef DLONG
|
63
|
+
MKL_Set_Interface_Layer(MKL_INTERFACE_ILP64);
|
64
|
+
#else
|
65
|
+
MKL_Set_Interface_Layer(MKL_INTERFACE_LP64);
|
66
|
+
#endif
|
67
|
+
*/
|
68
|
+
p->n = A->n;
|
69
|
+
p->m = A->m;
|
70
|
+
p->n_plus_m = p->n + p->m;
|
71
|
+
|
72
|
+
/* Even though we overwrite rhs with sol pardiso requires the memory */
|
73
|
+
p->sol = (scs_float *)scs_malloc(sizeof(scs_float) * p->n_plus_m);
|
74
|
+
p->diag_r_idxs = (scs_int *)scs_calloc(p->n_plus_m, sizeof(scs_int));
|
75
|
+
p->diag_p = (scs_float *)scs_calloc(p->n, sizeof(scs_float));
|
76
|
+
|
77
|
+
/* MKL pardiso requires upper triangular CSR matrices. The KKT matrix stuffed
|
78
|
+
* as CSC lower triangular is equivalent. Pass upper=0. */
|
79
|
+
p->kkt = SCS(form_kkt)(A, P, p->diag_p, diag_r, p->diag_r_idxs, 0);
|
80
|
+
if (!(p->kkt)) {
|
81
|
+
scs_printf("Error in forming KKT matrix");
|
82
|
+
scs_free_lin_sys_work(p);
|
83
|
+
return SCS_NULL;
|
84
|
+
}
|
85
|
+
|
86
|
+
for (i = 0; i < 64; i++) {
|
87
|
+
p->iparm[i] = 0; /* Setup Pardiso control parameters */
|
88
|
+
p->pt[i] = 0; /* Initialize the internal solver memory pointer */
|
89
|
+
}
|
90
|
+
|
91
|
+
/* Set Pardiso variables */
|
92
|
+
p->mtype = -2; /* Real symmetric indefinite matrix */
|
93
|
+
p->nrhs = 1; /* Number of right hand sides */
|
94
|
+
p->maxfct = 1; /* Maximum number of numerical factorizations */
|
95
|
+
p->mnum = 1; /* Which factorization to use */
|
96
|
+
p->error = 0; /* Initialize error flag */
|
97
|
+
p->msglvl = VERBOSITY; /* Printing information */
|
98
|
+
|
99
|
+
/* For all iparm vars see MKL documentation */
|
100
|
+
p->iparm[0] = 1; /* Parsido must inspect iparm */
|
101
|
+
p->iparm[1] = 3; /* Fill-in reordering from OpenMP */
|
102
|
+
p->iparm[5] = 1; /* Write solution into b */
|
103
|
+
p->iparm[7] = 0; /* Automatic iterative refinement calculation */
|
104
|
+
p->iparm[9] = 8; /* Perturb the pivot elements with 1E-8 */
|
105
|
+
p->iparm[34] = 1; /* Use C-style indexing for indices */
|
106
|
+
/* p->iparm[36] = -80; */ /* Form block sparse matrices */
|
107
|
+
|
108
|
+
#ifdef SFLOAT
|
109
|
+
p->iparm[27] = 1; /* 1 is single precision, 0 is double */
|
110
|
+
#endif
|
111
|
+
|
112
|
+
/* Permutation and symbolic factorization */
|
113
|
+
scs_int phase = PARDISO_SYMBOLIC;
|
114
|
+
_PARDISO(p->pt, &(p->maxfct), &(p->mnum), &(p->mtype), &phase, &(p->n_plus_m),
|
115
|
+
p->kkt->x, p->kkt->p, p->kkt->i, SCS_NULL, &(p->nrhs), p->iparm,
|
116
|
+
&(p->msglvl), SCS_NULL, SCS_NULL, &(p->error));
|
117
|
+
|
118
|
+
if (p->error != 0) {
|
119
|
+
scs_printf("Error during symbolic factorization: %d", (int)p->error);
|
120
|
+
scs_free_lin_sys_work(p);
|
121
|
+
return SCS_NULL;
|
122
|
+
}
|
123
|
+
|
124
|
+
/* Numerical factorization */
|
125
|
+
p->phase = PARDISO_NUMERIC;
|
126
|
+
_PARDISO(p->pt, &(p->maxfct), &(p->mnum), &(p->mtype), &(p->phase),
|
127
|
+
&(p->n_plus_m), p->kkt->x, p->kkt->p, p->kkt->i, SCS_NULL,
|
128
|
+
&(p->nrhs), p->iparm, &(p->msglvl), SCS_NULL, SCS_NULL, &(p->error));
|
129
|
+
|
130
|
+
if (p->error) {
|
131
|
+
scs_printf("Error during numerical factorization: %d", (int)p->error);
|
132
|
+
scs_free_lin_sys_work(p);
|
133
|
+
return SCS_NULL;
|
134
|
+
}
|
135
|
+
|
136
|
+
if (p->iparm[21] < p->n) {
|
137
|
+
scs_printf("KKT matrix has < n positive eigenvalues. P not PSD.");
|
138
|
+
return SCS_NULL;
|
139
|
+
}
|
140
|
+
|
141
|
+
return p;
|
142
|
+
}
|
143
|
+
|
144
|
+
/* Returns solution to linear system Ax = b with solution stored in b */
|
145
|
+
scs_int scs_solve_lin_sys(ScsLinSysWork *p, scs_float *b, const scs_float *ws,
|
146
|
+
scs_float tol) {
|
147
|
+
/* Back substitution and iterative refinement */
|
148
|
+
p->phase = PARDISO_SOLVE;
|
149
|
+
_PARDISO(p->pt, &(p->maxfct), &(p->mnum), &(p->mtype), &(p->phase),
|
150
|
+
&(p->n_plus_m), p->kkt->x, p->kkt->p, p->kkt->i, SCS_NULL,
|
151
|
+
&(p->nrhs), p->iparm, &(p->msglvl), b, p->sol, &(p->error));
|
152
|
+
if (p->error != 0) {
|
153
|
+
scs_printf("Error during linear system solution: %d", (int)p->error);
|
154
|
+
}
|
155
|
+
return p->error;
|
156
|
+
}
|
157
|
+
|
158
|
+
/* Update factorization when R changes */
|
159
|
+
void scs_update_lin_sys_diag_r(ScsLinSysWork *p, const scs_float *diag_r) {
|
160
|
+
scs_int i;
|
161
|
+
|
162
|
+
for (i = 0; i < p->n; ++i) {
|
163
|
+
/* top left is R_x + P, bottom right is -R_y */
|
164
|
+
p->kkt->x[p->diag_r_idxs[i]] = p->diag_p[i] + diag_r[i];
|
165
|
+
}
|
166
|
+
for (i = p->n; i < p->n + p->m; ++i) {
|
167
|
+
/* top left is R_x + P, bottom right is -R_y */
|
168
|
+
p->kkt->x[p->diag_r_idxs[i]] = -diag_r[i];
|
169
|
+
}
|
170
|
+
|
171
|
+
/* Perform numerical factorization */
|
172
|
+
p->phase = PARDISO_NUMERIC;
|
173
|
+
_PARDISO(p->pt, &(p->maxfct), &(p->mnum), &(p->mtype), &(p->phase),
|
174
|
+
&(p->n_plus_m), p->kkt->x, p->kkt->p, p->kkt->i, SCS_NULL,
|
175
|
+
&(p->nrhs), p->iparm, &(p->msglvl), SCS_NULL, SCS_NULL, &(p->error));
|
176
|
+
|
177
|
+
if (p->error != 0) {
|
178
|
+
scs_printf("Error in PARDISO factorization when updating: %d.\n",
|
179
|
+
(int)p->error);
|
180
|
+
scs_free_lin_sys_work(p);
|
181
|
+
}
|
182
|
+
}
|