scs 0.2.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/CHANGELOG.md +3 -0
- data/LICENSE.txt +22 -0
- data/README.md +84 -0
- data/ext/scs/Rakefile +11 -0
- data/lib/scs/ffi.rb +117 -0
- data/lib/scs/solver.rb +178 -0
- data/lib/scs/version.rb +3 -0
- data/lib/scs.rb +17 -0
- data/vendor/scs/LICENSE.txt +21 -0
- data/vendor/scs/Makefile +164 -0
- data/vendor/scs/README.md +220 -0
- data/vendor/scs/include/aa.h +56 -0
- data/vendor/scs/include/cones.h +46 -0
- data/vendor/scs/include/ctrlc.h +33 -0
- data/vendor/scs/include/glbopts.h +177 -0
- data/vendor/scs/include/linalg.h +26 -0
- data/vendor/scs/include/linsys.h +64 -0
- data/vendor/scs/include/normalize.h +18 -0
- data/vendor/scs/include/rw.h +17 -0
- data/vendor/scs/include/scs.h +161 -0
- data/vendor/scs/include/scs_blas.h +51 -0
- data/vendor/scs/include/util.h +65 -0
- data/vendor/scs/linsys/amatrix.c +305 -0
- data/vendor/scs/linsys/amatrix.h +36 -0
- data/vendor/scs/linsys/amatrix.o +0 -0
- data/vendor/scs/linsys/cpu/direct/private.c +366 -0
- data/vendor/scs/linsys/cpu/direct/private.h +26 -0
- data/vendor/scs/linsys/cpu/direct/private.o +0 -0
- data/vendor/scs/linsys/cpu/indirect/private.c +256 -0
- data/vendor/scs/linsys/cpu/indirect/private.h +31 -0
- data/vendor/scs/linsys/cpu/indirect/private.o +0 -0
- data/vendor/scs/linsys/external/amd/LICENSE.txt +934 -0
- data/vendor/scs/linsys/external/amd/SuiteSparse_config.c +469 -0
- data/vendor/scs/linsys/external/amd/SuiteSparse_config.h +254 -0
- data/vendor/scs/linsys/external/amd/SuiteSparse_config.o +0 -0
- data/vendor/scs/linsys/external/amd/amd.h +400 -0
- data/vendor/scs/linsys/external/amd/amd_1.c +180 -0
- data/vendor/scs/linsys/external/amd/amd_1.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_2.c +1842 -0
- data/vendor/scs/linsys/external/amd/amd_2.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_aat.c +184 -0
- data/vendor/scs/linsys/external/amd/amd_aat.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_control.c +64 -0
- data/vendor/scs/linsys/external/amd/amd_control.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_defaults.c +37 -0
- data/vendor/scs/linsys/external/amd/amd_defaults.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_dump.c +179 -0
- data/vendor/scs/linsys/external/amd/amd_dump.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_global.c +16 -0
- data/vendor/scs/linsys/external/amd/amd_global.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_info.c +119 -0
- data/vendor/scs/linsys/external/amd/amd_info.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_internal.h +304 -0
- data/vendor/scs/linsys/external/amd/amd_order.c +199 -0
- data/vendor/scs/linsys/external/amd/amd_order.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_post_tree.c +120 -0
- data/vendor/scs/linsys/external/amd/amd_post_tree.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_postorder.c +206 -0
- data/vendor/scs/linsys/external/amd/amd_postorder.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_preprocess.c +118 -0
- data/vendor/scs/linsys/external/amd/amd_preprocess.o +0 -0
- data/vendor/scs/linsys/external/amd/amd_valid.c +92 -0
- data/vendor/scs/linsys/external/amd/amd_valid.o +0 -0
- data/vendor/scs/linsys/external/amd/changes +11 -0
- data/vendor/scs/linsys/external/qdldl/LICENSE +201 -0
- data/vendor/scs/linsys/external/qdldl/README.md +120 -0
- data/vendor/scs/linsys/external/qdldl/changes +4 -0
- data/vendor/scs/linsys/external/qdldl/qdldl.c +298 -0
- data/vendor/scs/linsys/external/qdldl/qdldl.h +177 -0
- data/vendor/scs/linsys/external/qdldl/qdldl.o +0 -0
- data/vendor/scs/linsys/external/qdldl/qdldl_types.h +21 -0
- data/vendor/scs/linsys/gpu/gpu.c +41 -0
- data/vendor/scs/linsys/gpu/gpu.h +85 -0
- data/vendor/scs/linsys/gpu/indirect/private.c +304 -0
- data/vendor/scs/linsys/gpu/indirect/private.h +36 -0
- data/vendor/scs/scs.mk +181 -0
- data/vendor/scs/src/aa.c +224 -0
- data/vendor/scs/src/aa.o +0 -0
- data/vendor/scs/src/cones.c +802 -0
- data/vendor/scs/src/cones.o +0 -0
- data/vendor/scs/src/ctrlc.c +77 -0
- data/vendor/scs/src/ctrlc.o +0 -0
- data/vendor/scs/src/linalg.c +84 -0
- data/vendor/scs/src/linalg.o +0 -0
- data/vendor/scs/src/normalize.c +93 -0
- data/vendor/scs/src/normalize.o +0 -0
- data/vendor/scs/src/rw.c +167 -0
- data/vendor/scs/src/rw.o +0 -0
- data/vendor/scs/src/scs.c +975 -0
- data/vendor/scs/src/scs.o +0 -0
- data/vendor/scs/src/scs_version.c +5 -0
- data/vendor/scs/src/scs_version.o +0 -0
- data/vendor/scs/src/util.c +196 -0
- data/vendor/scs/src/util.o +0 -0
- data/vendor/scs/test/data/small_random_socp +0 -0
- data/vendor/scs/test/minunit.h +13 -0
- data/vendor/scs/test/problem_utils.h +93 -0
- data/vendor/scs/test/problems/rob_gauss_cov_est.h +85 -0
- data/vendor/scs/test/problems/small_lp.h +50 -0
- data/vendor/scs/test/problems/small_random_socp.h +33 -0
- data/vendor/scs/test/random_socp_prob.c +171 -0
- data/vendor/scs/test/run_from_file.c +69 -0
- data/vendor/scs/test/run_tests +2 -0
- data/vendor/scs/test/run_tests.c +32 -0
- metadata +203 -0
@@ -0,0 +1,120 @@
|
|
1
|
+
# QDLDL
|
2
|
+
A free LDL factorisation routine for quasi-definite linear systems: `Ax=b`
|
3
|
+
|
4
|
+
[![Build Status](https://travis-ci.org/oxfordcontrol/qdldl.svg?branch=master)](https://travis-ci.org/oxfordcontrol/qdldl)
|
5
|
+
[![Build status](https://ci.appveyor.com/api/projects/status/ns4br7v6y3i5stai/branch/master?svg=true)](https://ci.appveyor.com/project/bstellato/qdldl-8q1mv/branch/master)
|
6
|
+
[![Coverage Status](https://coveralls.io/repos/github/oxfordcontrol/qdldl/badge.svg)](https://coveralls.io/github/oxfordcontrol/qdldl)
|
7
|
+
|
8
|
+
|
9
|
+
## Getting started
|
10
|
+
To start using QDLDL, first clone the repository
|
11
|
+
|
12
|
+
```bash
|
13
|
+
git clone https://github.com/oxfordcontrol/qdldl.git
|
14
|
+
```
|
15
|
+
|
16
|
+
### Build
|
17
|
+
|
18
|
+
To build QDLDL, you need to install [cmake](https://cmake.org/) and run
|
19
|
+
|
20
|
+
```bash
|
21
|
+
mkdir build
|
22
|
+
cd build
|
23
|
+
cmake ..
|
24
|
+
cmake --build .
|
25
|
+
```
|
26
|
+
|
27
|
+
This will generate an `out/` folder with contents:
|
28
|
+
|
29
|
+
- `qdldl_example`: a **code example** from [`examples/c/example.c`](./examples/c/example.c)
|
30
|
+
- `libqdldl`: a static and a dynamic versions of the library.
|
31
|
+
|
32
|
+
You can include an addition option `-DUNITTESTS=ON` when calling `cmake`, which will result in an additional executable `qdldl_tester` being built in the `out/` folder to test QDLDL on a variety of problems, including those with rank deficient or otherwise ill-formatted inputs.
|
33
|
+
|
34
|
+
**N.B.** All files will have file extensions appropriate to your operating system.
|
35
|
+
|
36
|
+
|
37
|
+
### Install/Uninstall
|
38
|
+
|
39
|
+
To install (uninstall) the libraries and headers you can simply run `make install` (`make uninstall`) after running the `cmake` command above.
|
40
|
+
|
41
|
+
|
42
|
+
## Calling QDLDL
|
43
|
+
|
44
|
+
### Main API
|
45
|
+
|
46
|
+
The QDLDL API consists of 5 functions documented in [`include/qdldl.h`](./include/qdldl.h).
|
47
|
+
For more details and a working example see [`examples/c/example.c`](./examples/c/example.c).
|
48
|
+
|
49
|
+
**N.B.** There is **no memory allocation** performed in these routines. The user is assumed to have the working vectors already allocated.
|
50
|
+
|
51
|
+
Here is a brief summary.
|
52
|
+
|
53
|
+
* `QDLDL_etree`: compute the elimination tree for the quasidefinite matrix factorization `A = LDL'`
|
54
|
+
* `QDLDL_factor`: return the factors `L`, `D` and `Dinv = 1./D`
|
55
|
+
* `QDLDL_solve`: solves the linear system `LDL'x = b`
|
56
|
+
* `QDLDL_Lsolve`: solves `Lx = b`
|
57
|
+
* `QDLDL_Ltsolve`: solves `L'x = b`
|
58
|
+
|
59
|
+
In the above function calls the matrices `A` and `L` are stored in compressed sparse column (CSC) format. The matrix `A` is assumed to be symmetric and only the upper triangular portion of A should be passed to the API. The factor `L` is lower triangular with implicit ones on the diagonal (i.e. the diagonal of L is not stored as part of the CSC formatted data.)
|
60
|
+
|
61
|
+
The matrices `D` and `Dinv` are both diagonal matrices, with the diagonal values stored in an array.
|
62
|
+
|
63
|
+
The matrix input `A` should be quasidefinite. The API provides some (non-comprehensive) error checking to protect against non-quasidefinite or non-upper triangular inputs.
|
64
|
+
|
65
|
+
### Custom types for integer, floats and booleans
|
66
|
+
QDLDL uses its own internal types for integers, floats and booleans (`QDLDL_int, QDLDL_float, QDLDL_bool`. They can be specified using the cmake options:
|
67
|
+
|
68
|
+
- `DFLOAT` (default false): uses float numbers instead of doubles
|
69
|
+
- `DLONG` (default true): uses long integers for indexing (for large matrices)
|
70
|
+
|
71
|
+
The `QDLDL_bool` is internally defined as `unsigned char`.
|
72
|
+
|
73
|
+
|
74
|
+
## Linking QDLDL
|
75
|
+
|
76
|
+
### Basic Example
|
77
|
+
A basic example appears in [`examples/c/example.c`](./examples/c/example.c) and is compiled using cmake and the `CMakeLists.txt` file in the root folder.
|
78
|
+
|
79
|
+
|
80
|
+
### Including in a cmake project
|
81
|
+
|
82
|
+
You can include QDLDL in a cmake project `foo` by adding the subdirectory as
|
83
|
+
```
|
84
|
+
# Add project
|
85
|
+
add_subdirectory(qdldl)
|
86
|
+
```
|
87
|
+
|
88
|
+
QDLDL can be linked using a static or dynamic linker
|
89
|
+
```
|
90
|
+
# Link static library
|
91
|
+
target_link_libraries (foo qdldlstatic)
|
92
|
+
|
93
|
+
# Link shared library
|
94
|
+
target_link_libraries (foo qdldl)
|
95
|
+
```
|
96
|
+
for dynamic linking the shared library should be available in your path.
|
97
|
+
|
98
|
+
There is also the option to include QDLDL as an object library in your project.
|
99
|
+
The current `CMakeLists.txt` file creates an object library called `qdldlobject`.
|
100
|
+
This can be added to your project by adding it after your sources.
|
101
|
+
For example, when creating a library `foo` you can add
|
102
|
+
|
103
|
+
```
|
104
|
+
add_library(foo foo.c foo.h $<TARGET_OBJECTS:qdldlobject>)
|
105
|
+
```
|
106
|
+
for more details see the [cmake documentation](https://cmake.org/cmake/help/latest/command/add_library.html#object-libraries).
|
107
|
+
|
108
|
+
|
109
|
+
## The algorithm
|
110
|
+
|
111
|
+
The algorithm is an independent implementation of the elimination tree and factorisation procedures outlined in
|
112
|
+
|
113
|
+
> T. A Davis. [Algorithm 849: a concise sparse cholesky factorization package](https://dl.acm.org/citation.cfm?id=1114277). ACM Trans. Math. Softw., 31(4):587–591, 2005.
|
114
|
+
|
115
|
+
|
116
|
+
## Credits
|
117
|
+
|
118
|
+
- [Paul Goulart](http://users.ox.ac.uk/~engs1373/): main development
|
119
|
+
- [Bartolomeo Stellato](https://stellato.io/): code refactoring and testing
|
120
|
+
- [Goran Banjac](http://control.ee.ethz.ch/~gbanjac/)
|
@@ -0,0 +1,298 @@
|
|
1
|
+
#include "qdldl.h"
|
2
|
+
#include "ctrlc.h"
|
3
|
+
|
4
|
+
#define QDLDL_UNKNOWN (-1)
|
5
|
+
#define QDLDL_USED (1)
|
6
|
+
#define QDLDL_UNUSED (0)
|
7
|
+
|
8
|
+
// //DEBUG
|
9
|
+
// #include <stdio.h>
|
10
|
+
// void qdprint_arrayi(const QDLDL_int* data, QDLDL_int n,char* varName){
|
11
|
+
|
12
|
+
// QDLDL_int i;
|
13
|
+
// printf("%s = [",varName);
|
14
|
+
// for(i=0; i< n; i++){
|
15
|
+
// printf("%lli,",data[i]);
|
16
|
+
// }
|
17
|
+
// printf("]\n");
|
18
|
+
|
19
|
+
// }
|
20
|
+
|
21
|
+
// void qdprint_arrayf(const QDLDL_float* data, QDLDL_int n, char* varName){
|
22
|
+
|
23
|
+
// QDLDL_int i;
|
24
|
+
// printf("%s = [",varName);
|
25
|
+
// for(i=0; i< n; i++){
|
26
|
+
// printf("%.3g,",data[i]);
|
27
|
+
// }
|
28
|
+
// printf("]\n");
|
29
|
+
|
30
|
+
// }
|
31
|
+
// // END DEBUG
|
32
|
+
|
33
|
+
/* Compute the elimination tree for a quasidefinite matrix
|
34
|
+
in compressed sparse column form.
|
35
|
+
*/
|
36
|
+
|
37
|
+
QDLDL_int QDLDL_etree(const QDLDL_int n,
|
38
|
+
const QDLDL_int* Ap,
|
39
|
+
const QDLDL_int* Ai,
|
40
|
+
QDLDL_int* work,
|
41
|
+
QDLDL_int* Lnz,
|
42
|
+
QDLDL_int* etree){
|
43
|
+
|
44
|
+
QDLDL_int sumLnz = 0;
|
45
|
+
QDLDL_int i,j,p;
|
46
|
+
|
47
|
+
|
48
|
+
for(i = 0; i < n; i++){
|
49
|
+
// zero out Lnz and work. Set all etree values to unknown
|
50
|
+
work[i] = 0;
|
51
|
+
Lnz[i] = 0;
|
52
|
+
etree[i] = QDLDL_UNKNOWN;
|
53
|
+
|
54
|
+
//Abort if A doesn't have at least one entry
|
55
|
+
//one entry in every column
|
56
|
+
if(Ap[i] == Ap[i+1]){
|
57
|
+
return -1;
|
58
|
+
}
|
59
|
+
|
60
|
+
}
|
61
|
+
|
62
|
+
for(j = 0; j < n; j++){
|
63
|
+
work[j] = j;
|
64
|
+
for(p = Ap[j]; p < Ap[j+1]; p++){
|
65
|
+
i = Ai[p];
|
66
|
+
if(i > j){return -1;}; //abort if entries on lower triangle
|
67
|
+
while(work[i] != j){
|
68
|
+
if(etree[i] == QDLDL_UNKNOWN){
|
69
|
+
etree[i] = j;
|
70
|
+
}
|
71
|
+
Lnz[i]++; //nonzeros in this column
|
72
|
+
work[i] = j;
|
73
|
+
i = etree[i];
|
74
|
+
}
|
75
|
+
}
|
76
|
+
}
|
77
|
+
|
78
|
+
//compute the total nonzeros in L. This much
|
79
|
+
//space is required to store Li and Lx
|
80
|
+
for(i = 0; i < n; i++){sumLnz += Lnz[i];}
|
81
|
+
|
82
|
+
return sumLnz;
|
83
|
+
}
|
84
|
+
|
85
|
+
|
86
|
+
|
87
|
+
QDLDL_int QDLDL_factor(const QDLDL_int n,
|
88
|
+
const QDLDL_int* Ap,
|
89
|
+
const QDLDL_int* Ai,
|
90
|
+
const QDLDL_float* Ax,
|
91
|
+
QDLDL_int* Lp,
|
92
|
+
QDLDL_int* Li,
|
93
|
+
QDLDL_float* Lx,
|
94
|
+
QDLDL_float* D,
|
95
|
+
QDLDL_float* Dinv,
|
96
|
+
const QDLDL_int* Lnz,
|
97
|
+
const QDLDL_int* etree,
|
98
|
+
QDLDL_bool* bwork,
|
99
|
+
QDLDL_int* iwork,
|
100
|
+
QDLDL_float* fwork){
|
101
|
+
|
102
|
+
QDLDL_int i,j,k,nnzY, bidx, cidx, nextIdx, nnzE, tmpIdx;
|
103
|
+
QDLDL_int *yIdx, *elimBuffer, *LNextSpaceInCol;
|
104
|
+
QDLDL_float *yVals;
|
105
|
+
QDLDL_float yVals_cidx;
|
106
|
+
QDLDL_bool *yMarkers;
|
107
|
+
QDLDL_int positiveValuesInD = 0;
|
108
|
+
|
109
|
+
//partition working memory into pieces
|
110
|
+
yMarkers = bwork;
|
111
|
+
yIdx = iwork;
|
112
|
+
elimBuffer = iwork + n;
|
113
|
+
LNextSpaceInCol = iwork + n*2;
|
114
|
+
yVals = fwork;
|
115
|
+
|
116
|
+
|
117
|
+
Lp[0] = 0; //first column starts at index zero
|
118
|
+
|
119
|
+
for(i = 0; i < n; i++){
|
120
|
+
|
121
|
+
//compute L column indices
|
122
|
+
Lp[i+1] = Lp[i] + Lnz[i]; //cumsum, total at the end
|
123
|
+
|
124
|
+
// set all Yidx to be 'unused' initially
|
125
|
+
//in each column of L, the next available space
|
126
|
+
//to start is just the first space in the column
|
127
|
+
yMarkers[i] = QDLDL_UNUSED;
|
128
|
+
yVals[i] = 0.0;
|
129
|
+
D[i] = 0.0;
|
130
|
+
LNextSpaceInCol[i] = Lp[i];
|
131
|
+
}
|
132
|
+
|
133
|
+
// First element of the diagonal D.
|
134
|
+
D[0] = Ax[0];
|
135
|
+
if(D[0] == 0.0){return -1;}
|
136
|
+
if(D[0] > 0.0){positiveValuesInD++;}
|
137
|
+
Dinv[0] = 1/D[0];
|
138
|
+
|
139
|
+
//Start from 1 here. The upper LH corner is trivially 0
|
140
|
+
//in L b/c we are only computing the subdiagonal elements
|
141
|
+
for(k = 1; k < n; k++){
|
142
|
+
if(scs_is_interrupted()) {
|
143
|
+
scs_printf("interrupt detected in factorization\n");
|
144
|
+
return -1;
|
145
|
+
}
|
146
|
+
|
147
|
+
//NB : For each k, we compute a solution to
|
148
|
+
//y = L(0:(k-1),0:k-1))\b, where b is the kth
|
149
|
+
//column of A that sits above the diagonal.
|
150
|
+
//The solution y is then the kth row of L,
|
151
|
+
//with an implied '1' at the diagonal entry.
|
152
|
+
|
153
|
+
//number of nonzeros in this row of L
|
154
|
+
nnzY = 0; //number of elements in this row
|
155
|
+
|
156
|
+
//This loop determines where nonzeros
|
157
|
+
//will go in the kth row of L, but doesn't
|
158
|
+
//compute the actual values
|
159
|
+
tmpIdx = Ap[k+1];
|
160
|
+
|
161
|
+
for(i = Ap[k]; i < tmpIdx; i++){
|
162
|
+
|
163
|
+
bidx = Ai[i]; // we are working on this element of b
|
164
|
+
|
165
|
+
//Initialize D[k] as the element of this column
|
166
|
+
//corresponding to the diagonal place. Don't use
|
167
|
+
//this element as part of the elimination step
|
168
|
+
//that computes the k^th row of L
|
169
|
+
if(bidx == k){
|
170
|
+
D[k] = Ax[i];
|
171
|
+
continue;
|
172
|
+
}
|
173
|
+
|
174
|
+
yVals[bidx] = Ax[i]; // initialise y(bidx) = b(bidx)
|
175
|
+
|
176
|
+
// use the forward elimination tree to figure
|
177
|
+
// out which elements must be eliminated after
|
178
|
+
// this element of b
|
179
|
+
nextIdx = bidx;
|
180
|
+
|
181
|
+
if(yMarkers[nextIdx] == QDLDL_UNUSED){ //this y term not already visited
|
182
|
+
|
183
|
+
yMarkers[nextIdx] = QDLDL_USED; //I touched this one
|
184
|
+
elimBuffer[0] = nextIdx; // It goes at the start of the current list
|
185
|
+
nnzE = 1; //length of unvisited elimination path from here
|
186
|
+
|
187
|
+
nextIdx = etree[bidx];
|
188
|
+
|
189
|
+
while(nextIdx != QDLDL_UNKNOWN && nextIdx < k){
|
190
|
+
if(yMarkers[nextIdx] == QDLDL_USED) break;
|
191
|
+
|
192
|
+
yMarkers[nextIdx] = QDLDL_USED; //I touched this one
|
193
|
+
elimBuffer[nnzE] = nextIdx; //It goes in the current list
|
194
|
+
nnzE++; //the list is one longer than before
|
195
|
+
nextIdx = etree[nextIdx]; //one step further along tree
|
196
|
+
|
197
|
+
} //end while
|
198
|
+
|
199
|
+
// now I put the buffered elimination list into
|
200
|
+
// my current ordering in reverse order
|
201
|
+
while(nnzE){
|
202
|
+
yIdx[nnzY++] = elimBuffer[--nnzE];
|
203
|
+
} //end while
|
204
|
+
} //end if
|
205
|
+
|
206
|
+
} //end for i
|
207
|
+
|
208
|
+
//This for loop places nonzeros values in the k^th row
|
209
|
+
for(i = (nnzY-1); i >=0; i--){
|
210
|
+
|
211
|
+
//which column are we working on?
|
212
|
+
cidx = yIdx[i];
|
213
|
+
|
214
|
+
// loop along the elements in this
|
215
|
+
// column of L and subtract to solve to y
|
216
|
+
tmpIdx = LNextSpaceInCol[cidx];
|
217
|
+
yVals_cidx = yVals[cidx];
|
218
|
+
for(j = Lp[cidx]; j < tmpIdx; j++){
|
219
|
+
yVals[Li[j]] -= Lx[j]*yVals_cidx;
|
220
|
+
}
|
221
|
+
|
222
|
+
//Now I have the cidx^th element of y = L\b.
|
223
|
+
//so compute the corresponding element of
|
224
|
+
//this row of L and put it into the right place
|
225
|
+
Li[tmpIdx] = k;
|
226
|
+
Lx[tmpIdx] = yVals_cidx *Dinv[cidx];
|
227
|
+
|
228
|
+
//D[k] -= yVals[cidx]*yVals[cidx]*Dinv[cidx];
|
229
|
+
D[k] -= yVals_cidx*Lx[tmpIdx];
|
230
|
+
LNextSpaceInCol[cidx]++;
|
231
|
+
|
232
|
+
//reset the yvalues and indices back to zero and QDLDL_UNUSED
|
233
|
+
//once I'm done with them
|
234
|
+
yVals[cidx] = 0.0;
|
235
|
+
yMarkers[cidx] = QDLDL_UNUSED;
|
236
|
+
|
237
|
+
} //end for i
|
238
|
+
|
239
|
+
//Maintain a count of the positive entries
|
240
|
+
//in D. If we hit a zero, we can't factor
|
241
|
+
//this matrix, so abort
|
242
|
+
if(D[k] == 0.0){return -1;}
|
243
|
+
if(D[k] > 0.0){positiveValuesInD++;}
|
244
|
+
|
245
|
+
//compute the inverse of the diagonal
|
246
|
+
Dinv[k]= 1/D[k];
|
247
|
+
|
248
|
+
} //end for k
|
249
|
+
|
250
|
+
return positiveValuesInD;
|
251
|
+
|
252
|
+
}
|
253
|
+
|
254
|
+
// Solves (L+I)x = b
|
255
|
+
void QDLDL_Lsolve(const QDLDL_int n,
|
256
|
+
const QDLDL_int* Lp,
|
257
|
+
const QDLDL_int* Li,
|
258
|
+
const QDLDL_float* Lx,
|
259
|
+
QDLDL_float* x){
|
260
|
+
|
261
|
+
QDLDL_int i,j;
|
262
|
+
for(i = 0; i < n; i++){
|
263
|
+
for(j = Lp[i]; j < Lp[i+1]; j++){
|
264
|
+
x[Li[j]] -= Lx[j]*x[i];
|
265
|
+
}
|
266
|
+
}
|
267
|
+
}
|
268
|
+
|
269
|
+
// Solves (L+I)'x = b
|
270
|
+
void QDLDL_Ltsolve(const QDLDL_int n,
|
271
|
+
const QDLDL_int* Lp,
|
272
|
+
const QDLDL_int* Li,
|
273
|
+
const QDLDL_float* Lx,
|
274
|
+
QDLDL_float* x){
|
275
|
+
|
276
|
+
QDLDL_int i,j;
|
277
|
+
for(i = n-1; i>=0; i--){
|
278
|
+
for(j = Lp[i]; j < Lp[i+1]; j++){
|
279
|
+
x[i] -= Lx[j]*x[Li[j]];
|
280
|
+
}
|
281
|
+
}
|
282
|
+
}
|
283
|
+
|
284
|
+
// Solves Ax = b where A has given LDL factors
|
285
|
+
void QDLDL_solve(const QDLDL_int n,
|
286
|
+
const QDLDL_int* Lp,
|
287
|
+
const QDLDL_int* Li,
|
288
|
+
const QDLDL_float* Lx,
|
289
|
+
const QDLDL_float* Dinv,
|
290
|
+
QDLDL_float* x){
|
291
|
+
|
292
|
+
QDLDL_int i;
|
293
|
+
|
294
|
+
QDLDL_Lsolve(n,Lp,Li,Lx,x);
|
295
|
+
for(i = 0; i < n; i++) x[i] *= Dinv[i];
|
296
|
+
QDLDL_Ltsolve(n,Lp,Li,Lx,x);
|
297
|
+
|
298
|
+
}
|
@@ -0,0 +1,177 @@
|
|
1
|
+
#ifndef QDLDL_H
|
2
|
+
#define QDLDL_H
|
3
|
+
|
4
|
+
// Include qdldl type options
|
5
|
+
#include "qdldl_types.h"
|
6
|
+
|
7
|
+
# ifdef __cplusplus
|
8
|
+
extern "C" {
|
9
|
+
# endif // ifdef __cplusplus
|
10
|
+
|
11
|
+
/**
|
12
|
+
* Compute the elimination tree for a quasidefinite matrix
|
13
|
+
* in compressed sparse column form, where the input matrix is
|
14
|
+
* assumed to contain data for the upper triangular part of A only,
|
15
|
+
* and there are no duplicate indices.
|
16
|
+
*
|
17
|
+
* Returns an elimination tree for the factorization A = LDL^T and a
|
18
|
+
* count of the nonzeros in each column of L that are strictly below the
|
19
|
+
* diagonal.
|
20
|
+
*
|
21
|
+
* Does not use MALLOC. It is assumed that the arrays work, Lnz, and
|
22
|
+
* etree will be allocated with a number of elements equal to n.
|
23
|
+
*
|
24
|
+
* The data in (n,Ap,Ai) are from a square matrix A in CSC format, and
|
25
|
+
* should include the upper triangular part of A only.
|
26
|
+
*
|
27
|
+
* This function is only intended for factorisation of QD matrices specified
|
28
|
+
* by their upper triangular part. An error is returned if any column has
|
29
|
+
* data below the diagonal or s completely empty.
|
30
|
+
*
|
31
|
+
* For matrices with a non-empty column but a zero on the corresponding diagonal,
|
32
|
+
* this function will *not* return an error, as it may still be possible to factor
|
33
|
+
* such a matrix in LDL form. No promises are made in this case though...
|
34
|
+
*
|
35
|
+
* @param n number of columns in CSC matrix A (assumed square)
|
36
|
+
* @param Ap column pointers (size n+1) for columns of A
|
37
|
+
* @param Ai row indices of A. Has Ap[n] elements
|
38
|
+
* @param work work vector (size n) (no meaning on return)
|
39
|
+
* @param Lnz count of nonzeros in each column of L (size n) below diagonal
|
40
|
+
* @param etree elimination tree (size n)
|
41
|
+
* @return total sum of Lnz (i.e. total nonzeros in L below diagonal). Returns
|
42
|
+
* -1 if the input does not have triu structure or has an empty
|
43
|
+
* column.
|
44
|
+
*
|
45
|
+
*
|
46
|
+
*/
|
47
|
+
|
48
|
+
QDLDL_int QDLDL_etree(const QDLDL_int n,
|
49
|
+
const QDLDL_int* Ap,
|
50
|
+
const QDLDL_int* Ai,
|
51
|
+
QDLDL_int* work,
|
52
|
+
QDLDL_int* Lnz,
|
53
|
+
QDLDL_int* etree);
|
54
|
+
|
55
|
+
/**
|
56
|
+
* Compute an LDL decomposition for a quasidefinite matrix
|
57
|
+
* in compressed sparse column form, where the input matrix is
|
58
|
+
* assumed to contain data for the upper triangular part of A only,
|
59
|
+
* and there are no duplicate indices.
|
60
|
+
*
|
61
|
+
* Returns factors L, D and Dinv = 1./D.
|
62
|
+
*
|
63
|
+
* Does not use MALLOC. It is assumed that L will be a compressed
|
64
|
+
* sparse column matrix with data (Ln,Lp,Li) with sufficient space
|
65
|
+
* allocated, with a number of nonzeros equal to the count given
|
66
|
+
* as a return value by osqp_ldl_etree
|
67
|
+
*
|
68
|
+
* @param n number of columns in L and A (both square)
|
69
|
+
* @param Ap column pointers (size n+1) for columns of A
|
70
|
+
* @param Ai row indices of A. Has Ap[n] elements
|
71
|
+
* @param Ln number of columns in CSC matrix L
|
72
|
+
* @param Lp column pointers (size Ln+1) for columns of L
|
73
|
+
* @param Li row indices of L. Has Lp[Ln] elements
|
74
|
+
* @param D vectorized factor D. Length is n
|
75
|
+
* @param Dinv reciprocal of D. Length is n
|
76
|
+
* @param Lnz count of nonzeros in each column of L below diagonal,
|
77
|
+
* as given by osqp_ldl_etree (not modified)
|
78
|
+
* @param etree elimination tree as as given by osqp_ldl_etree (not modified)
|
79
|
+
* @param bwork working array of bools. Length is n
|
80
|
+
* @param iwork working array of integers. Length is 3*n
|
81
|
+
* @param fwork working array of floats. Length is n
|
82
|
+
* @return Returns a count of the number of positive elements
|
83
|
+
* in D. Returns -1 and exits immediately if any element
|
84
|
+
* of D evaluates exactly to zero (matrix is not quasidefinite
|
85
|
+
* or otherwise LDL factorisable)
|
86
|
+
*
|
87
|
+
*/
|
88
|
+
|
89
|
+
|
90
|
+
QDLDL_int QDLDL_factor(const QDLDL_int n,
|
91
|
+
const QDLDL_int* Ap,
|
92
|
+
const QDLDL_int* Ai,
|
93
|
+
const QDLDL_float* Ax,
|
94
|
+
QDLDL_int* Lp,
|
95
|
+
QDLDL_int* Li,
|
96
|
+
QDLDL_float* Lx,
|
97
|
+
QDLDL_float* D,
|
98
|
+
QDLDL_float* Dinv,
|
99
|
+
const QDLDL_int* Lnz,
|
100
|
+
const QDLDL_int* etree,
|
101
|
+
QDLDL_bool* bwork,
|
102
|
+
QDLDL_int* iwork,
|
103
|
+
QDLDL_float* fwork);
|
104
|
+
|
105
|
+
|
106
|
+
/**
|
107
|
+
* Solves LDL'x = b
|
108
|
+
*
|
109
|
+
* It is assumed that L will be a compressed
|
110
|
+
* sparse column matrix with data (Ln,Lp,Li).
|
111
|
+
*
|
112
|
+
* @param n number of columns in L (both square)
|
113
|
+
* @param Ln number of columns in CSC matrix L
|
114
|
+
* @param Lp column pointers (size Ln+1) for columns of L
|
115
|
+
* @param Li row indices of L. Has Lp[Ln] elements
|
116
|
+
* @param Dinv reciprocal of D. Length is n
|
117
|
+
* @param x initialized to b. Equal to x on return
|
118
|
+
*
|
119
|
+
*
|
120
|
+
*/
|
121
|
+
void QDLDL_solve(const QDLDL_int n,
|
122
|
+
const QDLDL_int* Lp,
|
123
|
+
const QDLDL_int* Li,
|
124
|
+
const QDLDL_float* Lx,
|
125
|
+
const QDLDL_float* Dinv,
|
126
|
+
QDLDL_float* x);
|
127
|
+
|
128
|
+
|
129
|
+
/**
|
130
|
+
* Solves (L+I)x = b
|
131
|
+
*
|
132
|
+
* It is assumed that L will be a compressed
|
133
|
+
* sparse column matrix with data (Ln,Lp,Li).
|
134
|
+
*
|
135
|
+
* @param n number of columns in L (both square)
|
136
|
+
* @param Ln number of columns in CSC matrix L
|
137
|
+
* @param Lp column pointers (size Ln+1) for columns of L
|
138
|
+
* @param Li row indices of L. Has Lp[Ln] elements
|
139
|
+
* @param Dinv reciprocal of D. Length is n
|
140
|
+
* @param x initialized to b. Equal to x on return
|
141
|
+
*
|
142
|
+
*
|
143
|
+
*/
|
144
|
+
|
145
|
+
void QDLDL_Lsolve(const QDLDL_int n,
|
146
|
+
const QDLDL_int* Lp,
|
147
|
+
const QDLDL_int* Li,
|
148
|
+
const QDLDL_float* Lx,
|
149
|
+
QDLDL_float* x);
|
150
|
+
|
151
|
+
/**
|
152
|
+
* Solves (L+I)'x = b
|
153
|
+
*
|
154
|
+
* It is assumed that L will be a compressed
|
155
|
+
* sparse column matrix with data (Ln,Lp,Li).
|
156
|
+
*
|
157
|
+
* @param n number of columns in L (both square)
|
158
|
+
* @param Ln number of columns in CSC matrix L
|
159
|
+
* @param Lp column pointers (size Ln+1) for columns of L
|
160
|
+
* @param Li row indices of L. Has Lp[Ln] elements
|
161
|
+
* @param Dinv reciprocal of D. Length is n
|
162
|
+
* @param x initialized to b. Equal to x on return
|
163
|
+
*
|
164
|
+
*
|
165
|
+
*/
|
166
|
+
|
167
|
+
void QDLDL_Ltsolve(const QDLDL_int n,
|
168
|
+
const QDLDL_int* Lp,
|
169
|
+
const QDLDL_int* Li,
|
170
|
+
const QDLDL_float* Lx,
|
171
|
+
QDLDL_float* x);
|
172
|
+
|
173
|
+
# ifdef __cplusplus
|
174
|
+
}
|
175
|
+
# endif // ifdef __cplusplus
|
176
|
+
|
177
|
+
#endif // ifndef QDLDL_H
|
Binary file
|
@@ -0,0 +1,21 @@
|
|
1
|
+
#ifndef QDLDL_TYPES_H
|
2
|
+
# define QDLDL_TYPES_H
|
3
|
+
|
4
|
+
#include "glbopts.h"
|
5
|
+
|
6
|
+
# ifdef __cplusplus
|
7
|
+
extern "C" {
|
8
|
+
# endif /* ifdef __cplusplus */
|
9
|
+
|
10
|
+
// QDLDL integer and float types
|
11
|
+
|
12
|
+
#define QDLDL_int scs_int
|
13
|
+
#define QDLDL_float scs_float
|
14
|
+
#define QDLDL_bool scs_int
|
15
|
+
|
16
|
+
# ifdef __cplusplus
|
17
|
+
}
|
18
|
+
# endif /* ifdef __cplusplus */
|
19
|
+
|
20
|
+
#endif /* ifndef QDLDL_TYPES_H */
|
21
|
+
|
@@ -0,0 +1,41 @@
|
|
1
|
+
#include "gpu.h"
|
2
|
+
|
3
|
+
void SCS(_accum_by_atrans_gpu)(const ScsGpuMatrix *Ag, const scs_float *x,
|
4
|
+
scs_float *y, cusparseHandle_t cusparse_handle) {
|
5
|
+
/* y += A'*x
|
6
|
+
x and y MUST be on GPU already
|
7
|
+
*/
|
8
|
+
const scs_float onef = 1.0;
|
9
|
+
CUSPARSE(csrmv)
|
10
|
+
(cusparse_handle, CUSPARSE_OPERATION_NON_TRANSPOSE, Ag->n, Ag->m, Ag->Annz,
|
11
|
+
&onef, Ag->descr, Ag->x, Ag->p, Ag->i, x, &onef, y);
|
12
|
+
}
|
13
|
+
|
14
|
+
void SCS(_accum_by_a_gpu)(const ScsGpuMatrix *Ag, const scs_float *x,
|
15
|
+
scs_float *y, cusparseHandle_t cusparse_handle) {
|
16
|
+
/* y += A*x
|
17
|
+
x and y MUST be on GPU already
|
18
|
+
*/
|
19
|
+
const scs_float onef = 1.0;
|
20
|
+
/* The A matrix idx pointers must be ORDERED */
|
21
|
+
CUSPARSE(csrmv)
|
22
|
+
(cusparse_handle, CUSPARSE_OPERATION_TRANSPOSE, Ag->n, Ag->m, Ag->Annz, &onef,
|
23
|
+
Ag->descr, Ag->x, Ag->p, Ag->i, x, &onef, y);
|
24
|
+
}
|
25
|
+
|
26
|
+
void SCS(free_gpu_matrix)(ScsGpuMatrix *A) {
|
27
|
+
cudaFree(A->x);
|
28
|
+
cudaFree(A->i);
|
29
|
+
cudaFree(A->p);
|
30
|
+
cusparseDestroyMatDescr(A->descr);
|
31
|
+
}
|
32
|
+
|
33
|
+
void SCS(normalize_a)(ScsMatrix *A, const ScsSettings *stgs, const ScsCone *k,
|
34
|
+
ScsScaling *scal) {
|
35
|
+
SCS(_normalize_a)(A, stgs, k, scal);
|
36
|
+
}
|
37
|
+
|
38
|
+
void SCS(un_normalize_a)(ScsMatrix *A, const ScsSettings *stgs,
|
39
|
+
const ScsScaling *scal) {
|
40
|
+
SCS(_un_normalize_a)(A, stgs, scal);
|
41
|
+
}
|