1 | """Proof of concept sparse matrix code |
---|
2 | """ |
---|
3 | |
---|
4 | import numpy as num |
---|
5 | |
---|
6 | |
---|
7 | class Sparse: |
---|
8 | |
---|
9 | def __init__(self, *args): |
---|
10 | """Create sparse matrix. |
---|
11 | There are two construction forms |
---|
12 | Usage: |
---|
13 | |
---|
14 | Sparse(A) #Creates sparse matrix from dense matrix A |
---|
15 | Sparse(M, N) #Creates empty MxN sparse matrix |
---|
16 | """ |
---|
17 | |
---|
18 | self.Data = {} |
---|
19 | |
---|
20 | if len(args) == 1: |
---|
21 | try: |
---|
22 | A = num.array(args[0]) |
---|
23 | except: |
---|
24 | raise 'Input must be convertable to a numeric array' |
---|
25 | |
---|
26 | assert len(A.shape) == 2, 'Input must be a 2d matrix' |
---|
27 | |
---|
28 | self.M, self.N = A.shape |
---|
29 | for i in range(self.M): |
---|
30 | for j in range(self.N): |
---|
31 | if A[i, j] != 0.0: |
---|
32 | self.Data[i, j] = A[i, j] |
---|
33 | |
---|
34 | |
---|
35 | elif len(args) == 2: |
---|
36 | self.M = args[0] |
---|
37 | self.N = args[1] |
---|
38 | else: |
---|
39 | raise 'Invalid construction' |
---|
40 | |
---|
41 | self.shape = (self.M, self.N) |
---|
42 | |
---|
43 | |
---|
44 | def __repr__(self): |
---|
45 | return '%d X %d sparse matrix:\n' %(self.M, self.N) + `self.Data` |
---|
46 | |
---|
47 | def __len__(self): |
---|
48 | """Return number of nonzeros of A |
---|
49 | """ |
---|
50 | return len(self.Data) |
---|
51 | |
---|
52 | def nonzeros(self): |
---|
53 | """Return number of nonzeros of A |
---|
54 | """ |
---|
55 | return len(self) |
---|
56 | |
---|
57 | def __setitem__(self, key, x): |
---|
58 | |
---|
59 | i,j = key |
---|
60 | # removing these asserts will not speed things up |
---|
61 | assert 0 <= i < self.M |
---|
62 | assert 0 <= j < self.N |
---|
63 | |
---|
64 | if x != 0: |
---|
65 | self.Data[key] = float(x) |
---|
66 | else: |
---|
67 | if self.Data.has_key( key ): |
---|
68 | del self.Data[key] |
---|
69 | |
---|
70 | def __getitem__(self, key): |
---|
71 | |
---|
72 | i,j = key |
---|
73 | # removing these asserts will not speed things up |
---|
74 | assert 0 <= i < self.M |
---|
75 | assert 0 <= j < self.N |
---|
76 | |
---|
77 | if self.Data.has_key( key ): |
---|
78 | return self.Data[ key ] |
---|
79 | else: |
---|
80 | return 0.0 |
---|
81 | |
---|
82 | def copy(self): |
---|
83 | #FIXME: Use the copy module instead |
---|
84 | new = Sparse(self.M,self.N) |
---|
85 | |
---|
86 | for key in self.Data.keys(): |
---|
87 | i, j = key |
---|
88 | |
---|
89 | new[i,j] = self.Data[i,j] |
---|
90 | |
---|
91 | return new |
---|
92 | |
---|
93 | |
---|
94 | def todense(self): |
---|
95 | D = num.zeros( (self.M, self.N), num.float) |
---|
96 | |
---|
97 | for i in range(self.M): |
---|
98 | for j in range(self.N): |
---|
99 | if self.Data.has_key( (i,j) ): |
---|
100 | D[i, j] = self.Data[ (i,j) ] |
---|
101 | return D |
---|
102 | |
---|
103 | |
---|
104 | |
---|
105 | def __mul__(self, other): |
---|
106 | """Multiply this matrix onto 'other' which can either be |
---|
107 | a numeric vector, a numeric matrix or another sparse matrix. |
---|
108 | """ |
---|
109 | |
---|
110 | try: |
---|
111 | B = num.array(other) |
---|
112 | except: |
---|
113 | msg = 'FIXME: Only numeric types implemented so far' |
---|
114 | raise msg |
---|
115 | |
---|
116 | |
---|
117 | # Assume numeric types from now on |
---|
118 | |
---|
119 | if len(B.shape) == 0: |
---|
120 | # Scalar - use __rmul__ method |
---|
121 | R = B*self |
---|
122 | |
---|
123 | elif len(B.shape) == 1: |
---|
124 | # Vector |
---|
125 | msg = 'Mismatching dimensions: You cannot multiply (%d x %d) matrix onto %d-vector'\ |
---|
126 | %(self.M, self.N, B.shape[0]) |
---|
127 | assert B.shape[0] == self.N, msg |
---|
128 | |
---|
129 | R = num.zeros(self.M, num.float) #Result |
---|
130 | |
---|
131 | # Multiply nonzero elements |
---|
132 | for key in self.Data.keys(): |
---|
133 | i, j = key |
---|
134 | |
---|
135 | R[i] += self.Data[key]*B[j] |
---|
136 | elif len(B.shape) == 2: |
---|
137 | |
---|
138 | |
---|
139 | R = num.zeros((self.M, B.shape[1]), num.float) #Result matrix |
---|
140 | |
---|
141 | # Multiply nonzero elements |
---|
142 | for col in range(R.shape[1]): |
---|
143 | # For each column |
---|
144 | |
---|
145 | for key in self.Data.keys(): |
---|
146 | i, j = key |
---|
147 | |
---|
148 | R[i, col] += self.Data[key]*B[j, col] |
---|
149 | |
---|
150 | |
---|
151 | else: |
---|
152 | raise ValueError, 'Dimension too high: d=%d' %len(B.shape) |
---|
153 | |
---|
154 | return R |
---|
155 | |
---|
156 | |
---|
157 | def __add__(self, other): |
---|
158 | """Add this matrix onto 'other' |
---|
159 | """ |
---|
160 | |
---|
161 | new = other.copy() |
---|
162 | for key in self.Data.keys(): |
---|
163 | i, j = key |
---|
164 | |
---|
165 | new[i,j] += self.Data[key] |
---|
166 | |
---|
167 | return new |
---|
168 | |
---|
169 | |
---|
170 | def __rmul__(self, other): |
---|
171 | """Right multiply this matrix with scalar |
---|
172 | """ |
---|
173 | |
---|
174 | try: |
---|
175 | other = float(other) |
---|
176 | except: |
---|
177 | msg = 'Sparse matrix can only "right-multiply" onto a scalar' |
---|
178 | raise TypeError, msg |
---|
179 | else: |
---|
180 | new = self.copy() |
---|
181 | #Multiply nonzero elements |
---|
182 | for key in new.Data.keys(): |
---|
183 | i, j = key |
---|
184 | |
---|
185 | new.Data[key] = other*new.Data[key] |
---|
186 | |
---|
187 | return new |
---|
188 | |
---|
189 | |
---|
190 | def trans_mult(self, other): |
---|
191 | """Multiply the transpose of matrix with 'other' which can be |
---|
192 | a numeric vector. |
---|
193 | """ |
---|
194 | |
---|
195 | try: |
---|
196 | B = num.array(other) |
---|
197 | except: |
---|
198 | print 'FIXME: Only numeric types implemented so far' |
---|
199 | |
---|
200 | |
---|
201 | #Assume numeric types from now on |
---|
202 | if len(B.shape) == 1: |
---|
203 | #Vector |
---|
204 | |
---|
205 | assert B.shape[0] == self.M, 'Mismatching dimensions' |
---|
206 | |
---|
207 | R = num.zeros((self.N,), num.float) #Result |
---|
208 | |
---|
209 | #Multiply nonzero elements |
---|
210 | for key in self.Data.keys(): |
---|
211 | i, j = key |
---|
212 | |
---|
213 | R[j] += self.Data[key]*B[i] |
---|
214 | |
---|
215 | else: |
---|
216 | raise 'Can only multiply with 1d array' |
---|
217 | |
---|
218 | return R |
---|
219 | |
---|
220 | class Sparse_CSR: |
---|
221 | |
---|
222 | def __init__(self, A): |
---|
223 | """Create sparse matrix in csr format. |
---|
224 | |
---|
225 | Sparse_CSR(A) #creates csr sparse matrix from sparse matrix |
---|
226 | Matrices are not built using this format, since it's painful to |
---|
227 | add values to an existing sparse_CSR instance (hence there are no |
---|
228 | objects to do this.) |
---|
229 | |
---|
230 | Rather, build a matrix, and convert it to this format for a speed |
---|
231 | increase. |
---|
232 | |
---|
233 | data - a 1D array of the data |
---|
234 | Colind - The ith item in this 1D array is the column index of the |
---|
235 | ith data in the data array |
---|
236 | rowptr - 1D array, with the index representing the row of the matrix. |
---|
237 | The item in the row represents the index into colind of the |
---|
238 | first data value of this row. |
---|
239 | Regard it as a pointer into the colind array, for the ith row. |
---|
240 | |
---|
241 | |
---|
242 | """ |
---|
243 | |
---|
244 | if isinstance(A,Sparse): |
---|
245 | |
---|
246 | keys = A.Data.keys() |
---|
247 | keys.sort() |
---|
248 | nnz = len(keys) |
---|
249 | data = num.zeros ( (nnz,), num.float) |
---|
250 | colind = num.zeros ( (nnz,), num.int) |
---|
251 | row_ptr = num.zeros ( (A.M+1,), num.int) |
---|
252 | current_row = -1 |
---|
253 | k = 0 |
---|
254 | for key in keys: |
---|
255 | ikey0 = int(key[0]) |
---|
256 | ikey1 = int(key[1]) |
---|
257 | if ikey0 != current_row: |
---|
258 | current_row = ikey0 |
---|
259 | row_ptr[ikey0] = k |
---|
260 | data[k] = A.Data[key] |
---|
261 | colind[k] = ikey1 |
---|
262 | k += 1 |
---|
263 | for row in range(current_row+1, A.M+1): |
---|
264 | row_ptr[row] = nnz |
---|
265 | #row_ptr[-1] = nnz |
---|
266 | |
---|
267 | self.data = data |
---|
268 | self.colind = colind |
---|
269 | self.row_ptr = row_ptr |
---|
270 | self.M = A.M |
---|
271 | self.N = A.N |
---|
272 | else: |
---|
273 | raise ValueError, "Sparse_CSR(A) expects A == Sparse Matrix" |
---|
274 | |
---|
275 | def __repr__(self): |
---|
276 | return '%d X %d sparse matrix:\n' %(self.M, self.N) + `self.data` |
---|
277 | |
---|
278 | def __len__(self): |
---|
279 | """Return number of nonzeros of A |
---|
280 | """ |
---|
281 | return self.row_ptr[-1] |
---|
282 | |
---|
283 | def nonzeros(self): |
---|
284 | """Return number of nonzeros of A |
---|
285 | """ |
---|
286 | return len(self) |
---|
287 | |
---|
288 | def todense(self): |
---|
289 | D = num.zeros( (self.M, self.N), num.float) |
---|
290 | |
---|
291 | for i in range(self.M): |
---|
292 | for ckey in range(self.row_ptr[i],self.row_ptr[i+1]): |
---|
293 | j = self.colind[ckey] |
---|
294 | D[i, j] = self.data[ckey] |
---|
295 | return D |
---|
296 | |
---|
297 | def __mul__(self, other): |
---|
298 | """Multiply this matrix onto 'other' which can either be |
---|
299 | a numeric vector. |
---|
300 | """ |
---|
301 | |
---|
302 | try: |
---|
303 | B = num.array(other) |
---|
304 | except: |
---|
305 | raise ValueError, 'FIXME: Only numeric types implemented so far' |
---|
306 | |
---|
307 | |
---|
308 | # Assume numeric types from now on |
---|
309 | |
---|
310 | if len(B.shape) == 0: |
---|
311 | # Scalar - use __rmul__ method |
---|
312 | #R = B*self |
---|
313 | raise TypeError, 'Sparse matrix X scalar not implemented' |
---|
314 | |
---|
315 | elif len(B.shape) == 1: |
---|
316 | # Vector |
---|
317 | msg = 'Mismatching dimensions: You cannot multiply (%d x %d) matrix onto %d-vector'\ |
---|
318 | %(self.M, self.N, B.shape[0]) |
---|
319 | assert B.shape[0] == self.N, msg |
---|
320 | |
---|
321 | R = csr_mv(self,B) |
---|
322 | |
---|
323 | |
---|
324 | elif len(B.shape) == 2: |
---|
325 | raise TypeError, 'Sparse matrix X matrix multiplication not implemented' |
---|
326 | |
---|
327 | else: |
---|
328 | raise ValueError, 'Dimension too high: d=%d' %len(B.shape) |
---|
329 | |
---|
330 | return R |
---|
331 | |
---|
332 | |
---|
333 | |
---|
334 | # Setup for C extensions |
---|
335 | from anuga.utilities import compile |
---|
336 | if compile.can_use_C_extension('sparse_ext.c'): |
---|
337 | # Access underlying c implementations |
---|
338 | from sparse_ext import csr_mv |
---|
339 | |
---|
340 | |
---|
341 | if __name__ == '__main__': |
---|
342 | # A little selftest |
---|
343 | |
---|
344 | A = Sparse(3,3) |
---|
345 | |
---|
346 | A[1,1] = 4 |
---|
347 | |
---|
348 | |
---|
349 | print A |
---|
350 | print A.todense() |
---|
351 | |
---|
352 | A[1,1] = 0 |
---|
353 | |
---|
354 | print A |
---|
355 | print A.todense() |
---|
356 | |
---|
357 | A[1,2] = 0 |
---|
358 | |
---|
359 | |
---|
360 | A[0,0] = 3 |
---|
361 | A[1,1] = 2 |
---|
362 | A[1,2] = 2 |
---|
363 | A[2,2] = 1 |
---|
364 | |
---|
365 | print A |
---|
366 | print A.todense() |
---|
367 | |
---|
368 | |
---|
369 | #Right hand side vector |
---|
370 | v = [2,3,4] |
---|
371 | |
---|
372 | u = A*v |
---|
373 | print u |
---|
374 | assert num.allclose(u, [6,14,4]) |
---|
375 | |
---|
376 | u = A.trans_mult(v) |
---|
377 | print u |
---|
378 | assert num.allclose(u, [6,6,10]) |
---|
379 | |
---|
380 | #Right hand side column |
---|
381 | v = num.array([[2,4],[3,4],[4,4]]) |
---|
382 | |
---|
383 | u = A*v[:,0] |
---|
384 | assert num.allclose(u, [6,14,4]) |
---|
385 | |
---|
386 | #u = A*v[:,1] |
---|
387 | #print u |
---|
388 | print A.shape |
---|
389 | |
---|
390 | B = 3*A |
---|
391 | print B.todense() |
---|
392 | |
---|
393 | B[1,0] = 2 |
---|
394 | |
---|
395 | C = A+B |
---|
396 | |
---|
397 | print C.todense() |
---|
398 | |
---|
399 | C = Sparse_CSR(C) |
---|
400 | |
---|
401 | y = C*[6,14,4] |
---|
402 | |
---|
403 | print y |
---|
404 | |
---|
405 | y2 = C*[[6,4],[4,28],[4,8]] |
---|
406 | |
---|
407 | print y2 |
---|