nmatrix 0.0.1 → 0.0.2
Sign up to get free protection for your applications and to get access to all the features.
- data/.gitignore +27 -0
- data/.rspec +2 -0
- data/Gemfile +3 -5
- data/Guardfile +6 -0
- data/History.txt +33 -0
- data/Manifest.txt +41 -38
- data/README.rdoc +88 -11
- data/Rakefile +35 -53
- data/ext/nmatrix/data/complex.h +372 -0
- data/ext/nmatrix/data/data.cpp +275 -0
- data/ext/nmatrix/data/data.h +707 -0
- data/ext/nmatrix/data/rational.h +421 -0
- data/ext/nmatrix/data/ruby_object.h +446 -0
- data/ext/nmatrix/extconf.rb +101 -51
- data/ext/nmatrix/new_extconf.rb +56 -0
- data/ext/nmatrix/nmatrix.cpp +1609 -0
- data/ext/nmatrix/nmatrix.h +265 -849
- data/ext/nmatrix/ruby_constants.cpp +134 -0
- data/ext/nmatrix/ruby_constants.h +103 -0
- data/ext/nmatrix/storage/common.cpp +70 -0
- data/ext/nmatrix/storage/common.h +170 -0
- data/ext/nmatrix/storage/dense.cpp +665 -0
- data/ext/nmatrix/storage/dense.h +116 -0
- data/ext/nmatrix/storage/list.cpp +1088 -0
- data/ext/nmatrix/storage/list.h +129 -0
- data/ext/nmatrix/storage/storage.cpp +658 -0
- data/ext/nmatrix/storage/storage.h +99 -0
- data/ext/nmatrix/storage/yale.cpp +1601 -0
- data/ext/nmatrix/storage/yale.h +208 -0
- data/ext/nmatrix/ttable_helper.rb +126 -0
- data/ext/nmatrix/{yale/smmp1_header.template.c → types.h} +36 -9
- data/ext/nmatrix/util/io.cpp +295 -0
- data/ext/nmatrix/util/io.h +117 -0
- data/ext/nmatrix/util/lapack.h +1175 -0
- data/ext/nmatrix/util/math.cpp +557 -0
- data/ext/nmatrix/util/math.h +1363 -0
- data/ext/nmatrix/util/sl_list.cpp +475 -0
- data/ext/nmatrix/util/sl_list.h +255 -0
- data/ext/nmatrix/util/util.h +78 -0
- data/lib/nmatrix/blas.rb +70 -0
- data/lib/nmatrix/io/mat5_reader.rb +567 -0
- data/lib/nmatrix/io/mat_reader.rb +162 -0
- data/lib/{string.rb → nmatrix/monkeys.rb} +49 -2
- data/lib/nmatrix/nmatrix.rb +199 -0
- data/lib/nmatrix/nvector.rb +103 -0
- data/lib/nmatrix/version.rb +27 -0
- data/lib/nmatrix.rb +22 -230
- data/nmatrix.gemspec +59 -0
- data/scripts/mac-brew-gcc.sh +47 -0
- data/spec/4x4_sparse.mat +0 -0
- data/spec/4x5_dense.mat +0 -0
- data/spec/blas_spec.rb +47 -0
- data/spec/elementwise_spec.rb +164 -0
- data/spec/io_spec.rb +60 -0
- data/spec/lapack_spec.rb +52 -0
- data/spec/math_spec.rb +96 -0
- data/spec/nmatrix_spec.rb +93 -89
- data/spec/nmatrix_yale_spec.rb +52 -36
- data/spec/nvector_spec.rb +1 -1
- data/spec/slice_spec.rb +257 -0
- data/spec/spec_helper.rb +51 -0
- data/spec/utm5940.mtx +83844 -0
- metadata +113 -71
- data/.autotest +0 -23
- data/.gemtest +0 -0
- data/ext/nmatrix/cblas.c +0 -150
- data/ext/nmatrix/dense/blas_header.template.c +0 -52
- data/ext/nmatrix/dense/elementwise.template.c +0 -107
- data/ext/nmatrix/dense/gemm.template.c +0 -159
- data/ext/nmatrix/dense/gemv.template.c +0 -130
- data/ext/nmatrix/dense/rationalmath.template.c +0 -68
- data/ext/nmatrix/dense.c +0 -307
- data/ext/nmatrix/depend +0 -18
- data/ext/nmatrix/generator/syntax_tree.rb +0 -481
- data/ext/nmatrix/generator.rb +0 -594
- data/ext/nmatrix/list.c +0 -774
- data/ext/nmatrix/nmatrix.c +0 -1977
- data/ext/nmatrix/rational.c +0 -98
- data/ext/nmatrix/yale/complexmath.template.c +0 -71
- data/ext/nmatrix/yale/elementwise.template.c +0 -46
- data/ext/nmatrix/yale/elementwise_op.template.c +0 -73
- data/ext/nmatrix/yale/numbmm.template.c +0 -94
- data/ext/nmatrix/yale/smmp1.template.c +0 -21
- data/ext/nmatrix/yale/smmp2.template.c +0 -43
- data/ext/nmatrix/yale/smmp2_header.template.c +0 -46
- data/ext/nmatrix/yale/sort_columns.template.c +0 -56
- data/ext/nmatrix/yale/symbmm.template.c +0 -54
- data/ext/nmatrix/yale/transp.template.c +0 -68
- data/ext/nmatrix/yale.c +0 -726
- data/lib/array.rb +0 -67
- data/spec/syntax_tree_spec.rb +0 -46
@@ -0,0 +1,1601 @@
|
|
1
|
+
/////////////////////////////////////////////////////////////////////
|
2
|
+
// = NMatrix
|
3
|
+
//
|
4
|
+
// A linear algebra library for scientific computation in Ruby.
|
5
|
+
// NMatrix is part of SciRuby.
|
6
|
+
//
|
7
|
+
// NMatrix was originally inspired by and derived from NArray, by
|
8
|
+
// Masahiro Tanaka: http://narray.rubyforge.org
|
9
|
+
//
|
10
|
+
// == Copyright Information
|
11
|
+
//
|
12
|
+
// SciRuby is Copyright (c) 2010 - 2012, Ruby Science Foundation
|
13
|
+
// NMatrix is Copyright (c) 2012, Ruby Science Foundation
|
14
|
+
//
|
15
|
+
// Please see LICENSE.txt for additional copyright notices.
|
16
|
+
//
|
17
|
+
// == Contributing
|
18
|
+
//
|
19
|
+
// By contributing source code to SciRuby, you agree to be bound by
|
20
|
+
// our Contributor Agreement:
|
21
|
+
//
|
22
|
+
// * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
|
23
|
+
//
|
24
|
+
// == yale.c
|
25
|
+
//
|
26
|
+
// "new yale" storage format for 2D matrices (like yale, but with
|
27
|
+
// the diagonal pulled out for O(1) access).
|
28
|
+
//
|
29
|
+
// Specifications:
|
30
|
+
// * dtype and index dtype must necessarily differ
|
31
|
+
// * index dtype is defined by whatever unsigned type can store
|
32
|
+
// max(rows,cols)
|
33
|
+
// * that means vector ija stores only index dtype, but a stores
|
34
|
+
// dtype
|
35
|
+
// * vectors must be able to grow as necessary
|
36
|
+
// * maximum size is rows*cols+1
|
37
|
+
|
38
|
+
/*
|
39
|
+
* Standard Includes
|
40
|
+
*/
|
41
|
+
|
42
|
+
#include <ruby.h>
|
43
|
+
#include <algorithm> // std::min
|
44
|
+
#include <cstdio> // std::fprintf
|
45
|
+
|
46
|
+
/*
|
47
|
+
* Project Includes
|
48
|
+
*/
|
49
|
+
|
50
|
+
// #include "types.h"
|
51
|
+
#include "util/math.h"
|
52
|
+
|
53
|
+
#include "data/data.h"
|
54
|
+
|
55
|
+
#include "common.h"
|
56
|
+
#include "yale.h"
|
57
|
+
|
58
|
+
#include "nmatrix.h"
|
59
|
+
#include "ruby_constants.h"
|
60
|
+
|
61
|
+
/*
|
62
|
+
* Macros
|
63
|
+
*/
|
64
|
+
#ifndef NM_MAX
|
65
|
+
#define NM_MAX(a,b) (((a)>(b))?(a):(b))
|
66
|
+
#define NM_MIN(a,b) (((a)<(b))?(a):(b))
|
67
|
+
#endif
|
68
|
+
|
69
|
+
|
70
|
+
/*
|
71
|
+
* Global Variables
|
72
|
+
*/
|
73
|
+
|
74
|
+
/*
|
75
|
+
* Forward Declarations
|
76
|
+
*/
|
77
|
+
|
78
|
+
extern "C" {
|
79
|
+
static YALE_STORAGE* nm_copy_alloc_struct(const YALE_STORAGE* rhs, const dtype_t new_dtype, const size_t new_capacity, const size_t new_size);
|
80
|
+
static YALE_STORAGE* alloc(dtype_t dtype, size_t* shape, size_t dim);
|
81
|
+
|
82
|
+
/* Ruby-accessible functions */
|
83
|
+
static VALUE nm_size(VALUE self);
|
84
|
+
static VALUE nm_a(VALUE self);
|
85
|
+
static VALUE nm_d(VALUE self);
|
86
|
+
static VALUE nm_lu(VALUE self);
|
87
|
+
static VALUE nm_ia(VALUE self);
|
88
|
+
static VALUE nm_ja(VALUE self);
|
89
|
+
static VALUE nm_ija(VALUE self);
|
90
|
+
|
91
|
+
} // end extern "C" block
|
92
|
+
|
93
|
+
|
94
|
+
|
95
|
+
namespace nm { namespace yale_storage {
|
96
|
+
|
97
|
+
template <typename DType, typename IType>
|
98
|
+
static bool ndrow_is_empty(const YALE_STORAGE* s, IType ija, const IType ija_next);
|
99
|
+
|
100
|
+
template <typename LDType, typename RDType, typename IType>
|
101
|
+
static bool ndrow_eqeq_ndrow(const YALE_STORAGE* l, const YALE_STORAGE* r, IType l_ija, const IType l_ija_next, IType r_ija, const IType r_ija_next);
|
102
|
+
|
103
|
+
template <typename LDType, typename RDType, typename IType>
|
104
|
+
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right);
|
105
|
+
|
106
|
+
template <typename IType>
|
107
|
+
static YALE_STORAGE* copy_alloc_struct(const YALE_STORAGE* rhs, const dtype_t new_dtype, const size_t new_capacity, const size_t new_size);
|
108
|
+
|
109
|
+
template <typename IType>
|
110
|
+
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, IType n);
|
111
|
+
|
112
|
+
template <typename IType>
|
113
|
+
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool* found);
|
114
|
+
|
115
|
+
template <typename IType>
|
116
|
+
static inline size_t get_size(const YALE_STORAGE* storage);
|
117
|
+
|
118
|
+
template <typename DType, typename IType>
|
119
|
+
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, DType* val, size_t n, bool struct_only);
|
120
|
+
|
121
|
+
template <typename DType, typename IType>
|
122
|
+
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only);
|
123
|
+
|
124
|
+
template <typename nm::ewop_t op, typename IType, typename DType>
|
125
|
+
YALE_STORAGE* ew_op(const YALE_STORAGE* left, const YALE_STORAGE* right, dtype_t dtype);
|
126
|
+
|
127
|
+
/*
|
128
|
+
* Functions
|
129
|
+
*/
|
130
|
+
|
131
|
+
/*
|
132
|
+
* Create Yale storage from IA, JA, and A vectors given in Old Yale format (probably from a file, since NMatrix only uses
|
133
|
+
* new Yale for its storage).
|
134
|
+
*
|
135
|
+
* This function is needed for Matlab .MAT v5 IO.
|
136
|
+
*/
|
137
|
+
template <typename LDType, typename RDType, typename IType>
|
138
|
+
YALE_STORAGE* create_from_old_yale(dtype_t dtype, size_t* shape, void* r_ia, void* r_ja, void* r_a) {
|
139
|
+
IType* ir = reinterpret_cast<IType*>(r_ia);
|
140
|
+
IType* jr = reinterpret_cast<IType*>(r_ja);
|
141
|
+
RDType* ar = reinterpret_cast<RDType*>(r_a);
|
142
|
+
|
143
|
+
// Read through ia and ja and figure out the ndnz (non-diagonal non-zeros) count.
|
144
|
+
size_t ndnz = 0, i, p, p_next;
|
145
|
+
|
146
|
+
for (i = 0; i < shape[0]; ++i) { // Walk down rows
|
147
|
+
for (p = ir[i], p_next = ir[i+1]; p < p_next; ++p) { // Now walk through columns
|
148
|
+
|
149
|
+
if (i != jr[p]) ++ndnz; // entry is non-diagonal and probably nonzero
|
150
|
+
|
151
|
+
}
|
152
|
+
}
|
153
|
+
|
154
|
+
// Having walked through the matrix, we now go about allocating the space for it.
|
155
|
+
YALE_STORAGE* s = alloc(dtype, shape, 2);
|
156
|
+
|
157
|
+
s->capacity = shape[0] + ndnz + 1;
|
158
|
+
s->ndnz = ndnz;
|
159
|
+
|
160
|
+
// Setup IJA and A arrays
|
161
|
+
s->ija = ALLOC_N( IType, s->capacity );
|
162
|
+
s->a = ALLOC_N( LDType, s->capacity );
|
163
|
+
IType* ijl = reinterpret_cast<IType*>(s->ija);
|
164
|
+
LDType* al = reinterpret_cast<LDType*>(s->a);
|
165
|
+
|
166
|
+
// set the diagonal to zero -- this prevents uninitialized values from popping up.
|
167
|
+
for (size_t index = 0; index < shape[0]; ++index) {
|
168
|
+
al[index] = 0;
|
169
|
+
}
|
170
|
+
|
171
|
+
// Figure out where to start writing JA in IJA:
|
172
|
+
size_t pp = s->shape[0]+1;
|
173
|
+
|
174
|
+
// Find beginning of first row
|
175
|
+
p = ir[0];
|
176
|
+
|
177
|
+
// Now fill the arrays
|
178
|
+
for (i = 0; i < s->shape[0]; ++i) {
|
179
|
+
|
180
|
+
// Set the beginning of the row (of output)
|
181
|
+
ijl[i] = pp;
|
182
|
+
|
183
|
+
// Now walk through columns, starting at end of row (of input)
|
184
|
+
for (size_t p_next = ir[i+1]; p < p_next; ++p, ++pp) {
|
185
|
+
|
186
|
+
if (i == jr[p]) { // diagonal
|
187
|
+
|
188
|
+
al[i] = ar[p];
|
189
|
+
--pp;
|
190
|
+
|
191
|
+
} else { // nondiagonal
|
192
|
+
|
193
|
+
ijl[pp] = jr[p];
|
194
|
+
al[pp] = ar[p];
|
195
|
+
|
196
|
+
}
|
197
|
+
}
|
198
|
+
}
|
199
|
+
|
200
|
+
ijl[i] = pp; // Set the end of the last row
|
201
|
+
|
202
|
+
// Set the zero position for our output matrix
|
203
|
+
al[i] = 0;
|
204
|
+
|
205
|
+
return s;
|
206
|
+
}
|
207
|
+
|
208
|
+
|
209
|
+
/*
|
210
|
+
* Take two Yale storages and merge them into a new Yale storage.
|
211
|
+
*
|
212
|
+
* Uses the left as a template for the creation of a new one.
|
213
|
+
*/
|
214
|
+
template <typename DType, typename IType>
|
215
|
+
YALE_STORAGE* create_merged(const YALE_STORAGE* left, const YALE_STORAGE* right) {
|
216
|
+
char ins_type;
|
217
|
+
|
218
|
+
size_t size = get_size<IType>(left);
|
219
|
+
|
220
|
+
// s represents the resulting storage
|
221
|
+
YALE_STORAGE* s = copy_alloc_struct<IType>(left, left->dtype, NM_MAX(left->capacity, right->capacity), size);
|
222
|
+
|
223
|
+
IType* sija = reinterpret_cast<IType*>(s->ija);
|
224
|
+
IType* rija = reinterpret_cast<IType*>(right->ija);
|
225
|
+
|
226
|
+
// set the element between D and LU (the boundary in A), which should be 0.
|
227
|
+
reinterpret_cast<DType*>(s->a)[s->shape[0]] = reinterpret_cast<DType*>(left->a)[left->shape[0]];
|
228
|
+
|
229
|
+
if (right && right != left) {
|
230
|
+
// some operations are unary and don't need this; others are x+x and don't need this
|
231
|
+
|
232
|
+
for (IType i = 0; i < s->shape[0]; ++i) {
|
233
|
+
|
234
|
+
IType ija = sija[i];
|
235
|
+
IType ija_next = sija[i+1];
|
236
|
+
|
237
|
+
for (IType r_ija = rija[i]; r_ija < rija[i+1]; ++r_ija) {
|
238
|
+
|
239
|
+
size_t ja = sija[ija]; // insert expects a size_t
|
240
|
+
|
241
|
+
if (ija == ija_next) {
|
242
|
+
// destination row is empty
|
243
|
+
ins_type = vector_insert<DType,IType>(s, ija, &ja, NULL, 1, true);
|
244
|
+
increment_ia_after<IType>(s, s->shape[0], i, 1);
|
245
|
+
++(s->ndnz);
|
246
|
+
++ija;
|
247
|
+
|
248
|
+
if (ins_type == 'i') ++ija_next;
|
249
|
+
|
250
|
+
} else {
|
251
|
+
bool found;
|
252
|
+
|
253
|
+
// merge positions into destination row
|
254
|
+
IType pos = insert_search<IType>(s, ija, ija_next-1, sija[ija], &found);
|
255
|
+
|
256
|
+
if (!found) {
|
257
|
+
vector_insert<DType,IType>(s, pos, &ja, NULL, 1, true);
|
258
|
+
increment_ia_after<IType>(s, s->shape[0], i, 1);
|
259
|
+
++(s->ndnz);
|
260
|
+
|
261
|
+
if (ins_type == 'i') ++ija_next;
|
262
|
+
}
|
263
|
+
|
264
|
+
// can now set a left boundary for the next search
|
265
|
+
ija = pos + 1;
|
266
|
+
}
|
267
|
+
}
|
268
|
+
}
|
269
|
+
}
|
270
|
+
|
271
|
+
return s;
|
272
|
+
}
|
273
|
+
|
274
|
+
|
275
|
+
/*
|
276
|
+
* Empty the matrix by initializing the IJA vector and setting the diagonal to 0.
|
277
|
+
*
|
278
|
+
* Called when most YALE_STORAGE objects are created.
|
279
|
+
*/
|
280
|
+
template <typename DType, typename IType>
|
281
|
+
void init(YALE_STORAGE* s) {
|
282
|
+
IType IA_INIT = s->shape[0] + 1;
|
283
|
+
|
284
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
285
|
+
// clear out IJA vector
|
286
|
+
for (IType i = 0; i < IA_INIT; ++i) {
|
287
|
+
ija[i] = IA_INIT; // set initial values for IJA
|
288
|
+
}
|
289
|
+
|
290
|
+
clear_diagonal_and_zero<DType>(s);
|
291
|
+
}
|
292
|
+
|
293
|
+
size_t max_size(YALE_STORAGE* s) {
|
294
|
+
size_t result = s->shape[0]*s->shape[1] + 1;
|
295
|
+
if (s->shape[0] > s->shape[1])
|
296
|
+
result += s->shape[0] - s->shape[1];
|
297
|
+
|
298
|
+
return result;
|
299
|
+
}
|
300
|
+
///////////////
|
301
|
+
// Accessors //
|
302
|
+
///////////////
|
303
|
+
|
304
|
+
/*
|
305
|
+
* Returns a slice of YALE_STORAGE object by coppy
|
306
|
+
*
|
307
|
+
* Slicing-related.
|
308
|
+
*/
|
309
|
+
template <typename DType,typename IType>
|
310
|
+
void* get(YALE_STORAGE* storage, SLICE* slice) {
|
311
|
+
|
312
|
+
size_t *offset = slice->coords;
|
313
|
+
// Copy shape for yale construction
|
314
|
+
size_t* shape = ALLOC_N(size_t, 2);
|
315
|
+
shape[0] = slice->lengths[0];
|
316
|
+
shape[1] = slice->lengths[1];
|
317
|
+
|
318
|
+
IType *src_ija = reinterpret_cast<IType*>(storage->ija);
|
319
|
+
DType *src_a = reinterpret_cast<DType*>(storage->a);
|
320
|
+
|
321
|
+
// Calc ndnz
|
322
|
+
size_t ndnz = 0;
|
323
|
+
size_t i,j; // indexes of destination matrix
|
324
|
+
size_t k,l; // indexes of source matrix
|
325
|
+
for (i = 0; i < shape[0]; i++) {
|
326
|
+
k = i + offset[0];
|
327
|
+
for (j = 0; j < shape[1]; j++) {
|
328
|
+
l = j + offset[1];
|
329
|
+
|
330
|
+
if (j == i) continue;
|
331
|
+
|
332
|
+
if (k == l && src_a[k] != 0) ndnz++; // for diagonal element of source
|
333
|
+
else { // for non-diagonal element
|
334
|
+
for (size_t c = src_ija[k]; c < src_ija[k+1]; c++)
|
335
|
+
if (src_ija[c] == l) { ndnz++; break; }
|
336
|
+
}
|
337
|
+
|
338
|
+
}
|
339
|
+
}
|
340
|
+
|
341
|
+
size_t request_capacity = shape[0] + ndnz + 1;
|
342
|
+
YALE_STORAGE* ns = nm_yale_storage_create(storage->dtype, shape, 2, request_capacity);
|
343
|
+
|
344
|
+
if (ns->capacity < request_capacity)
|
345
|
+
rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %ld requested, max allowable is %ld", request_capacity, ns->capacity);
|
346
|
+
|
347
|
+
// Initialize the A and IJA arrays
|
348
|
+
init<DType,IType>(ns);
|
349
|
+
IType *dst_ija = reinterpret_cast<IType*>(ns->ija);
|
350
|
+
DType *dst_a = reinterpret_cast<DType*>(ns->a);
|
351
|
+
|
352
|
+
size_t ija = shape[0] + 1;
|
353
|
+
DType val;
|
354
|
+
for (i = 0; i < shape[0]; ++i) {
|
355
|
+
k = i + offset[0];
|
356
|
+
for (j = 0; j < shape[1]; ++j) {
|
357
|
+
l = j + offset[1];
|
358
|
+
|
359
|
+
// Get value from source matrix
|
360
|
+
if (k == l) val = src_a[k];
|
361
|
+
else {
|
362
|
+
// copy non-diagonal element
|
363
|
+
for (size_t c = src_ija[k]; c < src_ija[k+1]; ++c) {
|
364
|
+
if (src_ija[c] == l) val = src_a[c];
|
365
|
+
}
|
366
|
+
}
|
367
|
+
|
368
|
+
// Set value to destination matrix
|
369
|
+
if (i == j) dst_a[i] = val;
|
370
|
+
else {
|
371
|
+
// copy non-diagonal element
|
372
|
+
dst_ija[ija] = j;
|
373
|
+
dst_a[ija] = val;
|
374
|
+
|
375
|
+
++ija;
|
376
|
+
for (size_t c = i + 1; c <= shape[0]; ++c) {
|
377
|
+
dst_ija[c] = ija;
|
378
|
+
}
|
379
|
+
}
|
380
|
+
}
|
381
|
+
}
|
382
|
+
|
383
|
+
dst_ija[shape[0]] = ija; // indicate the end of the last row
|
384
|
+
ns->ndnz = ndnz;
|
385
|
+
return ns;
|
386
|
+
}
|
387
|
+
/*
|
388
|
+
* Returns a pointer to the correct location in the A vector of a YALE_STORAGE object, given some set of coordinates
|
389
|
+
* (the coordinates are stored in slice).
|
390
|
+
*/
|
391
|
+
template <typename DType,typename IType>
|
392
|
+
void* ref(YALE_STORAGE* storage, SLICE* slice) {
|
393
|
+
size_t* coords = slice->coords;
|
394
|
+
|
395
|
+
if (!slice->single) rb_raise(rb_eNotImpError, "This type slicing not supported yet.");
|
396
|
+
|
397
|
+
DType* a = reinterpret_cast<DType*>(storage->a);
|
398
|
+
IType* ija = reinterpret_cast<IType*>(storage->ija);
|
399
|
+
|
400
|
+
if (coords[0] == coords[1])
|
401
|
+
return &(a[ coords[0] ]); // return diagonal entry
|
402
|
+
|
403
|
+
if (ija[coords[0]] == ija[coords[0]+1])
|
404
|
+
return &(a[ storage->shape[0] ]); // return zero pointer
|
405
|
+
|
406
|
+
// binary search for the column's location
|
407
|
+
int pos = binary_search<IType>(storage,
|
408
|
+
ija[coords[0]],
|
409
|
+
ija[coords[0]+1]-1,
|
410
|
+
coords[1]);
|
411
|
+
|
412
|
+
if (pos != -1 && ija[pos] == coords[1])
|
413
|
+
return &(a[pos]); // found exact value
|
414
|
+
|
415
|
+
return &(a[ storage->shape[0] ]); // return a pointer that happens to be zero
|
416
|
+
}
|
417
|
+
|
418
|
+
/*
|
419
|
+
* Attempt to set some cell in a YALE_STORAGE object. Must supply coordinates and a pointer to a value (which will be
|
420
|
+
* copied into the storage object).
|
421
|
+
*/
|
422
|
+
template <typename DType, typename IType>
|
423
|
+
char set(YALE_STORAGE* storage, SLICE* slice, void* value) {
|
424
|
+
DType* v = reinterpret_cast<DType*>(value);
|
425
|
+
size_t* coords = slice->coords;
|
426
|
+
|
427
|
+
bool found = false;
|
428
|
+
char ins_type;
|
429
|
+
|
430
|
+
if (coords[0] == coords[1]) {
|
431
|
+
reinterpret_cast<DType*>(storage->a)[coords[0]] = *v; // set diagonal
|
432
|
+
return 'r';
|
433
|
+
}
|
434
|
+
|
435
|
+
// Get IJA positions of the beginning and end of the row
|
436
|
+
if (reinterpret_cast<IType*>(storage->ija)[coords[0]] == reinterpret_cast<IType*>(storage->ija)[coords[0]+1]) {
|
437
|
+
// empty row
|
438
|
+
ins_type = vector_insert<DType,IType>(storage, reinterpret_cast<IType*>(storage->ija)[coords[0]], &(coords[1]), v, 1, false);
|
439
|
+
increment_ia_after<IType>(storage, storage->shape[0], coords[0], 1);
|
440
|
+
storage->ndnz++;
|
441
|
+
|
442
|
+
return ins_type;
|
443
|
+
}
|
444
|
+
|
445
|
+
// non-empty row. search for coords[1] in the IJA array, between ija and ija_next
|
446
|
+
// (including ija, not including ija_next)
|
447
|
+
//ija_size = get_size<IType>(storage);
|
448
|
+
|
449
|
+
// Do a binary search for the column
|
450
|
+
size_t pos = insert_search<IType>(storage,
|
451
|
+
reinterpret_cast<IType*>(storage->ija)[coords[0]],
|
452
|
+
reinterpret_cast<IType*>(storage->ija)[coords[0]+1]-1,
|
453
|
+
coords[1], &found);
|
454
|
+
|
455
|
+
if (found) { // replace
|
456
|
+
reinterpret_cast<IType*>(storage->ija)[pos] = coords[1];
|
457
|
+
reinterpret_cast<DType*>(storage->a)[pos] = *v;
|
458
|
+
return 'r';
|
459
|
+
}
|
460
|
+
|
461
|
+
ins_type = vector_insert<DType,IType>(storage, pos, &(coords[1]), v, 1, false);
|
462
|
+
increment_ia_after<IType>(storage, storage->shape[0], coords[0], 1);
|
463
|
+
storage->ndnz++;
|
464
|
+
|
465
|
+
return ins_type;
|
466
|
+
}
|
467
|
+
|
468
|
+
///////////
|
469
|
+
// Tests //
|
470
|
+
///////////
|
471
|
+
|
472
|
+
/*
|
473
|
+
* Yale eql? -- for whole-matrix comparison returning a single value.
|
474
|
+
*/
|
475
|
+
template <typename LDType, typename RDType, typename IType>
|
476
|
+
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right) {
|
477
|
+
LDType* la = reinterpret_cast<LDType*>(left->a);
|
478
|
+
RDType* ra = reinterpret_cast<RDType*>(right->a);
|
479
|
+
|
480
|
+
// Compare the diagonals first.
|
481
|
+
for (size_t index = 0; index < left->shape[0]; ++index) {
|
482
|
+
if (la[index] != ra[index]) return false;
|
483
|
+
}
|
484
|
+
|
485
|
+
IType* lij = reinterpret_cast<IType*>(left->ija);
|
486
|
+
IType* rij = reinterpret_cast<IType*>(right->ija);
|
487
|
+
|
488
|
+
for (IType i = 0; i < left->shape[0]; ++i) {
|
489
|
+
|
490
|
+
// Get start and end positions of row
|
491
|
+
IType l_ija = lij[i],
|
492
|
+
l_ija_next = lij[i+1],
|
493
|
+
r_ija = rij[i],
|
494
|
+
r_ija_next = rij[i+1];
|
495
|
+
|
496
|
+
// Check to see if one row is empty and the other isn't.
|
497
|
+
if (ndrow_is_empty<LDType,IType>(left, l_ija, l_ija_next)) {
|
498
|
+
if (!ndrow_is_empty<RDType,IType>(right, r_ija, r_ija_next)) {
|
499
|
+
return false;
|
500
|
+
}
|
501
|
+
|
502
|
+
} else if (ndrow_is_empty<RDType,IType>(right, r_ija, r_ija_next)) {
|
503
|
+
// one is empty but the other isn't
|
504
|
+
return false;
|
505
|
+
|
506
|
+
} else if (!ndrow_eqeq_ndrow<LDType,RDType,IType>(left, right, l_ija, l_ija_next, r_ija, r_ija_next)) {
|
507
|
+
// Neither row is empty. Must compare the rows directly.
|
508
|
+
return false;
|
509
|
+
}
|
510
|
+
|
511
|
+
}
|
512
|
+
|
513
|
+
return true;
|
514
|
+
}
|
515
|
+
|
516
|
+
/*
|
517
|
+
* Are two non-diagonal rows the same? We already know.
|
518
|
+
*/
|
519
|
+
template <typename LDType, typename RDType, typename IType>
|
520
|
+
static bool ndrow_eqeq_ndrow(const YALE_STORAGE* l, const YALE_STORAGE* r, IType l_ija, const IType l_ija_next, IType r_ija, const IType r_ija_next) {
|
521
|
+
bool l_no_more = false, r_no_more = false;
|
522
|
+
|
523
|
+
IType *lij = reinterpret_cast<IType*>(l->ija),
|
524
|
+
*rij = reinterpret_cast<IType*>(r->ija);
|
525
|
+
|
526
|
+
LDType* la = reinterpret_cast<LDType*>(l->a);
|
527
|
+
RDType* ra = reinterpret_cast<RDType*>(r->a);
|
528
|
+
|
529
|
+
IType l_ja = lij[l_ija],
|
530
|
+
r_ja = rij[r_ija];
|
531
|
+
|
532
|
+
IType ja = std::min(l_ja, r_ja);
|
533
|
+
|
534
|
+
while (!(l_no_more && r_no_more)) {
|
535
|
+
if (l_ja == r_ja) {
|
536
|
+
|
537
|
+
if (ra[r_ija] != la[l_ija]) return false; // Direct comparison
|
538
|
+
|
539
|
+
++l_ija;
|
540
|
+
++r_ija;
|
541
|
+
|
542
|
+
if (l_ija < l_ija_next) {
|
543
|
+
l_ja = lij[l_ija];
|
544
|
+
|
545
|
+
} else {
|
546
|
+
l_no_more = true;
|
547
|
+
}
|
548
|
+
|
549
|
+
if (r_ija < r_ija_next) {
|
550
|
+
r_ja = rij[r_ija];
|
551
|
+
|
552
|
+
} else {
|
553
|
+
r_no_more = true;
|
554
|
+
}
|
555
|
+
|
556
|
+
ja = std::min(l_ja, r_ja);
|
557
|
+
|
558
|
+
} else if (l_no_more || ja < l_ja) {
|
559
|
+
|
560
|
+
if (ra[r_ija] != 0) return false;
|
561
|
+
|
562
|
+
++r_ija;
|
563
|
+
if (r_ija < r_ija_next) {
|
564
|
+
// get next column
|
565
|
+
r_ja = rij[r_ija];
|
566
|
+
ja = std::min(l_ja, r_ja);
|
567
|
+
|
568
|
+
} else {
|
569
|
+
l_no_more = true;
|
570
|
+
}
|
571
|
+
|
572
|
+
} else if (r_no_more || ja < r_ja) {
|
573
|
+
|
574
|
+
if (la[l_ija] != 0) return false;
|
575
|
+
|
576
|
+
++l_ija;
|
577
|
+
if (l_ija < l_ija_next) {
|
578
|
+
// get next column
|
579
|
+
l_ja = lij[l_ija];
|
580
|
+
ja = std::min(l_ja, r_ja);
|
581
|
+
} else {
|
582
|
+
l_no_more = true;
|
583
|
+
}
|
584
|
+
|
585
|
+
} else {
|
586
|
+
std::fprintf(stderr, "Unhandled in eqeq: l_ja=%d, r_ja=%d\n", (int)l_ja, (int)r_ja);
|
587
|
+
}
|
588
|
+
}
|
589
|
+
|
590
|
+
// every item matched
|
591
|
+
return true;
|
592
|
+
}
|
593
|
+
|
594
|
+
/*
|
595
|
+
* Is the non-diagonal portion of the row empty?
|
596
|
+
*/
|
597
|
+
template <typename DType, typename IType>
|
598
|
+
static bool ndrow_is_empty(const YALE_STORAGE* s, IType ija, const IType ija_next) {
|
599
|
+
if (ija == ija_next) return true;
|
600
|
+
|
601
|
+
DType* a = reinterpret_cast<DType*>(s->a);
|
602
|
+
|
603
|
+
// do all the entries = zero?
|
604
|
+
for (; ija < ija_next; ++ija) {
|
605
|
+
if (a[ija] != 0) return false;
|
606
|
+
}
|
607
|
+
|
608
|
+
return true;
|
609
|
+
}
|
610
|
+
|
611
|
+
//////////
|
612
|
+
// Math //
|
613
|
+
//////////
|
614
|
+
|
615
|
+
#define YALE_IA(s) (reinterpret_cast<IType*>(s->ija))
|
616
|
+
#define YALE_IJ(s) (reinterpret_cast<IType*>(s->ija) + s->shape[0] + 1)
|
617
|
+
#define YALE_COUNT(yale) (yale->ndnz + yale->shape[0])
|
618
|
+
|
619
|
+
template <typename nm::ewop_t op, typename IType, typename DType>
|
620
|
+
YALE_STORAGE* ew_op(const YALE_STORAGE* left, const YALE_STORAGE* right, dtype_t dtype) {
|
621
|
+
size_t init_capacity;
|
622
|
+
size_t* new_shape;
|
623
|
+
|
624
|
+
unsigned int da_index,
|
625
|
+
la_index,
|
626
|
+
ra_index,
|
627
|
+
|
628
|
+
a_index_offset,
|
629
|
+
|
630
|
+
la_row_max,
|
631
|
+
ra_row_max,
|
632
|
+
|
633
|
+
row_index;
|
634
|
+
|
635
|
+
DType tmp_result;
|
636
|
+
|
637
|
+
DType * la = reinterpret_cast<DType*> (left->a),
|
638
|
+
* ra = reinterpret_cast<DType*>(right->a),
|
639
|
+
* da;
|
640
|
+
|
641
|
+
YALE_STORAGE* dest;
|
642
|
+
|
643
|
+
new_shape = reinterpret_cast<size_t*>(calloc(2, sizeof(size_t)));
|
644
|
+
new_shape[0] = left->shape[0];
|
645
|
+
new_shape[1] = left->shape[1];
|
646
|
+
|
647
|
+
init_capacity = std::min(left->ndnz + right->ndnz + new_shape[0], new_shape[0] * new_shape[1]);
|
648
|
+
|
649
|
+
dest = nm_yale_storage_create(dtype, new_shape, 2, init_capacity);
|
650
|
+
da = reinterpret_cast<DType*>(dest->a);
|
651
|
+
|
652
|
+
// Calculate diagonal values.
|
653
|
+
for (da_index = 0; da_index < dest->shape[0]; ++da_index) {
|
654
|
+
da[da_index] = ew_op_switch<op, DType, DType>(la[da_index], ra[da_index]);
|
655
|
+
}
|
656
|
+
|
657
|
+
// Set the zero representation seperator.
|
658
|
+
da[da_index] = typeid(DType) == typeid(RubyObject) ? INT2FIX(0) : 0;
|
659
|
+
|
660
|
+
/*
|
661
|
+
* Calculate the offset between start of the A arrays and the non-diagonal
|
662
|
+
* entries.
|
663
|
+
*/
|
664
|
+
a_index_offset = dest->shape[0] + 1;
|
665
|
+
|
666
|
+
// Re-base the A arrays.
|
667
|
+
la = la + a_index_offset;
|
668
|
+
ra = ra + a_index_offset;
|
669
|
+
da = da + a_index_offset;
|
670
|
+
|
671
|
+
// Initialize our A array indices.
|
672
|
+
la_index = ra_index = da_index = 0;
|
673
|
+
|
674
|
+
// Calculate the non-diagonal values.
|
675
|
+
for (row_index = 0; row_index < dest->shape[0]; ++row_index) {
|
676
|
+
/*
|
677
|
+
* Each row.
|
678
|
+
*/
|
679
|
+
|
680
|
+
printf("Row %d\n", row_index);
|
681
|
+
|
682
|
+
// Get row bounds.
|
683
|
+
la_row_max = YALE_IA( left)[row_index + 1] - a_index_offset;
|
684
|
+
ra_row_max = YALE_IA(right)[row_index + 1] - a_index_offset;
|
685
|
+
|
686
|
+
printf("Left : Row Start: %d - Row End %d\n", la_index + a_index_offset, la_row_max + a_index_offset);
|
687
|
+
printf("Right : Row Start: %d - Row End %d\n", ra_index + a_index_offset, ra_row_max + a_index_offset);
|
688
|
+
|
689
|
+
/*
|
690
|
+
* Set this row's left bound (which is also the previous row's right
|
691
|
+
* bound).
|
692
|
+
*/
|
693
|
+
YALE_IA(dest)[row_index] = da_index + a_index_offset;
|
694
|
+
|
695
|
+
printf("Left bound of row %d in destination: %d\n", (int)row_index, (int)YALE_IA(dest)[row_index]);
|
696
|
+
|
697
|
+
// Iterate over non-diagonal entries in this row.
|
698
|
+
while (la_index < la_row_max and ra_index < ra_row_max) {
|
699
|
+
/*
|
700
|
+
* Elements are present on both the left- and right-hand side.
|
701
|
+
*/
|
702
|
+
|
703
|
+
printf("Marker 0\n");
|
704
|
+
|
705
|
+
if (YALE_IJ(left)[la_index] == YALE_IJ(right)[ra_index]) {
|
706
|
+
/*
|
707
|
+
* Current left- and right-hand values are in the same row and
|
708
|
+
* column.
|
709
|
+
*/
|
710
|
+
|
711
|
+
printf("Calculating value for [%d, %d].\n", (int)row_index, (int)YALE_IJ(left)[la_index]);
|
712
|
+
|
713
|
+
tmp_result = ew_op_switch<op, DType, DType>(la[la_index], ra[ra_index]);
|
714
|
+
|
715
|
+
if (tmp_result != 0) {
|
716
|
+
printf("Setting value for [%d, %d] at index %d in destination's A array.\n", (int)row_index, (int)YALE_IJ(left)[la_index], (int)(da_index + a_index_offset));
|
717
|
+
|
718
|
+
da[da_index] = tmp_result;
|
719
|
+
YALE_IJ(dest)[da_index] = YALE_IJ(left)[la_index];
|
720
|
+
|
721
|
+
++da_index;
|
722
|
+
|
723
|
+
} else {
|
724
|
+
printf("Result was 0. Skipping.\n");
|
725
|
+
}
|
726
|
+
|
727
|
+
++la_index;
|
728
|
+
++ra_index;
|
729
|
+
|
730
|
+
} else if (YALE_IJ(left)[la_index] < YALE_IJ(right)[ra_index]) {
|
731
|
+
/*
|
732
|
+
* The right-hand index is ahead of the left-hand index.
|
733
|
+
*/
|
734
|
+
|
735
|
+
if (op != EW_MUL) {
|
736
|
+
// If this is multiplion there is no point in doing the operation.
|
737
|
+
|
738
|
+
tmp_result = ew_op_switch<op, DType, DType>(la[la_index], typeid(DType) == typeid(RubyObject) ? INT2FIX(0) : 0);
|
739
|
+
|
740
|
+
printf("Setting value for [%d, %d].\n", (int)row_index, (int)YALE_IJ(left)[la_index]);
|
741
|
+
|
742
|
+
if (tmp_result != 0) {
|
743
|
+
da[da_index] = tmp_result;
|
744
|
+
YALE_IJ(dest)[da_index] = YALE_IJ(left)[la_index];
|
745
|
+
|
746
|
+
++da_index;
|
747
|
+
}
|
748
|
+
}
|
749
|
+
|
750
|
+
++la_index;
|
751
|
+
|
752
|
+
} else {
|
753
|
+
/*
|
754
|
+
* The left-hand index is ahead of the right-hand index.
|
755
|
+
*/
|
756
|
+
|
757
|
+
if (op != EW_MUL) {
|
758
|
+
// If this is multiplion there is no point in doing the operation.
|
759
|
+
|
760
|
+
tmp_result = ew_op_switch<op, DType, DType>(typeid(DType) == typeid(RubyObject) ? INT2FIX(0) : 0, ra[ra_index]);
|
761
|
+
|
762
|
+
printf("Setting value for [%d, %d].\n", (int)row_index, (int)YALE_IJ(right)[ra_index]);
|
763
|
+
|
764
|
+
if (tmp_result != 0) {
|
765
|
+
da[da_index] = tmp_result;
|
766
|
+
YALE_IJ(dest)[da_index] = YALE_IJ(right)[ra_index];
|
767
|
+
|
768
|
+
++da_index;
|
769
|
+
}
|
770
|
+
}
|
771
|
+
|
772
|
+
++ra_index;
|
773
|
+
}
|
774
|
+
}
|
775
|
+
|
776
|
+
if (op != EW_MUL) {
|
777
|
+
/*
|
778
|
+
* Process the remaining elements on the left- or right-hand side. One or
|
779
|
+
* the other, or neither, of the following loops may execute, but not
|
780
|
+
* both.
|
781
|
+
*
|
782
|
+
* If we are doing multiplication this is unnecessary as all remaining
|
783
|
+
* operations will produce a zero value.
|
784
|
+
*/
|
785
|
+
|
786
|
+
while (la_index < la_row_max) {
|
787
|
+
/*
|
788
|
+
* Process the remaining elements on the left-hand side.
|
789
|
+
*/
|
790
|
+
|
791
|
+
printf("Marker 1\n");
|
792
|
+
|
793
|
+
tmp_result = ew_op_switch<op, DType, DType>(la[la_index], typeid(DType) == typeid(RubyObject) ? INT2FIX(0) : 0);
|
794
|
+
|
795
|
+
printf("Setting value for [%d, %d].\n", (int)row_index, (int)YALE_IJ(left)[la_index]);
|
796
|
+
|
797
|
+
if (tmp_result != 0) {
|
798
|
+
da[da_index] = tmp_result;
|
799
|
+
YALE_IJ(dest)[da_index] = YALE_IJ(left)[la_index];
|
800
|
+
|
801
|
+
++da_index;
|
802
|
+
}
|
803
|
+
|
804
|
+
++la_index;
|
805
|
+
}
|
806
|
+
|
807
|
+
while (ra_index < ra_row_max) {
|
808
|
+
/*
|
809
|
+
* Process the remaining elements on the right-hand side.
|
810
|
+
*/
|
811
|
+
|
812
|
+
printf("Marker 2\n");
|
813
|
+
|
814
|
+
tmp_result = ew_op_switch<op, DType, DType>(typeid(DType) == typeid(RubyObject) ? INT2FIX(0) : 0, ra[ra_index]);
|
815
|
+
|
816
|
+
printf("Setting value for [%d, %d].\n", (int)row_index, (int)YALE_IJ(right)[ra_index]);
|
817
|
+
|
818
|
+
if (tmp_result != 0) {
|
819
|
+
da[da_index] = tmp_result;
|
820
|
+
YALE_IJ(dest)[da_index] = YALE_IJ(right)[ra_index];
|
821
|
+
|
822
|
+
++da_index;
|
823
|
+
}
|
824
|
+
|
825
|
+
++ra_index;
|
826
|
+
}
|
827
|
+
}
|
828
|
+
|
829
|
+
// Advance the row indices.
|
830
|
+
la_index = la_row_max;
|
831
|
+
ra_index = ra_row_max;
|
832
|
+
|
833
|
+
printf("End of row %d\n\n", row_index);
|
834
|
+
}
|
835
|
+
|
836
|
+
// Set the last row's right bound.
|
837
|
+
YALE_IA(dest)[row_index] = da_index + a_index_offset;
|
838
|
+
|
839
|
+
printf("Right bound of row %d in destination: %d\n", row_index - 1, da_index + a_index_offset);
|
840
|
+
|
841
|
+
// Set the number of non-diagonal non-zero entries in the destination matrix.
|
842
|
+
dest->ndnz = da_index;
|
843
|
+
|
844
|
+
printf("Number of non-diagonal non-zero entires: %ld\n\n", (unsigned long)(dest->ndnz));
|
845
|
+
|
846
|
+
// Set the capacity of the destination matrix.
|
847
|
+
dest->capacity = dest->shape[0] + dest->ndnz + 1;
|
848
|
+
|
849
|
+
// Resize the destination matrix.
|
850
|
+
dest->a = realloc(dest->a, sizeof(DType) * dest->capacity);
|
851
|
+
dest->ija = realloc(dest->ija, sizeof(IType) * dest->capacity);
|
852
|
+
|
853
|
+
return dest;
|
854
|
+
}
|
855
|
+
|
856
|
+
/////////////
|
857
|
+
// Utility //
|
858
|
+
/////////////
|
859
|
+
|
860
|
+
/*
|
861
|
+
* Binary search for returning stored values. Returns a non-negative position, or -1 for not found.
|
862
|
+
*/
|
863
|
+
template <typename IType>
|
864
|
+
int binary_search(YALE_STORAGE* s, IType left, IType right, IType key) {
|
865
|
+
|
866
|
+
if (left > right) return -1;
|
867
|
+
|
868
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
869
|
+
|
870
|
+
IType mid = (left + right)/2;
|
871
|
+
IType mid_j = ija[mid];
|
872
|
+
|
873
|
+
if (mid_j == key)
|
874
|
+
return mid;
|
875
|
+
|
876
|
+
else if (mid_j > key)
|
877
|
+
return binary_search<IType>(s, left, mid - 1, key);
|
878
|
+
|
879
|
+
else
|
880
|
+
return binary_search<IType>(s, mid + 1, right, key);
|
881
|
+
}
|
882
|
+
|
883
|
+
|
884
|
+
|
885
|
+
/*
|
886
|
+
* Resize yale storage vectors A and IJA in preparation for an insertion.
|
887
|
+
*/
|
888
|
+
template <typename DType, typename IType>
|
889
|
+
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only) {
|
890
|
+
// Determine the new capacity for the IJA and A vectors.
|
891
|
+
size_t new_capacity = s->capacity * GROWTH_CONSTANT;
|
892
|
+
size_t max_capacity = max_size(s);
|
893
|
+
|
894
|
+
if (new_capacity > max_capacity) {
|
895
|
+
new_capacity = max_capacity;
|
896
|
+
|
897
|
+
if (current_size + n > max_capacity) rb_raise(rb_eNoMemError, "insertion size exceeded maximum yale matrix size");
|
898
|
+
}
|
899
|
+
|
900
|
+
if (new_capacity < current_size + n)
|
901
|
+
new_capacity = current_size + n;
|
902
|
+
|
903
|
+
// Allocate the new vectors.
|
904
|
+
IType* new_ija = ALLOC_N( IType, new_capacity );
|
905
|
+
NM_CHECK_ALLOC(new_ija);
|
906
|
+
|
907
|
+
DType* new_a = ALLOC_N( DType, new_capacity );
|
908
|
+
NM_CHECK_ALLOC(new_a);
|
909
|
+
|
910
|
+
IType* old_ija = reinterpret_cast<IType*>(s->ija);
|
911
|
+
DType* old_a = reinterpret_cast<DType*>(s->a);
|
912
|
+
|
913
|
+
// Copy all values prior to the insertion site to the new IJA and new A
|
914
|
+
if (struct_only) {
|
915
|
+
for (size_t i = 0; i < pos; ++i) {
|
916
|
+
new_ija[i] = old_ija[i];
|
917
|
+
}
|
918
|
+
} else {
|
919
|
+
for (size_t i = 0; i < pos; ++i) {
|
920
|
+
new_ija[i] = old_ija[i];
|
921
|
+
new_a[i] = old_a[i];
|
922
|
+
}
|
923
|
+
}
|
924
|
+
|
925
|
+
|
926
|
+
// Copy all values subsequent to the insertion site to the new IJA and new A, leaving room (size n) for insertion.
|
927
|
+
if (struct_only) {
|
928
|
+
for (size_t i = pos; i < current_size - pos + n - 1; ++i) {
|
929
|
+
new_ija[i+n] = old_ija[i];
|
930
|
+
}
|
931
|
+
} else {
|
932
|
+
for (size_t i = pos; i < current_size - pos + n - 1; ++i) {
|
933
|
+
new_ija[i+n] = old_ija[i];
|
934
|
+
new_a[i+n] = old_a[i];
|
935
|
+
}
|
936
|
+
}
|
937
|
+
|
938
|
+
s->capacity = new_capacity;
|
939
|
+
|
940
|
+
free(s->ija);
|
941
|
+
free(s->a);
|
942
|
+
|
943
|
+
s->ija = reinterpret_cast<void*>(new_ija);
|
944
|
+
s->a = reinterpret_cast<void*>(new_a);
|
945
|
+
|
946
|
+
return 'i';
|
947
|
+
}
|
948
|
+
|
949
|
+
/*
|
950
|
+
* Insert a value or contiguous values in the ija and a vectors (after ja and
|
951
|
+
* diag). Does not free anything; you are responsible!
|
952
|
+
*
|
953
|
+
* TODO: Improve this so it can handle non-contiguous element insertions
|
954
|
+
* efficiently. For now, we can just sort the elements in the row in
|
955
|
+
* question.)
|
956
|
+
*/
|
957
|
+
template <typename DType, typename IType>
|
958
|
+
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, DType* val, size_t n, bool struct_only) {
|
959
|
+
if (pos < s->shape[0]) {
|
960
|
+
rb_raise(rb_eArgError, "vector insert pos is before beginning of ja; this should not happen");
|
961
|
+
}
|
962
|
+
|
963
|
+
size_t size = get_size<IType>(s);
|
964
|
+
|
965
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
966
|
+
DType* a = reinterpret_cast<DType*>(s->a);
|
967
|
+
|
968
|
+
if (size + n > s->capacity) {
|
969
|
+
vector_insert_resize<DType,IType>(s, size, pos, j, n, struct_only);
|
970
|
+
|
971
|
+
// Need to get the new locations for ija and a.
|
972
|
+
ija = reinterpret_cast<IType*>(s->ija);
|
973
|
+
a = reinterpret_cast<DType*>(s->a);
|
974
|
+
|
975
|
+
} else {
|
976
|
+
/*
|
977
|
+
* No resize required:
|
978
|
+
* easy (but somewhat slow), just copy elements to the tail, starting at
|
979
|
+
* the end, one element at a time.
|
980
|
+
*
|
981
|
+
* TODO: This can be made slightly more efficient, but only after the tests
|
982
|
+
* are written.
|
983
|
+
*/
|
984
|
+
|
985
|
+
if (struct_only) {
|
986
|
+
for (size_t i = 0; i < size - pos; ++i) {
|
987
|
+
ija[size+n-1-i] = ija[size-1-i];
|
988
|
+
}
|
989
|
+
} else {
|
990
|
+
for (size_t i = 0; i < size - pos; ++i) {
|
991
|
+
ija[size+n-1-i] = ija[size-1-i];
|
992
|
+
a[size+n-1-i] = a[size-1-i];
|
993
|
+
}
|
994
|
+
}
|
995
|
+
}
|
996
|
+
|
997
|
+
// Now insert the new values.
|
998
|
+
if (struct_only) {
|
999
|
+
for (size_t i = 0; i < n; ++i) {
|
1000
|
+
ija[pos+i] = j[i];
|
1001
|
+
}
|
1002
|
+
} else {
|
1003
|
+
for (size_t i = 0; i < n; ++i) {
|
1004
|
+
ija[pos+i] = j[i];
|
1005
|
+
a[pos+i] = val[i];
|
1006
|
+
}
|
1007
|
+
}
|
1008
|
+
|
1009
|
+
return 'i';
|
1010
|
+
}
|
1011
|
+
|
1012
|
+
/*
|
1013
|
+
* If we add n items to row i, we need to increment ija[i+1] and onward.
|
1014
|
+
*/
|
1015
|
+
template <typename IType>
|
1016
|
+
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, IType n) {
|
1017
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
1018
|
+
|
1019
|
+
++i;
|
1020
|
+
for (; i <= ija_size; ++i) {
|
1021
|
+
ija[i] += n;
|
1022
|
+
}
|
1023
|
+
}
|
1024
|
+
|
1025
|
+
/*
|
1026
|
+
* Binary search for returning insertion points.
|
1027
|
+
*/
|
1028
|
+
template <typename IType>
|
1029
|
+
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool* found) {
|
1030
|
+
|
1031
|
+
if (left > right) {
|
1032
|
+
*found = false;
|
1033
|
+
return left;
|
1034
|
+
}
|
1035
|
+
|
1036
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
1037
|
+
IType mid = (left + right)/2;
|
1038
|
+
IType mid_j = ija[mid];
|
1039
|
+
|
1040
|
+
if (mid_j == key) {
|
1041
|
+
*found = true;
|
1042
|
+
return mid;
|
1043
|
+
|
1044
|
+
} else if (mid_j > key) {
|
1045
|
+
return insert_search<IType>(s, left, mid-1, key, found);
|
1046
|
+
|
1047
|
+
} else {
|
1048
|
+
return insert_search<IType>(s, mid+1, right, key, found);
|
1049
|
+
}
|
1050
|
+
}
|
1051
|
+
|
1052
|
+
/////////////////////////
|
1053
|
+
// Copying and Casting //
|
1054
|
+
/////////////////////////
|
1055
|
+
|
1056
|
+
/*
|
1057
|
+
* Templated copy constructor for changing dtypes.
|
1058
|
+
*/
|
1059
|
+
template <typename LDType, typename RDType, typename IType>
|
1060
|
+
YALE_STORAGE* cast_copy(const YALE_STORAGE* rhs, dtype_t new_dtype) {
|
1061
|
+
|
1062
|
+
// Allocate a new structure
|
1063
|
+
size_t size = get_size<IType>(rhs);
|
1064
|
+
YALE_STORAGE* lhs = copy_alloc_struct<IType>(rhs, new_dtype, rhs->capacity, size);
|
1065
|
+
|
1066
|
+
if (rhs->dtype == new_dtype) { // FIXME: Test if this condition is actually faster; second condition should work just as well.
|
1067
|
+
|
1068
|
+
memcpy(lhs->a, rhs->a, size * DTYPE_SIZES[new_dtype]);
|
1069
|
+
|
1070
|
+
} else {
|
1071
|
+
|
1072
|
+
LDType* la = reinterpret_cast<LDType*>(lhs->a);
|
1073
|
+
RDType* ra = reinterpret_cast<RDType*>(rhs->a);
|
1074
|
+
|
1075
|
+
for (size_t index = 0; index < size; ++index) {
|
1076
|
+
la[index] = ra[index];
|
1077
|
+
}
|
1078
|
+
|
1079
|
+
}
|
1080
|
+
|
1081
|
+
return lhs;
|
1082
|
+
}
|
1083
|
+
|
1084
|
+
/*
|
1085
|
+
* Template access for getting the size of Yale storage.
|
1086
|
+
*/
|
1087
|
+
template <typename IType>
|
1088
|
+
static inline size_t get_size(const YALE_STORAGE* storage) {
|
1089
|
+
return static_cast<size_t>(reinterpret_cast<IType*>(storage->ija)[ storage->shape[0] ]);
|
1090
|
+
}
|
1091
|
+
|
1092
|
+
/*
|
1093
|
+
* Allocate for a copy or copy-cast operation, and copy the IJA portion of the
|
1094
|
+
* matrix (the structure).
|
1095
|
+
*/
|
1096
|
+
template <typename IType>
|
1097
|
+
static YALE_STORAGE* copy_alloc_struct(const YALE_STORAGE* rhs, const dtype_t new_dtype, const size_t new_capacity, const size_t new_size) {
|
1098
|
+
YALE_STORAGE* lhs = ALLOC( YALE_STORAGE );
|
1099
|
+
lhs->dim = rhs->dim;
|
1100
|
+
lhs->shape = ALLOC_N( size_t, lhs->dim );
|
1101
|
+
memcpy(lhs->shape, rhs->shape, lhs->dim * sizeof(size_t));
|
1102
|
+
lhs->itype = rhs->itype;
|
1103
|
+
lhs->capacity = new_capacity;
|
1104
|
+
lhs->dtype = new_dtype;
|
1105
|
+
lhs->ndnz = rhs->ndnz;
|
1106
|
+
|
1107
|
+
lhs->ija = ALLOC_N( IType, lhs->capacity );
|
1108
|
+
lhs->a = ALLOC_N( char, DTYPE_SIZES[new_dtype] * lhs->capacity );
|
1109
|
+
|
1110
|
+
// Now copy the contents -- but only within the boundaries set by the size. Leave
|
1111
|
+
// the rest uninitialized.
|
1112
|
+
for (size_t i = 0; i < get_size<IType>(rhs); ++i)
|
1113
|
+
reinterpret_cast<IType*>(lhs->ija)[i] = reinterpret_cast<IType*>(rhs->ija)[i]; // copy indices
|
1114
|
+
|
1115
|
+
return lhs;
|
1116
|
+
}
|
1117
|
+
|
1118
|
+
template <typename DType, typename IType>
|
1119
|
+
static STORAGE* matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector) {
|
1120
|
+
YALE_STORAGE *left = (YALE_STORAGE*)(casted_storage.left),
|
1121
|
+
*right = (YALE_STORAGE*)(casted_storage.right);
|
1122
|
+
|
1123
|
+
// We can safely get dtype from the casted matrices; post-condition of binary_storage_cast_alloc is that dtype is the
|
1124
|
+
// same for left and right.
|
1125
|
+
// int8_t dtype = left->dtype;
|
1126
|
+
|
1127
|
+
// Create result storage.
|
1128
|
+
YALE_STORAGE* result = nm_yale_storage_create(left->dtype, resulting_shape, 2, left->capacity + right->capacity);
|
1129
|
+
init<DType,IType>(result);
|
1130
|
+
|
1131
|
+
IType* ijl = reinterpret_cast<IType*>(left->ija);
|
1132
|
+
IType* ijr = reinterpret_cast<IType*>(right->ija);
|
1133
|
+
IType* ija = reinterpret_cast<IType*>(result->ija);
|
1134
|
+
|
1135
|
+
// Symbolic multiplication step (build the structure)
|
1136
|
+
nm::math::symbmm<IType>(result->shape[0], result->shape[1], ijl, ijl, true, ijr, ijr, true, ija, true);
|
1137
|
+
|
1138
|
+
// Numeric multiplication step (fill in the elements)
|
1139
|
+
nm::math::numbmm<DType,IType>(result->shape[0], result->shape[1],
|
1140
|
+
ijl, ijl, reinterpret_cast<DType*>(left->a), true,
|
1141
|
+
ijr, ijr, reinterpret_cast<DType*>(right->a), true,
|
1142
|
+
ija, ija, reinterpret_cast<DType*>(result->a), true);
|
1143
|
+
|
1144
|
+
// Sort the columns
|
1145
|
+
nm::math::smmp_sort_columns<DType,IType>(result->shape[0], ija, ija, reinterpret_cast<DType*>(result->a));
|
1146
|
+
|
1147
|
+
return reinterpret_cast<STORAGE*>(result);
|
1148
|
+
}
|
1149
|
+
|
1150
|
+
}} // end of namespace nm::yale_storage.
|
1151
|
+
|
1152
|
+
///////////////////
|
1153
|
+
// Ruby Bindings //
|
1154
|
+
///////////////////
|
1155
|
+
|
1156
|
+
/* These bindings are mostly only for debugging Yale. They are called from Init_nmatrix. */
|
1157
|
+
|
1158
|
+
extern "C" {
|
1159
|
+
|
1160
|
+
void nm_init_yale_functions() {
|
1161
|
+
cNMatrix_YaleFunctions = rb_define_module_under(cNMatrix, "YaleFunctions");
|
1162
|
+
|
1163
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ija", (METHOD)nm_ija, 0);
|
1164
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_a", (METHOD)nm_a, 0);
|
1165
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_size", (METHOD)nm_size, 0);
|
1166
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ia", (METHOD)nm_ia, 0);
|
1167
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ja", (METHOD)nm_ja, 0);
|
1168
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_d", (METHOD)nm_d, 0);
|
1169
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_lu", (METHOD)nm_lu, 0);
|
1170
|
+
rb_define_const(cNMatrix_YaleFunctions, "YALE_GROWTH_CONSTANT", rb_float_new(nm::yale_storage::GROWTH_CONSTANT));
|
1171
|
+
}
|
1172
|
+
|
1173
|
+
|
1174
|
+
/////////////////
|
1175
|
+
// C ACCESSORS //
|
1176
|
+
/////////////////
|
1177
|
+
|
1178
|
+
/*
|
1179
|
+
* C accessor for inserting some value in a matrix (or replacing an existing cell).
|
1180
|
+
*/
|
1181
|
+
char nm_yale_storage_set(STORAGE* storage, SLICE* slice, void* v) {
|
1182
|
+
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::set, char, YALE_STORAGE* storage, SLICE* slice, void* value);
|
1183
|
+
|
1184
|
+
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1185
|
+
|
1186
|
+
return ttable[casted_storage->dtype][casted_storage->itype](casted_storage, slice, v);
|
1187
|
+
}
|
1188
|
+
|
1189
|
+
/*
|
1190
|
+
* C accessor for yale_storage::get, which returns a slice of YALE_STORAGE object by coppy
|
1191
|
+
*
|
1192
|
+
* Slicing-related.
|
1193
|
+
*/
|
1194
|
+
void* nm_yale_storage_get(STORAGE* storage, SLICE* slice) {
|
1195
|
+
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::get, void*, YALE_STORAGE* storage, SLICE* slice);
|
1196
|
+
YALE_STORAGE* s = (YALE_STORAGE*)storage;
|
1197
|
+
|
1198
|
+
|
1199
|
+
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1200
|
+
return ttable[casted_storage->dtype][casted_storage->itype](casted_storage, slice);
|
1201
|
+
}
|
1202
|
+
|
1203
|
+
/*
|
1204
|
+
* C accessor for yale_storage::ref, which returns a pointer to the correct location in a YALE_STORAGE object
|
1205
|
+
* for some set of coordinates.
|
1206
|
+
*/
|
1207
|
+
void* nm_yale_storage_ref(STORAGE* storage, SLICE* slice) {
|
1208
|
+
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::ref, void*, YALE_STORAGE* storage, SLICE* slice);
|
1209
|
+
|
1210
|
+
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1211
|
+
return ttable[casted_storage->dtype][casted_storage->itype](casted_storage, slice);
|
1212
|
+
}
|
1213
|
+
|
1214
|
+
/*
|
1215
|
+
* C accessor for determining whether two YALE_STORAGE objects have the same contents.
|
1216
|
+
*
|
1217
|
+
* FIXME: Is this for element-wise or whole-matrix equality?
|
1218
|
+
*/
|
1219
|
+
bool nm_yale_storage_eqeq(const STORAGE* left, const STORAGE* right) {
|
1220
|
+
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::eqeq, bool, const YALE_STORAGE* left, const YALE_STORAGE* right);
|
1221
|
+
|
1222
|
+
const YALE_STORAGE* casted_left = reinterpret_cast<const YALE_STORAGE*>(left);
|
1223
|
+
|
1224
|
+
return ttable[casted_left->dtype][right->dtype][casted_left->itype](casted_left, (const YALE_STORAGE*)right);
|
1225
|
+
}
|
1226
|
+
|
1227
|
+
/*
|
1228
|
+
* Copy constructor for changing dtypes. (C accessor)
|
1229
|
+
*/
|
1230
|
+
STORAGE* nm_yale_storage_cast_copy(const STORAGE* rhs, dtype_t new_dtype) {
|
1231
|
+
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::cast_copy, YALE_STORAGE*, const YALE_STORAGE* rhs, dtype_t new_dtype);
|
1232
|
+
|
1233
|
+
const YALE_STORAGE* casted_rhs = reinterpret_cast<const YALE_STORAGE*>(rhs);
|
1234
|
+
|
1235
|
+
return (STORAGE*)ttable[new_dtype][casted_rhs->dtype][casted_rhs->itype](casted_rhs, new_dtype);
|
1236
|
+
}
|
1237
|
+
|
1238
|
+
/*
|
1239
|
+
* Returns size of Yale storage as a size_t (no matter what the itype is). (C accessor)
|
1240
|
+
*/
|
1241
|
+
inline size_t nm_yale_storage_get_size(const YALE_STORAGE* storage) {
|
1242
|
+
NAMED_ITYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::get_size, size_t, const YALE_STORAGE* storage);
|
1243
|
+
|
1244
|
+
return ttable[storage->itype](storage);
|
1245
|
+
}
|
1246
|
+
|
1247
|
+
/*
|
1248
|
+
* C accessor for allocating a yale storage object for cast-copying. Copies the IJA vector, does not copy the A vector.
|
1249
|
+
*/
|
1250
|
+
static YALE_STORAGE* nm_copy_alloc_struct(const YALE_STORAGE* rhs, const dtype_t new_dtype, const size_t new_capacity, const size_t new_size) {
|
1251
|
+
NAMED_ITYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::copy_alloc_struct, YALE_STORAGE*, const YALE_STORAGE* rhs, const dtype_t new_dtype, const size_t new_capacity, const size_t new_size);
|
1252
|
+
|
1253
|
+
return ttable[rhs->itype](rhs, new_dtype, new_capacity, new_size);
|
1254
|
+
}
|
1255
|
+
|
1256
|
+
/*
|
1257
|
+
* Transposing copy constructor.
|
1258
|
+
*/
|
1259
|
+
STORAGE* nm_yale_storage_copy_transposed(const STORAGE* rhs_base) {
|
1260
|
+
YALE_STORAGE* rhs = (YALE_STORAGE*)rhs_base;
|
1261
|
+
|
1262
|
+
size_t* shape = ALLOC_N(size_t, 2);
|
1263
|
+
shape[0] = rhs->shape[1];
|
1264
|
+
shape[1] = rhs->shape[0];
|
1265
|
+
|
1266
|
+
size_t size = nm_yale_storage_get_size(rhs);
|
1267
|
+
|
1268
|
+
YALE_STORAGE* lhs = nm_yale_storage_create(rhs->dtype, shape, 2, size);
|
1269
|
+
nm_yale_storage_init(lhs);
|
1270
|
+
|
1271
|
+
NAMED_LI_DTYPE_TEMPLATE_TABLE(transp, nm::math::transpose_yale, void, const size_t n, const size_t m, const void* ia_, const void* ja_, const void* a_, const bool diaga, void* ib_, void* jb_, void* b_, const bool move);
|
1272
|
+
|
1273
|
+
transp[lhs->dtype][lhs->itype](rhs->shape[0], rhs->shape[1], rhs->ija, rhs->ija, rhs->a, true, lhs->ija, lhs->ija, lhs->a, true);
|
1274
|
+
|
1275
|
+
return (STORAGE*)lhs;
|
1276
|
+
}
|
1277
|
+
|
1278
|
+
/*
|
1279
|
+
* C accessor for multiplying two YALE_STORAGE matrices, which have already been casted to the same dtype.
|
1280
|
+
*
|
1281
|
+
* FIXME: What happens if the two matrices have different itypes?
|
1282
|
+
*/
|
1283
|
+
STORAGE* nm_yale_storage_matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector) {
|
1284
|
+
LI_DTYPE_TEMPLATE_TABLE(nm::yale_storage::matrix_multiply, STORAGE*, const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector);
|
1285
|
+
|
1286
|
+
YALE_STORAGE* storage_access = (YALE_STORAGE*)(casted_storage.left);
|
1287
|
+
|
1288
|
+
return ttable[storage_access->dtype][storage_access->itype](casted_storage, resulting_shape, vector);
|
1289
|
+
}
|
1290
|
+
|
1291
|
+
/*
|
1292
|
+
* Documentation goes here.
|
1293
|
+
*/
|
1294
|
+
STORAGE* nm_yale_storage_ew_op(nm::ewop_t op, const STORAGE* left, const STORAGE* right) {
|
1295
|
+
OP_ITYPE_DTYPE_TEMPLATE_TABLE(nm::yale_storage::ew_op, YALE_STORAGE*, const YALE_STORAGE*, const YALE_STORAGE*, dtype_t);
|
1296
|
+
|
1297
|
+
YALE_STORAGE* new_l = NULL, * new_r = NULL;
|
1298
|
+
YALE_STORAGE* result;
|
1299
|
+
|
1300
|
+
const YALE_STORAGE* casted_l, * casted_r;
|
1301
|
+
|
1302
|
+
dtype_t new_dtype;
|
1303
|
+
|
1304
|
+
if (left->dtype != right->dtype) {
|
1305
|
+
|
1306
|
+
new_dtype = Upcast[left->dtype][right->dtype];
|
1307
|
+
|
1308
|
+
if (left->dtype != new_dtype) {
|
1309
|
+
new_l = reinterpret_cast<YALE_STORAGE*>(nm_yale_storage_cast_copy( left, new_dtype));
|
1310
|
+
}
|
1311
|
+
|
1312
|
+
if (right->dtype != new_dtype) {
|
1313
|
+
new_r = reinterpret_cast<YALE_STORAGE*>(nm_yale_storage_cast_copy(right, new_dtype));
|
1314
|
+
}
|
1315
|
+
|
1316
|
+
if (static_cast<uint8_t>(op) < nm::NUM_NONCOMP_EWOPS) {
|
1317
|
+
result = ttable[op][new_l->itype][new_dtype]( left->dtype == new_dtype ?
|
1318
|
+
reinterpret_cast<const YALE_STORAGE*>( left) :
|
1319
|
+
reinterpret_cast<const YALE_STORAGE*>(new_l),
|
1320
|
+
|
1321
|
+
right->dtype == new_dtype ?
|
1322
|
+
reinterpret_cast<const YALE_STORAGE*>(right) :
|
1323
|
+
reinterpret_cast<const YALE_STORAGE*>(new_r),
|
1324
|
+
|
1325
|
+
new_dtype);
|
1326
|
+
|
1327
|
+
} else {
|
1328
|
+
rb_raise(rb_eNotImpError, "Elementwise comparison is not yet implemented for the Yale storage class.");
|
1329
|
+
}
|
1330
|
+
|
1331
|
+
if (new_l != NULL) {
|
1332
|
+
nm_yale_storage_delete(new_l);
|
1333
|
+
}
|
1334
|
+
|
1335
|
+
if (new_r != NULL) {
|
1336
|
+
nm_yale_storage_delete(new_r);
|
1337
|
+
}
|
1338
|
+
|
1339
|
+
return result;
|
1340
|
+
|
1341
|
+
} else {
|
1342
|
+
|
1343
|
+
casted_l = reinterpret_cast<const YALE_STORAGE*>( left);
|
1344
|
+
casted_r = reinterpret_cast<const YALE_STORAGE*>(right);
|
1345
|
+
|
1346
|
+
if (static_cast<uint8_t>(op) < nm::NUM_NONCOMP_EWOPS) {
|
1347
|
+
|
1348
|
+
return ttable[op][casted_l->itype][casted_l->dtype](casted_l, casted_r, casted_l->dtype);
|
1349
|
+
|
1350
|
+
} else {
|
1351
|
+
rb_raise(rb_eNotImpError, "Elementwise comparison is not yet implemented for the Yale storage class.");
|
1352
|
+
}
|
1353
|
+
}
|
1354
|
+
}
|
1355
|
+
|
1356
|
+
///////////////
|
1357
|
+
// Lifecycle //
|
1358
|
+
///////////////
|
1359
|
+
|
1360
|
+
/*
|
1361
|
+
* C accessor function for creating a YALE_STORAGE object. Prior to calling this function, you MUST
|
1362
|
+
* allocate shape (should be size_t * 2) -- don't use use a regular size_t array!
|
1363
|
+
*
|
1364
|
+
* For this type, dim must always be 2. The final argument is the initial capacity with which to
|
1365
|
+
* create the storage.
|
1366
|
+
*/
|
1367
|
+
|
1368
|
+
YALE_STORAGE* nm_yale_storage_create(dtype_t dtype, size_t* shape, size_t dim, size_t init_capacity) {
|
1369
|
+
YALE_STORAGE* s;
|
1370
|
+
size_t max_capacity;
|
1371
|
+
|
1372
|
+
// FIXME: This error should be handled in the nmatrix.c file.
|
1373
|
+
if (dim != 2) {
|
1374
|
+
rb_raise(rb_eNotImpError, "Can only support 2D matrices");
|
1375
|
+
}
|
1376
|
+
|
1377
|
+
s = alloc(dtype, shape, dim);
|
1378
|
+
max_capacity = nm::yale_storage::max_size(s);
|
1379
|
+
|
1380
|
+
// Set matrix capacity (and ensure its validity)
|
1381
|
+
if (init_capacity < NM_YALE_MINIMUM(s)) {
|
1382
|
+
s->capacity = NM_YALE_MINIMUM(s);
|
1383
|
+
|
1384
|
+
} else if (init_capacity > max_capacity) {
|
1385
|
+
// Don't allow storage to be created larger than necessary
|
1386
|
+
s->capacity = max_capacity;
|
1387
|
+
|
1388
|
+
} else {
|
1389
|
+
s->capacity = init_capacity;
|
1390
|
+
|
1391
|
+
}
|
1392
|
+
|
1393
|
+
s->ija = ALLOC_N( char, ITYPE_SIZES[s->itype] * s->capacity );
|
1394
|
+
s->a = ALLOC_N( char, DTYPE_SIZES[s->dtype] * s->capacity );
|
1395
|
+
|
1396
|
+
return s;
|
1397
|
+
}
|
1398
|
+
|
1399
|
+
/*
|
1400
|
+
* Destructor for yale storage (C-accessible).
|
1401
|
+
*/
|
1402
|
+
void nm_yale_storage_delete(STORAGE* s) {
|
1403
|
+
if (s) {
|
1404
|
+
YALE_STORAGE* storage = (YALE_STORAGE*)s;
|
1405
|
+
free(storage->shape);
|
1406
|
+
free(storage->ija);
|
1407
|
+
free(storage->a);
|
1408
|
+
free(storage);
|
1409
|
+
}
|
1410
|
+
}
|
1411
|
+
|
1412
|
+
/*
|
1413
|
+
* C accessor for yale_storage::init, a templated function.
|
1414
|
+
*
|
1415
|
+
* Initializes the IJA vector of the YALE_STORAGE matrix.
|
1416
|
+
*/
|
1417
|
+
void nm_yale_storage_init(YALE_STORAGE* s) {
|
1418
|
+
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::init, void, YALE_STORAGE* s);
|
1419
|
+
|
1420
|
+
ttable[s->dtype][s->itype](s);
|
1421
|
+
}
|
1422
|
+
|
1423
|
+
/*
|
1424
|
+
* Ruby GC mark function for YALE_STORAGE. C accessible.
|
1425
|
+
*/
|
1426
|
+
void nm_yale_storage_mark(void* storage_base) {
|
1427
|
+
YALE_STORAGE* storage = (YALE_STORAGE*)storage_base;
|
1428
|
+
size_t i;
|
1429
|
+
|
1430
|
+
if (storage && storage->dtype == RUBYOBJ) {
|
1431
|
+
for (i = storage->capacity; i-- > 0;) {
|
1432
|
+
rb_gc_mark(*((VALUE*)((char*)(storage->a) + i*DTYPE_SIZES[RUBYOBJ])));
|
1433
|
+
}
|
1434
|
+
}
|
1435
|
+
}
|
1436
|
+
|
1437
|
+
/*
|
1438
|
+
* Allocates and initializes the basic struct (but not the IJA or A vectors).
|
1439
|
+
*/
|
1440
|
+
static YALE_STORAGE* alloc(dtype_t dtype, size_t* shape, size_t dim) {
|
1441
|
+
YALE_STORAGE* s;
|
1442
|
+
|
1443
|
+
s = ALLOC( YALE_STORAGE );
|
1444
|
+
|
1445
|
+
s->ndnz = 0;
|
1446
|
+
s->dtype = dtype;
|
1447
|
+
s->shape = shape;
|
1448
|
+
s->dim = dim;
|
1449
|
+
s->itype = nm_yale_storage_itype_by_shape(shape);
|
1450
|
+
|
1451
|
+
return s;
|
1452
|
+
}
|
1453
|
+
|
1454
|
+
YALE_STORAGE* nm_yale_storage_create_from_old_yale(dtype_t dtype, size_t* shape, void* ia, void* ja, void* a, dtype_t from_dtype) {
|
1455
|
+
|
1456
|
+
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::create_from_old_yale, YALE_STORAGE*, dtype_t dtype, size_t* shape, void* r_ia, void* r_ja, void* r_a);
|
1457
|
+
|
1458
|
+
// With C++ templates, we don't want to have a 4-parameter template. That would be LDType, RDType, LIType, RIType.
|
1459
|
+
// We can prevent that by copying ia and ja into the correct itype (if necessary) before passing them to the yale
|
1460
|
+
// copy constructor.
|
1461
|
+
itype_t to_itype = nm_yale_storage_itype_by_shape(shape);
|
1462
|
+
|
1463
|
+
return ttable[dtype][from_dtype][to_itype](dtype, shape, ia, ja, a);
|
1464
|
+
|
1465
|
+
}
|
1466
|
+
|
1467
|
+
//////////////////////////////////////////////
|
1468
|
+
// YALE-SPECIFIC FUNCTIONS (RUBY ACCESSORS) //
|
1469
|
+
//////////////////////////////////////////////
|
1470
|
+
|
1471
|
+
/*
|
1472
|
+
* Get the size of a Yale matrix (the number of elements actually stored).
|
1473
|
+
*
|
1474
|
+
* For capacity (the maximum number of elements that can be stored without a resize), use capacity instead.
|
1475
|
+
*/
|
1476
|
+
static VALUE nm_size(VALUE self) {
|
1477
|
+
YALE_STORAGE* s = (YALE_STORAGE*)NM_STORAGE(self);
|
1478
|
+
|
1479
|
+
return rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*(s->shape[0]), s->itype).rval;
|
1480
|
+
}
|
1481
|
+
|
1482
|
+
|
1483
|
+
/*
|
1484
|
+
* Get the A array of a Yale matrix (which stores the diagonal and the LU portions of the matrix).
|
1485
|
+
*/
|
1486
|
+
static VALUE nm_a(VALUE self) {
|
1487
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1488
|
+
|
1489
|
+
size_t size = nm_yale_storage_get_size(s);
|
1490
|
+
VALUE* vals = ALLOCA_N(VALUE, size);
|
1491
|
+
|
1492
|
+
for (size_t i = 0; i < size; ++i) {
|
1493
|
+
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1494
|
+
}
|
1495
|
+
VALUE ary = rb_ary_new4(size, vals);
|
1496
|
+
|
1497
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1498
|
+
rb_ary_push(ary, Qnil);
|
1499
|
+
|
1500
|
+
return ary;
|
1501
|
+
}
|
1502
|
+
|
1503
|
+
|
1504
|
+
/*
|
1505
|
+
* Get the diagonal ("D") portion of the A array of a Yale matrix.
|
1506
|
+
*/
|
1507
|
+
static VALUE nm_d(VALUE self) {
|
1508
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1509
|
+
|
1510
|
+
VALUE* vals = ALLOCA_N(VALUE, s->shape[0]);
|
1511
|
+
|
1512
|
+
for (size_t i = 0; i < s->shape[0]; ++i) {
|
1513
|
+
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1514
|
+
}
|
1515
|
+
return rb_ary_new4(s->shape[0], vals);
|
1516
|
+
}
|
1517
|
+
|
1518
|
+
/*
|
1519
|
+
* Get the non-diagonal ("LU") portion of the A array of a Yale matrix.
|
1520
|
+
*/
|
1521
|
+
static VALUE nm_lu(VALUE self) {
|
1522
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1523
|
+
|
1524
|
+
size_t size = nm_yale_storage_get_size(s);
|
1525
|
+
|
1526
|
+
VALUE* vals = ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
1527
|
+
|
1528
|
+
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
1529
|
+
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*(s->shape[0] + 1 + i), s->dtype).rval;
|
1530
|
+
}
|
1531
|
+
|
1532
|
+
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
1533
|
+
|
1534
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1535
|
+
rb_ary_push(ary, Qnil);
|
1536
|
+
|
1537
|
+
return ary;
|
1538
|
+
}
|
1539
|
+
|
1540
|
+
/*
|
1541
|
+
* Get the IA portion of the IJA array of a Yale matrix. This gives the start and end positions of rows in the
|
1542
|
+
* JA and LU portions of the IJA and A arrays, respectively.
|
1543
|
+
*/
|
1544
|
+
static VALUE nm_ia(VALUE self) {
|
1545
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1546
|
+
|
1547
|
+
VALUE* vals = ALLOCA_N(VALUE, s->shape[0] + 1);
|
1548
|
+
|
1549
|
+
for (size_t i = 0; i < s->shape[0] + 1; ++i) {
|
1550
|
+
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*i, s->itype).rval;
|
1551
|
+
}
|
1552
|
+
|
1553
|
+
return rb_ary_new4(s->shape[0]+1, vals);
|
1554
|
+
}
|
1555
|
+
|
1556
|
+
/*
|
1557
|
+
* Get the JA portion of the IJA array of a Yale matrix. This gives the column indices for entries in corresponding
|
1558
|
+
* positions in the LU portion of the A array.
|
1559
|
+
*/
|
1560
|
+
static VALUE nm_ja(VALUE self) {
|
1561
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1562
|
+
|
1563
|
+
size_t size = nm_yale_storage_get_size(s);
|
1564
|
+
|
1565
|
+
VALUE* vals = ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
1566
|
+
|
1567
|
+
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
1568
|
+
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*(s->shape[0] + 1 + i), s->itype).rval;
|
1569
|
+
}
|
1570
|
+
|
1571
|
+
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
1572
|
+
|
1573
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1574
|
+
rb_ary_push(ary, Qnil);
|
1575
|
+
|
1576
|
+
return ary;
|
1577
|
+
}
|
1578
|
+
|
1579
|
+
/*
|
1580
|
+
* Get the IJA array of a Yale matrix.
|
1581
|
+
*/
|
1582
|
+
static VALUE nm_ija(VALUE self) {
|
1583
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1584
|
+
|
1585
|
+
size_t size = nm_yale_storage_get_size(s);
|
1586
|
+
|
1587
|
+
VALUE* vals = ALLOCA_N(VALUE, size);
|
1588
|
+
|
1589
|
+
for (size_t i = 0; i < size; ++i) {
|
1590
|
+
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*i, s->itype).rval;
|
1591
|
+
}
|
1592
|
+
|
1593
|
+
VALUE ary = rb_ary_new4(size, vals);
|
1594
|
+
|
1595
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1596
|
+
rb_ary_push(ary, Qnil);
|
1597
|
+
|
1598
|
+
return ary;
|
1599
|
+
}
|
1600
|
+
|
1601
|
+
} // end of extern "C" block
|