nmatrix 0.0.8 → 0.0.9
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/.gitignore +3 -8
- data/.rspec +1 -1
- data/.travis.yml +12 -0
- data/CONTRIBUTING.md +27 -12
- data/Gemfile +1 -0
- data/History.txt +38 -0
- data/Manifest.txt +15 -15
- data/README.rdoc +7 -6
- data/Rakefile +40 -5
- data/ext/nmatrix/data/data.cpp +2 -37
- data/ext/nmatrix/data/data.h +19 -121
- data/ext/nmatrix/data/meta.h +70 -0
- data/ext/nmatrix/extconf.rb +40 -12
- data/ext/nmatrix/math/math.h +13 -103
- data/ext/nmatrix/nmatrix.cpp +10 -2018
- data/ext/nmatrix/nmatrix.h +16 -13
- data/ext/nmatrix/ruby_constants.cpp +12 -1
- data/ext/nmatrix/ruby_constants.h +7 -1
- data/ext/nmatrix/ruby_nmatrix.c +2169 -0
- data/ext/nmatrix/storage/dense.cpp +123 -14
- data/ext/nmatrix/storage/dense.h +10 -4
- data/ext/nmatrix/storage/list.cpp +265 -48
- data/ext/nmatrix/storage/list.h +6 -9
- data/ext/nmatrix/storage/storage.cpp +44 -54
- data/ext/nmatrix/storage/storage.h +2 -2
- data/ext/nmatrix/storage/yale/class.h +1070 -0
- data/ext/nmatrix/storage/yale/iterators/base.h +142 -0
- data/ext/nmatrix/storage/yale/iterators/iterator.h +130 -0
- data/ext/nmatrix/storage/yale/iterators/row.h +449 -0
- data/ext/nmatrix/storage/yale/iterators/row_stored.h +139 -0
- data/ext/nmatrix/storage/yale/iterators/row_stored_nd.h +167 -0
- data/ext/nmatrix/storage/yale/iterators/stored_diagonal.h +123 -0
- data/ext/nmatrix/storage/yale/math/transpose.h +110 -0
- data/ext/nmatrix/storage/yale/yale.cpp +1785 -0
- data/ext/nmatrix/storage/{yale.h → yale/yale.h} +23 -55
- data/ext/nmatrix/types.h +2 -0
- data/ext/nmatrix/util/io.cpp +27 -45
- data/ext/nmatrix/util/io.h +0 -2
- data/ext/nmatrix/util/sl_list.cpp +169 -28
- data/ext/nmatrix/util/sl_list.h +9 -3
- data/lib/nmatrix/blas.rb +20 -20
- data/lib/nmatrix/enumerate.rb +1 -1
- data/lib/nmatrix/io/mat5_reader.rb +8 -14
- data/lib/nmatrix/lapack.rb +3 -3
- data/lib/nmatrix/math.rb +3 -3
- data/lib/nmatrix/nmatrix.rb +19 -5
- data/lib/nmatrix/nvector.rb +2 -0
- data/lib/nmatrix/shortcuts.rb +90 -125
- data/lib/nmatrix/version.rb +1 -1
- data/nmatrix.gemspec +7 -8
- data/spec/{nmatrix_spec.rb → 00_nmatrix_spec.rb} +45 -208
- data/spec/01_enum_spec.rb +184 -0
- data/spec/{slice_spec.rb → 02_slice_spec.rb} +55 -39
- data/spec/blas_spec.rb +22 -54
- data/spec/elementwise_spec.rb +9 -8
- data/spec/io_spec.rb +6 -4
- data/spec/lapack_spec.rb +26 -26
- data/spec/math_spec.rb +9 -5
- data/spec/nmatrix_yale_spec.rb +29 -61
- data/spec/shortcuts_spec.rb +34 -22
- data/spec/slice_set_spec.rb +157 -0
- data/spec/spec_helper.rb +42 -2
- data/spec/stat_spec.rb +192 -0
- metadata +52 -55
- data/ext/nmatrix/storage/yale.cpp +0 -2284
- data/spec/nmatrix_list_spec.rb +0 -113
- data/spec/nvector_spec.rb +0 -112
@@ -1,2284 +0,0 @@
|
|
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 - 2013, Ruby Science Foundation
|
13
|
-
// NMatrix is Copyright (c) 2013, 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
|
-
#include <iostream>
|
46
|
-
#include <array>
|
47
|
-
#include <typeinfo>
|
48
|
-
|
49
|
-
#define RB_P(OBJ) \
|
50
|
-
rb_funcall(rb_stderr, rb_intern("print"), 1, rb_funcall(OBJ, rb_intern("object_id"), 0)); \
|
51
|
-
rb_funcall(rb_stderr, rb_intern("puts"), 1, rb_funcall(OBJ, rb_intern("inspect"), 0));
|
52
|
-
|
53
|
-
/*
|
54
|
-
* Project Includes
|
55
|
-
*/
|
56
|
-
|
57
|
-
// #include "types.h"
|
58
|
-
#include "data/data.h"
|
59
|
-
#include "math/math.h"
|
60
|
-
|
61
|
-
#include "common.h"
|
62
|
-
#include "yale.h"
|
63
|
-
|
64
|
-
#include "nmatrix.h"
|
65
|
-
#include "ruby_constants.h"
|
66
|
-
|
67
|
-
/*
|
68
|
-
* Macros
|
69
|
-
*/
|
70
|
-
|
71
|
-
#ifndef NM_MAX
|
72
|
-
#define NM_MAX(a,b) (((a)>(b))?(a):(b))
|
73
|
-
#define NM_MIN(a,b) (((a)<(b))?(a):(b))
|
74
|
-
#endif
|
75
|
-
|
76
|
-
#ifndef NM_MAX_ITYPE
|
77
|
-
#define NM_MAX_ITYPE(a,b) ((static_cast<int8_t>(a) > static_cast<int8_t>(b)) ? static_cast<nm::itype_t>(a) : static_cast<nm::itype_t>(b))
|
78
|
-
#define NM_MIN_ITYPE(a,b) ((static_cast<int8_t>(a) < static_cast<int8_t>(b)) ? static_cast<nm::itype_t>(a) : static_cast<nm::itype_t>(b))
|
79
|
-
#endif
|
80
|
-
|
81
|
-
/*
|
82
|
-
* Forward Declarations
|
83
|
-
*/
|
84
|
-
|
85
|
-
extern "C" {
|
86
|
-
static YALE_STORAGE* nm_copy_alloc_struct(const YALE_STORAGE* rhs, const nm::dtype_t new_dtype, const size_t new_capacity, const size_t new_size);
|
87
|
-
static YALE_STORAGE* alloc(nm::dtype_t dtype, size_t* shape, size_t dim, nm::itype_t min_itype);
|
88
|
-
|
89
|
-
static size_t yale_count_slice_copy_ndnz(const YALE_STORAGE* s, size_t*, size_t*);
|
90
|
-
|
91
|
-
static void* default_value_ptr(const YALE_STORAGE* s);
|
92
|
-
static VALUE default_value(const YALE_STORAGE* s);
|
93
|
-
static VALUE obj_at(YALE_STORAGE* s, size_t k);
|
94
|
-
|
95
|
-
/* Ruby-accessible functions */
|
96
|
-
static VALUE nm_size(VALUE self);
|
97
|
-
static VALUE nm_a(int argc, VALUE* argv, VALUE self);
|
98
|
-
static VALUE nm_d(int argc, VALUE* argv, VALUE self);
|
99
|
-
static VALUE nm_lu(VALUE self);
|
100
|
-
static VALUE nm_ia(VALUE self);
|
101
|
-
static VALUE nm_ja(VALUE self);
|
102
|
-
static VALUE nm_ija(int argc, VALUE* argv, VALUE self);
|
103
|
-
|
104
|
-
static VALUE nm_nd_row(int argc, VALUE* argv, VALUE self);
|
105
|
-
|
106
|
-
static inline size_t src_ndnz(const YALE_STORAGE* s) {
|
107
|
-
return reinterpret_cast<YALE_STORAGE*>(s->src)->ndnz;
|
108
|
-
}
|
109
|
-
|
110
|
-
} // end extern "C" block
|
111
|
-
|
112
|
-
namespace nm { namespace yale_storage {
|
113
|
-
|
114
|
-
template <typename DType, typename IType>
|
115
|
-
static bool ndrow_is_empty(const YALE_STORAGE* s, IType ija, const IType ija_next);
|
116
|
-
|
117
|
-
template <typename LDType, typename RDType, typename IType>
|
118
|
-
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);
|
119
|
-
|
120
|
-
template <typename LDType, typename RDType, typename IType>
|
121
|
-
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right);
|
122
|
-
|
123
|
-
template <typename LDType, typename RDType, typename IType>
|
124
|
-
static bool eqeq_different_defaults(const YALE_STORAGE* s, const LDType& s_init, const YALE_STORAGE* t, const RDType& t_init);
|
125
|
-
|
126
|
-
template <typename IType>
|
127
|
-
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);
|
128
|
-
|
129
|
-
template <typename IType>
|
130
|
-
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, IType n);
|
131
|
-
|
132
|
-
template <typename IType>
|
133
|
-
static void c_increment_ia_after(YALE_STORAGE* s, size_t ija_size, size_t i, size_t n) {
|
134
|
-
increment_ia_after<IType>(s, ija_size, i, n);
|
135
|
-
}
|
136
|
-
|
137
|
-
template <typename IType>
|
138
|
-
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool* found);
|
139
|
-
|
140
|
-
template <typename DType, typename IType>
|
141
|
-
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, void* val_, size_t n, bool struct_only);
|
142
|
-
|
143
|
-
template <typename DType, typename IType>
|
144
|
-
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only);
|
145
|
-
|
146
|
-
|
147
|
-
/*
|
148
|
-
* Functions
|
149
|
-
*/
|
150
|
-
|
151
|
-
/*
|
152
|
-
* Copy a vector from one IType or DType to another.
|
153
|
-
*/
|
154
|
-
template <typename LType, typename RType>
|
155
|
-
static inline void copy_recast_vector(const void* in_, void* out_, size_t length) {
|
156
|
-
const RType* in = reinterpret_cast<const RType*>(in_);
|
157
|
-
LType* out = reinterpret_cast<LType*>(out_);
|
158
|
-
for (size_t i = 0; i < length; ++i) {
|
159
|
-
out[i] = in[i];
|
160
|
-
}
|
161
|
-
out;
|
162
|
-
}
|
163
|
-
|
164
|
-
|
165
|
-
static inline void copy_recast_itype_vector(const void* in, nm::itype_t in_itype, void* out, nm::itype_t out_itype, size_t length) {
|
166
|
-
NAMED_LR_ITYPE_TEMPLATE_TABLE(ttable, copy_recast_vector, void, const void* in_, void* out_, size_t length);
|
167
|
-
|
168
|
-
ttable[out_itype][in_itype](in, out, length);
|
169
|
-
}
|
170
|
-
|
171
|
-
|
172
|
-
/*
|
173
|
-
* Create Yale storage from IA, JA, and A vectors given in Old Yale format (probably from a file, since NMatrix only uses
|
174
|
-
* new Yale for its storage).
|
175
|
-
*
|
176
|
-
* This function is needed for Matlab .MAT v5 IO.
|
177
|
-
*/
|
178
|
-
template <typename LDType, typename RDType, typename IType>
|
179
|
-
YALE_STORAGE* create_from_old_yale(dtype_t dtype, size_t* shape, void* r_ia, void* r_ja, void* r_a) {
|
180
|
-
IType* ir = reinterpret_cast<IType*>(r_ia);
|
181
|
-
IType* jr = reinterpret_cast<IType*>(r_ja);
|
182
|
-
RDType* ar = reinterpret_cast<RDType*>(r_a);
|
183
|
-
|
184
|
-
// Read through ia and ja and figure out the ndnz (non-diagonal non-zeros) count.
|
185
|
-
size_t ndnz = 0, i, p, p_next;
|
186
|
-
|
187
|
-
for (i = 0; i < shape[0]; ++i) { // Walk down rows
|
188
|
-
for (p = ir[i], p_next = ir[i+1]; p < p_next; ++p) { // Now walk through columns
|
189
|
-
|
190
|
-
if (i != jr[p]) ++ndnz; // entry is non-diagonal and probably nonzero
|
191
|
-
|
192
|
-
}
|
193
|
-
}
|
194
|
-
|
195
|
-
// Having walked through the matrix, we now go about allocating the space for it.
|
196
|
-
YALE_STORAGE* s = alloc(dtype, shape, 2, UINT8);
|
197
|
-
|
198
|
-
s->capacity = shape[0] + ndnz + 1;
|
199
|
-
s->ndnz = ndnz;
|
200
|
-
|
201
|
-
// Setup IJA and A arrays
|
202
|
-
s->ija = ALLOC_N( IType, s->capacity );
|
203
|
-
s->a = ALLOC_N( LDType, s->capacity );
|
204
|
-
IType* ijl = reinterpret_cast<IType*>(s->ija);
|
205
|
-
LDType* al = reinterpret_cast<LDType*>(s->a);
|
206
|
-
|
207
|
-
// set the diagonal to zero -- this prevents uninitialized values from popping up.
|
208
|
-
for (size_t index = 0; index < shape[0]; ++index) {
|
209
|
-
al[index] = 0;
|
210
|
-
}
|
211
|
-
|
212
|
-
// Figure out where to start writing JA in IJA:
|
213
|
-
size_t pp = s->shape[0]+1;
|
214
|
-
|
215
|
-
// Find beginning of first row
|
216
|
-
p = ir[0];
|
217
|
-
|
218
|
-
// Now fill the arrays
|
219
|
-
for (i = 0; i < s->shape[0]; ++i) {
|
220
|
-
|
221
|
-
// Set the beginning of the row (of output)
|
222
|
-
ijl[i] = pp;
|
223
|
-
|
224
|
-
// Now walk through columns, starting at end of row (of input)
|
225
|
-
for (size_t p_next = ir[i+1]; p < p_next; ++p, ++pp) {
|
226
|
-
|
227
|
-
if (i == jr[p]) { // diagonal
|
228
|
-
|
229
|
-
al[i] = ar[p];
|
230
|
-
--pp;
|
231
|
-
|
232
|
-
} else { // nondiagonal
|
233
|
-
|
234
|
-
ijl[pp] = jr[p];
|
235
|
-
al[pp] = ar[p];
|
236
|
-
|
237
|
-
}
|
238
|
-
}
|
239
|
-
}
|
240
|
-
|
241
|
-
ijl[i] = pp; // Set the end of the last row
|
242
|
-
|
243
|
-
// Set the zero position for our output matrix
|
244
|
-
al[i] = 0;
|
245
|
-
|
246
|
-
return s;
|
247
|
-
}
|
248
|
-
|
249
|
-
|
250
|
-
/*
|
251
|
-
* Empty the matrix by initializing the IJA vector and setting the diagonal to 0.
|
252
|
-
*
|
253
|
-
* Called when most YALE_STORAGE objects are created.
|
254
|
-
*/
|
255
|
-
template <typename DType, typename IType>
|
256
|
-
void init(YALE_STORAGE* s, void* init_val) {
|
257
|
-
IType IA_INIT = s->shape[0] + 1;
|
258
|
-
|
259
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
260
|
-
// clear out IJA vector
|
261
|
-
for (IType i = 0; i < IA_INIT; ++i) {
|
262
|
-
ija[i] = IA_INIT; // set initial values for IJA
|
263
|
-
}
|
264
|
-
|
265
|
-
clear_diagonal_and_zero<DType>(s, init_val);
|
266
|
-
}
|
267
|
-
|
268
|
-
size_t max_size(YALE_STORAGE* s) {
|
269
|
-
size_t result = s->shape[0]*s->shape[1] + 1;
|
270
|
-
if (s->shape[0] > s->shape[1])
|
271
|
-
result += s->shape[0] - s->shape[1];
|
272
|
-
|
273
|
-
return result;
|
274
|
-
}
|
275
|
-
|
276
|
-
|
277
|
-
///////////////
|
278
|
-
// Accessors //
|
279
|
-
///////////////
|
280
|
-
|
281
|
-
|
282
|
-
/*
|
283
|
-
* Determine the number of non-diagonal non-zeros in a not-yet-created copy of a slice or matrix.
|
284
|
-
*/
|
285
|
-
template <typename DType, typename IType>
|
286
|
-
static size_t count_slice_copy_ndnz(const YALE_STORAGE* s, size_t* offset, size_t* shape) {
|
287
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
288
|
-
DType* a = reinterpret_cast<DType*>(s->a);
|
289
|
-
|
290
|
-
DType ZERO(*reinterpret_cast<DType*>(default_value_ptr(s)));
|
291
|
-
|
292
|
-
// Calc ndnz for the destination
|
293
|
-
size_t ndnz = 0;
|
294
|
-
size_t i, j; // indexes of destination matrix
|
295
|
-
size_t k, l; // indexes of source matrix
|
296
|
-
for (i = 0; i < shape[0]; i++) {
|
297
|
-
k = i + offset[0];
|
298
|
-
for (j = 0; j < shape[1]; j++) {
|
299
|
-
l = j + offset[1];
|
300
|
-
|
301
|
-
if (j == i) continue;
|
302
|
-
|
303
|
-
if (k == l) { // for diagonal element of source
|
304
|
-
if (a[k] != ZERO) ++ndnz;
|
305
|
-
} else { // for non-diagonal element
|
306
|
-
for (size_t c = ija[k]; c < ija[k+1]; c++) {
|
307
|
-
if (ija[c] == l) {
|
308
|
-
++ndnz;
|
309
|
-
break;
|
310
|
-
}
|
311
|
-
}
|
312
|
-
}
|
313
|
-
}
|
314
|
-
}
|
315
|
-
|
316
|
-
return ndnz;
|
317
|
-
}
|
318
|
-
|
319
|
-
|
320
|
-
|
321
|
-
/*
|
322
|
-
* Copy some portion of a matrix into a new matrix.
|
323
|
-
*/
|
324
|
-
template <typename LDType, typename RDType, typename IType>
|
325
|
-
static void slice_copy(YALE_STORAGE* ns, const YALE_STORAGE* s, size_t* offset, size_t* lengths, dtype_t new_dtype) {
|
326
|
-
|
327
|
-
IType* src_ija = reinterpret_cast<IType*>(s->ija);
|
328
|
-
RDType* src_a = reinterpret_cast<RDType*>(s->a);
|
329
|
-
|
330
|
-
RDType RZERO(*reinterpret_cast<RDType*>(default_value_ptr(s)));
|
331
|
-
|
332
|
-
// Initialize the A and IJA arrays
|
333
|
-
LDType val(RZERO); // need default value for init. Can't use ns default value because it's not initialized yet
|
334
|
-
init<LDType,IType>(ns, &val);
|
335
|
-
IType* dst_ija = reinterpret_cast<IType*>(ns->ija);
|
336
|
-
LDType* dst_a = reinterpret_cast<LDType*>(ns->a);
|
337
|
-
|
338
|
-
size_t ija = lengths[0] + 1;
|
339
|
-
|
340
|
-
size_t i, j; // indexes of destination matrix
|
341
|
-
size_t k, l; // indexes of source matrix
|
342
|
-
|
343
|
-
for (i = 0; i < lengths[0]; ++i) {
|
344
|
-
k = i + offset[0];
|
345
|
-
for (j = 0; j < lengths[1]; ++j) {
|
346
|
-
bool found = false;
|
347
|
-
l = j + offset[1];
|
348
|
-
|
349
|
-
// Get value from source matrix
|
350
|
-
if (k == l) { // source diagonal
|
351
|
-
if (src_a[k] != RZERO) { // don't bother copying non-zero values from the diagonal
|
352
|
-
val = src_a[k];
|
353
|
-
found = true;
|
354
|
-
}
|
355
|
-
} else {
|
356
|
-
// copy one non-diagonal element
|
357
|
-
for (size_t c = src_ija[k]; !found && c < src_ija[k+1]; ++c) {
|
358
|
-
if (src_ija[c] == l) {
|
359
|
-
val = src_a[c];
|
360
|
-
found = true;
|
361
|
-
}
|
362
|
-
}
|
363
|
-
}
|
364
|
-
|
365
|
-
if (found) {
|
366
|
-
// Set value in destination matrix
|
367
|
-
if (i == j) {
|
368
|
-
dst_a[i] = val;
|
369
|
-
} else {
|
370
|
-
// copy non-diagonal element
|
371
|
-
dst_ija[ija] = j;
|
372
|
-
dst_a[ija] = val;
|
373
|
-
++ija;
|
374
|
-
for (size_t c = i + 1; c <= lengths[0]; ++c) {
|
375
|
-
dst_ija[c] = ija;
|
376
|
-
}
|
377
|
-
}
|
378
|
-
}
|
379
|
-
}
|
380
|
-
}
|
381
|
-
|
382
|
-
dst_ija[lengths[0]] = ija; // indicate the end of the last row
|
383
|
-
ns->ndnz = ija - lengths[0] - 1; // update ndnz count
|
384
|
-
}
|
385
|
-
|
386
|
-
|
387
|
-
/*
|
388
|
-
* Get a single element of a yale storage object
|
389
|
-
*/
|
390
|
-
template <typename DType, typename IType>
|
391
|
-
static void* get_single(YALE_STORAGE* storage, SLICE* slice) {
|
392
|
-
|
393
|
-
DType* a = reinterpret_cast<DType*>(storage->a);
|
394
|
-
IType* ija = reinterpret_cast<IType*>(storage->ija);
|
395
|
-
|
396
|
-
size_t coord0 = storage->offset[0] + slice->coords[0];
|
397
|
-
size_t coord1 = storage->offset[1] + slice->coords[1];
|
398
|
-
|
399
|
-
if (coord0 == coord1)
|
400
|
-
return &(a[ coord0 ]); // return diagonal entry
|
401
|
-
|
402
|
-
if (ija[coord0] == ija[coord0+1])
|
403
|
-
return &(a[ storage->src->shape[0] ]); // return zero pointer
|
404
|
-
|
405
|
-
// binary search for the column's location
|
406
|
-
int pos = binary_search<IType>(storage, ija[coord0], ija[coord0+1]-1, coord1);
|
407
|
-
|
408
|
-
if (pos != -1 && ija[pos] == coord1)
|
409
|
-
return &(a[pos]); // found exact value
|
410
|
-
|
411
|
-
return &(a[ storage->src->shape[0] ]); // return a pointer that happens to be zero
|
412
|
-
}
|
413
|
-
|
414
|
-
|
415
|
-
/*
|
416
|
-
* Returns a pointer to the correct location in the A vector of a YALE_STORAGE object, given some set of coordinates
|
417
|
-
* (the coordinates are stored in slice).
|
418
|
-
*/
|
419
|
-
template <typename DType,typename IType>
|
420
|
-
void* ref(YALE_STORAGE* s, SLICE* slice) {
|
421
|
-
|
422
|
-
YALE_STORAGE* ns = ALLOC( YALE_STORAGE );
|
423
|
-
|
424
|
-
ns->dim = s->dim;
|
425
|
-
ns->offset = ALLOC_N(size_t, ns->dim);
|
426
|
-
ns->shape = ALLOC_N(size_t, ns->dim);
|
427
|
-
|
428
|
-
for (size_t i = 0; i < ns->dim; ++i) {
|
429
|
-
ns->offset[i] = slice->coords[i] + s->offset[i];
|
430
|
-
ns->shape[i] = slice->lengths[i];
|
431
|
-
}
|
432
|
-
|
433
|
-
ns->dtype = s->dtype;
|
434
|
-
ns->itype = s->itype; // or should we go by shape?
|
435
|
-
|
436
|
-
ns->a = s->a;
|
437
|
-
ns->ija = s->ija;
|
438
|
-
|
439
|
-
ns->src = s->src;
|
440
|
-
s->src->count++;
|
441
|
-
|
442
|
-
ns->ndnz = 0;
|
443
|
-
ns->capacity= 0;
|
444
|
-
|
445
|
-
return ns;
|
446
|
-
|
447
|
-
}
|
448
|
-
|
449
|
-
/*
|
450
|
-
* Attempt to set some cell in a YALE_STORAGE object. Must supply coordinates and a pointer to a value (which will be
|
451
|
-
* copied into the storage object).
|
452
|
-
*/
|
453
|
-
template <typename DType, typename IType>
|
454
|
-
char set(YALE_STORAGE* storage, SLICE* slice, void* value) {
|
455
|
-
DType* v = reinterpret_cast<DType*>(value);
|
456
|
-
size_t coord0 = storage->offset[0] + slice->coords[0],
|
457
|
-
coord1 = storage->offset[1] + slice->coords[1];
|
458
|
-
|
459
|
-
bool found = false;
|
460
|
-
char ins_type;
|
461
|
-
|
462
|
-
if (coord0 == coord1) {
|
463
|
-
reinterpret_cast<DType*>(storage->a)[coord0] = *v; // set diagonal
|
464
|
-
return 'r';
|
465
|
-
}
|
466
|
-
|
467
|
-
// Get IJA positions of the beginning and end of the row
|
468
|
-
if (reinterpret_cast<IType*>(storage->ija)[coord0] == reinterpret_cast<IType*>(storage->ija)[coord0+1]) {
|
469
|
-
// empty row
|
470
|
-
ins_type = vector_insert<DType,IType>(storage, reinterpret_cast<IType*>(storage->ija)[coord0], &(coord1), v, 1, false);
|
471
|
-
increment_ia_after<IType>(storage, storage->shape[0], coord0, 1);
|
472
|
-
reinterpret_cast<YALE_STORAGE*>(storage->src)->ndnz++;
|
473
|
-
|
474
|
-
return ins_type;
|
475
|
-
}
|
476
|
-
|
477
|
-
// non-empty row. search for coords[1] in the IJA array, between ija and ija_next
|
478
|
-
// (including ija, not including ija_next)
|
479
|
-
//ija_size = get_size<IType>(storage);
|
480
|
-
|
481
|
-
// Do a binary search for the column
|
482
|
-
size_t pos = insert_search<IType>(storage,
|
483
|
-
reinterpret_cast<IType*>(storage->ija)[coord0],
|
484
|
-
reinterpret_cast<IType*>(storage->ija)[coord0+1]-1,
|
485
|
-
coord1, &found);
|
486
|
-
|
487
|
-
if (found) { // replace
|
488
|
-
reinterpret_cast<IType*>(storage->ija)[pos] = coord1;
|
489
|
-
reinterpret_cast<DType*>(storage->a)[pos] = *v;
|
490
|
-
return 'r';
|
491
|
-
}
|
492
|
-
|
493
|
-
ins_type = vector_insert<DType,IType>(storage, pos, &(coord1), v, 1, false);
|
494
|
-
increment_ia_after<IType>(storage, storage->shape[0], coord0, 1);
|
495
|
-
reinterpret_cast<YALE_STORAGE*>(storage->src)->ndnz++;
|
496
|
-
|
497
|
-
return ins_type;
|
498
|
-
}
|
499
|
-
|
500
|
-
///////////
|
501
|
-
// Tests //
|
502
|
-
///////////
|
503
|
-
|
504
|
-
/*
|
505
|
-
* Yale eql? -- for whole-matrix comparison returning a single value.
|
506
|
-
*/
|
507
|
-
template <typename LDType, typename RDType, typename IType>
|
508
|
-
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right) {
|
509
|
-
LDType l_init = reinterpret_cast<LDType*>(left->a )[left->shape[0] ];
|
510
|
-
RDType r_init = reinterpret_cast<RDType*>(right->a)[right->shape[0]];
|
511
|
-
|
512
|
-
// If the defaults are different between the two matrices, or if slicing is involved, use this other function instead:
|
513
|
-
if (l_init != r_init || left->src != left || right->src != right)
|
514
|
-
return eqeq_different_defaults<LDType,RDType,IType>(left, l_init, right, r_init);
|
515
|
-
|
516
|
-
LDType* la = reinterpret_cast<LDType*>(left->a);
|
517
|
-
RDType* ra = reinterpret_cast<RDType*>(right->a);
|
518
|
-
|
519
|
-
// Compare the diagonals first.
|
520
|
-
for (size_t index = 0; index < left->shape[0]; ++index) {
|
521
|
-
if (la[index] != ra[index]) return false;
|
522
|
-
}
|
523
|
-
|
524
|
-
IType* lij = reinterpret_cast<IType*>(left->ija);
|
525
|
-
IType* rij = reinterpret_cast<IType*>(right->ija);
|
526
|
-
|
527
|
-
for (IType i = 0; i < left->shape[0]; ++i) {
|
528
|
-
|
529
|
-
// Get start and end positions of row
|
530
|
-
IType l_ija = lij[i],
|
531
|
-
l_ija_next = lij[i+1],
|
532
|
-
r_ija = rij[i],
|
533
|
-
r_ija_next = rij[i+1];
|
534
|
-
|
535
|
-
// Check to see if one row is empty and the other isn't.
|
536
|
-
if (ndrow_is_empty<LDType,IType>(left, l_ija, l_ija_next)) {
|
537
|
-
if (!ndrow_is_empty<RDType,IType>(right, r_ija, r_ija_next)) {
|
538
|
-
return false;
|
539
|
-
}
|
540
|
-
|
541
|
-
} else if (ndrow_is_empty<RDType,IType>(right, r_ija, r_ija_next)) {
|
542
|
-
// one is empty but the other isn't
|
543
|
-
return false;
|
544
|
-
|
545
|
-
} else if (!ndrow_eqeq_ndrow<LDType,RDType,IType>(left, right, l_ija, l_ija_next, r_ija, r_ija_next)) {
|
546
|
-
// Neither row is empty. Must compare the rows directly.
|
547
|
-
return false;
|
548
|
-
}
|
549
|
-
|
550
|
-
}
|
551
|
-
|
552
|
-
return true;
|
553
|
-
}
|
554
|
-
|
555
|
-
|
556
|
-
|
557
|
-
/*
|
558
|
-
* Are two non-diagonal rows the same? We already know.
|
559
|
-
*/
|
560
|
-
template <typename LDType, typename RDType, typename IType>
|
561
|
-
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) {
|
562
|
-
bool l_no_more = false, r_no_more = false;
|
563
|
-
|
564
|
-
IType *lij = reinterpret_cast<IType*>(l->ija),
|
565
|
-
*rij = reinterpret_cast<IType*>(r->ija);
|
566
|
-
|
567
|
-
LDType* la = reinterpret_cast<LDType*>(l->a);
|
568
|
-
RDType* ra = reinterpret_cast<RDType*>(r->a);
|
569
|
-
|
570
|
-
IType l_ja = lij[l_ija],
|
571
|
-
r_ja = rij[r_ija];
|
572
|
-
|
573
|
-
IType ja = std::min(l_ja, r_ja);
|
574
|
-
|
575
|
-
LDType LZERO = la[l->shape[0]];
|
576
|
-
RDType RZERO = ra[r->shape[0]];
|
577
|
-
|
578
|
-
while (!(l_no_more && r_no_more)) {
|
579
|
-
if (l_ja == r_ja) {
|
580
|
-
|
581
|
-
if (ra[r_ija] != la[l_ija]) return false; // Direct comparison
|
582
|
-
|
583
|
-
++l_ija;
|
584
|
-
++r_ija;
|
585
|
-
|
586
|
-
if (l_ija < l_ija_next) {
|
587
|
-
l_ja = lij[l_ija];
|
588
|
-
|
589
|
-
} else {
|
590
|
-
l_no_more = true;
|
591
|
-
}
|
592
|
-
|
593
|
-
if (r_ija < r_ija_next) {
|
594
|
-
r_ja = rij[r_ija];
|
595
|
-
|
596
|
-
} else {
|
597
|
-
r_no_more = true;
|
598
|
-
}
|
599
|
-
|
600
|
-
ja = std::min(l_ja, r_ja);
|
601
|
-
|
602
|
-
} else if (l_no_more || ja < l_ja) {
|
603
|
-
|
604
|
-
if (ra[r_ija] != RZERO) return false;
|
605
|
-
|
606
|
-
++r_ija;
|
607
|
-
if (r_ija < r_ija_next) {
|
608
|
-
// get next column
|
609
|
-
r_ja = rij[r_ija];
|
610
|
-
ja = std::min(l_ja, r_ja);
|
611
|
-
|
612
|
-
} else {
|
613
|
-
l_no_more = true;
|
614
|
-
}
|
615
|
-
|
616
|
-
} else if (r_no_more || ja < r_ja) {
|
617
|
-
|
618
|
-
if (la[l_ija] != LZERO) return false;
|
619
|
-
|
620
|
-
++l_ija;
|
621
|
-
if (l_ija < l_ija_next) {
|
622
|
-
// get next column
|
623
|
-
l_ja = lij[l_ija];
|
624
|
-
ja = std::min(l_ja, r_ja);
|
625
|
-
} else {
|
626
|
-
l_no_more = true;
|
627
|
-
}
|
628
|
-
|
629
|
-
} else {
|
630
|
-
std::fprintf(stderr, "Unhandled in eqeq: l_ja=%d, r_ja=%d\n", (int)l_ja, (int)r_ja);
|
631
|
-
}
|
632
|
-
}
|
633
|
-
|
634
|
-
// every item matched
|
635
|
-
return true;
|
636
|
-
}
|
637
|
-
|
638
|
-
/*
|
639
|
-
* Is the non-diagonal portion of the row empty?
|
640
|
-
*/
|
641
|
-
template <typename DType, typename IType>
|
642
|
-
static bool ndrow_is_empty(const YALE_STORAGE* s, IType ija, const IType ija_next) {
|
643
|
-
if (ija == ija_next) return true;
|
644
|
-
|
645
|
-
DType* a = reinterpret_cast<DType*>(s->a);
|
646
|
-
|
647
|
-
// do all the entries = zero?
|
648
|
-
for (; ija < ija_next; ++ija) {
|
649
|
-
if (a[ija] != 0) return false;
|
650
|
-
}
|
651
|
-
|
652
|
-
return true;
|
653
|
-
}
|
654
|
-
|
655
|
-
//////////
|
656
|
-
// Math //
|
657
|
-
//////////
|
658
|
-
|
659
|
-
#define YALE_IA(s) (reinterpret_cast<IType*>(s->ija))
|
660
|
-
#define YALE_IJ(s) (reinterpret_cast<IType*>(s->ija) + s->shape[0] + 1)
|
661
|
-
#define YALE_COUNT(yale) (yale->ndnz + yale->shape[0])
|
662
|
-
|
663
|
-
/////////////
|
664
|
-
// Utility //
|
665
|
-
/////////////
|
666
|
-
|
667
|
-
|
668
|
-
/*
|
669
|
-
* Binary search for finding the beginning of a slice. Returns the position of the first element which is larger than
|
670
|
-
* bound.
|
671
|
-
*/
|
672
|
-
template <typename IType>
|
673
|
-
IType binary_search_left_boundary(const YALE_STORAGE* s, IType left, IType right, IType bound) {
|
674
|
-
if (left > right) return -1;
|
675
|
-
|
676
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
677
|
-
|
678
|
-
if (ija[left] >= bound) return left; // shortcut
|
679
|
-
|
680
|
-
IType mid = (left + right) / 2;
|
681
|
-
IType mid_j = ija[mid];
|
682
|
-
|
683
|
-
if (mid_j == bound)
|
684
|
-
return mid;
|
685
|
-
else if (mid_j > bound) { // eligible! don't exclude it.
|
686
|
-
return binary_search_left_boundary<IType>(s, left, mid, bound);
|
687
|
-
} else // (mid_j < bound)
|
688
|
-
return binary_search_left_boundary<IType>(s, mid + 1, right, bound);
|
689
|
-
}
|
690
|
-
|
691
|
-
|
692
|
-
/*
|
693
|
-
* Binary search for returning stored values. Returns a non-negative position, or -1 for not found.
|
694
|
-
*/
|
695
|
-
template <typename IType>
|
696
|
-
int binary_search(YALE_STORAGE* s, IType left, IType right, IType key) {
|
697
|
-
|
698
|
-
if (left > right) return -1;
|
699
|
-
|
700
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
701
|
-
|
702
|
-
IType mid = (left + right)/2;
|
703
|
-
IType mid_j = ija[mid];
|
704
|
-
|
705
|
-
if (mid_j == key)
|
706
|
-
return mid;
|
707
|
-
|
708
|
-
else if (mid_j > key)
|
709
|
-
return binary_search<IType>(s, left, mid - 1, key);
|
710
|
-
|
711
|
-
else
|
712
|
-
return binary_search<IType>(s, mid + 1, right, key);
|
713
|
-
}
|
714
|
-
|
715
|
-
|
716
|
-
/*
|
717
|
-
* Resize yale storage vectors A and IJA, copying values.
|
718
|
-
*/
|
719
|
-
static void vector_grow(YALE_STORAGE* s) {
|
720
|
-
if (s->src != s) throw; // need to correct this quickly.
|
721
|
-
|
722
|
-
size_t new_capacity = s->capacity * GROWTH_CONSTANT;
|
723
|
-
size_t max_capacity = max_size(s);
|
724
|
-
|
725
|
-
if (new_capacity > max_capacity) new_capacity = max_capacity;
|
726
|
-
|
727
|
-
void* new_ija = ALLOC_N(char, ITYPE_SIZES[s->itype] * new_capacity);
|
728
|
-
NM_CHECK_ALLOC(new_ija);
|
729
|
-
|
730
|
-
void* new_a = ALLOC_N(char, DTYPE_SIZES[s->dtype] * new_capacity);
|
731
|
-
NM_CHECK_ALLOC(new_a);
|
732
|
-
|
733
|
-
void* old_ija = s->ija;
|
734
|
-
void* old_a = s->a;
|
735
|
-
|
736
|
-
memcpy(new_ija, old_ija, s->capacity * ITYPE_SIZES[s->itype]);
|
737
|
-
memcpy(new_a, old_a, s->capacity * DTYPE_SIZES[s->dtype]);
|
738
|
-
|
739
|
-
s->capacity = new_capacity;
|
740
|
-
|
741
|
-
xfree(old_ija);
|
742
|
-
xfree(old_a);
|
743
|
-
|
744
|
-
s->ija = new_ija;
|
745
|
-
s->a = new_a;
|
746
|
-
}
|
747
|
-
|
748
|
-
|
749
|
-
/*
|
750
|
-
* Resize yale storage vectors A and IJA in preparation for an insertion.
|
751
|
-
*/
|
752
|
-
template <typename DType, typename IType>
|
753
|
-
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only) {
|
754
|
-
if (s != s->src) throw;
|
755
|
-
|
756
|
-
// Determine the new capacity for the IJA and A vectors.
|
757
|
-
size_t new_capacity = s->capacity * GROWTH_CONSTANT;
|
758
|
-
size_t max_capacity = max_size(s);
|
759
|
-
|
760
|
-
if (new_capacity > max_capacity) {
|
761
|
-
new_capacity = max_capacity;
|
762
|
-
|
763
|
-
if (current_size + n > max_capacity) rb_raise(rb_eNoMemError, "insertion size exceeded maximum yale matrix size");
|
764
|
-
}
|
765
|
-
|
766
|
-
if (new_capacity < current_size + n)
|
767
|
-
new_capacity = current_size + n;
|
768
|
-
|
769
|
-
// Allocate the new vectors.
|
770
|
-
IType* new_ija = ALLOC_N( IType, new_capacity );
|
771
|
-
NM_CHECK_ALLOC(new_ija);
|
772
|
-
|
773
|
-
DType* new_a = ALLOC_N( DType, new_capacity );
|
774
|
-
NM_CHECK_ALLOC(new_a);
|
775
|
-
|
776
|
-
IType* old_ija = reinterpret_cast<IType*>(s->ija);
|
777
|
-
DType* old_a = reinterpret_cast<DType*>(s->a);
|
778
|
-
|
779
|
-
// Copy all values prior to the insertion site to the new IJA and new A
|
780
|
-
if (struct_only) {
|
781
|
-
for (size_t i = 0; i < pos; ++i) {
|
782
|
-
new_ija[i] = old_ija[i];
|
783
|
-
}
|
784
|
-
} else {
|
785
|
-
for (size_t i = 0; i < pos; ++i) {
|
786
|
-
new_ija[i] = old_ija[i];
|
787
|
-
new_a[i] = old_a[i];
|
788
|
-
}
|
789
|
-
}
|
790
|
-
|
791
|
-
|
792
|
-
// Copy all values subsequent to the insertion site to the new IJA and new A, leaving room (size n) for insertion.
|
793
|
-
if (struct_only) {
|
794
|
-
for (size_t i = pos; i < current_size; ++i) {
|
795
|
-
new_ija[i+n] = old_ija[i];
|
796
|
-
}
|
797
|
-
} else {
|
798
|
-
for (size_t i = pos; i < current_size; ++i) {
|
799
|
-
new_ija[i+n] = old_ija[i];
|
800
|
-
new_a[i+n] = old_a[i];
|
801
|
-
}
|
802
|
-
}
|
803
|
-
|
804
|
-
s->capacity = new_capacity;
|
805
|
-
|
806
|
-
xfree(s->ija);
|
807
|
-
xfree(s->a);
|
808
|
-
|
809
|
-
s->ija = reinterpret_cast<void*>(new_ija);
|
810
|
-
s->a = reinterpret_cast<void*>(new_a);
|
811
|
-
|
812
|
-
return 'i';
|
813
|
-
}
|
814
|
-
|
815
|
-
/*
|
816
|
-
* Insert a value or contiguous values in the ija and a vectors (after ja and
|
817
|
-
* diag). Does not free anything; you are responsible!
|
818
|
-
*
|
819
|
-
* TODO: Improve this so it can handle non-contiguous element insertions
|
820
|
-
* efficiently. For now, we can just sort the elements in the row in
|
821
|
-
* question.)
|
822
|
-
*/
|
823
|
-
template <typename DType, typename IType>
|
824
|
-
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, void* val_, size_t n, bool struct_only) {
|
825
|
-
if (pos < s->shape[0]) {
|
826
|
-
rb_raise(rb_eArgError, "vector insert pos (%d) is before beginning of ja (%d); this should not happen", pos, s->shape[0]);
|
827
|
-
}
|
828
|
-
|
829
|
-
DType* val = reinterpret_cast<DType*>(val_);
|
830
|
-
|
831
|
-
size_t size = get_size<IType>(s);
|
832
|
-
|
833
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
834
|
-
DType* a = reinterpret_cast<DType*>(s->a);
|
835
|
-
|
836
|
-
if (size + n > s->capacity) {
|
837
|
-
vector_insert_resize<DType,IType>(s, size, pos, j, n, struct_only);
|
838
|
-
|
839
|
-
// Need to get the new locations for ija and a.
|
840
|
-
ija = reinterpret_cast<IType*>(s->ija);
|
841
|
-
a = reinterpret_cast<DType*>(s->a);
|
842
|
-
|
843
|
-
} else {
|
844
|
-
/*
|
845
|
-
* No resize required:
|
846
|
-
* easy (but somewhat slow), just copy elements to the tail, starting at
|
847
|
-
* the end, one element at a time.
|
848
|
-
*
|
849
|
-
* TODO: This can be made slightly more efficient, but only after the tests
|
850
|
-
* are written.
|
851
|
-
*/
|
852
|
-
|
853
|
-
if (struct_only) {
|
854
|
-
for (size_t i = 0; i < size - pos; ++i) {
|
855
|
-
ija[size+n-1-i] = ija[size-1-i];
|
856
|
-
}
|
857
|
-
} else {
|
858
|
-
for (size_t i = 0; i < size - pos; ++i) {
|
859
|
-
ija[size+n-1-i] = ija[size-1-i];
|
860
|
-
a[size+n-1-i] = a[size-1-i];
|
861
|
-
}
|
862
|
-
}
|
863
|
-
}
|
864
|
-
|
865
|
-
// Now insert the new values.
|
866
|
-
if (struct_only) {
|
867
|
-
for (size_t i = 0; i < n; ++i) {
|
868
|
-
ija[pos+i] = j[i];
|
869
|
-
}
|
870
|
-
} else {
|
871
|
-
for (size_t i = 0; i < n; ++i) {
|
872
|
-
ija[pos+i] = j[i];
|
873
|
-
a[pos+i] = val[i];
|
874
|
-
}
|
875
|
-
}
|
876
|
-
|
877
|
-
return 'i';
|
878
|
-
}
|
879
|
-
|
880
|
-
/*
|
881
|
-
* If we add n items to row i, we need to increment ija[i+1] and onward.
|
882
|
-
*/
|
883
|
-
template <typename IType>
|
884
|
-
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, IType n) {
|
885
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
886
|
-
|
887
|
-
++i;
|
888
|
-
for (; i <= ija_size; ++i) {
|
889
|
-
ija[i] += n;
|
890
|
-
}
|
891
|
-
}
|
892
|
-
|
893
|
-
/*
|
894
|
-
* Binary search for returning insertion points.
|
895
|
-
*/
|
896
|
-
template <typename IType>
|
897
|
-
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool* found) {
|
898
|
-
|
899
|
-
if (left > right) {
|
900
|
-
*found = false;
|
901
|
-
return left;
|
902
|
-
}
|
903
|
-
|
904
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
905
|
-
IType mid = (left + right)/2;
|
906
|
-
IType mid_j = ija[mid];
|
907
|
-
|
908
|
-
if (mid_j == key) {
|
909
|
-
*found = true;
|
910
|
-
return mid;
|
911
|
-
|
912
|
-
} else if (mid_j > key) {
|
913
|
-
return insert_search<IType>(s, left, mid-1, key, found);
|
914
|
-
|
915
|
-
} else {
|
916
|
-
return insert_search<IType>(s, mid+1, right, key, found);
|
917
|
-
}
|
918
|
-
}
|
919
|
-
|
920
|
-
/////////////////////////
|
921
|
-
// Copying and Casting //
|
922
|
-
/////////////////////////
|
923
|
-
|
924
|
-
/*
|
925
|
-
* Templated copy constructor for changing dtypes.
|
926
|
-
*/
|
927
|
-
template <typename LDType, typename RDType, typename IType>
|
928
|
-
YALE_STORAGE* cast_copy(const YALE_STORAGE* rhs, dtype_t new_dtype) {
|
929
|
-
|
930
|
-
YALE_STORAGE* lhs;
|
931
|
-
|
932
|
-
if (rhs->src != rhs) { // copy the reference
|
933
|
-
// Copy shape for yale construction
|
934
|
-
size_t* shape = ALLOC_N(size_t, 2);
|
935
|
-
shape[0] = rhs->shape[0];
|
936
|
-
shape[1] = rhs->shape[1];
|
937
|
-
size_t ndnz = src_ndnz(rhs);
|
938
|
-
if (shape[0] != rhs->src->shape[0] || shape[1] != rhs->src->shape[1])
|
939
|
-
ndnz = count_slice_copy_ndnz<RDType,IType>(rhs, rhs->offset, rhs->shape); // expensive, avoid if possible
|
940
|
-
size_t request_capacity = shape[0] + ndnz + 1;
|
941
|
-
// FIXME: Should we use a different itype? Or same?
|
942
|
-
lhs = nm_yale_storage_create(new_dtype, shape, 2, request_capacity, rhs->itype);
|
943
|
-
|
944
|
-
// This check probably isn't necessary.
|
945
|
-
if (lhs->capacity < request_capacity)
|
946
|
-
rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %ld requested, max allowable is %ld", request_capacity, lhs->capacity);
|
947
|
-
|
948
|
-
slice_copy<LDType, RDType, IType>(lhs, rhs, rhs->offset, rhs->shape, new_dtype);
|
949
|
-
} else { // regular copy
|
950
|
-
|
951
|
-
// Allocate a new structure
|
952
|
-
size_t size = get_size<IType>(rhs);
|
953
|
-
lhs = copy_alloc_struct<IType>(rhs, new_dtype, rhs->capacity, size);
|
954
|
-
|
955
|
-
LDType* la = reinterpret_cast<LDType*>(lhs->a);
|
956
|
-
RDType* ra = reinterpret_cast<RDType*>(rhs->a);
|
957
|
-
|
958
|
-
for (size_t index = 0; index < size; ++index) {
|
959
|
-
la[index] = ra[index];
|
960
|
-
}
|
961
|
-
}
|
962
|
-
|
963
|
-
return lhs;
|
964
|
-
}
|
965
|
-
|
966
|
-
/*
|
967
|
-
* Template access for getting the size of Yale storage.
|
968
|
-
*/
|
969
|
-
template <typename IType>
|
970
|
-
static inline size_t get_size(const YALE_STORAGE* storage) {
|
971
|
-
return static_cast<size_t>(reinterpret_cast<IType*>(storage->ija)[ storage->shape[0] ]);
|
972
|
-
}
|
973
|
-
|
974
|
-
|
975
|
-
/*
|
976
|
-
* Allocate for a copy or copy-cast operation, and copy the IJA portion of the
|
977
|
-
* matrix (the structure).
|
978
|
-
*/
|
979
|
-
template <typename IType>
|
980
|
-
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) {
|
981
|
-
YALE_STORAGE* lhs = ALLOC( YALE_STORAGE );
|
982
|
-
lhs->dim = rhs->dim;
|
983
|
-
lhs->shape = ALLOC_N( size_t, lhs->dim );
|
984
|
-
lhs->offset = ALLOC_N( size_t, lhs->dim );
|
985
|
-
memcpy(lhs->shape, rhs->shape, lhs->dim * sizeof(size_t));
|
986
|
-
//memcpy(lhs->offset, rhs->offset, lhs->dim * sizeof(size_t));
|
987
|
-
lhs->offset[0] = 0;
|
988
|
-
lhs->offset[1] = 0;
|
989
|
-
|
990
|
-
lhs->itype = rhs->itype;
|
991
|
-
lhs->capacity = new_capacity;
|
992
|
-
lhs->dtype = new_dtype;
|
993
|
-
lhs->ndnz = rhs->ndnz;
|
994
|
-
|
995
|
-
lhs->ija = ALLOC_N( IType, lhs->capacity );
|
996
|
-
lhs->a = ALLOC_N( char, DTYPE_SIZES[new_dtype] * lhs->capacity );
|
997
|
-
lhs->src = lhs;
|
998
|
-
lhs->count = 1;
|
999
|
-
|
1000
|
-
// Now copy the contents -- but only within the boundaries set by the size. Leave
|
1001
|
-
// the rest uninitialized.
|
1002
|
-
if (!rhs->offset[0] && !rhs->offset[1]) {
|
1003
|
-
for (size_t i = 0; i < get_size<IType>(rhs); ++i)
|
1004
|
-
reinterpret_cast<IType*>(lhs->ija)[i] = reinterpret_cast<IType*>(rhs->ija)[i]; // copy indices
|
1005
|
-
} else {
|
1006
|
-
rb_raise(rb_eNotImpError, "cannot copy struct due to different offsets");
|
1007
|
-
}
|
1008
|
-
return lhs;
|
1009
|
-
}
|
1010
|
-
|
1011
|
-
template <typename DType, typename IType>
|
1012
|
-
static STORAGE* matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector, nm::itype_t result_itype) {
|
1013
|
-
YALE_STORAGE *left = (YALE_STORAGE*)(casted_storage.left),
|
1014
|
-
*right = (YALE_STORAGE*)(casted_storage.right);
|
1015
|
-
|
1016
|
-
// We can safely get dtype from the casted matrices; post-condition of binary_storage_cast_alloc is that dtype is the
|
1017
|
-
// same for left and right.
|
1018
|
-
// int8_t dtype = left->dtype;
|
1019
|
-
|
1020
|
-
// Massage the IType arrays into the correct form.
|
1021
|
-
|
1022
|
-
IType* ijl;
|
1023
|
-
if (left->itype == result_itype) ijl = reinterpret_cast<IType*>(left->ija);
|
1024
|
-
else { // make a temporary copy of the IJA vector for L with the correct itype
|
1025
|
-
size_t length = nm_yale_storage_get_size(left);
|
1026
|
-
ijl = ALLOCA_N(IType, length);
|
1027
|
-
copy_recast_itype_vector(reinterpret_cast<void*>(left->ija), left->itype, reinterpret_cast<void*>(ijl), result_itype, length);
|
1028
|
-
}
|
1029
|
-
|
1030
|
-
IType* ijr;
|
1031
|
-
if (right->itype == result_itype) ijr = reinterpret_cast<IType*>(right->ija);
|
1032
|
-
else { // make a temporary copy of the IJA vector for R with the correct itype
|
1033
|
-
size_t length = nm_yale_storage_get_size(right);
|
1034
|
-
ijr = ALLOCA_N(IType, length);
|
1035
|
-
copy_recast_itype_vector(reinterpret_cast<void*>(right->ija), right->itype, reinterpret_cast<void*>(ijr), result_itype, length);
|
1036
|
-
}
|
1037
|
-
|
1038
|
-
// First, count the ndnz of the result.
|
1039
|
-
// TODO: This basically requires running symbmm twice to get the exact ndnz size. That's frustrating. Are there simple
|
1040
|
-
// cases where we can avoid running it?
|
1041
|
-
size_t result_ndnz = nm::math::symbmm<IType>(resulting_shape[0], left->shape[1], resulting_shape[1], ijl, ijl, true, ijr, ijr, true, NULL, true);
|
1042
|
-
|
1043
|
-
// Create result storage.
|
1044
|
-
YALE_STORAGE* result = nm_yale_storage_create(left->dtype, resulting_shape, 2, result_ndnz, result_itype);
|
1045
|
-
init<DType,IType>(result, NULL);
|
1046
|
-
IType* ija = reinterpret_cast<IType*>(result->ija);
|
1047
|
-
|
1048
|
-
// Symbolic multiplication step (build the structure)
|
1049
|
-
nm::math::symbmm<IType>(resulting_shape[0], left->shape[1], resulting_shape[1], ijl, ijl, true, ijr, ijr, true, ija, true);
|
1050
|
-
|
1051
|
-
// Numeric multiplication step (fill in the elements)
|
1052
|
-
|
1053
|
-
nm::math::numbmm<DType,IType>(result->shape[0], left->shape[1], result->shape[1],
|
1054
|
-
ijl, ijl, reinterpret_cast<DType*>(left->a), true,
|
1055
|
-
ijr, ijr, reinterpret_cast<DType*>(right->a), true,
|
1056
|
-
ija, ija, reinterpret_cast<DType*>(result->a), true);
|
1057
|
-
|
1058
|
-
|
1059
|
-
// Sort the columns
|
1060
|
-
nm::math::smmp_sort_columns<DType,IType>(result->shape[0], ija, ija, reinterpret_cast<DType*>(result->a));
|
1061
|
-
|
1062
|
-
return reinterpret_cast<STORAGE*>(result);
|
1063
|
-
}
|
1064
|
-
|
1065
|
-
|
1066
|
-
/*
|
1067
|
-
* Get the sum of offsets from the original matrix (for sliced iteration).
|
1068
|
-
*/
|
1069
|
-
static std::array<size_t,2> get_offsets(YALE_STORAGE* x) {
|
1070
|
-
std::array<size_t, 2> offsets{ {0,0} };
|
1071
|
-
while (x != x->src) {
|
1072
|
-
offsets[0] += x->offset[0];
|
1073
|
-
offsets[1] += x->offset[1];
|
1074
|
-
x = reinterpret_cast<YALE_STORAGE*>(x->src);
|
1075
|
-
}
|
1076
|
-
return offsets;
|
1077
|
-
}
|
1078
|
-
|
1079
|
-
|
1080
|
-
template <typename IType>
|
1081
|
-
class IJAManager {
|
1082
|
-
protected:
|
1083
|
-
bool needs_free;
|
1084
|
-
|
1085
|
-
public:
|
1086
|
-
IType* ija;
|
1087
|
-
|
1088
|
-
IJAManager(YALE_STORAGE* s, itype_t temp_itype) : needs_free(false), ija(reinterpret_cast<IType*>(s->ija)) {
|
1089
|
-
if (s->itype != temp_itype) {
|
1090
|
-
size_t len = nm_yale_storage_get_size(s);
|
1091
|
-
needs_free = true;
|
1092
|
-
ija = ALLOC_N(IType, len);
|
1093
|
-
copy_recast_itype_vector(s->ija, s->itype, reinterpret_cast<void*>(ija), temp_itype, len);
|
1094
|
-
}
|
1095
|
-
}
|
1096
|
-
|
1097
|
-
~IJAManager() {
|
1098
|
-
if (needs_free) xfree(ija);
|
1099
|
-
}
|
1100
|
-
};
|
1101
|
-
|
1102
|
-
|
1103
|
-
template <typename IType>
|
1104
|
-
class RowIterator {
|
1105
|
-
protected:
|
1106
|
-
YALE_STORAGE* s;
|
1107
|
-
IType* ija;
|
1108
|
-
void* a;
|
1109
|
-
IType i, k, k_end;
|
1110
|
-
size_t j_offset, j_shape;
|
1111
|
-
bool diag, End;
|
1112
|
-
VALUE init;
|
1113
|
-
public:
|
1114
|
-
RowIterator(YALE_STORAGE* s_, IType* ija_, IType i_, size_t j_shape_, size_t j_offset_ = 0)
|
1115
|
-
: s(s_),
|
1116
|
-
ija(ija_),
|
1117
|
-
a(s->a),
|
1118
|
-
i(i_),
|
1119
|
-
k(ija[i]),
|
1120
|
-
k_end(ija[i+1]),
|
1121
|
-
j_offset(j_offset_),
|
1122
|
-
j_shape(j_shape_),
|
1123
|
-
diag(row_has_no_nd() || diag_is_first()),
|
1124
|
-
End(false),
|
1125
|
-
init(default_value(s))
|
1126
|
-
{ }
|
1127
|
-
|
1128
|
-
RowIterator(YALE_STORAGE* s_, IType i_, size_t j_shape_, size_t j_offset_ = 0)
|
1129
|
-
: s(s_),
|
1130
|
-
ija(reinterpret_cast<IType*>(s->ija)),
|
1131
|
-
a(s->a),
|
1132
|
-
i(i_),
|
1133
|
-
k(ija[i]),
|
1134
|
-
k_end(ija[i+1]),
|
1135
|
-
j_offset(j_offset_),
|
1136
|
-
j_shape(j_shape_),
|
1137
|
-
diag(row_has_no_nd() || diag_is_first()),
|
1138
|
-
End(false),
|
1139
|
-
init(default_value(s))
|
1140
|
-
{ }
|
1141
|
-
|
1142
|
-
RowIterator(const RowIterator& rhs) : s(rhs.s), ija(rhs.ija), a(s->a), i(rhs.i), k(rhs.k), k_end(rhs.k_end), j_offset(rhs.j_offset), j_shape(rhs.j_shape), diag(rhs.diag), End(rhs.End), init(rhs.init) { }
|
1143
|
-
|
1144
|
-
VALUE obj() const {
|
1145
|
-
return diag ? obj_at(s, i) : obj_at(s, k);
|
1146
|
-
}
|
1147
|
-
|
1148
|
-
template <typename T>
|
1149
|
-
T cobj() const {
|
1150
|
-
if (typeid(T) == typeid(RubyObject)) return obj();
|
1151
|
-
return diag ? reinterpret_cast<T*>(s->a)[i] : reinterpret_cast<T*>(s->a)[k];
|
1152
|
-
}
|
1153
|
-
|
1154
|
-
inline IType proper_j() const {
|
1155
|
-
return diag ? i : ija[k];
|
1156
|
-
}
|
1157
|
-
|
1158
|
-
inline IType offset_j() const {
|
1159
|
-
return proper_j() - j_offset;
|
1160
|
-
}
|
1161
|
-
|
1162
|
-
/* Returns true if an additional value is inserted, false if it goes on the diagonal */
|
1163
|
-
bool insert(IType j, VALUE v) {
|
1164
|
-
if (j == i) { // insert regardless on diagonal
|
1165
|
-
reinterpret_cast<VALUE*>(a)[j] = v;
|
1166
|
-
return false;
|
1167
|
-
|
1168
|
-
} else {
|
1169
|
-
if (rb_funcall(v, rb_intern("!="), 1, init) == Qtrue) {
|
1170
|
-
if (k >= s->capacity) {
|
1171
|
-
vector_grow(s);
|
1172
|
-
ija = reinterpret_cast<IType*>(s->ija);
|
1173
|
-
a = s->a;
|
1174
|
-
}
|
1175
|
-
reinterpret_cast<VALUE*>(a)[k] = v;
|
1176
|
-
ija[k] = j;
|
1177
|
-
k++;
|
1178
|
-
return true;
|
1179
|
-
}
|
1180
|
-
return false;
|
1181
|
-
}
|
1182
|
-
}
|
1183
|
-
|
1184
|
-
void update_row_end() {
|
1185
|
-
ija[i+1] = k;
|
1186
|
-
k_end = k;
|
1187
|
-
}
|
1188
|
-
|
1189
|
-
/* Past the j_shape? */
|
1190
|
-
inline bool end() const {
|
1191
|
-
if (End) return true;
|
1192
|
-
//if (diag) return i - j_offset >= j_shape;
|
1193
|
-
//else return k >= s->capacity || ija[k] - j_offset >= j_shape;
|
1194
|
-
return (diag ? i : ija[k]) - j_offset >= j_shape;
|
1195
|
-
}
|
1196
|
-
|
1197
|
-
inline bool row_has_no_nd() const { return ija[i] == k_end; /* k_start == k_end */ }
|
1198
|
-
inline bool diag_is_first() const { return i < ija[ija[i]]; }
|
1199
|
-
inline bool diag_is_last() const { return i > ija[k_end-1]; } // only works if !row_has_no_nd()
|
1200
|
-
inline bool k_is_last_nd() const { return k == k_end-1; }
|
1201
|
-
inline bool k_is_last() const { return k_is_last_nd() && !diag_is_last(); }
|
1202
|
-
inline bool diag_is_ahead() const { return i > ija[k]; }
|
1203
|
-
inline bool row_has_diag() const { return i < s->shape[1]; }
|
1204
|
-
inline bool diag_is_next() const { // assumes we've already tested for diag, row_has_no_nd(), diag_is_first()
|
1205
|
-
if (i == ija[k]+1) return true; // definite next
|
1206
|
-
else if (k+1 < k_end && i >= ija[k+1]+1) return false; // at least one item before it
|
1207
|
-
else return true;
|
1208
|
-
}
|
1209
|
-
|
1210
|
-
RowIterator<IType>& operator++() {
|
1211
|
-
if (diag) { // we're at the diagonal
|
1212
|
-
if (row_has_no_nd() || diag_is_last()) End = true; // and there are no non-diagonals (or none still to visit)
|
1213
|
-
diag = false;
|
1214
|
-
} else if (!row_has_diag()) { // row has no diagonal entries
|
1215
|
-
if (row_has_no_nd() || k_is_last_nd()) End = true; // row is totally empty, or we're at last entry
|
1216
|
-
else k++; // still entries to visit
|
1217
|
-
} else { // not at diag but it exists somewhere in the row, and row has at least one nd entry
|
1218
|
-
if (diag_is_ahead()) { // diag is ahead
|
1219
|
-
if (k_is_last_nd()) diag = true; // diag is next and last
|
1220
|
-
else if (diag_is_next()) { // diag is next and not last
|
1221
|
-
diag = true;
|
1222
|
-
k++;
|
1223
|
-
} else k++; // diag is not next
|
1224
|
-
} else { // diag is past
|
1225
|
-
if (k_is_last_nd()) End = true; // and we're at the end
|
1226
|
-
else k++; // and we're not at the end
|
1227
|
-
}
|
1228
|
-
}
|
1229
|
-
|
1230
|
-
return *this;
|
1231
|
-
}
|
1232
|
-
|
1233
|
-
|
1234
|
-
RowIterator<IType> operator++(int unused) {
|
1235
|
-
RowIterator<IType> x(*this);
|
1236
|
-
++(*this);
|
1237
|
-
return x;
|
1238
|
-
}
|
1239
|
-
};
|
1240
|
-
|
1241
|
-
|
1242
|
-
|
1243
|
-
template <typename IType>
|
1244
|
-
static VALUE map_stored(VALUE self) {
|
1245
|
-
|
1246
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1247
|
-
|
1248
|
-
size_t* shape = ALLOC_N(size_t, 2);
|
1249
|
-
shape[0] = s->shape[0];
|
1250
|
-
shape[1] = s->shape[1];
|
1251
|
-
|
1252
|
-
std::array<size_t,2> s_offsets = get_offsets(s);
|
1253
|
-
|
1254
|
-
RETURN_SIZED_ENUMERATOR(self, 0, 0, nm_yale_enumerator_length);
|
1255
|
-
VALUE init = rb_yield(default_value(s));
|
1256
|
-
|
1257
|
-
// Try to find a reasonable capacity to request when creating the matrix
|
1258
|
-
size_t ndnz = src_ndnz(s);
|
1259
|
-
if (s->src != s) // need to guess capacity
|
1260
|
-
ndnz = yale_count_slice_copy_ndnz(s, s->offset, s->shape);
|
1261
|
-
size_t request_capacity = s->shape[0] + ndnz + 1;
|
1262
|
-
|
1263
|
-
YALE_STORAGE* r = nm_yale_storage_create(nm::RUBYOBJ, shape, 2, request_capacity, NM_ITYPE(self));
|
1264
|
-
if (r->capacity < request_capacity)
|
1265
|
-
rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %ld requested, max allowable is %ld", request_capacity, r->capacity);
|
1266
|
-
nm_yale_storage_init(r, &init);
|
1267
|
-
|
1268
|
-
for (IType ri = 0; ri < shape[0]; ++ri) {
|
1269
|
-
RowIterator<IType> sit(s, ri + s_offsets[0], shape[1], s_offsets[1]);
|
1270
|
-
RowIterator<IType> rit(r, ri, shape[1]);
|
1271
|
-
|
1272
|
-
while (!sit.end()) {
|
1273
|
-
VALUE rv = rb_yield(sit.obj());
|
1274
|
-
VALUE rj = sit.offset_j();
|
1275
|
-
rit.insert(rj, rv);
|
1276
|
-
++sit;
|
1277
|
-
}
|
1278
|
-
// Update the row end information.
|
1279
|
-
rit.update_row_end();
|
1280
|
-
}
|
1281
|
-
|
1282
|
-
NMATRIX* m = nm_create(nm::YALE_STORE, reinterpret_cast<STORAGE*>(r));
|
1283
|
-
return Data_Wrap_Struct(CLASS_OF(self), nm_yale_storage_mark, nm_delete, m);
|
1284
|
-
}
|
1285
|
-
|
1286
|
-
|
1287
|
-
/*
|
1288
|
-
* eqeq function for slicing and different defaults.
|
1289
|
-
*/
|
1290
|
-
template <typename LDType, typename RDType, typename IType>
|
1291
|
-
static bool eqeq_different_defaults(const YALE_STORAGE* s, const LDType& s_init, const YALE_STORAGE* t, const RDType& t_init) {
|
1292
|
-
|
1293
|
-
std::array<size_t,2> s_offsets = get_offsets(const_cast<YALE_STORAGE*>(s)),
|
1294
|
-
t_offsets = get_offsets(const_cast<YALE_STORAGE*>(t));
|
1295
|
-
|
1296
|
-
for (IType ri = 0; ri < s->shape[0]; ++ri) {
|
1297
|
-
RowIterator<IType> sit(const_cast<YALE_STORAGE*>(s), reinterpret_cast<IType*>(s->ija), ri + s_offsets[0], s->shape[1], s_offsets[1]);
|
1298
|
-
RowIterator<IType> tit(const_cast<YALE_STORAGE*>(t), reinterpret_cast<IType*>(t->ija), ri + t_offsets[0], s->shape[1], t_offsets[1]);
|
1299
|
-
|
1300
|
-
while (!sit.end() || !tit.end()) {
|
1301
|
-
|
1302
|
-
// Perform the computation. Use a default value if the matrix doesn't have some value stored.
|
1303
|
-
if (tit.end() || (!sit.end() && sit.offset_j() < tit.offset_j())) {
|
1304
|
-
if (sit.template cobj<LDType>() != t_init) return false;
|
1305
|
-
++sit;
|
1306
|
-
|
1307
|
-
} else if (sit.end() || (!tit.end() && sit.offset_j() > tit.offset_j())) {
|
1308
|
-
if (s_init != tit.template cobj<RDType>()) return false;
|
1309
|
-
++tit;
|
1310
|
-
|
1311
|
-
} else { // same index
|
1312
|
-
if (sit.template cobj<LDType>() != tit.template cobj<RDType>()) return false;
|
1313
|
-
++sit;
|
1314
|
-
++tit;
|
1315
|
-
}
|
1316
|
-
}
|
1317
|
-
}
|
1318
|
-
return true;
|
1319
|
-
}
|
1320
|
-
|
1321
|
-
|
1322
|
-
template <typename IType>
|
1323
|
-
static VALUE map_merged_stored(VALUE left, VALUE right, VALUE init, nm::itype_t itype) {
|
1324
|
-
|
1325
|
-
YALE_STORAGE *s = NM_STORAGE_YALE(left),
|
1326
|
-
*t = NM_STORAGE_YALE(right);
|
1327
|
-
|
1328
|
-
size_t* shape = ALLOC_N(size_t, 2);
|
1329
|
-
shape[0] = s->shape[0];
|
1330
|
-
shape[1] = s->shape[1];
|
1331
|
-
|
1332
|
-
std::array<size_t,2> s_offsets = get_offsets(s),
|
1333
|
-
t_offsets = get_offsets(t);
|
1334
|
-
|
1335
|
-
VALUE s_init = default_value(s),
|
1336
|
-
t_init = default_value(t);
|
1337
|
-
|
1338
|
-
RETURN_SIZED_ENUMERATOR(left, 0, 0, 0);
|
1339
|
-
|
1340
|
-
if (init == Qnil)
|
1341
|
-
init = rb_yield_values(2, s_init, t_init);
|
1342
|
-
|
1343
|
-
// Make a reasonable approximation of the resulting capacity
|
1344
|
-
size_t s_ndnz = src_ndnz(s), t_ndnz = src_ndnz(t);
|
1345
|
-
if (s->src != s) s_ndnz = yale_count_slice_copy_ndnz(s, s->offset, s->shape);
|
1346
|
-
if (t->src != t) t_ndnz = yale_count_slice_copy_ndnz(t, t->offset, t->shape);
|
1347
|
-
size_t request_capacity = shape[0] + NM_MAX(s_ndnz, t_ndnz) + 1;
|
1348
|
-
|
1349
|
-
YALE_STORAGE* r = nm_yale_storage_create(nm::RUBYOBJ, shape, 2, request_capacity, itype);
|
1350
|
-
if (r->capacity < request_capacity)
|
1351
|
-
rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %ld requested, max allowable is %ld", request_capacity, r->capacity);
|
1352
|
-
|
1353
|
-
nm_yale_storage_init(r, &init);
|
1354
|
-
|
1355
|
-
IJAManager<IType> sm(s, itype),
|
1356
|
-
tm(t, itype);
|
1357
|
-
|
1358
|
-
for (IType ri = 0; ri < shape[0]; ++ri) {
|
1359
|
-
RowIterator<IType> sit(s, sm.ija, ri + s_offsets[0], shape[1], s_offsets[1]);
|
1360
|
-
RowIterator<IType> tit(t, tm.ija, ri + t_offsets[0], shape[1], t_offsets[1]);
|
1361
|
-
|
1362
|
-
RowIterator<IType> rit(r, reinterpret_cast<IType*>(r->ija), ri, shape[1]);
|
1363
|
-
while (!sit.end() || !tit.end()) {
|
1364
|
-
VALUE rv;
|
1365
|
-
IType rj;
|
1366
|
-
|
1367
|
-
// Perform the computation. Use a default value if the matrix doesn't have some value stored.
|
1368
|
-
if (tit.end() || (!sit.end() && sit.offset_j() < tit.offset_j())) {
|
1369
|
-
rv = rb_yield_values(2, sit.obj(), t_init);
|
1370
|
-
rj = sit.offset_j();
|
1371
|
-
++sit;
|
1372
|
-
|
1373
|
-
} else if (sit.end() || (!tit.end() && sit.offset_j() > tit.offset_j())) {
|
1374
|
-
rv = rb_yield_values(2, s_init, tit.obj());
|
1375
|
-
rj = tit.offset_j();
|
1376
|
-
++tit;
|
1377
|
-
|
1378
|
-
} else { // same index
|
1379
|
-
rv = rb_yield_values(2, sit.obj(), tit.obj());
|
1380
|
-
rj = sit.offset_j();
|
1381
|
-
++sit;
|
1382
|
-
++tit;
|
1383
|
-
}
|
1384
|
-
|
1385
|
-
rit.insert(rj, rv); // handles increment (and testing for default, etc)
|
1386
|
-
|
1387
|
-
}
|
1388
|
-
|
1389
|
-
// Update the row end information.
|
1390
|
-
rit.update_row_end();
|
1391
|
-
}
|
1392
|
-
|
1393
|
-
NMATRIX* m = nm_create(nm::YALE_STORE, reinterpret_cast<STORAGE*>(r));
|
1394
|
-
return Data_Wrap_Struct(CLASS_OF(left), nm_yale_storage_mark, nm_delete, m);
|
1395
|
-
}
|
1396
|
-
|
1397
|
-
|
1398
|
-
/*
|
1399
|
-
* This function and the two helper structs enable us to use partial template specialization.
|
1400
|
-
* See also: http://stackoverflow.com/questions/6623375/c-template-specialization-on-functions
|
1401
|
-
*/
|
1402
|
-
template <typename DType, typename IType>
|
1403
|
-
static VALUE each_stored_with_indices(VALUE nm) {
|
1404
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1405
|
-
DType* a = reinterpret_cast<DType*>(s->a);
|
1406
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
1407
|
-
|
1408
|
-
// If we don't have a block, return an enumerator.
|
1409
|
-
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_stored_enumerator_length);
|
1410
|
-
|
1411
|
-
// Iterate along diagonal
|
1412
|
-
for (size_t sk = NM_MAX(s->offset[0], s->offset[1]); sk < NM_MIN(s->shape[0] + s->offset[0], s->shape[1] + s->offset[1]); ++sk) {
|
1413
|
-
VALUE ii = LONG2NUM(sk - s->offset[0]),
|
1414
|
-
jj = LONG2NUM(sk - s->offset[1]);
|
1415
|
-
|
1416
|
-
rb_yield_values(3, obj_at(s, sk), ii, jj);
|
1417
|
-
}
|
1418
|
-
|
1419
|
-
// Iterate through non-diagonal elements, row by row
|
1420
|
-
for (long ri = 0; ri < s->shape[0]; ++ri) {
|
1421
|
-
long si = ri + s->offset[0];
|
1422
|
-
IType p = ija[si],
|
1423
|
-
next_p = ija[si+1];
|
1424
|
-
|
1425
|
-
// if this is a reference to another matrix, we should find the left boundary of the slice
|
1426
|
-
if (s != s->src && p < next_p)
|
1427
|
-
p = binary_search_left_boundary<IType>(s, p, next_p-1, s->offset[1]);
|
1428
|
-
|
1429
|
-
for (; p < next_p; ++p) {
|
1430
|
-
long sj = static_cast<long>(ija[p]),
|
1431
|
-
rj = sj - s->offset[1];
|
1432
|
-
if (rj < 0) continue;
|
1433
|
-
|
1434
|
-
if (rj >= s->shape[1]) break;
|
1435
|
-
|
1436
|
-
rb_yield_values(3, obj_at(s, p), LONG2NUM(ri), LONG2NUM(rj));
|
1437
|
-
}
|
1438
|
-
}
|
1439
|
-
|
1440
|
-
return nm;
|
1441
|
-
}
|
1442
|
-
|
1443
|
-
template <typename DType, typename IType>
|
1444
|
-
static VALUE each_with_indices(VALUE nm) {
|
1445
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1446
|
-
DType* a = reinterpret_cast<DType*>(s->a);
|
1447
|
-
IType* ija = reinterpret_cast<IType*>(s->ija);
|
1448
|
-
|
1449
|
-
// If we don't have a block, return an enumerator.
|
1450
|
-
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_enumerator_length);
|
1451
|
-
|
1452
|
-
// Iterate in two dimensions.
|
1453
|
-
// s stands for src, r stands for ref (for ri, rj, si, sj)
|
1454
|
-
for (long ri = 0; ri < s->shape[0]; ++ri) {
|
1455
|
-
long si = ri + s->offset[0];
|
1456
|
-
VALUE ii = LONG2NUM(ri + s->offset[0]);
|
1457
|
-
|
1458
|
-
IType k = ija[si], k_next = ija[si+1];
|
1459
|
-
|
1460
|
-
for (long rj = 0; rj < s->shape[1]; ++rj) {
|
1461
|
-
long sj = rj + s->offset[1];
|
1462
|
-
VALUE v, jj = LONG2NUM(rj);
|
1463
|
-
|
1464
|
-
// zero is stored in s->shape[0]
|
1465
|
-
if (si == sj) {
|
1466
|
-
v = obj_at(s, si);
|
1467
|
-
} else {
|
1468
|
-
// Walk through the row until we find the correct location.
|
1469
|
-
while (ija[k] < sj && k < k_next) ++k;
|
1470
|
-
if (k < k_next && ija[k] == sj) {
|
1471
|
-
v = obj_at(s, k);
|
1472
|
-
++k;
|
1473
|
-
} else v = default_value(s); // rubyobj_from_cval(&(a[s->shape[0]]), NM_DTYPE(nm)).rval;
|
1474
|
-
}
|
1475
|
-
rb_yield_values(3, v, ii, jj);
|
1476
|
-
}
|
1477
|
-
}
|
1478
|
-
|
1479
|
-
return nm;
|
1480
|
-
}
|
1481
|
-
|
1482
|
-
|
1483
|
-
} // end of namespace nm::yale_storage
|
1484
|
-
|
1485
|
-
|
1486
|
-
// Helper function used only for the RETURN_SIZED_ENUMERATOR macro. Returns the length of
|
1487
|
-
// the matrix's storage.
|
1488
|
-
static VALUE nm_yale_stored_enumerator_length(VALUE nmatrix) {
|
1489
|
-
long len = nm_yale_storage_get_size(NM_STORAGE_YALE(nmatrix));
|
1490
|
-
return LONG2NUM(len);
|
1491
|
-
}
|
1492
|
-
|
1493
|
-
|
1494
|
-
} // end of namespace nm.
|
1495
|
-
|
1496
|
-
///////////////////
|
1497
|
-
// Ruby Bindings //
|
1498
|
-
///////////////////
|
1499
|
-
|
1500
|
-
/* These bindings are mostly only for debugging Yale. They are called from Init_nmatrix. */
|
1501
|
-
|
1502
|
-
extern "C" {
|
1503
|
-
|
1504
|
-
void nm_init_yale_functions() {
|
1505
|
-
/*
|
1506
|
-
* This module stores methods that are useful for debugging Yale matrices,
|
1507
|
-
* i.e. the ones with +:yale+ stype.
|
1508
|
-
*/
|
1509
|
-
cNMatrix_YaleFunctions = rb_define_module_under(cNMatrix, "YaleFunctions");
|
1510
|
-
|
1511
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_ija", (METHOD)nm_ija, -1);
|
1512
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_a", (METHOD)nm_a, -1);
|
1513
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_size", (METHOD)nm_size, 0);
|
1514
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_ia", (METHOD)nm_ia, 0);
|
1515
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_ja", (METHOD)nm_ja, 0);
|
1516
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_d", (METHOD)nm_d, -1);
|
1517
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_lu", (METHOD)nm_lu, 0);
|
1518
|
-
|
1519
|
-
rb_define_method(cNMatrix_YaleFunctions, "yale_nd_row", (METHOD)nm_nd_row, -1);
|
1520
|
-
|
1521
|
-
rb_define_const(cNMatrix_YaleFunctions, "YALE_GROWTH_CONSTANT", rb_float_new(nm::yale_storage::GROWTH_CONSTANT));
|
1522
|
-
}
|
1523
|
-
|
1524
|
-
|
1525
|
-
/////////////////
|
1526
|
-
// C ACCESSORS //
|
1527
|
-
/////////////////
|
1528
|
-
|
1529
|
-
|
1530
|
-
/* C interface for NMatrix#each_with_indices (Yale) */
|
1531
|
-
VALUE nm_yale_each_with_indices(VALUE nmatrix) {
|
1532
|
-
nm::dtype_t d = NM_DTYPE(nmatrix);
|
1533
|
-
nm::itype_t i = NM_ITYPE(nmatrix);
|
1534
|
-
|
1535
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::each_with_indices, VALUE, VALUE)
|
1536
|
-
|
1537
|
-
return ttable[d][i](nmatrix);
|
1538
|
-
}
|
1539
|
-
|
1540
|
-
|
1541
|
-
/* C interface for NMatrix#each_stored_with_indices (Yale) */
|
1542
|
-
VALUE nm_yale_each_stored_with_indices(VALUE nmatrix) {
|
1543
|
-
nm::dtype_t d = NM_DTYPE(nmatrix);
|
1544
|
-
nm::itype_t i = NM_ITYPE(nmatrix);
|
1545
|
-
|
1546
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::each_stored_with_indices, VALUE, VALUE)
|
1547
|
-
|
1548
|
-
return ttable[d][i](nmatrix);
|
1549
|
-
}
|
1550
|
-
|
1551
|
-
|
1552
|
-
|
1553
|
-
/*
|
1554
|
-
* C accessor for inserting some value in a matrix (or replacing an existing cell).
|
1555
|
-
*/
|
1556
|
-
char nm_yale_storage_set(STORAGE* storage, SLICE* slice, void* v) {
|
1557
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::set, char, YALE_STORAGE* storage, SLICE* slice, void* value);
|
1558
|
-
|
1559
|
-
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1560
|
-
|
1561
|
-
return ttable[casted_storage->dtype][casted_storage->itype](casted_storage, slice, v);
|
1562
|
-
}
|
1563
|
-
|
1564
|
-
|
1565
|
-
/*
|
1566
|
-
* Determine the number of non-diagonal non-zeros in a not-yet-created copy of a slice or matrix.
|
1567
|
-
*/
|
1568
|
-
static size_t yale_count_slice_copy_ndnz(const YALE_STORAGE* s, size_t* offset, size_t* shape) {
|
1569
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::count_slice_copy_ndnz, size_t, const YALE_STORAGE*, size_t*, size_t*)
|
1570
|
-
|
1571
|
-
return ttable[s->dtype][s->itype](s, offset, shape);
|
1572
|
-
}
|
1573
|
-
|
1574
|
-
|
1575
|
-
/*
|
1576
|
-
* C accessor for yale_storage::get, which returns a slice of YALE_STORAGE object by copy
|
1577
|
-
*
|
1578
|
-
* Slicing-related.
|
1579
|
-
*/
|
1580
|
-
void* nm_yale_storage_get(STORAGE* storage, SLICE* slice) {
|
1581
|
-
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1582
|
-
|
1583
|
-
if (slice->single) {
|
1584
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(elem_copy_table, nm::yale_storage::get_single, void*, YALE_STORAGE*, SLICE*)
|
1585
|
-
|
1586
|
-
return elem_copy_table[casted_storage->dtype][casted_storage->itype](casted_storage, slice);
|
1587
|
-
} else {
|
1588
|
-
// Copy shape for yale construction
|
1589
|
-
size_t* shape = ALLOC_N(size_t, 2);
|
1590
|
-
shape[0] = slice->lengths[0];
|
1591
|
-
shape[1] = slice->lengths[1];
|
1592
|
-
|
1593
|
-
// only count ndnz if our slice is smaller, otherwise use the given value
|
1594
|
-
size_t ndnz = src_ndnz(casted_storage);
|
1595
|
-
if (shape[0] != casted_storage->shape[0] || shape[1] != casted_storage->shape[1])
|
1596
|
-
ndnz = yale_count_slice_copy_ndnz(casted_storage, slice->coords, shape); // expensive operation
|
1597
|
-
|
1598
|
-
size_t request_capacity = shape[0] + ndnz + 1; // capacity of new matrix
|
1599
|
-
YALE_STORAGE* ns = nm_yale_storage_create(casted_storage->dtype, shape, 2, request_capacity, casted_storage->itype);
|
1600
|
-
|
1601
|
-
// This check probably isn't necessary.
|
1602
|
-
if (ns->capacity < request_capacity)
|
1603
|
-
rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %ld requested, max allowable is %ld", request_capacity, ns->capacity);
|
1604
|
-
|
1605
|
-
NAMED_LRI_DTYPE_TEMPLATE_TABLE(slice_copy_table, nm::yale_storage::slice_copy, void, YALE_STORAGE* ns, const YALE_STORAGE* s, size_t*, size_t*, nm::dtype_t)
|
1606
|
-
|
1607
|
-
slice_copy_table[ns->dtype][casted_storage->dtype][casted_storage->itype](ns, casted_storage, slice->coords, slice->lengths, casted_storage->dtype);
|
1608
|
-
|
1609
|
-
return ns;
|
1610
|
-
}
|
1611
|
-
}
|
1612
|
-
|
1613
|
-
/*
|
1614
|
-
* C accessor for yale_storage::vector_insert
|
1615
|
-
*/
|
1616
|
-
static char nm_yale_storage_vector_insert(YALE_STORAGE* s, size_t pos, size_t* js, void* vals, size_t n, bool struct_only, nm::dtype_t dtype, nm::itype_t itype) {
|
1617
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::vector_insert, char, YALE_STORAGE*, size_t, size_t*, void*, size_t, bool);
|
1618
|
-
|
1619
|
-
return ttable[dtype][itype](s, pos, js, vals, n, struct_only);
|
1620
|
-
}
|
1621
|
-
|
1622
|
-
/*
|
1623
|
-
* C accessor for yale_storage::increment_ia_after, typically called after ::vector_insert
|
1624
|
-
*/
|
1625
|
-
static void nm_yale_storage_increment_ia_after(YALE_STORAGE* s, size_t ija_size, size_t i, size_t n, nm::itype_t itype) {
|
1626
|
-
NAMED_ITYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::c_increment_ia_after, void, YALE_STORAGE*, size_t, size_t, size_t);
|
1627
|
-
|
1628
|
-
ttable[itype](s, ija_size, i, n);
|
1629
|
-
}
|
1630
|
-
|
1631
|
-
|
1632
|
-
/*
|
1633
|
-
* C accessor for yale_storage::ref, which returns a pointer to the correct location in a YALE_STORAGE object
|
1634
|
-
* for some set of coordinates.
|
1635
|
-
*/
|
1636
|
-
void* nm_yale_storage_ref(STORAGE* storage, SLICE* slice) {
|
1637
|
-
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1638
|
-
|
1639
|
-
if (slice->single) {
|
1640
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(elem_copy_table, nm::yale_storage::get_single, void*, YALE_STORAGE*, SLICE*)
|
1641
|
-
return elem_copy_table[casted_storage->dtype][casted_storage->itype](casted_storage, slice);
|
1642
|
-
} else {
|
1643
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ref_table, nm::yale_storage::ref, void*, YALE_STORAGE* storage, SLICE* slice)
|
1644
|
-
return ref_table[casted_storage->dtype][casted_storage->itype](casted_storage, slice);
|
1645
|
-
}
|
1646
|
-
}
|
1647
|
-
|
1648
|
-
|
1649
|
-
/*
|
1650
|
-
* C accessor for determining whether two YALE_STORAGE objects have the same contents.
|
1651
|
-
*/
|
1652
|
-
bool nm_yale_storage_eqeq(const STORAGE* left, const STORAGE* right) {
|
1653
|
-
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::eqeq, bool, const YALE_STORAGE* left, const YALE_STORAGE* right);
|
1654
|
-
|
1655
|
-
const YALE_STORAGE* casted_left = reinterpret_cast<const YALE_STORAGE*>(left);
|
1656
|
-
|
1657
|
-
return ttable[casted_left->dtype][right->dtype][casted_left->itype](casted_left, (const YALE_STORAGE*)right);
|
1658
|
-
}
|
1659
|
-
|
1660
|
-
|
1661
|
-
/*
|
1662
|
-
* Copy constructor for changing dtypes. (C accessor)
|
1663
|
-
*/
|
1664
|
-
STORAGE* nm_yale_storage_cast_copy(const STORAGE* rhs, nm::dtype_t new_dtype, void* dummy) {
|
1665
|
-
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::cast_copy, YALE_STORAGE*, const YALE_STORAGE* rhs, nm::dtype_t new_dtype);
|
1666
|
-
|
1667
|
-
const YALE_STORAGE* casted_rhs = reinterpret_cast<const YALE_STORAGE*>(rhs);
|
1668
|
-
|
1669
|
-
return (STORAGE*)ttable[new_dtype][casted_rhs->dtype][casted_rhs->itype](casted_rhs, new_dtype);
|
1670
|
-
}
|
1671
|
-
|
1672
|
-
|
1673
|
-
/*
|
1674
|
-
* Returns size of Yale storage as a size_t (no matter what the itype is). (C accessor)
|
1675
|
-
*/
|
1676
|
-
size_t nm_yale_storage_get_size(const YALE_STORAGE* storage) {
|
1677
|
-
NAMED_ITYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::get_size, size_t, const YALE_STORAGE* storage);
|
1678
|
-
|
1679
|
-
return ttable[storage->itype](storage);
|
1680
|
-
}
|
1681
|
-
|
1682
|
-
|
1683
|
-
|
1684
|
-
/*
|
1685
|
-
* Return a void pointer to the matrix's default value entry.
|
1686
|
-
*/
|
1687
|
-
static void* default_value_ptr(const YALE_STORAGE* s) {
|
1688
|
-
return reinterpret_cast<void*>(reinterpret_cast<char*>(s->a) + (s->src->shape[0] * DTYPE_SIZES[s->dtype]));
|
1689
|
-
}
|
1690
|
-
|
1691
|
-
/*
|
1692
|
-
* Return the Ruby object at a given location in storage.
|
1693
|
-
*/
|
1694
|
-
static VALUE obj_at(YALE_STORAGE* s, size_t k) {
|
1695
|
-
if (s->dtype == nm::RUBYOBJ) return reinterpret_cast<VALUE*>(s->a)[k];
|
1696
|
-
else return rubyobj_from_cval(reinterpret_cast<void*>(reinterpret_cast<char*>(s->a) + k * DTYPE_SIZES[s->dtype]), s->dtype).rval;
|
1697
|
-
}
|
1698
|
-
|
1699
|
-
|
1700
|
-
/*
|
1701
|
-
* Return the matrix's default value as a Ruby VALUE.
|
1702
|
-
*/
|
1703
|
-
static VALUE default_value(const YALE_STORAGE* s) {
|
1704
|
-
if (s->dtype == nm::RUBYOBJ) return *reinterpret_cast<VALUE*>(default_value_ptr(s));
|
1705
|
-
else return rubyobj_from_cval(default_value_ptr(s), s->dtype).rval;
|
1706
|
-
}
|
1707
|
-
|
1708
|
-
|
1709
|
-
/*
|
1710
|
-
* Check to see if a default value is some form of zero. Easy for non-Ruby object matrices, which should always be 0.
|
1711
|
-
*/
|
1712
|
-
static bool default_value_is_numeric_zero(const YALE_STORAGE* s) {
|
1713
|
-
return rb_funcall(default_value(s), rb_intern("=="), 1, INT2FIX(0)) == Qtrue;
|
1714
|
-
}
|
1715
|
-
|
1716
|
-
|
1717
|
-
/*
|
1718
|
-
* C accessor for allocating a yale storage object for cast-copying. Copies the IJA vector, does not copy the A vector.
|
1719
|
-
*/
|
1720
|
-
static YALE_STORAGE* nm_copy_alloc_struct(const YALE_STORAGE* rhs, const nm::dtype_t new_dtype, const size_t new_capacity, const size_t new_size) {
|
1721
|
-
NAMED_ITYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::copy_alloc_struct, YALE_STORAGE*, const YALE_STORAGE* rhs, const nm::dtype_t new_dtype, const size_t new_capacity, const size_t new_size);
|
1722
|
-
|
1723
|
-
return ttable[rhs->itype](rhs, new_dtype, new_capacity, new_size);
|
1724
|
-
}
|
1725
|
-
|
1726
|
-
/*
|
1727
|
-
* Transposing copy constructor.
|
1728
|
-
*/
|
1729
|
-
STORAGE* nm_yale_storage_copy_transposed(const STORAGE* rhs_base) {
|
1730
|
-
YALE_STORAGE* rhs = (YALE_STORAGE*)rhs_base;
|
1731
|
-
|
1732
|
-
size_t* shape = ALLOC_N(size_t, 2);
|
1733
|
-
shape[0] = rhs->shape[1];
|
1734
|
-
shape[1] = rhs->shape[0];
|
1735
|
-
|
1736
|
-
size_t size = nm_yale_storage_get_size(rhs);
|
1737
|
-
|
1738
|
-
YALE_STORAGE* lhs = nm_yale_storage_create(rhs->dtype, shape, 2, size, rhs->itype);
|
1739
|
-
nm_yale_storage_init(lhs, default_value_ptr(rhs));
|
1740
|
-
|
1741
|
-
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);
|
1742
|
-
|
1743
|
-
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);
|
1744
|
-
|
1745
|
-
return (STORAGE*)lhs;
|
1746
|
-
}
|
1747
|
-
|
1748
|
-
/*
|
1749
|
-
* C accessor for multiplying two YALE_STORAGE matrices, which have already been casted to the same dtype.
|
1750
|
-
*
|
1751
|
-
* FIXME: There should be some mathematical way to determine the worst-case IType based on the input ITypes. Right now
|
1752
|
-
* it just uses the default.
|
1753
|
-
*/
|
1754
|
-
STORAGE* nm_yale_storage_matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector) {
|
1755
|
-
LI_DTYPE_TEMPLATE_TABLE(nm::yale_storage::matrix_multiply, STORAGE*, const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector, nm::itype_t resulting_itype);
|
1756
|
-
|
1757
|
-
YALE_STORAGE* left = reinterpret_cast<YALE_STORAGE*>(casted_storage.left);
|
1758
|
-
YALE_STORAGE* right = reinterpret_cast<YALE_STORAGE*>(casted_storage.right);
|
1759
|
-
|
1760
|
-
if (!default_value_is_numeric_zero(left) || !default_value_is_numeric_zero(right)) {
|
1761
|
-
rb_raise(rb_eNotImpError, "matrix default value must be some form of zero (not false or nil) for multiplication");
|
1762
|
-
return NULL;
|
1763
|
-
}
|
1764
|
-
|
1765
|
-
// Determine the itype for the matrix that will be returned.
|
1766
|
-
nm::itype_t itype = nm_yale_storage_itype_by_shape(resulting_shape),
|
1767
|
-
max_itype = NM_MAX_ITYPE(left->itype, right->itype);
|
1768
|
-
if (static_cast<int8_t>(itype) < static_cast<int8_t>(max_itype)) itype = max_itype;
|
1769
|
-
|
1770
|
-
return ttable[left->dtype][itype](casted_storage, resulting_shape, vector, itype);
|
1771
|
-
}
|
1772
|
-
|
1773
|
-
|
1774
|
-
///////////////
|
1775
|
-
// Lifecycle //
|
1776
|
-
///////////////
|
1777
|
-
|
1778
|
-
/*
|
1779
|
-
* C accessor function for creating a YALE_STORAGE object. Prior to calling this function, you MUST
|
1780
|
-
* allocate shape (should be size_t * 2) -- don't use use a regular size_t array!
|
1781
|
-
*
|
1782
|
-
* For this type, dim must always be 2. The final argument is the initial capacity with which to
|
1783
|
-
* create the storage.
|
1784
|
-
*/
|
1785
|
-
|
1786
|
-
YALE_STORAGE* nm_yale_storage_create(nm::dtype_t dtype, size_t* shape, size_t dim, size_t init_capacity, nm::itype_t min_itype) {
|
1787
|
-
YALE_STORAGE* s;
|
1788
|
-
size_t max_capacity;
|
1789
|
-
|
1790
|
-
// FIXME: This error should be handled in the nmatrix.c file.
|
1791
|
-
if (dim != 2) {
|
1792
|
-
rb_raise(rb_eNotImpError, "Can only support 2D matrices");
|
1793
|
-
}
|
1794
|
-
|
1795
|
-
s = alloc(dtype, shape, dim, min_itype);
|
1796
|
-
max_capacity = nm::yale_storage::max_size(s);
|
1797
|
-
|
1798
|
-
// Set matrix capacity (and ensure its validity)
|
1799
|
-
if (init_capacity < NM_YALE_MINIMUM(s)) {
|
1800
|
-
s->capacity = NM_YALE_MINIMUM(s);
|
1801
|
-
|
1802
|
-
} else if (init_capacity > max_capacity) {
|
1803
|
-
// Don't allow storage to be created larger than necessary
|
1804
|
-
s->capacity = max_capacity;
|
1805
|
-
|
1806
|
-
} else {
|
1807
|
-
s->capacity = init_capacity;
|
1808
|
-
|
1809
|
-
}
|
1810
|
-
|
1811
|
-
s->ija = ALLOC_N( char, ITYPE_SIZES[s->itype] * s->capacity );
|
1812
|
-
s->a = ALLOC_N( char, DTYPE_SIZES[s->dtype] * s->capacity );
|
1813
|
-
|
1814
|
-
return s;
|
1815
|
-
}
|
1816
|
-
|
1817
|
-
/*
|
1818
|
-
* Destructor for yale storage (C-accessible).
|
1819
|
-
*/
|
1820
|
-
void nm_yale_storage_delete(STORAGE* s) {
|
1821
|
-
if (s) {
|
1822
|
-
YALE_STORAGE* storage = (YALE_STORAGE*)s;
|
1823
|
-
if (storage->count-- == 1) {
|
1824
|
-
xfree(storage->shape);
|
1825
|
-
xfree(storage->offset);
|
1826
|
-
xfree(storage->ija);
|
1827
|
-
xfree(storage->a);
|
1828
|
-
xfree(storage);
|
1829
|
-
}
|
1830
|
-
}
|
1831
|
-
}
|
1832
|
-
|
1833
|
-
/*
|
1834
|
-
* Destructor for the yale storage ref
|
1835
|
-
*/
|
1836
|
-
void nm_yale_storage_delete_ref(STORAGE* s) {
|
1837
|
-
if (s) {
|
1838
|
-
YALE_STORAGE* storage = (YALE_STORAGE*)s;
|
1839
|
-
nm_yale_storage_delete( reinterpret_cast<STORAGE*>(storage->src) );
|
1840
|
-
xfree(storage->shape);
|
1841
|
-
xfree(storage->offset);
|
1842
|
-
xfree(s);
|
1843
|
-
}
|
1844
|
-
}
|
1845
|
-
|
1846
|
-
/*
|
1847
|
-
* C accessor for yale_storage::init, a templated function.
|
1848
|
-
*
|
1849
|
-
* Initializes the IJA vector of the YALE_STORAGE matrix.
|
1850
|
-
*/
|
1851
|
-
void nm_yale_storage_init(YALE_STORAGE* s, void* init_val) {
|
1852
|
-
NAMED_LI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::init, void, YALE_STORAGE*, void*);
|
1853
|
-
|
1854
|
-
ttable[s->dtype][s->itype](s, init_val);
|
1855
|
-
}
|
1856
|
-
|
1857
|
-
|
1858
|
-
/*
|
1859
|
-
* Ruby GC mark function for YALE_STORAGE. C accessible.
|
1860
|
-
*/
|
1861
|
-
void nm_yale_storage_mark(void* storage_base) {
|
1862
|
-
YALE_STORAGE* storage = (YALE_STORAGE*)storage_base;
|
1863
|
-
size_t i;
|
1864
|
-
|
1865
|
-
if (storage && storage->dtype == nm::RUBYOBJ) {
|
1866
|
-
for (i = storage->capacity; i-- > 0;) {
|
1867
|
-
rb_gc_mark(*((VALUE*)((char*)(storage->a) + i*DTYPE_SIZES[nm::RUBYOBJ])));
|
1868
|
-
}
|
1869
|
-
}
|
1870
|
-
}
|
1871
|
-
|
1872
|
-
|
1873
|
-
/*
|
1874
|
-
* Allocates and initializes the basic struct (but not the IJA or A vectors).
|
1875
|
-
*/
|
1876
|
-
static YALE_STORAGE* alloc(nm::dtype_t dtype, size_t* shape, size_t dim, nm::itype_t min_itype) {
|
1877
|
-
YALE_STORAGE* s;
|
1878
|
-
|
1879
|
-
s = ALLOC( YALE_STORAGE );
|
1880
|
-
|
1881
|
-
s->ndnz = 0;
|
1882
|
-
s->dtype = dtype;
|
1883
|
-
s->shape = shape;
|
1884
|
-
s->offset = ALLOC_N(size_t, dim);
|
1885
|
-
for (size_t i = 0; i < dim; ++i)
|
1886
|
-
s->offset[i] = 0;
|
1887
|
-
s->dim = dim;
|
1888
|
-
s->itype = nm_yale_storage_itype_by_shape(shape);
|
1889
|
-
s->src = reinterpret_cast<STORAGE*>(s);
|
1890
|
-
s->count = 1;
|
1891
|
-
|
1892
|
-
// See if a higher itype has been requested.
|
1893
|
-
if (static_cast<int8_t>(s->itype) < static_cast<int8_t>(min_itype))
|
1894
|
-
s->itype = min_itype;
|
1895
|
-
|
1896
|
-
return s;
|
1897
|
-
}
|
1898
|
-
|
1899
|
-
YALE_STORAGE* nm_yale_storage_create_from_old_yale(nm::dtype_t dtype, size_t* shape, void* ia, void* ja, void* a, nm::dtype_t from_dtype) {
|
1900
|
-
|
1901
|
-
NAMED_LRI_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::create_from_old_yale, YALE_STORAGE*, nm::dtype_t dtype, size_t* shape, void* r_ia, void* r_ja, void* r_a);
|
1902
|
-
|
1903
|
-
// With C++ templates, we don't want to have a 4-parameter template. That would be LDType, RDType, LIType, RIType.
|
1904
|
-
// We can prevent that by copying ia and ja into the correct itype (if necessary) before passing them to the yale
|
1905
|
-
// copy constructor.
|
1906
|
-
nm::itype_t to_itype = nm_yale_storage_itype_by_shape(shape);
|
1907
|
-
|
1908
|
-
return ttable[dtype][from_dtype][to_itype](dtype, shape, ia, ja, a);
|
1909
|
-
|
1910
|
-
}
|
1911
|
-
|
1912
|
-
//////////////////////////////////////////////
|
1913
|
-
// YALE-SPECIFIC FUNCTIONS (RUBY ACCESSORS) //
|
1914
|
-
//////////////////////////////////////////////
|
1915
|
-
|
1916
|
-
/*
|
1917
|
-
* call-seq:
|
1918
|
-
* yale_size -> Integer
|
1919
|
-
*
|
1920
|
-
* Get the size of a Yale matrix (the number of elements actually stored).
|
1921
|
-
*
|
1922
|
-
* For capacity (the maximum number of elements that can be stored without a resize), use capacity instead.
|
1923
|
-
*/
|
1924
|
-
static VALUE nm_size(VALUE self) {
|
1925
|
-
YALE_STORAGE* s = (YALE_STORAGE*)NM_STORAGE(self);
|
1926
|
-
|
1927
|
-
return rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*(s->shape[0]), s->itype).rval;
|
1928
|
-
}
|
1929
|
-
|
1930
|
-
|
1931
|
-
/*
|
1932
|
-
* call-seq:
|
1933
|
-
* yale_a -> Array
|
1934
|
-
* yale_d(index) -> ...
|
1935
|
-
*
|
1936
|
-
* Get the A array of a Yale matrix (which stores the diagonal and the LU portions of the matrix).
|
1937
|
-
*/
|
1938
|
-
static VALUE nm_a(int argc, VALUE* argv, VALUE self) {
|
1939
|
-
VALUE idx;
|
1940
|
-
rb_scan_args(argc, argv, "01", &idx);
|
1941
|
-
|
1942
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1943
|
-
size_t size = nm_yale_storage_get_size(s);
|
1944
|
-
|
1945
|
-
if (idx == Qnil) {
|
1946
|
-
VALUE* vals = ALLOCA_N(VALUE, size);
|
1947
|
-
|
1948
|
-
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
1949
|
-
for (size_t i = 0; i < size; ++i) {
|
1950
|
-
vals[i] = reinterpret_cast<VALUE*>(s->a)[i];
|
1951
|
-
}
|
1952
|
-
} else {
|
1953
|
-
for (size_t i = 0; i < size; ++i) {
|
1954
|
-
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1955
|
-
}
|
1956
|
-
}
|
1957
|
-
VALUE ary = rb_ary_new4(size, vals);
|
1958
|
-
|
1959
|
-
for (size_t i = size; i < reinterpret_cast<YALE_STORAGE*>(s->src)->capacity; ++i)
|
1960
|
-
rb_ary_push(ary, Qnil);
|
1961
|
-
|
1962
|
-
return ary;
|
1963
|
-
} else {
|
1964
|
-
size_t index = FIX2INT(idx);
|
1965
|
-
if (index >= size) rb_raise(rb_eRangeError, "out of range");
|
1966
|
-
|
1967
|
-
return rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype] * index, s->dtype).rval;
|
1968
|
-
}
|
1969
|
-
}
|
1970
|
-
|
1971
|
-
|
1972
|
-
/*
|
1973
|
-
* call-seq:
|
1974
|
-
* yale_d -> Array
|
1975
|
-
* yale_d(index) -> ...
|
1976
|
-
*
|
1977
|
-
* Get the diagonal ("D") portion of the A array of a Yale matrix.
|
1978
|
-
*/
|
1979
|
-
static VALUE nm_d(int argc, VALUE* argv, VALUE self) {
|
1980
|
-
VALUE idx;
|
1981
|
-
rb_scan_args(argc, argv, "01", &idx);
|
1982
|
-
|
1983
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1984
|
-
|
1985
|
-
if (idx == Qnil) {
|
1986
|
-
VALUE* vals = ALLOCA_N(VALUE, s->shape[0]);
|
1987
|
-
|
1988
|
-
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
1989
|
-
for (size_t i = 0; i < s->shape[0]; ++i) {
|
1990
|
-
vals[i] = reinterpret_cast<VALUE*>(s->a)[i];
|
1991
|
-
}
|
1992
|
-
} else {
|
1993
|
-
for (size_t i = 0; i < s->shape[0]; ++i) {
|
1994
|
-
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1995
|
-
}
|
1996
|
-
}
|
1997
|
-
|
1998
|
-
return rb_ary_new4(s->shape[0], vals);
|
1999
|
-
} else {
|
2000
|
-
size_t index = FIX2INT(idx);
|
2001
|
-
if (index >= s->shape[0]) rb_raise(rb_eRangeError, "out of range");
|
2002
|
-
|
2003
|
-
return rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype] * index, s->dtype).rval;
|
2004
|
-
}
|
2005
|
-
}
|
2006
|
-
|
2007
|
-
/*
|
2008
|
-
* call-seq:
|
2009
|
-
* yale_lu -> Array
|
2010
|
-
*
|
2011
|
-
* Get the non-diagonal ("LU") portion of the A array of a Yale matrix.
|
2012
|
-
*/
|
2013
|
-
static VALUE nm_lu(VALUE self) {
|
2014
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2015
|
-
|
2016
|
-
size_t size = nm_yale_storage_get_size(s);
|
2017
|
-
|
2018
|
-
VALUE* vals = ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
2019
|
-
|
2020
|
-
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
2021
|
-
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
2022
|
-
vals[i] = reinterpret_cast<VALUE*>(s->a)[s->shape[0] + 1 + i];
|
2023
|
-
}
|
2024
|
-
} else {
|
2025
|
-
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
2026
|
-
vals[i] = rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*(s->shape[0] + 1 + i), s->dtype).rval;
|
2027
|
-
}
|
2028
|
-
}
|
2029
|
-
|
2030
|
-
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
2031
|
-
|
2032
|
-
for (size_t i = size; i < reinterpret_cast<YALE_STORAGE*>(s->src)->capacity; ++i)
|
2033
|
-
rb_ary_push(ary, Qnil);
|
2034
|
-
|
2035
|
-
return ary;
|
2036
|
-
}
|
2037
|
-
|
2038
|
-
/*
|
2039
|
-
* call-seq:
|
2040
|
-
* yale_ia -> Array
|
2041
|
-
*
|
2042
|
-
* Get the IA portion of the IJA array of a Yale matrix. This gives the start and end positions of rows in the
|
2043
|
-
* JA and LU portions of the IJA and A arrays, respectively.
|
2044
|
-
*/
|
2045
|
-
static VALUE nm_ia(VALUE self) {
|
2046
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2047
|
-
|
2048
|
-
VALUE* vals = ALLOCA_N(VALUE, s->shape[0] + 1);
|
2049
|
-
|
2050
|
-
for (size_t i = 0; i < s->shape[0] + 1; ++i) {
|
2051
|
-
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*i, s->itype).rval;
|
2052
|
-
}
|
2053
|
-
|
2054
|
-
return rb_ary_new4(s->shape[0]+1, vals);
|
2055
|
-
}
|
2056
|
-
|
2057
|
-
/*
|
2058
|
-
* call-seq:
|
2059
|
-
* yale_ja -> Array
|
2060
|
-
*
|
2061
|
-
* Get the JA portion of the IJA array of a Yale matrix. This gives the column indices for entries in corresponding
|
2062
|
-
* positions in the LU portion of the A array.
|
2063
|
-
*/
|
2064
|
-
static VALUE nm_ja(VALUE self) {
|
2065
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2066
|
-
|
2067
|
-
size_t size = nm_yale_storage_get_size(s);
|
2068
|
-
|
2069
|
-
VALUE* vals = ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
2070
|
-
|
2071
|
-
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
2072
|
-
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*(s->shape[0] + 1 + i), s->itype).rval;
|
2073
|
-
}
|
2074
|
-
|
2075
|
-
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
2076
|
-
|
2077
|
-
for (size_t i = size; i < reinterpret_cast<YALE_STORAGE*>(s->src)->capacity; ++i)
|
2078
|
-
rb_ary_push(ary, Qnil);
|
2079
|
-
|
2080
|
-
return ary;
|
2081
|
-
}
|
2082
|
-
|
2083
|
-
/*
|
2084
|
-
* call-seq:
|
2085
|
-
* yale_ija -> Array
|
2086
|
-
* yale_ija(index) -> ...
|
2087
|
-
*
|
2088
|
-
* Get the IJA array of a Yale matrix (or a component of the IJA array).
|
2089
|
-
*/
|
2090
|
-
static VALUE nm_ija(int argc, VALUE* argv, VALUE self) {
|
2091
|
-
VALUE idx;
|
2092
|
-
rb_scan_args(argc, argv, "01", &idx);
|
2093
|
-
|
2094
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2095
|
-
size_t size = nm_yale_storage_get_size(s);
|
2096
|
-
|
2097
|
-
if (idx == Qnil) {
|
2098
|
-
|
2099
|
-
VALUE* vals = ALLOCA_N(VALUE, size);
|
2100
|
-
|
2101
|
-
for (size_t i = 0; i < size; ++i) {
|
2102
|
-
vals[i] = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*i, s->itype).rval;
|
2103
|
-
}
|
2104
|
-
|
2105
|
-
VALUE ary = rb_ary_new4(size, vals);
|
2106
|
-
|
2107
|
-
for (size_t i = size; i < reinterpret_cast<YALE_STORAGE*>(s->src)->capacity; ++i)
|
2108
|
-
rb_ary_push(ary, Qnil);
|
2109
|
-
|
2110
|
-
return ary;
|
2111
|
-
|
2112
|
-
} else {
|
2113
|
-
size_t index = FIX2INT(idx);
|
2114
|
-
if (index >= size) rb_raise(rb_eRangeError, "out of range");
|
2115
|
-
|
2116
|
-
return rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype] * index, s->itype).rval;
|
2117
|
-
}
|
2118
|
-
}
|
2119
|
-
|
2120
|
-
|
2121
|
-
/*
|
2122
|
-
* call-seq:
|
2123
|
-
* yale_nd_row -> ...
|
2124
|
-
*
|
2125
|
-
* This function gets the non-diagonal contents of a Yale matrix row.
|
2126
|
-
* The first argument should be the row index. The optional second argument may be :hash or :keys, but defaults
|
2127
|
-
* to :hash. If :keys is given, it will only return the Hash keys (the column indices).
|
2128
|
-
*
|
2129
|
-
* This function is meant to accomplish its purpose as efficiently as possible. It does not check for appropriate
|
2130
|
-
* range.
|
2131
|
-
*/
|
2132
|
-
static VALUE nm_nd_row(int argc, VALUE* argv, VALUE self) {
|
2133
|
-
VALUE i_, as;
|
2134
|
-
rb_scan_args(argc, argv, "11", &i_, &as);
|
2135
|
-
|
2136
|
-
bool keys = false;
|
2137
|
-
if (as != Qnil && rb_to_id(as) != nm_rb_hash) keys = true;
|
2138
|
-
|
2139
|
-
size_t i = FIX2INT(i_);
|
2140
|
-
|
2141
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2142
|
-
nm::dtype_t dtype = NM_DTYPE(self);
|
2143
|
-
nm::itype_t itype = NM_ITYPE(self);
|
2144
|
-
|
2145
|
-
// get the position as a size_t
|
2146
|
-
// TODO: Come up with a faster way to get this than transforming to a Ruby object first.
|
2147
|
-
size_t pos = FIX2INT(rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[itype]*i, itype).rval);
|
2148
|
-
size_t nextpos = FIX2INT(rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[itype]*(i+1), itype).rval);
|
2149
|
-
size_t diff = nextpos - pos;
|
2150
|
-
|
2151
|
-
VALUE ret;
|
2152
|
-
if (keys) {
|
2153
|
-
ret = rb_ary_new3(diff);
|
2154
|
-
|
2155
|
-
for (size_t idx = pos; idx < nextpos; ++idx) {
|
2156
|
-
rb_ary_store(ret, idx - pos, rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*idx, s->itype).rval);
|
2157
|
-
}
|
2158
|
-
|
2159
|
-
} else {
|
2160
|
-
ret = rb_hash_new();
|
2161
|
-
|
2162
|
-
for (size_t idx = pos; idx < nextpos; ++idx) {
|
2163
|
-
rb_hash_aset(ret, rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[s->itype]*idx, s->itype).rval,
|
2164
|
-
rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*idx, s->dtype).rval);
|
2165
|
-
}
|
2166
|
-
}
|
2167
|
-
|
2168
|
-
return ret;
|
2169
|
-
}
|
2170
|
-
|
2171
|
-
/*
|
2172
|
-
* call-seq:
|
2173
|
-
* yale_vector_set(i, column_index_array, cell_contents_array, pos) -> Fixnum
|
2174
|
-
*
|
2175
|
-
* Insert at position pos an array of non-diagonal elements with column indices given. Note that the column indices and values
|
2176
|
-
* must be storage-contiguous -- that is, you can't insert them around existing elements in some row, only amid some
|
2177
|
-
* elements in some row. You *can* insert them around a diagonal element, since this is stored separately. This function
|
2178
|
-
* may not be used for the insertion of diagonal elements in most cases, as these are already present in the data
|
2179
|
-
* structure and are typically modified by replacement rather than insertion.
|
2180
|
-
*
|
2181
|
-
* The last argument, pos, may be nil if you want to insert at the beginning of a row. Otherwise it needs to be provided.
|
2182
|
-
* Don't expect this function to know the difference. It really does very little checking, because its goal is to make
|
2183
|
-
* multiple contiguous insertion as quick as possible.
|
2184
|
-
*
|
2185
|
-
* You should also not attempt to insert values which are the default (0). These are not supposed to be stored, and may
|
2186
|
-
* lead to undefined behavior.
|
2187
|
-
*
|
2188
|
-
* Example:
|
2189
|
-
* m.yale_vector_set(3, [0,3,4], [1,1,1], 15)
|
2190
|
-
*
|
2191
|
-
* The example above inserts the values 1, 1, and 1 in columns 0, 3, and 4, assumed to be located at position 15 (which
|
2192
|
-
* corresponds to row 3).
|
2193
|
-
*
|
2194
|
-
* Example:
|
2195
|
-
* next = m.yale_vector_set(3, [0,3,4], [1,1,1])
|
2196
|
-
*
|
2197
|
-
* This example determines that i=3 is at position 15 automatically. The value returned, next, is the position where the
|
2198
|
-
* next value(s) should be inserted.
|
2199
|
-
*/
|
2200
|
-
VALUE nm_vector_set(int argc, VALUE* argv, VALUE self) { //, VALUE i_, VALUE jv, VALUE vv, VALUE pos_) {
|
2201
|
-
|
2202
|
-
// i, jv, vv are mandatory; pos is optional; thus "31"
|
2203
|
-
VALUE i_, jv, vv, pos_;
|
2204
|
-
rb_scan_args(argc, argv, "31", &i_, &jv, &vv, &pos_);
|
2205
|
-
|
2206
|
-
size_t len = RARRAY_LEN(jv); // need length in order to read the arrays in
|
2207
|
-
size_t vvlen = RARRAY_LEN(vv);
|
2208
|
-
if (len != vvlen)
|
2209
|
-
rb_raise(rb_eArgError, "lengths must match between j array (%d) and value array (%d)", len, vvlen);
|
2210
|
-
|
2211
|
-
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
2212
|
-
nm::dtype_t dtype = NM_DTYPE(self);
|
2213
|
-
nm::itype_t itype = NM_ITYPE(self);
|
2214
|
-
|
2215
|
-
size_t i = FIX2INT(i_); // get the row
|
2216
|
-
|
2217
|
-
// get the position as a size_t
|
2218
|
-
// TODO: Come up with a faster way to get this than transforming to a Ruby object first.
|
2219
|
-
if (pos_ == Qnil) pos_ = rubyobj_from_cval_by_itype((char*)(s->ija) + ITYPE_SIZES[itype]*i, itype).rval;
|
2220
|
-
size_t pos = FIX2INT(pos_);
|
2221
|
-
|
2222
|
-
// Allocate the j array and the values array
|
2223
|
-
size_t* j = ALLOCA_N(size_t, len);
|
2224
|
-
void* vals = ALLOCA_N(char, DTYPE_SIZES[dtype] * len);
|
2225
|
-
|
2226
|
-
// Copy array contents
|
2227
|
-
for (size_t idx = 0; idx < len; ++idx) {
|
2228
|
-
j[idx] = FIX2INT(rb_ary_entry(jv, idx));
|
2229
|
-
rubyval_to_cval(rb_ary_entry(vv, idx), dtype, (char*)vals + idx * DTYPE_SIZES[dtype]);
|
2230
|
-
}
|
2231
|
-
|
2232
|
-
char ins_type = nm_yale_storage_vector_insert(s, pos, j, vals, len, false, dtype, itype);
|
2233
|
-
nm_yale_storage_increment_ia_after(s, s->shape[0], i, len, itype);
|
2234
|
-
reinterpret_cast<YALE_STORAGE*>(s->src)->ndnz += len;
|
2235
|
-
|
2236
|
-
// Return the updated position
|
2237
|
-
pos += len;
|
2238
|
-
return INT2FIX(pos);
|
2239
|
-
}
|
2240
|
-
|
2241
|
-
|
2242
|
-
|
2243
|
-
|
2244
|
-
/*
|
2245
|
-
* call-seq:
|
2246
|
-
* __yale_default_value__ -> ...
|
2247
|
-
*
|
2248
|
-
* Get the default_value property from a yale matrix.
|
2249
|
-
*/
|
2250
|
-
VALUE nm_yale_default_value(VALUE self) {
|
2251
|
-
return default_value(NM_STORAGE_YALE(self));
|
2252
|
-
}
|
2253
|
-
|
2254
|
-
|
2255
|
-
/*
|
2256
|
-
* call-seq:
|
2257
|
-
* __yale_map_merged_stored__(right) -> Enumerator
|
2258
|
-
*
|
2259
|
-
* A map operation on two Yale matrices which only iterates across the stored indices.
|
2260
|
-
*/
|
2261
|
-
VALUE nm_yale_map_merged_stored(VALUE left, VALUE right, VALUE init) {
|
2262
|
-
YALE_STORAGE *s = NM_STORAGE_YALE(left),
|
2263
|
-
*t = NM_STORAGE_YALE(right);
|
2264
|
-
|
2265
|
-
ITYPE_TEMPLATE_TABLE(nm::yale_storage::map_merged_stored, VALUE, VALUE l, VALUE r, VALUE init, nm::itype_t)
|
2266
|
-
|
2267
|
-
nm::itype_t itype = NM_MAX_ITYPE(s->itype, t->itype);
|
2268
|
-
return ttable[itype](left, right, init, itype);
|
2269
|
-
}
|
2270
|
-
|
2271
|
-
|
2272
|
-
/*
|
2273
|
-
* call-seq:
|
2274
|
-
* __yale_map_stored__ -> Enumerator
|
2275
|
-
*
|
2276
|
-
* A map operation on two Yale matrices which only iterates across the stored indices.
|
2277
|
-
*/
|
2278
|
-
VALUE nm_yale_map_stored(VALUE self) {
|
2279
|
-
ITYPE_TEMPLATE_TABLE(nm::yale_storage::map_stored, VALUE, VALUE)
|
2280
|
-
|
2281
|
-
return ttable[NM_ITYPE(self)](self);
|
2282
|
-
}
|
2283
|
-
|
2284
|
-
} // end of extern "C" block
|