pnmatrix 1.2.4
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +7 -0
- data/ext/nmatrix/binary_format.txt +53 -0
- data/ext/nmatrix/data/complex.h +388 -0
- data/ext/nmatrix/data/data.cpp +274 -0
- data/ext/nmatrix/data/data.h +651 -0
- data/ext/nmatrix/data/meta.h +64 -0
- data/ext/nmatrix/data/ruby_object.h +386 -0
- data/ext/nmatrix/extconf.rb +70 -0
- data/ext/nmatrix/math/asum.h +99 -0
- data/ext/nmatrix/math/cblas_enums.h +36 -0
- data/ext/nmatrix/math/cblas_templates_core.h +507 -0
- data/ext/nmatrix/math/gemm.h +241 -0
- data/ext/nmatrix/math/gemv.h +178 -0
- data/ext/nmatrix/math/getrf.h +255 -0
- data/ext/nmatrix/math/getrs.h +121 -0
- data/ext/nmatrix/math/imax.h +82 -0
- data/ext/nmatrix/math/laswp.h +165 -0
- data/ext/nmatrix/math/long_dtype.h +62 -0
- data/ext/nmatrix/math/magnitude.h +54 -0
- data/ext/nmatrix/math/math.h +751 -0
- data/ext/nmatrix/math/nrm2.h +165 -0
- data/ext/nmatrix/math/rot.h +117 -0
- data/ext/nmatrix/math/rotg.h +106 -0
- data/ext/nmatrix/math/scal.h +71 -0
- data/ext/nmatrix/math/trsm.h +336 -0
- data/ext/nmatrix/math/util.h +162 -0
- data/ext/nmatrix/math.cpp +1368 -0
- data/ext/nmatrix/nm_memory.h +60 -0
- data/ext/nmatrix/nmatrix.cpp +285 -0
- data/ext/nmatrix/nmatrix.h +476 -0
- data/ext/nmatrix/ruby_constants.cpp +151 -0
- data/ext/nmatrix/ruby_constants.h +106 -0
- data/ext/nmatrix/ruby_nmatrix.c +3130 -0
- data/ext/nmatrix/storage/common.cpp +77 -0
- data/ext/nmatrix/storage/common.h +183 -0
- data/ext/nmatrix/storage/dense/dense.cpp +1096 -0
- data/ext/nmatrix/storage/dense/dense.h +129 -0
- data/ext/nmatrix/storage/list/list.cpp +1628 -0
- data/ext/nmatrix/storage/list/list.h +138 -0
- data/ext/nmatrix/storage/storage.cpp +730 -0
- data/ext/nmatrix/storage/storage.h +99 -0
- data/ext/nmatrix/storage/yale/class.h +1139 -0
- data/ext/nmatrix/storage/yale/iterators/base.h +143 -0
- data/ext/nmatrix/storage/yale/iterators/iterator.h +131 -0
- data/ext/nmatrix/storage/yale/iterators/row.h +450 -0
- data/ext/nmatrix/storage/yale/iterators/row_stored.h +140 -0
- data/ext/nmatrix/storage/yale/iterators/row_stored_nd.h +169 -0
- data/ext/nmatrix/storage/yale/iterators/stored_diagonal.h +124 -0
- data/ext/nmatrix/storage/yale/math/transpose.h +110 -0
- data/ext/nmatrix/storage/yale/yale.cpp +2074 -0
- data/ext/nmatrix/storage/yale/yale.h +203 -0
- data/ext/nmatrix/types.h +55 -0
- data/ext/nmatrix/util/io.cpp +279 -0
- data/ext/nmatrix/util/io.h +115 -0
- data/ext/nmatrix/util/sl_list.cpp +627 -0
- data/ext/nmatrix/util/sl_list.h +144 -0
- data/ext/nmatrix/util/util.h +78 -0
- data/lib/nmatrix/blas.rb +378 -0
- data/lib/nmatrix/cruby/math.rb +744 -0
- data/lib/nmatrix/enumerate.rb +253 -0
- data/lib/nmatrix/homogeneous.rb +241 -0
- data/lib/nmatrix/io/fortran_format.rb +138 -0
- data/lib/nmatrix/io/harwell_boeing.rb +221 -0
- data/lib/nmatrix/io/market.rb +263 -0
- data/lib/nmatrix/io/point_cloud.rb +189 -0
- data/lib/nmatrix/jruby/decomposition.rb +24 -0
- data/lib/nmatrix/jruby/enumerable.rb +13 -0
- data/lib/nmatrix/jruby/error.rb +4 -0
- data/lib/nmatrix/jruby/math.rb +501 -0
- data/lib/nmatrix/jruby/nmatrix_java.rb +840 -0
- data/lib/nmatrix/jruby/operators.rb +283 -0
- data/lib/nmatrix/jruby/slice.rb +264 -0
- data/lib/nmatrix/lapack_core.rb +181 -0
- data/lib/nmatrix/lapack_plugin.rb +44 -0
- data/lib/nmatrix/math.rb +953 -0
- data/lib/nmatrix/mkmf.rb +100 -0
- data/lib/nmatrix/monkeys.rb +137 -0
- data/lib/nmatrix/nmatrix.rb +1172 -0
- data/lib/nmatrix/rspec.rb +75 -0
- data/lib/nmatrix/shortcuts.rb +1163 -0
- data/lib/nmatrix/version.rb +39 -0
- data/lib/nmatrix/yale_functions.rb +118 -0
- data/lib/nmatrix.rb +28 -0
- data/spec/00_nmatrix_spec.rb +892 -0
- data/spec/01_enum_spec.rb +196 -0
- data/spec/02_slice_spec.rb +407 -0
- data/spec/03_nmatrix_monkeys_spec.rb +80 -0
- data/spec/2x2_dense_double.mat +0 -0
- data/spec/4x4_sparse.mat +0 -0
- data/spec/4x5_dense.mat +0 -0
- data/spec/blas_spec.rb +215 -0
- data/spec/elementwise_spec.rb +311 -0
- data/spec/homogeneous_spec.rb +100 -0
- data/spec/io/fortran_format_spec.rb +88 -0
- data/spec/io/harwell_boeing_spec.rb +98 -0
- data/spec/io/test.rua +9 -0
- data/spec/io_spec.rb +159 -0
- data/spec/lapack_core_spec.rb +482 -0
- data/spec/leakcheck.rb +16 -0
- data/spec/math_spec.rb +1363 -0
- data/spec/nmatrix_yale_resize_test_associations.yaml +2802 -0
- data/spec/nmatrix_yale_spec.rb +286 -0
- data/spec/rspec_monkeys.rb +56 -0
- data/spec/rspec_spec.rb +35 -0
- data/spec/shortcuts_spec.rb +474 -0
- data/spec/slice_set_spec.rb +162 -0
- data/spec/spec_helper.rb +172 -0
- data/spec/stat_spec.rb +214 -0
- data/spec/test.pcd +20 -0
- data/spec/utm5940.mtx +83844 -0
- metadata +295 -0
@@ -0,0 +1,2074 @@
|
|
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 - 2014, Ruby Science Foundation
|
13
|
+
// NMatrix is Copyright (c) 2012 - 2014, John Woods and the 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
|
+
#include <tuple>
|
49
|
+
#include <queue>
|
50
|
+
|
51
|
+
/*
|
52
|
+
* Project Includes
|
53
|
+
*/
|
54
|
+
|
55
|
+
// #include "types.h"
|
56
|
+
#include "../../data/data.h"
|
57
|
+
#include "../../math/math.h"
|
58
|
+
|
59
|
+
#include "../common.h"
|
60
|
+
|
61
|
+
#include "../../nmatrix.h"
|
62
|
+
#include "../../data/meta.h"
|
63
|
+
|
64
|
+
#include "iterators/base.h"
|
65
|
+
#include "iterators/stored_diagonal.h"
|
66
|
+
#include "iterators/row_stored_nd.h"
|
67
|
+
#include "iterators/row_stored.h"
|
68
|
+
#include "iterators/row.h"
|
69
|
+
#include "iterators/iterator.h"
|
70
|
+
#include "class.h"
|
71
|
+
#include "yale.h"
|
72
|
+
#include "../../ruby_constants.h"
|
73
|
+
|
74
|
+
/*
|
75
|
+
* Macros
|
76
|
+
*/
|
77
|
+
|
78
|
+
#ifndef NM_MAX
|
79
|
+
#define NM_MAX(a,b) (((a)>(b))?(a):(b))
|
80
|
+
#define NM_MIN(a,b) (((a)<(b))?(a):(b))
|
81
|
+
#endif
|
82
|
+
|
83
|
+
/*
|
84
|
+
* Forward Declarations
|
85
|
+
*/
|
86
|
+
|
87
|
+
extern "C" {
|
88
|
+
static YALE_STORAGE* alloc(nm::dtype_t dtype, size_t* shape, size_t dim);
|
89
|
+
|
90
|
+
static size_t yale_count_slice_copy_ndnz(const YALE_STORAGE* s, size_t*, size_t*);
|
91
|
+
|
92
|
+
static void* default_value_ptr(const YALE_STORAGE* s);
|
93
|
+
static VALUE default_value(const YALE_STORAGE* s);
|
94
|
+
static VALUE obj_at(YALE_STORAGE* s, size_t k);
|
95
|
+
|
96
|
+
/* Ruby-accessible functions */
|
97
|
+
static VALUE nm_size(VALUE self);
|
98
|
+
static VALUE nm_a(int argc, VALUE* argv, VALUE self);
|
99
|
+
static VALUE nm_d(int argc, VALUE* argv, VALUE self);
|
100
|
+
static VALUE nm_lu(VALUE self);
|
101
|
+
static VALUE nm_ia(VALUE self);
|
102
|
+
static VALUE nm_ja(VALUE self);
|
103
|
+
static VALUE nm_ija(int argc, VALUE* argv, VALUE self);
|
104
|
+
static VALUE nm_row_keys_intersection(VALUE m1, VALUE ii1, VALUE m2, VALUE ii2);
|
105
|
+
|
106
|
+
static VALUE nm_nd_row(int argc, VALUE* argv, VALUE self);
|
107
|
+
|
108
|
+
static inline size_t src_ndnz(const YALE_STORAGE* s) {
|
109
|
+
return reinterpret_cast<YALE_STORAGE*>(s->src)->ndnz;
|
110
|
+
}
|
111
|
+
|
112
|
+
} // end extern "C" block
|
113
|
+
|
114
|
+
namespace nm { namespace yale_storage {
|
115
|
+
|
116
|
+
template <typename LD, typename RD>
|
117
|
+
static VALUE map_merged_stored(VALUE left, VALUE right, VALUE init);
|
118
|
+
|
119
|
+
template <typename DType>
|
120
|
+
static bool ndrow_is_empty(const YALE_STORAGE* s, IType ija, const IType ija_next);
|
121
|
+
|
122
|
+
template <typename LDType, typename RDType>
|
123
|
+
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);
|
124
|
+
|
125
|
+
template <typename LDType, typename RDType>
|
126
|
+
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right);
|
127
|
+
|
128
|
+
template <typename LDType, typename RDType>
|
129
|
+
static bool eqeq_different_defaults(const YALE_STORAGE* s, const LDType& s_init, const YALE_STORAGE* t, const RDType& t_init);
|
130
|
+
|
131
|
+
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, long n);
|
132
|
+
|
133
|
+
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool& found);
|
134
|
+
|
135
|
+
template <typename DType>
|
136
|
+
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, void* val_, size_t n, bool struct_only);
|
137
|
+
|
138
|
+
template <typename DType>
|
139
|
+
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only);
|
140
|
+
|
141
|
+
template <typename DType>
|
142
|
+
static std::tuple<long,bool,std::queue<std::tuple<IType,IType,int> > > count_slice_set_ndnz_change(YALE_STORAGE* s, size_t* coords, size_t* lengths, DType* v, size_t v_size);
|
143
|
+
|
144
|
+
static inline IType* IJA(const YALE_STORAGE* s) {
|
145
|
+
return reinterpret_cast<YALE_STORAGE*>(s->src)->ija;
|
146
|
+
}
|
147
|
+
|
148
|
+
static inline IType IJA_SET(const YALE_STORAGE* s, size_t loc, IType val) {
|
149
|
+
return IJA(s)[loc] = val;
|
150
|
+
}
|
151
|
+
|
152
|
+
template <typename DType>
|
153
|
+
static inline DType* A(const YALE_STORAGE* s) {
|
154
|
+
return reinterpret_cast<DType*>(reinterpret_cast<YALE_STORAGE*>(s->src)->a);
|
155
|
+
}
|
156
|
+
|
157
|
+
template <typename DType>
|
158
|
+
static inline DType A_SET(const YALE_STORAGE* s, size_t loc, DType val) {
|
159
|
+
return A<DType>(s)[loc] = val;
|
160
|
+
}
|
161
|
+
|
162
|
+
|
163
|
+
/*
|
164
|
+
* Functions
|
165
|
+
*/
|
166
|
+
|
167
|
+
/*
|
168
|
+
* Copy a vector from one DType to another.
|
169
|
+
*/
|
170
|
+
template <typename LType, typename RType>
|
171
|
+
static inline void copy_recast_vector(const void* in_, void* out_, size_t length) {
|
172
|
+
const RType* in = reinterpret_cast<const RType*>(in_);
|
173
|
+
LType* out = reinterpret_cast<LType*>(out_);
|
174
|
+
for (size_t i = 0; i < length; ++i) {
|
175
|
+
out[i] = in[i];
|
176
|
+
}
|
177
|
+
out;
|
178
|
+
}
|
179
|
+
|
180
|
+
|
181
|
+
|
182
|
+
/*
|
183
|
+
* Create Yale storage from IA, JA, and A vectors given in Old Yale format (probably from a file, since NMatrix only uses
|
184
|
+
* new Yale for its storage).
|
185
|
+
*
|
186
|
+
* This function is needed for Matlab .MAT v5 IO.
|
187
|
+
*/
|
188
|
+
template <typename LDType, typename RDType>
|
189
|
+
YALE_STORAGE* create_from_old_yale(dtype_t dtype, size_t* shape, char* r_ia, char* r_ja, char* r_a) {
|
190
|
+
IType* ir = reinterpret_cast<IType*>(r_ia);
|
191
|
+
IType* jr = reinterpret_cast<IType*>(r_ja);
|
192
|
+
RDType* ar = reinterpret_cast<RDType*>(r_a);
|
193
|
+
|
194
|
+
// Read through ia and ja and figure out the ndnz (non-diagonal non-zeros) count.
|
195
|
+
size_t ndnz = 0, i, p, p_next;
|
196
|
+
|
197
|
+
for (i = 0; i < shape[0]; ++i) { // Walk down rows
|
198
|
+
for (p = ir[i], p_next = ir[i+1]; p < p_next; ++p) { // Now walk through columns
|
199
|
+
|
200
|
+
if (i != jr[p]) ++ndnz; // entry is non-diagonal and probably nonzero
|
201
|
+
|
202
|
+
}
|
203
|
+
}
|
204
|
+
|
205
|
+
// Having walked through the matrix, we now go about allocating the space for it.
|
206
|
+
YALE_STORAGE* s = alloc(dtype, shape, 2);
|
207
|
+
|
208
|
+
s->capacity = shape[0] + ndnz + 1;
|
209
|
+
s->ndnz = ndnz;
|
210
|
+
|
211
|
+
// Setup IJA and A arrays
|
212
|
+
s->ija = NM_ALLOC_N( IType, s->capacity );
|
213
|
+
s->a = NM_ALLOC_N( LDType, s->capacity );
|
214
|
+
IType* ijl = reinterpret_cast<IType*>(s->ija);
|
215
|
+
LDType* al = reinterpret_cast<LDType*>(s->a);
|
216
|
+
|
217
|
+
// set the diagonal to zero -- this prevents uninitialized values from popping up.
|
218
|
+
for (size_t index = 0; index < shape[0]; ++index) {
|
219
|
+
al[index] = 0;
|
220
|
+
}
|
221
|
+
|
222
|
+
// Figure out where to start writing JA in IJA:
|
223
|
+
size_t pp = s->shape[0]+1;
|
224
|
+
|
225
|
+
// Find beginning of first row
|
226
|
+
p = ir[0];
|
227
|
+
|
228
|
+
// Now fill the arrays
|
229
|
+
for (i = 0; i < s->shape[0]; ++i) {
|
230
|
+
|
231
|
+
// Set the beginning of the row (of output)
|
232
|
+
ijl[i] = pp;
|
233
|
+
|
234
|
+
// Now walk through columns, starting at end of row (of input)
|
235
|
+
for (size_t p_next = ir[i+1]; p < p_next; ++p, ++pp) {
|
236
|
+
|
237
|
+
if (i == jr[p]) { // diagonal
|
238
|
+
|
239
|
+
al[i] = ar[p];
|
240
|
+
--pp;
|
241
|
+
|
242
|
+
} else { // nondiagonal
|
243
|
+
|
244
|
+
ijl[pp] = jr[p];
|
245
|
+
al[pp] = ar[p];
|
246
|
+
|
247
|
+
}
|
248
|
+
}
|
249
|
+
}
|
250
|
+
|
251
|
+
ijl[i] = pp; // Set the end of the last row
|
252
|
+
|
253
|
+
// Set the zero position for our output matrix
|
254
|
+
al[i] = 0;
|
255
|
+
|
256
|
+
return s;
|
257
|
+
}
|
258
|
+
|
259
|
+
|
260
|
+
/*
|
261
|
+
* Empty the matrix by initializing the IJA vector and setting the diagonal to 0.
|
262
|
+
*
|
263
|
+
* Called when most YALE_STORAGE objects are created.
|
264
|
+
*
|
265
|
+
* Can't go inside of class YaleStorage because YaleStorage creation requires that
|
266
|
+
* IJA already be initialized.
|
267
|
+
*/
|
268
|
+
template <typename DType>
|
269
|
+
void init(YALE_STORAGE* s, void* init_val) {
|
270
|
+
IType IA_INIT = s->shape[0] + 1;
|
271
|
+
|
272
|
+
IType* ija = reinterpret_cast<IType*>(s->ija);
|
273
|
+
// clear out IJA vector
|
274
|
+
for (IType i = 0; i < IA_INIT; ++i) {
|
275
|
+
ija[i] = IA_INIT; // set initial values for IJA
|
276
|
+
}
|
277
|
+
|
278
|
+
clear_diagonal_and_zero<DType>(s, init_val);
|
279
|
+
}
|
280
|
+
|
281
|
+
|
282
|
+
template <typename LDType, typename RDType>
|
283
|
+
static YALE_STORAGE* slice_copy(YALE_STORAGE* s) {
|
284
|
+
YaleStorage<RDType> y(s);
|
285
|
+
return y.template alloc_copy<LDType, false>();
|
286
|
+
}
|
287
|
+
|
288
|
+
|
289
|
+
/*
|
290
|
+
* Template version of copy transposed. This could also, in theory, allow a map -- but transpose.h
|
291
|
+
* would need to be updated.
|
292
|
+
*
|
293
|
+
* TODO: Update for slicing? Update for different dtype in and out? We can cast rather easily without
|
294
|
+
* too much modification.
|
295
|
+
*/
|
296
|
+
template <typename D>
|
297
|
+
YALE_STORAGE* copy_transposed(YALE_STORAGE* rhs) {
|
298
|
+
YaleStorage<D> y(rhs);
|
299
|
+
return y.template alloc_copy_transposed<D, false>();
|
300
|
+
}
|
301
|
+
|
302
|
+
|
303
|
+
///////////////
|
304
|
+
// Accessors //
|
305
|
+
///////////////
|
306
|
+
|
307
|
+
|
308
|
+
/*
|
309
|
+
* Determine the number of non-diagonal non-zeros in a not-yet-created copy of a slice or matrix.
|
310
|
+
*/
|
311
|
+
template <typename DType>
|
312
|
+
static size_t count_slice_copy_ndnz(const YALE_STORAGE* s, size_t* offset, size_t* shape) {
|
313
|
+
IType* ija = s->ija;
|
314
|
+
DType* a = reinterpret_cast<DType*>(s->a);
|
315
|
+
|
316
|
+
DType ZERO(*reinterpret_cast<DType*>(default_value_ptr(s)));
|
317
|
+
|
318
|
+
// Calc ndnz for the destination
|
319
|
+
size_t ndnz = 0;
|
320
|
+
size_t i, j; // indexes of destination matrix
|
321
|
+
size_t k, l; // indexes of source matrix
|
322
|
+
for (i = 0; i < shape[0]; i++) {
|
323
|
+
k = i + offset[0];
|
324
|
+
for (j = 0; j < shape[1]; j++) {
|
325
|
+
l = j + offset[1];
|
326
|
+
|
327
|
+
if (j == i) continue;
|
328
|
+
|
329
|
+
if (k == l) { // for diagonal element of source
|
330
|
+
if (a[k] != ZERO) ++ndnz;
|
331
|
+
} else { // for non-diagonal element
|
332
|
+
for (size_t c = ija[k]; c < ija[k+1]; c++) {
|
333
|
+
if (ija[c] == l) {
|
334
|
+
++ndnz;
|
335
|
+
break;
|
336
|
+
}
|
337
|
+
}
|
338
|
+
}
|
339
|
+
}
|
340
|
+
}
|
341
|
+
|
342
|
+
return ndnz;
|
343
|
+
}
|
344
|
+
|
345
|
+
|
346
|
+
|
347
|
+
/*
|
348
|
+
* Get a single element of a yale storage object
|
349
|
+
*/
|
350
|
+
template <typename DType>
|
351
|
+
static void* get_single(YALE_STORAGE* storage, SLICE* slice) {
|
352
|
+
YaleStorage<DType> y(storage);
|
353
|
+
return reinterpret_cast<void*>(y.get_single_p(slice));
|
354
|
+
}
|
355
|
+
|
356
|
+
|
357
|
+
/*
|
358
|
+
* Returns a reference-slice of a matrix.
|
359
|
+
*/
|
360
|
+
template <typename DType>
|
361
|
+
YALE_STORAGE* ref(YALE_STORAGE* s, SLICE* slice) {
|
362
|
+
return YaleStorage<DType>(s).alloc_ref(slice);
|
363
|
+
}
|
364
|
+
|
365
|
+
|
366
|
+
/*
|
367
|
+
* Attempt to set a cell or cells in a Yale matrix.
|
368
|
+
*/
|
369
|
+
template <typename DType>
|
370
|
+
void set(VALUE left, SLICE* slice, VALUE right) {
|
371
|
+
YALE_STORAGE* storage = NM_STORAGE_YALE(left);
|
372
|
+
YaleStorage<DType> y(storage);
|
373
|
+
y.insert(slice, right);
|
374
|
+
}
|
375
|
+
|
376
|
+
///////////
|
377
|
+
// Tests //
|
378
|
+
///////////
|
379
|
+
|
380
|
+
/*
|
381
|
+
* Yale eql? -- for whole-matrix comparison returning a single value.
|
382
|
+
*/
|
383
|
+
template <typename LDType, typename RDType>
|
384
|
+
static bool eqeq(const YALE_STORAGE* left, const YALE_STORAGE* right) {
|
385
|
+
return YaleStorage<LDType>(left) == YaleStorage<RDType>(right);
|
386
|
+
}
|
387
|
+
|
388
|
+
|
389
|
+
//////////
|
390
|
+
// Math //
|
391
|
+
//////////
|
392
|
+
|
393
|
+
#define YALE_IA(s) (reinterpret_cast<IType*>(s->ija))
|
394
|
+
#define YALE_IJ(s) (reinterpret_cast<IType*>(s->ija) + s->shape[0] + 1)
|
395
|
+
#define YALE_COUNT(yale) (yale->ndnz + yale->shape[0])
|
396
|
+
|
397
|
+
/////////////
|
398
|
+
// Utility //
|
399
|
+
/////////////
|
400
|
+
|
401
|
+
|
402
|
+
/*
|
403
|
+
* Binary search for finding the beginning of a slice. Returns the position of the first element which is larger than
|
404
|
+
* bound.
|
405
|
+
*/
|
406
|
+
IType binary_search_left_boundary(const YALE_STORAGE* s, IType left, IType right, IType bound) {
|
407
|
+
if (left > right) return -1;
|
408
|
+
|
409
|
+
IType* ija = IJA(s);
|
410
|
+
|
411
|
+
if (ija[left] >= bound) return left; // shortcut
|
412
|
+
|
413
|
+
IType mid = (left + right) / 2;
|
414
|
+
IType mid_j = ija[mid];
|
415
|
+
|
416
|
+
if (mid_j == bound)
|
417
|
+
return mid;
|
418
|
+
else if (mid_j > bound) { // eligible! don't exclude it.
|
419
|
+
return binary_search_left_boundary(s, left, mid, bound);
|
420
|
+
} else // (mid_j < bound)
|
421
|
+
return binary_search_left_boundary(s, mid + 1, right, bound);
|
422
|
+
}
|
423
|
+
|
424
|
+
|
425
|
+
/*
|
426
|
+
* Binary search for returning stored values. Returns a non-negative position, or -1 for not found.
|
427
|
+
*/
|
428
|
+
int binary_search(YALE_STORAGE* s, IType left, IType right, IType key) {
|
429
|
+
if (s->src != s) throw; // need to fix this quickly
|
430
|
+
|
431
|
+
if (left > right) return -1;
|
432
|
+
|
433
|
+
IType* ija = s->ija;
|
434
|
+
|
435
|
+
IType mid = (left + right)/2;
|
436
|
+
IType mid_j = ija[mid];
|
437
|
+
|
438
|
+
if (mid_j == key)
|
439
|
+
return mid;
|
440
|
+
|
441
|
+
else if (mid_j > key)
|
442
|
+
return binary_search(s, left, mid - 1, key);
|
443
|
+
|
444
|
+
else
|
445
|
+
return binary_search(s, mid + 1, right, key);
|
446
|
+
}
|
447
|
+
|
448
|
+
|
449
|
+
/*
|
450
|
+
* Resize yale storage vectors A and IJA, copying values.
|
451
|
+
*/
|
452
|
+
static void vector_grow(YALE_STORAGE* s) {
|
453
|
+
if (s != s->src) {
|
454
|
+
throw; // need to correct this quickly.
|
455
|
+
}
|
456
|
+
nm_yale_storage_register(s);
|
457
|
+
size_t new_capacity = s->capacity * GROWTH_CONSTANT;
|
458
|
+
size_t max_capacity = YaleStorage<uint8_t>::max_size(s->shape);
|
459
|
+
|
460
|
+
if (new_capacity > max_capacity) new_capacity = max_capacity;
|
461
|
+
|
462
|
+
IType* new_ija = NM_ALLOC_N(IType, new_capacity);
|
463
|
+
void* new_a = NM_ALLOC_N(char, DTYPE_SIZES[s->dtype] * new_capacity);
|
464
|
+
|
465
|
+
IType* old_ija = s->ija;
|
466
|
+
void* old_a = s->a;
|
467
|
+
|
468
|
+
memcpy(new_ija, old_ija, s->capacity * sizeof(IType));
|
469
|
+
memcpy(new_a, old_a, s->capacity * DTYPE_SIZES[s->dtype]);
|
470
|
+
|
471
|
+
s->capacity = new_capacity;
|
472
|
+
|
473
|
+
if (s->dtype == nm::RUBYOBJ)
|
474
|
+
nm_yale_storage_register_a(new_a, s->capacity * DTYPE_SIZES[s->dtype]);
|
475
|
+
|
476
|
+
NM_FREE(old_ija);
|
477
|
+
nm_yale_storage_unregister(s);
|
478
|
+
NM_FREE(old_a);
|
479
|
+
if (s->dtype == nm::RUBYOBJ)
|
480
|
+
nm_yale_storage_unregister_a(new_a, s->capacity * DTYPE_SIZES[s->dtype]);
|
481
|
+
|
482
|
+
s->ija = new_ija;
|
483
|
+
s->a = new_a;
|
484
|
+
|
485
|
+
}
|
486
|
+
|
487
|
+
|
488
|
+
/*
|
489
|
+
* Resize yale storage vectors A and IJA in preparation for an insertion.
|
490
|
+
*/
|
491
|
+
template <typename DType>
|
492
|
+
static char vector_insert_resize(YALE_STORAGE* s, size_t current_size, size_t pos, size_t* j, size_t n, bool struct_only) {
|
493
|
+
if (s != s->src) throw;
|
494
|
+
|
495
|
+
// Determine the new capacity for the IJA and A vectors.
|
496
|
+
size_t new_capacity = s->capacity * GROWTH_CONSTANT;
|
497
|
+
size_t max_capacity = YaleStorage<DType>::max_size(s->shape);
|
498
|
+
|
499
|
+
if (new_capacity > max_capacity) {
|
500
|
+
new_capacity = max_capacity;
|
501
|
+
|
502
|
+
if (current_size + n > max_capacity) rb_raise(rb_eNoMemError, "insertion size exceeded maximum yale matrix size");
|
503
|
+
}
|
504
|
+
|
505
|
+
if (new_capacity < current_size + n)
|
506
|
+
new_capacity = current_size + n;
|
507
|
+
|
508
|
+
nm_yale_storage_register(s);
|
509
|
+
|
510
|
+
// Allocate the new vectors.
|
511
|
+
IType* new_ija = NM_ALLOC_N( IType, new_capacity );
|
512
|
+
NM_CHECK_ALLOC(new_ija);
|
513
|
+
|
514
|
+
DType* new_a = NM_ALLOC_N( DType, new_capacity );
|
515
|
+
NM_CHECK_ALLOC(new_a);
|
516
|
+
|
517
|
+
IType* old_ija = reinterpret_cast<IType*>(s->ija);
|
518
|
+
DType* old_a = reinterpret_cast<DType*>(s->a);
|
519
|
+
|
520
|
+
// Copy all values prior to the insertion site to the new IJA and new A
|
521
|
+
if (struct_only) {
|
522
|
+
for (size_t i = 0; i < pos; ++i) {
|
523
|
+
new_ija[i] = old_ija[i];
|
524
|
+
}
|
525
|
+
} else {
|
526
|
+
for (size_t i = 0; i < pos; ++i) {
|
527
|
+
new_ija[i] = old_ija[i];
|
528
|
+
new_a[i] = old_a[i];
|
529
|
+
}
|
530
|
+
}
|
531
|
+
|
532
|
+
|
533
|
+
// Copy all values subsequent to the insertion site to the new IJA and new A, leaving room (size n) for insertion.
|
534
|
+
if (struct_only) {
|
535
|
+
for (size_t i = pos; i < current_size; ++i) {
|
536
|
+
new_ija[i+n] = old_ija[i];
|
537
|
+
}
|
538
|
+
} else {
|
539
|
+
for (size_t i = pos; i < current_size; ++i) {
|
540
|
+
new_ija[i+n] = old_ija[i];
|
541
|
+
new_a[i+n] = old_a[i];
|
542
|
+
}
|
543
|
+
}
|
544
|
+
|
545
|
+
s->capacity = new_capacity;
|
546
|
+
if (s->dtype == nm::RUBYOBJ)
|
547
|
+
nm_yale_storage_register_a(new_a, new_capacity);
|
548
|
+
|
549
|
+
NM_FREE(s->ija);
|
550
|
+
nm_yale_storage_unregister(s);
|
551
|
+
NM_FREE(s->a);
|
552
|
+
|
553
|
+
if (s->dtype == nm::RUBYOBJ)
|
554
|
+
nm_yale_storage_unregister_a(new_a, new_capacity);
|
555
|
+
|
556
|
+
s->ija = new_ija;
|
557
|
+
s->a = reinterpret_cast<void*>(new_a);
|
558
|
+
|
559
|
+
return 'i';
|
560
|
+
}
|
561
|
+
|
562
|
+
/*
|
563
|
+
* Insert a value or contiguous values in the ija and a vectors (after ja and
|
564
|
+
* diag). Does not free anything; you are responsible!
|
565
|
+
*
|
566
|
+
* TODO: Improve this so it can handle non-contiguous element insertions
|
567
|
+
* efficiently. For now, we can just sort the elements in the row in
|
568
|
+
* question.)
|
569
|
+
*/
|
570
|
+
template <typename DType>
|
571
|
+
static char vector_insert(YALE_STORAGE* s, size_t pos, size_t* j, void* val_, size_t n, bool struct_only) {
|
572
|
+
|
573
|
+
if (pos < s->shape[0]) {
|
574
|
+
rb_raise(rb_eArgError, "vector insert pos (%lu) is before beginning of ja (%lu); this should not happen", pos, s->shape[0]);
|
575
|
+
}
|
576
|
+
|
577
|
+
DType* val = reinterpret_cast<DType*>(val_);
|
578
|
+
|
579
|
+
size_t size = s->ija[s->shape[0]];
|
580
|
+
|
581
|
+
IType* ija = s->ija;
|
582
|
+
DType* a = reinterpret_cast<DType*>(s->a);
|
583
|
+
|
584
|
+
if (size + n > s->capacity) {
|
585
|
+
vector_insert_resize<DType>(s, size, pos, j, n, struct_only);
|
586
|
+
|
587
|
+
// Need to get the new locations for ija and a.
|
588
|
+
ija = s->ija;
|
589
|
+
a = reinterpret_cast<DType*>(s->a);
|
590
|
+
} else {
|
591
|
+
/*
|
592
|
+
* No resize required:
|
593
|
+
* easy (but somewhat slow), just copy elements to the tail, starting at
|
594
|
+
* the end, one element at a time.
|
595
|
+
*
|
596
|
+
* TODO: This can be made slightly more efficient, but only after the tests
|
597
|
+
* are written.
|
598
|
+
*/
|
599
|
+
|
600
|
+
if (struct_only) {
|
601
|
+
for (size_t i = 0; i < size - pos; ++i) {
|
602
|
+
ija[size+n-1-i] = ija[size-1-i];
|
603
|
+
}
|
604
|
+
} else {
|
605
|
+
for (size_t i = 0; i < size - pos; ++i) {
|
606
|
+
ija[size+n-1-i] = ija[size-1-i];
|
607
|
+
a[size+n-1-i] = a[size-1-i];
|
608
|
+
}
|
609
|
+
}
|
610
|
+
}
|
611
|
+
|
612
|
+
// Now insert the new values.
|
613
|
+
if (struct_only) {
|
614
|
+
for (size_t i = 0; i < n; ++i) {
|
615
|
+
ija[pos+i] = j[i];
|
616
|
+
}
|
617
|
+
} else {
|
618
|
+
for (size_t i = 0; i < n; ++i) {
|
619
|
+
ija[pos+i] = j[i];
|
620
|
+
a[pos+i] = val[i];
|
621
|
+
}
|
622
|
+
}
|
623
|
+
|
624
|
+
return 'i';
|
625
|
+
}
|
626
|
+
|
627
|
+
/*
|
628
|
+
* If we add n items to row i, we need to increment ija[i+1] and onward.
|
629
|
+
*/
|
630
|
+
static void increment_ia_after(YALE_STORAGE* s, IType ija_size, IType i, long n) {
|
631
|
+
IType* ija = s->ija;
|
632
|
+
|
633
|
+
++i;
|
634
|
+
for (; i <= ija_size; ++i) {
|
635
|
+
ija[i] += n;
|
636
|
+
}
|
637
|
+
}
|
638
|
+
|
639
|
+
/*
|
640
|
+
* Binary search for returning insertion points.
|
641
|
+
*/
|
642
|
+
static IType insert_search(YALE_STORAGE* s, IType left, IType right, IType key, bool& found) {
|
643
|
+
|
644
|
+
if (left > right) {
|
645
|
+
found = false;
|
646
|
+
return left;
|
647
|
+
}
|
648
|
+
|
649
|
+
IType* ija = s->ija;
|
650
|
+
IType mid = (left + right)/2;
|
651
|
+
IType mid_j = ija[mid];
|
652
|
+
|
653
|
+
if (mid_j == key) {
|
654
|
+
found = true;
|
655
|
+
return mid;
|
656
|
+
|
657
|
+
} else if (mid_j > key) {
|
658
|
+
return insert_search(s, left, mid-1, key, found);
|
659
|
+
|
660
|
+
} else {
|
661
|
+
return insert_search(s, mid+1, right, key, found);
|
662
|
+
}
|
663
|
+
}
|
664
|
+
|
665
|
+
/////////////////////////
|
666
|
+
// Copying and Casting //
|
667
|
+
/////////////////////////
|
668
|
+
|
669
|
+
/*
|
670
|
+
* Templated copy constructor for changing dtypes.
|
671
|
+
*/
|
672
|
+
template <typename L, typename R>
|
673
|
+
YALE_STORAGE* cast_copy(const YALE_STORAGE* rhs) {
|
674
|
+
YaleStorage<R> y(rhs);
|
675
|
+
return y.template alloc_copy<L>();
|
676
|
+
}
|
677
|
+
|
678
|
+
/*
|
679
|
+
* Template access for getting the size of Yale storage.
|
680
|
+
*/
|
681
|
+
size_t get_size(const YALE_STORAGE* storage) {
|
682
|
+
return storage->ija[ storage->shape[0] ];
|
683
|
+
}
|
684
|
+
|
685
|
+
|
686
|
+
template <typename DType>
|
687
|
+
static STORAGE* matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector) {
|
688
|
+
YALE_STORAGE *left = (YALE_STORAGE*)(casted_storage.left),
|
689
|
+
*right = (YALE_STORAGE*)(casted_storage.right);
|
690
|
+
|
691
|
+
nm_yale_storage_register(left);
|
692
|
+
nm_yale_storage_register(right);
|
693
|
+
// We can safely get dtype from the casted matrices; post-condition of binary_storage_cast_alloc is that dtype is the
|
694
|
+
// same for left and right.
|
695
|
+
// int8_t dtype = left->dtype;
|
696
|
+
|
697
|
+
IType* ijl = left->ija;
|
698
|
+
IType* ijr = right->ija;
|
699
|
+
|
700
|
+
// First, count the ndnz of the result.
|
701
|
+
// TODO: This basically requires running symbmm twice to get the exact ndnz size. That's frustrating. Are there simple
|
702
|
+
// cases where we can avoid running it?
|
703
|
+
size_t result_ndnz = nm::math::symbmm(resulting_shape[0], left->shape[1], resulting_shape[1], ijl, ijl, true, ijr, ijr, true, NULL, true);
|
704
|
+
|
705
|
+
// Create result storage.
|
706
|
+
YALE_STORAGE* result = nm_yale_storage_create(left->dtype, resulting_shape, 2, result_ndnz);
|
707
|
+
init<DType>(result, NULL);
|
708
|
+
IType* ija = result->ija;
|
709
|
+
|
710
|
+
// Symbolic multiplication step (build the structure)
|
711
|
+
nm::math::symbmm(resulting_shape[0], left->shape[1], resulting_shape[1], ijl, ijl, true, ijr, ijr, true, ija, true);
|
712
|
+
|
713
|
+
// Numeric multiplication step (fill in the elements)
|
714
|
+
|
715
|
+
nm::math::numbmm<DType>(result->shape[0], left->shape[1], result->shape[1],
|
716
|
+
ijl, ijl, reinterpret_cast<DType*>(left->a), true,
|
717
|
+
ijr, ijr, reinterpret_cast<DType*>(right->a), true,
|
718
|
+
ija, ija, reinterpret_cast<DType*>(result->a), true);
|
719
|
+
|
720
|
+
|
721
|
+
// Sort the columns
|
722
|
+
nm::math::smmp_sort_columns<DType>(result->shape[0], ija, ija, reinterpret_cast<DType*>(result->a));
|
723
|
+
|
724
|
+
nm_yale_storage_unregister(right);
|
725
|
+
nm_yale_storage_unregister(left);
|
726
|
+
return reinterpret_cast<STORAGE*>(result);
|
727
|
+
}
|
728
|
+
|
729
|
+
|
730
|
+
/*
|
731
|
+
* Get the sum of offsets from the original matrix (for sliced iteration).
|
732
|
+
*/
|
733
|
+
static std::array<size_t,2> get_offsets(YALE_STORAGE* x) {
|
734
|
+
std::array<size_t, 2> offsets{ {0,0} };
|
735
|
+
while (x != x->src) {
|
736
|
+
offsets[0] += x->offset[0];
|
737
|
+
offsets[1] += x->offset[1];
|
738
|
+
x = reinterpret_cast<YALE_STORAGE*>(x->src);
|
739
|
+
}
|
740
|
+
return offsets;
|
741
|
+
}
|
742
|
+
|
743
|
+
|
744
|
+
class RowIterator {
|
745
|
+
protected:
|
746
|
+
YALE_STORAGE* s;
|
747
|
+
IType* ija;
|
748
|
+
void* a;
|
749
|
+
IType i, k, k_end;
|
750
|
+
size_t j_offset, j_shape;
|
751
|
+
bool diag, End;
|
752
|
+
VALUE init;
|
753
|
+
public:
|
754
|
+
RowIterator(YALE_STORAGE* s_, IType* ija_, IType i_, size_t j_shape_, size_t j_offset_ = 0)
|
755
|
+
: s(s_),
|
756
|
+
ija(ija_),
|
757
|
+
a(reinterpret_cast<YALE_STORAGE*>(s->src)->a),
|
758
|
+
i(i_),
|
759
|
+
k(ija[i]),
|
760
|
+
k_end(ija[i+1]),
|
761
|
+
j_offset(j_offset_),
|
762
|
+
j_shape(j_shape_),
|
763
|
+
diag(row_has_no_nd() || diag_is_first()),
|
764
|
+
End(false),
|
765
|
+
init(default_value(s))
|
766
|
+
{ }
|
767
|
+
|
768
|
+
RowIterator(YALE_STORAGE* s_, IType i_, size_t j_shape_, size_t j_offset_ = 0)
|
769
|
+
: s(s_),
|
770
|
+
ija(IJA(s)),
|
771
|
+
a(reinterpret_cast<YALE_STORAGE*>(s->src)->a),
|
772
|
+
i(i_),
|
773
|
+
k(ija[i]),
|
774
|
+
k_end(ija[i+1]),
|
775
|
+
j_offset(j_offset_),
|
776
|
+
j_shape(j_shape_),
|
777
|
+
diag(row_has_no_nd() || diag_is_first()),
|
778
|
+
End(false),
|
779
|
+
init(default_value(s))
|
780
|
+
{ }
|
781
|
+
|
782
|
+
RowIterator(const RowIterator& rhs) : s(rhs.s), ija(rhs.ija), a(reinterpret_cast<YALE_STORAGE*>(s->src)->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) { }
|
783
|
+
|
784
|
+
VALUE obj() const {
|
785
|
+
return diag ? obj_at(s, i) : obj_at(s, k);
|
786
|
+
}
|
787
|
+
|
788
|
+
template <typename T>
|
789
|
+
T cobj() const {
|
790
|
+
if (typeid(T) == typeid(RubyObject)) return obj();
|
791
|
+
return A<T>(s)[diag ? i : k];
|
792
|
+
}
|
793
|
+
|
794
|
+
inline IType proper_j() const {
|
795
|
+
return diag ? i : ija[k];
|
796
|
+
}
|
797
|
+
|
798
|
+
inline IType offset_j() const {
|
799
|
+
return proper_j() - j_offset;
|
800
|
+
}
|
801
|
+
|
802
|
+
inline size_t capacity() const {
|
803
|
+
return reinterpret_cast<YALE_STORAGE*>(s->src)->capacity;
|
804
|
+
}
|
805
|
+
|
806
|
+
inline void vector_grow() {
|
807
|
+
YALE_STORAGE* src = reinterpret_cast<YALE_STORAGE*>(s->src);
|
808
|
+
nm::yale_storage::vector_grow(src);
|
809
|
+
ija = reinterpret_cast<IType*>(src->ija);
|
810
|
+
a = src->a;
|
811
|
+
}
|
812
|
+
|
813
|
+
/* Returns true if an additional value is inserted, false if it goes on the diagonal */
|
814
|
+
bool insert(IType j, VALUE v) {
|
815
|
+
if (j == i) { // insert regardless on diagonal
|
816
|
+
reinterpret_cast<VALUE*>(a)[j] = v;
|
817
|
+
return false;
|
818
|
+
|
819
|
+
} else {
|
820
|
+
if (rb_funcall(v, rb_intern("!="), 1, init) == Qtrue) {
|
821
|
+
if (k >= capacity()) {
|
822
|
+
vector_grow();
|
823
|
+
}
|
824
|
+
reinterpret_cast<VALUE*>(a)[k] = v;
|
825
|
+
ija[k] = j;
|
826
|
+
k++;
|
827
|
+
return true;
|
828
|
+
}
|
829
|
+
return false;
|
830
|
+
}
|
831
|
+
}
|
832
|
+
|
833
|
+
void update_row_end() {
|
834
|
+
ija[i+1] = k;
|
835
|
+
k_end = k;
|
836
|
+
}
|
837
|
+
|
838
|
+
/* Past the j_shape? */
|
839
|
+
inline bool end() const {
|
840
|
+
if (End) return true;
|
841
|
+
//if (diag) return i - j_offset >= j_shape;
|
842
|
+
//else return k >= s->capacity || ija[k] - j_offset >= j_shape;
|
843
|
+
return (int)(diag ? i : ija[k]) - (int)(j_offset) >= (int)(j_shape);
|
844
|
+
}
|
845
|
+
|
846
|
+
inline bool row_has_no_nd() const { return ija[i] == k_end; /* k_start == k_end */ }
|
847
|
+
inline bool diag_is_first() const { return i < ija[ija[i]]; }
|
848
|
+
inline bool diag_is_last() const { return i > ija[k_end-1]; } // only works if !row_has_no_nd()
|
849
|
+
inline bool k_is_last_nd() const { return k == k_end-1; }
|
850
|
+
inline bool k_is_last() const { return k_is_last_nd() && !diag_is_last(); }
|
851
|
+
inline bool diag_is_ahead() const { return i > ija[k]; }
|
852
|
+
inline bool row_has_diag() const { return i < s->shape[1]; }
|
853
|
+
inline bool diag_is_next() const { // assumes we've already tested for diag, row_has_no_nd(), diag_is_first()
|
854
|
+
if (i == ija[k]+1) return true; // definite next
|
855
|
+
else if (k+1 < k_end && i >= ija[k+1]+1) return false; // at least one item before it
|
856
|
+
else return true;
|
857
|
+
}
|
858
|
+
|
859
|
+
RowIterator& operator++() {
|
860
|
+
if (diag) { // we're at the diagonal
|
861
|
+
if (row_has_no_nd() || diag_is_last()) End = true; // and there are no non-diagonals (or none still to visit)
|
862
|
+
diag = false;
|
863
|
+
} else if (!row_has_diag()) { // row has no diagonal entries
|
864
|
+
if (row_has_no_nd() || k_is_last_nd()) End = true; // row is totally empty, or we're at last entry
|
865
|
+
else k++; // still entries to visit
|
866
|
+
} else { // not at diag but it exists somewhere in the row, and row has at least one nd entry
|
867
|
+
if (diag_is_ahead()) { // diag is ahead
|
868
|
+
if (k_is_last_nd()) diag = true; // diag is next and last
|
869
|
+
else if (diag_is_next()) { // diag is next and not last
|
870
|
+
diag = true;
|
871
|
+
k++;
|
872
|
+
} else k++; // diag is not next
|
873
|
+
} else { // diag is past
|
874
|
+
if (k_is_last_nd()) End = true; // and we're at the end
|
875
|
+
else k++; // and we're not at the end
|
876
|
+
}
|
877
|
+
}
|
878
|
+
|
879
|
+
return *this;
|
880
|
+
}
|
881
|
+
|
882
|
+
|
883
|
+
RowIterator operator++(int unused) {
|
884
|
+
RowIterator x(*this);
|
885
|
+
++(*this);
|
886
|
+
return x;
|
887
|
+
}
|
888
|
+
};
|
889
|
+
|
890
|
+
|
891
|
+
// Helper function used only for the RETURN_SIZED_ENUMERATOR macro. Returns the length of
|
892
|
+
// the matrix's storage.
|
893
|
+
static VALUE nm_yale_stored_enumerator_length(VALUE nmatrix) {
|
894
|
+
NM_CONSERVATIVE(nm_register_value(&nmatrix));
|
895
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nmatrix);
|
896
|
+
YALE_STORAGE* src = s->src == s ? s : reinterpret_cast<YALE_STORAGE*>(s->src);
|
897
|
+
size_t ia_size = src->shape[0];
|
898
|
+
// FIXME: This needs to be corrected for slicing.
|
899
|
+
size_t len = std::min( s->shape[0] + s->offset[0], s->shape[1] + s->offset[1] ) + nm_yale_storage_get_size(src) - ia_size;
|
900
|
+
NM_CONSERVATIVE(nm_unregister_value(&nmatrix));
|
901
|
+
return INT2FIX(len);
|
902
|
+
}
|
903
|
+
|
904
|
+
|
905
|
+
// Helper function used only for the RETURN_SIZED_ENUMERATOR macro. Returns the length of
|
906
|
+
// the matrix's storage.
|
907
|
+
static VALUE nm_yale_stored_nondiagonal_enumerator_length(VALUE nmatrix) {
|
908
|
+
NM_CONSERVATIVE(nm_register_value(&nmatrix));
|
909
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nmatrix);
|
910
|
+
if (s->src != s) s = reinterpret_cast<YALE_STORAGE*>(s->src); // need to get the original storage shape
|
911
|
+
|
912
|
+
size_t ia_size = s->shape[0];
|
913
|
+
size_t len = nm_yale_storage_get_size(NM_STORAGE_YALE(nmatrix)) - ia_size;
|
914
|
+
NM_CONSERVATIVE(nm_unregister_value(&nmatrix));
|
915
|
+
return INT2FIX(len);
|
916
|
+
}
|
917
|
+
|
918
|
+
// Helper function for diagonal length.
|
919
|
+
static VALUE nm_yale_stored_diagonal_enumerator_length(VALUE nmatrix) {
|
920
|
+
NM_CONSERVATIVE(nm_register_value(&nmatrix));
|
921
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nmatrix);
|
922
|
+
size_t len = std::min( s->shape[0] + s->offset[0], s->shape[1] + s->offset[1] );
|
923
|
+
NM_CONSERVATIVE(nm_unregister_value(&nmatrix));
|
924
|
+
return INT2FIX(len);
|
925
|
+
}
|
926
|
+
|
927
|
+
|
928
|
+
// Helper function for full enumerator length.
|
929
|
+
static VALUE nm_yale_enumerator_length(VALUE nmatrix) {
|
930
|
+
NM_CONSERVATIVE(nm_register_value(&nmatrix));
|
931
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nmatrix);
|
932
|
+
size_t len = s->shape[0] * s->shape[1];
|
933
|
+
NM_CONSERVATIVE(nm_unregister_value(&nmatrix));
|
934
|
+
return INT2FIX(len);
|
935
|
+
}
|
936
|
+
|
937
|
+
|
938
|
+
/*
|
939
|
+
* Map the stored values of a matrix in storage order.
|
940
|
+
*/
|
941
|
+
template <typename D>
|
942
|
+
static VALUE map_stored(VALUE self) {
|
943
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
944
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
945
|
+
YaleStorage<D> y(s);
|
946
|
+
|
947
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
948
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
949
|
+
RETURN_SIZED_ENUMERATOR(self, 0, 0, nm_yale_stored_enumerator_length);
|
950
|
+
|
951
|
+
YALE_STORAGE* r = y.template alloc_copy<nm::RubyObject, true>();
|
952
|
+
nm_yale_storage_register(r);
|
953
|
+
NMATRIX* m = nm_create(nm::YALE_STORE, reinterpret_cast<STORAGE*>(r));
|
954
|
+
VALUE to_return = Data_Wrap_Struct(CLASS_OF(self), nm_mark, nm_delete, m);
|
955
|
+
nm_yale_storage_unregister(r);
|
956
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
957
|
+
return to_return;
|
958
|
+
}
|
959
|
+
|
960
|
+
|
961
|
+
/*
|
962
|
+
* map_stored which visits the stored entries of two matrices in order.
|
963
|
+
*/
|
964
|
+
template <typename LD, typename RD>
|
965
|
+
static VALUE map_merged_stored(VALUE left, VALUE right, VALUE init) {
|
966
|
+
nm::YaleStorage<LD> l(NM_STORAGE_YALE(left));
|
967
|
+
nm::YaleStorage<RD> r(NM_STORAGE_YALE(right));
|
968
|
+
VALUE to_return = l.map_merged_stored(CLASS_OF(left), r, init);
|
969
|
+
return to_return;
|
970
|
+
}
|
971
|
+
|
972
|
+
|
973
|
+
/*
|
974
|
+
* Iterate over the stored entries in Yale (diagonal and non-diagonal non-zeros)
|
975
|
+
*/
|
976
|
+
template <typename DType>
|
977
|
+
static VALUE each_stored_with_indices(VALUE nm) {
|
978
|
+
NM_CONSERVATIVE(nm_register_value(&nm));
|
979
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
980
|
+
YaleStorage<DType> y(s);
|
981
|
+
|
982
|
+
// If we don't have a block, return an enumerator.
|
983
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
984
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
985
|
+
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_stored_enumerator_length);
|
986
|
+
|
987
|
+
for (typename YaleStorage<DType>::const_stored_diagonal_iterator d = y.csdbegin(); d != y.csdend(); ++d) {
|
988
|
+
rb_yield_values(3, ~d, d.rb_i(), d.rb_j());
|
989
|
+
}
|
990
|
+
|
991
|
+
for (typename YaleStorage<DType>::const_row_iterator it = y.cribegin(); it != y.criend(); ++it) {
|
992
|
+
for (auto jt = it.ndbegin(); jt != it.ndend(); ++jt) {
|
993
|
+
rb_yield_values(3, ~jt, it.rb_i(), jt.rb_j());
|
994
|
+
}
|
995
|
+
}
|
996
|
+
|
997
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
998
|
+
|
999
|
+
return nm;
|
1000
|
+
}
|
1001
|
+
|
1002
|
+
|
1003
|
+
/*
|
1004
|
+
* Iterate over the stored diagonal entries in Yale.
|
1005
|
+
*/
|
1006
|
+
template <typename DType>
|
1007
|
+
static VALUE stored_diagonal_each_with_indices(VALUE nm) {
|
1008
|
+
NM_CONSERVATIVE(nm_register_value(&nm));
|
1009
|
+
|
1010
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1011
|
+
YaleStorage<DType> y(s);
|
1012
|
+
|
1013
|
+
// If we don't have a block, return an enumerator.
|
1014
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
1015
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1016
|
+
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_stored_diagonal_length); // FIXME: need diagonal length
|
1017
|
+
|
1018
|
+
for (typename YaleStorage<DType>::const_stored_diagonal_iterator d = y.csdbegin(); d != y.csdend(); ++d) {
|
1019
|
+
rb_yield_values(3, ~d, d.rb_i(), d.rb_j());
|
1020
|
+
}
|
1021
|
+
|
1022
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1023
|
+
|
1024
|
+
return nm;
|
1025
|
+
}
|
1026
|
+
|
1027
|
+
|
1028
|
+
/*
|
1029
|
+
* Iterate over the stored diagonal entries in Yale.
|
1030
|
+
*/
|
1031
|
+
template <typename DType>
|
1032
|
+
static VALUE stored_nondiagonal_each_with_indices(VALUE nm) {
|
1033
|
+
NM_CONSERVATIVE(nm_register_value(&nm));
|
1034
|
+
|
1035
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1036
|
+
YaleStorage<DType> y(s);
|
1037
|
+
|
1038
|
+
// If we don't have a block, return an enumerator.
|
1039
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
1040
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1041
|
+
RETURN_SIZED_ENUMERATOR(nm, 0, 0, 0); // FIXME: need diagonal length
|
1042
|
+
|
1043
|
+
for (typename YaleStorage<DType>::const_row_iterator it = y.cribegin(); it != y.criend(); ++it) {
|
1044
|
+
for (auto jt = it.ndbegin(); jt != it.ndend(); ++jt) {
|
1045
|
+
rb_yield_values(3, ~jt, it.rb_i(), jt.rb_j());
|
1046
|
+
}
|
1047
|
+
}
|
1048
|
+
|
1049
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1050
|
+
|
1051
|
+
return nm;
|
1052
|
+
}
|
1053
|
+
|
1054
|
+
|
1055
|
+
/*
|
1056
|
+
* Iterate over the stored entries in Yale in order of i,j. Visits every diagonal entry, even if it's the default.
|
1057
|
+
*/
|
1058
|
+
template <typename DType>
|
1059
|
+
static VALUE each_ordered_stored_with_indices(VALUE nm) {
|
1060
|
+
NM_CONSERVATIVE(nm_register_value(&nm));
|
1061
|
+
|
1062
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1063
|
+
YaleStorage<DType> y(s);
|
1064
|
+
|
1065
|
+
// If we don't have a block, return an enumerator.
|
1066
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
1067
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1068
|
+
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_stored_enumerator_length);
|
1069
|
+
|
1070
|
+
for (typename YaleStorage<DType>::const_row_iterator it = y.cribegin(); it != y.criend(); ++it) {
|
1071
|
+
for (auto jt = it.begin(); jt != it.end(); ++jt) {
|
1072
|
+
rb_yield_values(3, ~jt, it.rb_i(), jt.rb_j());
|
1073
|
+
}
|
1074
|
+
}
|
1075
|
+
|
1076
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1077
|
+
|
1078
|
+
return nm;
|
1079
|
+
}
|
1080
|
+
|
1081
|
+
|
1082
|
+
template <typename DType>
|
1083
|
+
static VALUE each_with_indices(VALUE nm) {
|
1084
|
+
NM_CONSERVATIVE(nm_register_value(&nm));
|
1085
|
+
|
1086
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(nm);
|
1087
|
+
YaleStorage<DType> y(s);
|
1088
|
+
|
1089
|
+
// If we don't have a block, return an enumerator.
|
1090
|
+
RETURN_SIZED_ENUMERATOR_PRE
|
1091
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1092
|
+
RETURN_SIZED_ENUMERATOR(nm, 0, 0, nm_yale_enumerator_length);
|
1093
|
+
|
1094
|
+
for (typename YaleStorage<DType>::const_iterator iter = y.cbegin(); iter != y.cend(); ++iter) {
|
1095
|
+
rb_yield_values(3, ~iter, iter.rb_i(), iter.rb_j());
|
1096
|
+
}
|
1097
|
+
|
1098
|
+
NM_CONSERVATIVE(nm_unregister_value(&nm));
|
1099
|
+
|
1100
|
+
return nm;
|
1101
|
+
}
|
1102
|
+
|
1103
|
+
template <typename D>
|
1104
|
+
static bool is_pos_default_value(YALE_STORAGE* s, size_t apos) {
|
1105
|
+
YaleStorage<D> y(s);
|
1106
|
+
return y.is_pos_default_value(apos);
|
1107
|
+
}
|
1108
|
+
|
1109
|
+
} // end of namespace nm::yale_storage
|
1110
|
+
|
1111
|
+
} // end of namespace nm.
|
1112
|
+
|
1113
|
+
///////////////////
|
1114
|
+
// Ruby Bindings //
|
1115
|
+
///////////////////
|
1116
|
+
|
1117
|
+
/* These bindings are mostly only for debugging Yale. They are called from Init_nmatrix. */
|
1118
|
+
|
1119
|
+
extern "C" {
|
1120
|
+
|
1121
|
+
void nm_init_yale_functions() {
|
1122
|
+
/*
|
1123
|
+
* This module stores methods that are useful for debugging Yale matrices,
|
1124
|
+
* i.e. the ones with +:yale+ stype.
|
1125
|
+
*/
|
1126
|
+
cNMatrix_YaleFunctions = rb_define_module_under(cNMatrix, "YaleFunctions");
|
1127
|
+
|
1128
|
+
// Expert recommendation. Eventually this should go in a separate gem, or at least a separate module.
|
1129
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_row_keys_intersection", (METHOD)nm_row_keys_intersection, 3);
|
1130
|
+
|
1131
|
+
// Debugging functions.
|
1132
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ija", (METHOD)nm_ija, -1);
|
1133
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_a", (METHOD)nm_a, -1);
|
1134
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_size", (METHOD)nm_size, 0);
|
1135
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ia", (METHOD)nm_ia, 0);
|
1136
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_ja", (METHOD)nm_ja, 0);
|
1137
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_d", (METHOD)nm_d, -1);
|
1138
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_lu", (METHOD)nm_lu, 0);
|
1139
|
+
|
1140
|
+
rb_define_method(cNMatrix_YaleFunctions, "yale_nd_row", (METHOD)nm_nd_row, -1);
|
1141
|
+
|
1142
|
+
/* Document-const:
|
1143
|
+
* Defines the growth rate of the sparse NMatrix's size. Default is 1.5.
|
1144
|
+
*/
|
1145
|
+
rb_define_const(cNMatrix_YaleFunctions, "YALE_GROWTH_CONSTANT", rb_float_new(nm::yale_storage::GROWTH_CONSTANT));
|
1146
|
+
|
1147
|
+
// This is so the user can easily check the IType size, mostly for debugging.
|
1148
|
+
size_t itype_size = sizeof(IType);
|
1149
|
+
VALUE itype_dtype;
|
1150
|
+
if (itype_size == sizeof(uint64_t)) {
|
1151
|
+
itype_dtype = ID2SYM(rb_intern("int64"));
|
1152
|
+
} else if (itype_size == sizeof(uint32_t)) {
|
1153
|
+
itype_dtype = ID2SYM(rb_intern("int32"));
|
1154
|
+
} else if (itype_size == sizeof(uint16_t)) {
|
1155
|
+
itype_dtype = ID2SYM(rb_intern("int16"));
|
1156
|
+
} else {
|
1157
|
+
rb_raise(rb_eStandardError, "unhandled length for sizeof(IType): %lu; note that IType is probably defined as size_t", sizeof(IType));
|
1158
|
+
}
|
1159
|
+
rb_define_const(cNMatrix, "INDEX_DTYPE", itype_dtype);
|
1160
|
+
}
|
1161
|
+
|
1162
|
+
/////////////////
|
1163
|
+
// C ACCESSORS //
|
1164
|
+
/////////////////
|
1165
|
+
|
1166
|
+
/* C interface for NMatrix#each_with_indices (Yale) */
|
1167
|
+
VALUE nm_yale_each_with_indices(VALUE nmatrix) {
|
1168
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::each_with_indices, VALUE, VALUE)
|
1169
|
+
|
1170
|
+
return ttable[ NM_DTYPE(nmatrix) ](nmatrix);
|
1171
|
+
}
|
1172
|
+
|
1173
|
+
|
1174
|
+
/* C interface for NMatrix#each_stored_with_indices (Yale) */
|
1175
|
+
VALUE nm_yale_each_stored_with_indices(VALUE nmatrix) {
|
1176
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::each_stored_with_indices, VALUE, VALUE)
|
1177
|
+
|
1178
|
+
return ttable[ NM_DTYPE(nmatrix) ](nmatrix);
|
1179
|
+
}
|
1180
|
+
|
1181
|
+
|
1182
|
+
/* Iterate along stored diagonal (not actual diagonal!) */
|
1183
|
+
VALUE nm_yale_stored_diagonal_each_with_indices(VALUE nmatrix) {
|
1184
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::stored_diagonal_each_with_indices, VALUE, VALUE)
|
1185
|
+
|
1186
|
+
return ttable[ NM_DTYPE(nmatrix) ](nmatrix);
|
1187
|
+
}
|
1188
|
+
|
1189
|
+
/* Iterate through stored nondiagonal (not actual diagonal!) */
|
1190
|
+
VALUE nm_yale_stored_nondiagonal_each_with_indices(VALUE nmatrix) {
|
1191
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::stored_nondiagonal_each_with_indices, VALUE, VALUE)
|
1192
|
+
|
1193
|
+
return ttable[ NM_DTYPE(nmatrix) ](nmatrix);
|
1194
|
+
}
|
1195
|
+
|
1196
|
+
|
1197
|
+
/* C interface for NMatrix#each_ordered_stored_with_indices (Yale) */
|
1198
|
+
VALUE nm_yale_each_ordered_stored_with_indices(VALUE nmatrix) {
|
1199
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::each_ordered_stored_with_indices, VALUE, VALUE)
|
1200
|
+
|
1201
|
+
return ttable[ NM_DTYPE(nmatrix) ](nmatrix);
|
1202
|
+
}
|
1203
|
+
|
1204
|
+
|
1205
|
+
|
1206
|
+
/*
|
1207
|
+
* C accessor for inserting some value in a matrix (or replacing an existing cell).
|
1208
|
+
*/
|
1209
|
+
void nm_yale_storage_set(VALUE left, SLICE* slice, VALUE right) {
|
1210
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::set, void, VALUE left, SLICE* slice, VALUE right);
|
1211
|
+
|
1212
|
+
ttable[NM_DTYPE(left)](left, slice, right);
|
1213
|
+
}
|
1214
|
+
|
1215
|
+
|
1216
|
+
/*
|
1217
|
+
* Determine the number of non-diagonal non-zeros in a not-yet-created copy of a slice or matrix.
|
1218
|
+
*/
|
1219
|
+
static size_t yale_count_slice_copy_ndnz(const YALE_STORAGE* s, size_t* offset, size_t* shape) {
|
1220
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::count_slice_copy_ndnz, size_t, const YALE_STORAGE*, size_t*, size_t*)
|
1221
|
+
|
1222
|
+
return ttable[s->dtype](s, offset, shape);
|
1223
|
+
}
|
1224
|
+
|
1225
|
+
|
1226
|
+
/*
|
1227
|
+
* C accessor for yale_storage::get, which returns a slice of YALE_STORAGE object by copy
|
1228
|
+
*
|
1229
|
+
* Slicing-related.
|
1230
|
+
*/
|
1231
|
+
void* nm_yale_storage_get(const STORAGE* storage, SLICE* slice) {
|
1232
|
+
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1233
|
+
|
1234
|
+
if (slice->single) {
|
1235
|
+
NAMED_DTYPE_TEMPLATE_TABLE(elem_copy_table, nm::yale_storage::get_single, void*, YALE_STORAGE*, SLICE*)
|
1236
|
+
|
1237
|
+
return elem_copy_table[casted_storage->dtype](casted_storage, slice);
|
1238
|
+
} else {
|
1239
|
+
nm_yale_storage_register(casted_storage);
|
1240
|
+
//return reinterpret_cast<void*>(nm::YaleStorage<nm::dtype_enum_T<storage->dtype>::type>(casted_storage).alloc_ref(slice));
|
1241
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ref_table, nm::yale_storage::ref, YALE_STORAGE*, YALE_STORAGE* storage, SLICE* slice)
|
1242
|
+
|
1243
|
+
YALE_STORAGE* ref = ref_table[casted_storage->dtype](casted_storage, slice);
|
1244
|
+
|
1245
|
+
NAMED_LR_DTYPE_TEMPLATE_TABLE(slice_copy_table, nm::yale_storage::slice_copy, YALE_STORAGE*, YALE_STORAGE*)
|
1246
|
+
|
1247
|
+
YALE_STORAGE* ns = slice_copy_table[casted_storage->dtype][casted_storage->dtype](ref);
|
1248
|
+
|
1249
|
+
NM_FREE(ref);
|
1250
|
+
|
1251
|
+
nm_yale_storage_unregister(casted_storage);
|
1252
|
+
|
1253
|
+
return ns;
|
1254
|
+
}
|
1255
|
+
}
|
1256
|
+
|
1257
|
+
/*
|
1258
|
+
* C accessor for yale_storage::vector_insert
|
1259
|
+
*/
|
1260
|
+
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) {
|
1261
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::vector_insert, char, YALE_STORAGE*, size_t, size_t*, void*, size_t, bool);
|
1262
|
+
|
1263
|
+
return ttable[dtype](s, pos, js, vals, n, struct_only);
|
1264
|
+
}
|
1265
|
+
|
1266
|
+
/*
|
1267
|
+
* C accessor for yale_storage::increment_ia_after, typically called after ::vector_insert
|
1268
|
+
*/
|
1269
|
+
static void nm_yale_storage_increment_ia_after(YALE_STORAGE* s, size_t ija_size, size_t i, long n) {
|
1270
|
+
nm::yale_storage::increment_ia_after(s, ija_size, i, n);
|
1271
|
+
}
|
1272
|
+
|
1273
|
+
|
1274
|
+
/*
|
1275
|
+
* C accessor for yale_storage::ref, which returns either a pointer to the correct location in a YALE_STORAGE object
|
1276
|
+
* for some set of coordinates, or a pointer to a single element.
|
1277
|
+
*/
|
1278
|
+
void* nm_yale_storage_ref(const STORAGE* storage, SLICE* slice) {
|
1279
|
+
YALE_STORAGE* casted_storage = (YALE_STORAGE*)storage;
|
1280
|
+
|
1281
|
+
if (slice->single) {
|
1282
|
+
//return reinterpret_cast<void*>(nm::YaleStorage<nm::dtype_enum_T<storage->dtype>::type>(casted_storage).get_single_p(slice));
|
1283
|
+
NAMED_DTYPE_TEMPLATE_TABLE(elem_copy_table, nm::yale_storage::get_single, void*, YALE_STORAGE*, SLICE*)
|
1284
|
+
return elem_copy_table[casted_storage->dtype](casted_storage, slice);
|
1285
|
+
} else {
|
1286
|
+
//return reinterpret_cast<void*>(nm::YaleStorage<nm::dtype_enum_T<storage->dtype>::type>(casted_storage).alloc_ref(slice));
|
1287
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ref_table, nm::yale_storage::ref, YALE_STORAGE*, YALE_STORAGE* storage, SLICE* slice)
|
1288
|
+
return reinterpret_cast<void*>(ref_table[casted_storage->dtype](casted_storage, slice));
|
1289
|
+
|
1290
|
+
}
|
1291
|
+
}
|
1292
|
+
|
1293
|
+
|
1294
|
+
/*
|
1295
|
+
* C accessor for determining whether two YALE_STORAGE objects have the same contents.
|
1296
|
+
*/
|
1297
|
+
bool nm_yale_storage_eqeq(const STORAGE* left, const STORAGE* right) {
|
1298
|
+
NAMED_LR_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::eqeq, bool, const YALE_STORAGE* left, const YALE_STORAGE* right);
|
1299
|
+
|
1300
|
+
const YALE_STORAGE* casted_left = reinterpret_cast<const YALE_STORAGE*>(left);
|
1301
|
+
|
1302
|
+
return ttable[casted_left->dtype][right->dtype](casted_left, (const YALE_STORAGE*)right);
|
1303
|
+
}
|
1304
|
+
|
1305
|
+
|
1306
|
+
/*
|
1307
|
+
* Copy constructor for changing dtypes. (C accessor)
|
1308
|
+
*/
|
1309
|
+
STORAGE* nm_yale_storage_cast_copy(const STORAGE* rhs, nm::dtype_t new_dtype, void* dummy) {
|
1310
|
+
NAMED_LR_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::cast_copy, YALE_STORAGE*, const YALE_STORAGE* rhs);
|
1311
|
+
|
1312
|
+
const YALE_STORAGE* casted_rhs = reinterpret_cast<const YALE_STORAGE*>(rhs);
|
1313
|
+
//return reinterpret_cast<STORAGE*>(nm::YaleStorage<nm::dtype_enum_T< rhs->dtype >::type>(rhs).alloc_copy<nm::dtype_enum_T< new_dtype >::type>());
|
1314
|
+
return (STORAGE*)ttable[new_dtype][casted_rhs->dtype](casted_rhs);
|
1315
|
+
}
|
1316
|
+
|
1317
|
+
|
1318
|
+
/*
|
1319
|
+
* Returns size of Yale storage as a size_t (no matter what the itype is). (C accessor)
|
1320
|
+
*/
|
1321
|
+
size_t nm_yale_storage_get_size(const YALE_STORAGE* storage) {
|
1322
|
+
return nm::yale_storage::get_size(storage);
|
1323
|
+
}
|
1324
|
+
|
1325
|
+
|
1326
|
+
|
1327
|
+
/*
|
1328
|
+
* Return a pointer to the matrix's default value entry.
|
1329
|
+
*/
|
1330
|
+
static void* default_value_ptr(const YALE_STORAGE* s) {
|
1331
|
+
return reinterpret_cast<void*>(reinterpret_cast<char*>(((YALE_STORAGE*)(s->src))->a) + (((YALE_STORAGE*)(s->src))->shape[0] * DTYPE_SIZES[s->dtype]));
|
1332
|
+
}
|
1333
|
+
|
1334
|
+
/*
|
1335
|
+
* Return the Ruby object at a given location in storage.
|
1336
|
+
*/
|
1337
|
+
static VALUE obj_at(YALE_STORAGE* s, size_t k) {
|
1338
|
+
if (s->dtype == nm::RUBYOBJ) return reinterpret_cast<VALUE*>(((YALE_STORAGE*)(s->src))->a)[k];
|
1339
|
+
else return nm::rubyobj_from_cval(reinterpret_cast<void*>(reinterpret_cast<char*>(((YALE_STORAGE*)(s->src))->a) + k * DTYPE_SIZES[s->dtype]), s->dtype).rval;
|
1340
|
+
}
|
1341
|
+
|
1342
|
+
|
1343
|
+
/*
|
1344
|
+
* Return the matrix's default value as a Ruby VALUE.
|
1345
|
+
*/
|
1346
|
+
static VALUE default_value(const YALE_STORAGE* s) {
|
1347
|
+
if (s->dtype == nm::RUBYOBJ) return *reinterpret_cast<VALUE*>(default_value_ptr(s));
|
1348
|
+
else return nm::rubyobj_from_cval(default_value_ptr(s), s->dtype).rval;
|
1349
|
+
}
|
1350
|
+
|
1351
|
+
|
1352
|
+
/*
|
1353
|
+
* Check to see if a default value is some form of zero. Easy for non-Ruby object matrices, which should always be 0.
|
1354
|
+
*/
|
1355
|
+
static bool default_value_is_numeric_zero(const YALE_STORAGE* s) {
|
1356
|
+
return rb_funcall(default_value(s), rb_intern("=="), 1, INT2FIX(0)) == Qtrue;
|
1357
|
+
}
|
1358
|
+
|
1359
|
+
|
1360
|
+
|
1361
|
+
/*
|
1362
|
+
* Transposing copy constructor.
|
1363
|
+
*/
|
1364
|
+
STORAGE* nm_yale_storage_copy_transposed(const STORAGE* rhs_base) {
|
1365
|
+
YALE_STORAGE* rhs = (YALE_STORAGE*)rhs_base;
|
1366
|
+
NAMED_DTYPE_TEMPLATE_TABLE(transp, nm::yale_storage::copy_transposed, YALE_STORAGE*, YALE_STORAGE*)
|
1367
|
+
return (STORAGE*)(transp[rhs->dtype](rhs));
|
1368
|
+
}
|
1369
|
+
|
1370
|
+
/*
|
1371
|
+
* C accessor for multiplying two YALE_STORAGE matrices, which have already been casted to the same dtype.
|
1372
|
+
*
|
1373
|
+
* FIXME: There should be some mathematical way to determine the worst-case IType based on the input ITypes. Right now
|
1374
|
+
* it just uses the default.
|
1375
|
+
*/
|
1376
|
+
STORAGE* nm_yale_storage_matrix_multiply(const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector) {
|
1377
|
+
DTYPE_TEMPLATE_TABLE(nm::yale_storage::matrix_multiply, STORAGE*, const STORAGE_PAIR& casted_storage, size_t* resulting_shape, bool vector);
|
1378
|
+
|
1379
|
+
YALE_STORAGE* left = reinterpret_cast<YALE_STORAGE*>(casted_storage.left);
|
1380
|
+
YALE_STORAGE* right = reinterpret_cast<YALE_STORAGE*>(casted_storage.right);
|
1381
|
+
|
1382
|
+
if (!default_value_is_numeric_zero(left) || !default_value_is_numeric_zero(right)) {
|
1383
|
+
rb_raise(rb_eNotImpError, "matrix default value must be some form of zero (not false or nil) for multiplication");
|
1384
|
+
return NULL;
|
1385
|
+
}
|
1386
|
+
|
1387
|
+
return ttable[left->dtype](casted_storage, resulting_shape, vector);
|
1388
|
+
}
|
1389
|
+
|
1390
|
+
|
1391
|
+
///////////////
|
1392
|
+
// Lifecycle //
|
1393
|
+
///////////////
|
1394
|
+
|
1395
|
+
/*
|
1396
|
+
* C accessor function for creating a YALE_STORAGE object. Prior to calling this function, you MUST
|
1397
|
+
* allocate shape (should be size_t * 2) -- don't use use a regular size_t array!
|
1398
|
+
*
|
1399
|
+
* For this type, dim must always be 2. The final argument is the initial capacity with which to
|
1400
|
+
* create the storage.
|
1401
|
+
*/
|
1402
|
+
|
1403
|
+
YALE_STORAGE* nm_yale_storage_create(nm::dtype_t dtype, size_t* shape, size_t dim, size_t init_capacity) {
|
1404
|
+
if (dim != 2) {
|
1405
|
+
rb_raise(nm_eStorageTypeError, "yale supports only 2-dimensional matrices");
|
1406
|
+
}
|
1407
|
+
DTYPE_OBJECT_STATIC_TABLE(nm::YaleStorage, create, YALE_STORAGE*, size_t* shape, size_t init_capacity)
|
1408
|
+
return ttable[dtype](shape, init_capacity);
|
1409
|
+
}
|
1410
|
+
|
1411
|
+
/*
|
1412
|
+
* Destructor for yale storage (C-accessible).
|
1413
|
+
*/
|
1414
|
+
void nm_yale_storage_delete(STORAGE* s) {
|
1415
|
+
if (s) {
|
1416
|
+
YALE_STORAGE* storage = (YALE_STORAGE*)s;
|
1417
|
+
if (storage->count-- == 1) {
|
1418
|
+
NM_FREE(storage->shape);
|
1419
|
+
NM_FREE(storage->offset);
|
1420
|
+
NM_FREE(storage->ija);
|
1421
|
+
NM_FREE(storage->a);
|
1422
|
+
NM_FREE(storage);
|
1423
|
+
}
|
1424
|
+
}
|
1425
|
+
}
|
1426
|
+
|
1427
|
+
/*
|
1428
|
+
* Destructor for the yale storage ref
|
1429
|
+
*/
|
1430
|
+
void nm_yale_storage_delete_ref(STORAGE* s) {
|
1431
|
+
if (s) {
|
1432
|
+
YALE_STORAGE* storage = (YALE_STORAGE*)s;
|
1433
|
+
nm_yale_storage_delete( reinterpret_cast<STORAGE*>(storage->src) );
|
1434
|
+
NM_FREE(storage->shape);
|
1435
|
+
NM_FREE(storage->offset);
|
1436
|
+
NM_FREE(s);
|
1437
|
+
}
|
1438
|
+
}
|
1439
|
+
|
1440
|
+
/*
|
1441
|
+
* C accessor for yale_storage::init, a templated function.
|
1442
|
+
*
|
1443
|
+
* Initializes the IJA vector of the YALE_STORAGE matrix.
|
1444
|
+
*/
|
1445
|
+
void nm_yale_storage_init(YALE_STORAGE* s, void* init_val) {
|
1446
|
+
DTYPE_TEMPLATE_TABLE(nm::yale_storage::init, void, YALE_STORAGE*, void*);
|
1447
|
+
|
1448
|
+
ttable[s->dtype](s, init_val);
|
1449
|
+
}
|
1450
|
+
|
1451
|
+
|
1452
|
+
/*
|
1453
|
+
* Ruby GC mark function for YALE_STORAGE. C accessible.
|
1454
|
+
*/
|
1455
|
+
void nm_yale_storage_mark(STORAGE* storage_base) {
|
1456
|
+
YALE_STORAGE* storage = (YALE_STORAGE*)storage_base;
|
1457
|
+
|
1458
|
+
if (storage && storage->dtype == nm::RUBYOBJ) {
|
1459
|
+
|
1460
|
+
VALUE* a = (VALUE*)(storage->a);
|
1461
|
+
rb_gc_mark_locations(a, &(a[storage->capacity-1]));
|
1462
|
+
}
|
1463
|
+
}
|
1464
|
+
|
1465
|
+
void nm_yale_storage_register_a(void* a, size_t size) {
|
1466
|
+
nm_register_values(reinterpret_cast<VALUE*>(a), size);
|
1467
|
+
}
|
1468
|
+
|
1469
|
+
void nm_yale_storage_unregister_a(void* a, size_t size) {
|
1470
|
+
nm_unregister_values(reinterpret_cast<VALUE*>(a), size);
|
1471
|
+
}
|
1472
|
+
|
1473
|
+
void nm_yale_storage_register(const STORAGE* s) {
|
1474
|
+
const YALE_STORAGE* y = reinterpret_cast<const YALE_STORAGE*>(s);
|
1475
|
+
if (y->dtype == nm::RUBYOBJ) {
|
1476
|
+
nm_register_values(reinterpret_cast<VALUE*>(y->a), nm::yale_storage::get_size(y));
|
1477
|
+
}
|
1478
|
+
}
|
1479
|
+
|
1480
|
+
void nm_yale_storage_unregister(const STORAGE* s) {
|
1481
|
+
const YALE_STORAGE* y = reinterpret_cast<const YALE_STORAGE*>(s);
|
1482
|
+
if (y->dtype == nm::RUBYOBJ) {
|
1483
|
+
nm_unregister_values(reinterpret_cast<VALUE*>(y->a), nm::yale_storage::get_size(y));
|
1484
|
+
}
|
1485
|
+
}
|
1486
|
+
|
1487
|
+
/*
|
1488
|
+
* Allocates and initializes the basic struct (but not the IJA or A vectors).
|
1489
|
+
*
|
1490
|
+
* This function is ONLY used when creating from old yale.
|
1491
|
+
*/
|
1492
|
+
static YALE_STORAGE* alloc(nm::dtype_t dtype, size_t* shape, size_t dim) {
|
1493
|
+
YALE_STORAGE* s;
|
1494
|
+
|
1495
|
+
s = NM_ALLOC( YALE_STORAGE );
|
1496
|
+
|
1497
|
+
s->ndnz = 0;
|
1498
|
+
s->dtype = dtype;
|
1499
|
+
s->shape = shape;
|
1500
|
+
s->offset = NM_ALLOC_N(size_t, dim);
|
1501
|
+
for (size_t i = 0; i < dim; ++i)
|
1502
|
+
s->offset[i] = 0;
|
1503
|
+
s->dim = dim;
|
1504
|
+
s->src = reinterpret_cast<STORAGE*>(s);
|
1505
|
+
s->count = 1;
|
1506
|
+
|
1507
|
+
return s;
|
1508
|
+
}
|
1509
|
+
|
1510
|
+
YALE_STORAGE* nm_yale_storage_create_from_old_yale(nm::dtype_t dtype, size_t* shape, char* ia, char* ja, char* a, nm::dtype_t from_dtype) {
|
1511
|
+
NAMED_LR_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::create_from_old_yale, YALE_STORAGE*, nm::dtype_t dtype, size_t* shape, char* r_ia, char* r_ja, char* r_a);
|
1512
|
+
|
1513
|
+
return ttable[dtype][from_dtype](dtype, shape, ia, ja, a);
|
1514
|
+
|
1515
|
+
}
|
1516
|
+
|
1517
|
+
//////////////////////////////////////////////
|
1518
|
+
// YALE-SPECIFIC FUNCTIONS (RUBY ACCESSORS) //
|
1519
|
+
//////////////////////////////////////////////
|
1520
|
+
|
1521
|
+
/*
|
1522
|
+
* call-seq:
|
1523
|
+
* yale_size -> Integer
|
1524
|
+
*
|
1525
|
+
* Get the size of a Yale matrix (the number of elements actually stored).
|
1526
|
+
*
|
1527
|
+
* For capacity (the maximum number of elements that can be stored without a resize), use capacity instead.
|
1528
|
+
*/
|
1529
|
+
static VALUE nm_size(VALUE self) {
|
1530
|
+
YALE_STORAGE* s = (YALE_STORAGE*)(NM_SRC(self));
|
1531
|
+
VALUE to_return = INT2FIX(nm::yale_storage::IJA(s)[s->shape[0]]);
|
1532
|
+
return to_return;
|
1533
|
+
}
|
1534
|
+
|
1535
|
+
|
1536
|
+
/*
|
1537
|
+
* Determine if some pos in the diagonal is the default. No bounds checking!
|
1538
|
+
*/
|
1539
|
+
static bool is_pos_default_value(YALE_STORAGE* s, size_t apos) {
|
1540
|
+
DTYPE_TEMPLATE_TABLE(nm::yale_storage::is_pos_default_value, bool, YALE_STORAGE*, size_t)
|
1541
|
+
return ttable[s->dtype](s, apos);
|
1542
|
+
}
|
1543
|
+
|
1544
|
+
|
1545
|
+
/*
|
1546
|
+
* call-seq:
|
1547
|
+
* yale_row_keys_intersection(i, m2, i2) -> Array
|
1548
|
+
*
|
1549
|
+
* This function is experimental.
|
1550
|
+
*
|
1551
|
+
* It finds the intersection of row i of the current matrix with row i2 of matrix m2.
|
1552
|
+
* Both matrices must be Yale. They may not be slices.
|
1553
|
+
*
|
1554
|
+
* Only checks the stored indices; does not care about matrix default value.
|
1555
|
+
*/
|
1556
|
+
static VALUE nm_row_keys_intersection(VALUE m1, VALUE ii1, VALUE m2, VALUE ii2) {
|
1557
|
+
|
1558
|
+
NM_CONSERVATIVE(nm_register_value(&m1));
|
1559
|
+
NM_CONSERVATIVE(nm_register_value(&m2));
|
1560
|
+
|
1561
|
+
if (NM_SRC(m1) != NM_STORAGE(m1) || NM_SRC(m2) != NM_STORAGE(m2)) {
|
1562
|
+
NM_CONSERVATIVE(nm_unregister_value(&m2));
|
1563
|
+
NM_CONSERVATIVE(nm_unregister_value(&m1));
|
1564
|
+
rb_raise(rb_eNotImpError, "must be called on a real matrix and not a slice");
|
1565
|
+
}
|
1566
|
+
|
1567
|
+
size_t i1 = FIX2INT(ii1),
|
1568
|
+
i2 = FIX2INT(ii2);
|
1569
|
+
|
1570
|
+
YALE_STORAGE *s = NM_STORAGE_YALE(m1),
|
1571
|
+
*t = NM_STORAGE_YALE(m2);
|
1572
|
+
|
1573
|
+
size_t pos1 = s->ija[i1],
|
1574
|
+
pos2 = t->ija[i2];
|
1575
|
+
|
1576
|
+
size_t nextpos1 = s->ija[i1+1],
|
1577
|
+
nextpos2 = t->ija[i2+1];
|
1578
|
+
|
1579
|
+
size_t diff1 = nextpos1 - pos1,
|
1580
|
+
diff2 = nextpos2 - pos2;
|
1581
|
+
|
1582
|
+
// Does the diagonal have a nonzero in it?
|
1583
|
+
bool diag1 = i1 < s->shape[0] && !is_pos_default_value(s, i1),
|
1584
|
+
diag2 = i2 < t->shape[0] && !is_pos_default_value(t, i2);
|
1585
|
+
|
1586
|
+
// Reserve max(diff1,diff2) space -- that's the max intersection possible.
|
1587
|
+
VALUE ret = rb_ary_new2(std::max(diff1,diff2)+1);
|
1588
|
+
nm_register_value(&ret);
|
1589
|
+
|
1590
|
+
// Handle once the special case where both have the diagonal in exactly
|
1591
|
+
// the same place.
|
1592
|
+
if (diag1 && diag2 && i1 == i2) {
|
1593
|
+
rb_ary_push(ret, INT2FIX(i1));
|
1594
|
+
diag1 = false; diag2 = false; // no need to deal with diagonals anymore.
|
1595
|
+
}
|
1596
|
+
|
1597
|
+
// Now find the intersection.
|
1598
|
+
size_t idx1 = pos1, idx2 = pos2;
|
1599
|
+
while (idx1 < nextpos1 && idx2 < nextpos2) {
|
1600
|
+
if (s->ija[idx1] == t->ija[idx2]) {
|
1601
|
+
rb_ary_push(ret, INT2FIX(s->ija[idx1]));
|
1602
|
+
++idx1; ++idx2;
|
1603
|
+
} else if (diag1 && i1 == t->ija[idx2]) {
|
1604
|
+
rb_ary_push(ret, INT2FIX(i1));
|
1605
|
+
diag1 = false;
|
1606
|
+
++idx2;
|
1607
|
+
} else if (diag2 && i2 == s->ija[idx1]) {
|
1608
|
+
rb_ary_push(ret, INT2FIX(i2));
|
1609
|
+
diag2 = false;
|
1610
|
+
++idx1;
|
1611
|
+
} else if (s->ija[idx1] < t->ija[idx2]) {
|
1612
|
+
++idx1;
|
1613
|
+
} else { // s->ija[idx1] > t->ija[idx2]
|
1614
|
+
++idx2;
|
1615
|
+
}
|
1616
|
+
}
|
1617
|
+
|
1618
|
+
// Past the end of row i2's stored entries; need to try to find diagonal
|
1619
|
+
if (diag2 && idx1 < nextpos1) {
|
1620
|
+
idx1 = nm::yale_storage::binary_search_left_boundary(s, idx1, nextpos1, i2);
|
1621
|
+
if (s->ija[idx1] == i2) rb_ary_push(ret, INT2FIX(i2));
|
1622
|
+
}
|
1623
|
+
|
1624
|
+
// Find the diagonal, if possible, in the other one.
|
1625
|
+
if (diag1 && idx2 < nextpos2) {
|
1626
|
+
idx2 = nm::yale_storage::binary_search_left_boundary(t, idx2, nextpos2, i1);
|
1627
|
+
if (t->ija[idx2] == i1) rb_ary_push(ret, INT2FIX(i1));
|
1628
|
+
}
|
1629
|
+
|
1630
|
+
nm_unregister_value(&ret);
|
1631
|
+
NM_CONSERVATIVE(nm_unregister_value(&m1));
|
1632
|
+
NM_CONSERVATIVE(nm_unregister_value(&m2));
|
1633
|
+
|
1634
|
+
return ret;
|
1635
|
+
}
|
1636
|
+
|
1637
|
+
|
1638
|
+
/*
|
1639
|
+
* call-seq:
|
1640
|
+
* yale_a -> Array
|
1641
|
+
* yale_d(index) -> ...
|
1642
|
+
*
|
1643
|
+
* Get the A array of a Yale matrix (which stores the diagonal and the LU portions of the matrix).
|
1644
|
+
*/
|
1645
|
+
static VALUE nm_a(int argc, VALUE* argv, VALUE self) {
|
1646
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1647
|
+
|
1648
|
+
VALUE idx;
|
1649
|
+
rb_scan_args(argc, argv, "01", &idx);
|
1650
|
+
NM_CONSERVATIVE(nm_register_value(&idx));
|
1651
|
+
|
1652
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1653
|
+
size_t size = nm_yale_storage_get_size(s);
|
1654
|
+
|
1655
|
+
if (idx == Qnil) {
|
1656
|
+
|
1657
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, size);
|
1658
|
+
|
1659
|
+
nm_register_values(vals, size);
|
1660
|
+
|
1661
|
+
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
1662
|
+
for (size_t i = 0; i < size; ++i) {
|
1663
|
+
vals[i] = reinterpret_cast<VALUE*>(s->a)[i];
|
1664
|
+
}
|
1665
|
+
} else {
|
1666
|
+
for (size_t i = 0; i < size; ++i) {
|
1667
|
+
vals[i] = nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1668
|
+
}
|
1669
|
+
}
|
1670
|
+
VALUE ary = rb_ary_new4(size, vals);
|
1671
|
+
|
1672
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1673
|
+
rb_ary_push(ary, Qnil);
|
1674
|
+
|
1675
|
+
nm_unregister_values(vals, size);
|
1676
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1677
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1678
|
+
return ary;
|
1679
|
+
} else {
|
1680
|
+
size_t index = FIX2INT(idx);
|
1681
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1682
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1683
|
+
if (index >= size) rb_raise(rb_eRangeError, "out of range");
|
1684
|
+
return nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype] * index, s->dtype).rval;
|
1685
|
+
}
|
1686
|
+
}
|
1687
|
+
|
1688
|
+
|
1689
|
+
/*
|
1690
|
+
* call-seq:
|
1691
|
+
* yale_d -> Array
|
1692
|
+
* yale_d(index) -> ...
|
1693
|
+
*
|
1694
|
+
* Get the diagonal ("D") portion of the A array of a Yale matrix.
|
1695
|
+
*/
|
1696
|
+
static VALUE nm_d(int argc, VALUE* argv, VALUE self) {
|
1697
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1698
|
+
VALUE idx;
|
1699
|
+
rb_scan_args(argc, argv, "01", &idx);
|
1700
|
+
NM_CONSERVATIVE(nm_register_value(&idx));
|
1701
|
+
|
1702
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1703
|
+
|
1704
|
+
if (idx == Qnil) {
|
1705
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, s->shape[0]);
|
1706
|
+
|
1707
|
+
nm_register_values(vals, s->shape[0]);
|
1708
|
+
|
1709
|
+
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
1710
|
+
for (size_t i = 0; i < s->shape[0]; ++i) {
|
1711
|
+
vals[i] = reinterpret_cast<VALUE*>(s->a)[i];
|
1712
|
+
}
|
1713
|
+
} else {
|
1714
|
+
for (size_t i = 0; i < s->shape[0]; ++i) {
|
1715
|
+
vals[i] = nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*i, s->dtype).rval;
|
1716
|
+
}
|
1717
|
+
}
|
1718
|
+
nm_unregister_values(vals, s->shape[0]);
|
1719
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1720
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1721
|
+
|
1722
|
+
return rb_ary_new4(s->shape[0], vals);
|
1723
|
+
} else {
|
1724
|
+
size_t index = FIX2INT(idx);
|
1725
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1726
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1727
|
+
if (index >= s->shape[0]) rb_raise(rb_eRangeError, "out of range");
|
1728
|
+
return nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype] * index, s->dtype).rval;
|
1729
|
+
}
|
1730
|
+
}
|
1731
|
+
|
1732
|
+
/*
|
1733
|
+
* call-seq:
|
1734
|
+
* yale_lu -> Array
|
1735
|
+
*
|
1736
|
+
* Get the non-diagonal ("LU") portion of the A array of a Yale matrix.
|
1737
|
+
*/
|
1738
|
+
static VALUE nm_lu(VALUE self) {
|
1739
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1740
|
+
|
1741
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1742
|
+
|
1743
|
+
size_t size = nm_yale_storage_get_size(s);
|
1744
|
+
|
1745
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
1746
|
+
|
1747
|
+
nm_register_values(vals, size - s->shape[0] - 1);
|
1748
|
+
|
1749
|
+
if (NM_DTYPE(self) == nm::RUBYOBJ) {
|
1750
|
+
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
1751
|
+
vals[i] = reinterpret_cast<VALUE*>(s->a)[s->shape[0] + 1 + i];
|
1752
|
+
}
|
1753
|
+
} else {
|
1754
|
+
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
1755
|
+
vals[i] = nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*(s->shape[0] + 1 + i), s->dtype).rval;
|
1756
|
+
}
|
1757
|
+
}
|
1758
|
+
|
1759
|
+
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
1760
|
+
|
1761
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1762
|
+
rb_ary_push(ary, Qnil);
|
1763
|
+
|
1764
|
+
nm_unregister_values(vals, size - s->shape[0] - 1);
|
1765
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1766
|
+
|
1767
|
+
return ary;
|
1768
|
+
}
|
1769
|
+
|
1770
|
+
/*
|
1771
|
+
* call-seq:
|
1772
|
+
* yale_ia -> Array
|
1773
|
+
*
|
1774
|
+
* Get the IA portion of the IJA array of a Yale matrix. This gives the start and end positions of rows in the
|
1775
|
+
* JA and LU portions of the IJA and A arrays, respectively.
|
1776
|
+
*/
|
1777
|
+
static VALUE nm_ia(VALUE self) {
|
1778
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1779
|
+
|
1780
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1781
|
+
|
1782
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, s->shape[0] + 1);
|
1783
|
+
|
1784
|
+
for (size_t i = 0; i < s->shape[0] + 1; ++i) {
|
1785
|
+
vals[i] = INT2FIX(s->ija[i]);
|
1786
|
+
}
|
1787
|
+
|
1788
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1789
|
+
|
1790
|
+
return rb_ary_new4(s->shape[0]+1, vals);
|
1791
|
+
}
|
1792
|
+
|
1793
|
+
/*
|
1794
|
+
* call-seq:
|
1795
|
+
* yale_ja -> Array
|
1796
|
+
*
|
1797
|
+
* Get the JA portion of the IJA array of a Yale matrix. This gives the column indices for entries in corresponding
|
1798
|
+
* positions in the LU portion of the A array.
|
1799
|
+
*/
|
1800
|
+
static VALUE nm_ja(VALUE self) {
|
1801
|
+
|
1802
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1803
|
+
|
1804
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1805
|
+
|
1806
|
+
size_t size = nm_yale_storage_get_size(s);
|
1807
|
+
|
1808
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, size - s->shape[0] - 1);
|
1809
|
+
|
1810
|
+
nm_register_values(vals, size - s->shape[0] - 1);
|
1811
|
+
|
1812
|
+
for (size_t i = 0; i < size - s->shape[0] - 1; ++i) {
|
1813
|
+
vals[i] = INT2FIX(s->ija[s->shape[0] + 1 + i]);
|
1814
|
+
}
|
1815
|
+
|
1816
|
+
VALUE ary = rb_ary_new4(size - s->shape[0] - 1, vals);
|
1817
|
+
|
1818
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1819
|
+
rb_ary_push(ary, Qnil);
|
1820
|
+
|
1821
|
+
nm_unregister_values(vals, size - s->shape[0] - 1);
|
1822
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1823
|
+
|
1824
|
+
return ary;
|
1825
|
+
}
|
1826
|
+
|
1827
|
+
/*
|
1828
|
+
* call-seq:
|
1829
|
+
* yale_ija -> Array
|
1830
|
+
* yale_ija(index) -> ...
|
1831
|
+
*
|
1832
|
+
* Get the IJA array of a Yale matrix (or a component of the IJA array).
|
1833
|
+
*/
|
1834
|
+
static VALUE nm_ija(int argc, VALUE* argv, VALUE self) {
|
1835
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1836
|
+
|
1837
|
+
VALUE idx;
|
1838
|
+
rb_scan_args(argc, argv, "01", &idx);
|
1839
|
+
NM_CONSERVATIVE(nm_register_value(&idx));
|
1840
|
+
|
1841
|
+
YALE_STORAGE* s = reinterpret_cast<YALE_STORAGE*>(NM_SRC(self));
|
1842
|
+
size_t size = nm_yale_storage_get_size(s);
|
1843
|
+
|
1844
|
+
if (idx == Qnil) {
|
1845
|
+
|
1846
|
+
VALUE* vals = NM_ALLOCA_N(VALUE, size);
|
1847
|
+
|
1848
|
+
nm_register_values(vals, size);
|
1849
|
+
|
1850
|
+
for (size_t i = 0; i < size; ++i) {
|
1851
|
+
vals[i] = INT2FIX(s->ija[i]);
|
1852
|
+
}
|
1853
|
+
|
1854
|
+
VALUE ary = rb_ary_new4(size, vals);
|
1855
|
+
|
1856
|
+
for (size_t i = size; i < s->capacity; ++i)
|
1857
|
+
rb_ary_push(ary, Qnil);
|
1858
|
+
|
1859
|
+
nm_unregister_values(vals, size);
|
1860
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1861
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1862
|
+
|
1863
|
+
return ary;
|
1864
|
+
|
1865
|
+
} else {
|
1866
|
+
size_t index = FIX2INT(idx);
|
1867
|
+
if (index >= size) rb_raise(rb_eRangeError, "out of range");
|
1868
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1869
|
+
NM_CONSERVATIVE(nm_unregister_value(&idx));
|
1870
|
+
return INT2FIX(s->ija[index]);
|
1871
|
+
}
|
1872
|
+
}
|
1873
|
+
|
1874
|
+
|
1875
|
+
/*
|
1876
|
+
* call-seq:
|
1877
|
+
* yale_nd_row -> ...
|
1878
|
+
*
|
1879
|
+
* This function gets the non-diagonal contents of a Yale matrix row.
|
1880
|
+
* The first argument should be the row index. The optional second argument may be :hash or :keys, but defaults
|
1881
|
+
* to :hash. If :keys is given, it will only return the Hash keys (the column indices).
|
1882
|
+
*
|
1883
|
+
* This function is meant to accomplish its purpose as efficiently as possible. It does not check for appropriate
|
1884
|
+
* range.
|
1885
|
+
*/
|
1886
|
+
static VALUE nm_nd_row(int argc, VALUE* argv, VALUE self) {
|
1887
|
+
|
1888
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1889
|
+
if (NM_SRC(self) != NM_STORAGE(self)) {
|
1890
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1891
|
+
rb_raise(rb_eNotImpError, "must be called on a real matrix and not a slice");
|
1892
|
+
}
|
1893
|
+
|
1894
|
+
VALUE i_, as;
|
1895
|
+
rb_scan_args(argc, argv, "11", &i_, &as);
|
1896
|
+
NM_CONSERVATIVE(nm_register_value(&as));
|
1897
|
+
NM_CONSERVATIVE(nm_register_value(&i_));
|
1898
|
+
|
1899
|
+
bool keys = false;
|
1900
|
+
if (as != Qnil && rb_to_id(as) != nm_rb_hash) keys = true;
|
1901
|
+
|
1902
|
+
size_t i = FIX2INT(i_);
|
1903
|
+
|
1904
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1905
|
+
//nm::dtype_t dtype = NM_DTYPE(self);
|
1906
|
+
|
1907
|
+
if (i >= s->shape[0]) {
|
1908
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1909
|
+
NM_CONSERVATIVE(nm_unregister_value(&as));
|
1910
|
+
NM_CONSERVATIVE(nm_unregister_value(&i_));
|
1911
|
+
rb_raise(rb_eRangeError, "out of range (%lu >= %lu)", i, s->shape[0]);
|
1912
|
+
}
|
1913
|
+
|
1914
|
+
size_t pos = s->ija[i];
|
1915
|
+
size_t nextpos = s->ija[i+1];
|
1916
|
+
size_t diff = nextpos - pos;
|
1917
|
+
|
1918
|
+
VALUE ret;
|
1919
|
+
if (keys) {
|
1920
|
+
ret = rb_ary_new3(diff);
|
1921
|
+
|
1922
|
+
for (size_t idx = pos; idx < nextpos; ++idx) {
|
1923
|
+
rb_ary_store(ret, idx - pos, INT2FIX(s->ija[idx]));
|
1924
|
+
}
|
1925
|
+
|
1926
|
+
} else {
|
1927
|
+
ret = rb_hash_new();
|
1928
|
+
|
1929
|
+
for (size_t idx = pos; idx < nextpos; ++idx) {
|
1930
|
+
rb_hash_aset(ret, INT2FIX(s->ija[idx]), nm::rubyobj_from_cval((char*)(s->a) + DTYPE_SIZES[s->dtype]*idx, s->dtype).rval);
|
1931
|
+
}
|
1932
|
+
}
|
1933
|
+
NM_CONSERVATIVE(nm_unregister_value(&as));
|
1934
|
+
NM_CONSERVATIVE(nm_unregister_value(&i_));
|
1935
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1936
|
+
return ret;
|
1937
|
+
}
|
1938
|
+
|
1939
|
+
/*
|
1940
|
+
* call-seq:
|
1941
|
+
* yale_vector_set(i, column_index_array, cell_contents_array, pos) -> Fixnum
|
1942
|
+
*
|
1943
|
+
* Insert at position pos an array of non-diagonal elements with column indices given. Note that the column indices and values
|
1944
|
+
* must be storage-contiguous -- that is, you can't insert them around existing elements in some row, only amid some
|
1945
|
+
* elements in some row. You *can* insert them around a diagonal element, since this is stored separately. This function
|
1946
|
+
* may not be used for the insertion of diagonal elements in most cases, as these are already present in the data
|
1947
|
+
* structure and are typically modified by replacement rather than insertion.
|
1948
|
+
*
|
1949
|
+
* The last argument, pos, may be nil if you want to insert at the beginning of a row. Otherwise it needs to be provided.
|
1950
|
+
* Don't expect this function to know the difference. It really does very little checking, because its goal is to make
|
1951
|
+
* multiple contiguous insertion as quick as possible.
|
1952
|
+
*
|
1953
|
+
* You should also not attempt to insert values which are the default (0). These are not supposed to be stored, and may
|
1954
|
+
* lead to undefined behavior.
|
1955
|
+
*
|
1956
|
+
* Example:
|
1957
|
+
* m.yale_vector_set(3, [0,3,4], [1,1,1], 15)
|
1958
|
+
*
|
1959
|
+
* The example above inserts the values 1, 1, and 1 in columns 0, 3, and 4, assumed to be located at position 15 (which
|
1960
|
+
* corresponds to row 3).
|
1961
|
+
*
|
1962
|
+
* Example:
|
1963
|
+
* next = m.yale_vector_set(3, [0,3,4], [1,1,1])
|
1964
|
+
*
|
1965
|
+
* This example determines that i=3 is at position 15 automatically. The value returned, next, is the position where the
|
1966
|
+
* next value(s) should be inserted.
|
1967
|
+
*/
|
1968
|
+
VALUE nm_vector_set(int argc, VALUE* argv, VALUE self) { //, VALUE i_, VALUE jv, VALUE vv, VALUE pos_) {
|
1969
|
+
|
1970
|
+
NM_CONSERVATIVE(nm_register_value(&self));
|
1971
|
+
|
1972
|
+
if (NM_SRC(self) != NM_STORAGE(self)) {
|
1973
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1974
|
+
rb_raise(rb_eNotImpError, "must be called on a real matrix and not a slice");
|
1975
|
+
}
|
1976
|
+
|
1977
|
+
// i, jv, vv are mandatory; pos is optional; thus "31"
|
1978
|
+
VALUE i_, jv, vv, pos_;
|
1979
|
+
rb_scan_args(argc, argv, "31", &i_, &jv, &vv, &pos_);
|
1980
|
+
NM_CONSERVATIVE(nm_register_value(&i_));
|
1981
|
+
NM_CONSERVATIVE(nm_register_value(&jv));
|
1982
|
+
NM_CONSERVATIVE(nm_register_value(&vv));
|
1983
|
+
NM_CONSERVATIVE(nm_register_value(&pos_));
|
1984
|
+
|
1985
|
+
size_t len = RARRAY_LEN(jv); // need length in order to read the arrays in
|
1986
|
+
size_t vvlen = RARRAY_LEN(vv);
|
1987
|
+
|
1988
|
+
if (len != vvlen) {
|
1989
|
+
NM_CONSERVATIVE(nm_unregister_value(&pos_));
|
1990
|
+
NM_CONSERVATIVE(nm_unregister_value(&vv));
|
1991
|
+
NM_CONSERVATIVE(nm_unregister_value(&jv));
|
1992
|
+
NM_CONSERVATIVE(nm_unregister_value(&i_));
|
1993
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
1994
|
+
rb_raise(rb_eArgError, "lengths must match between j array (%lu) and value array (%lu)", len, vvlen);
|
1995
|
+
}
|
1996
|
+
|
1997
|
+
YALE_STORAGE* s = NM_STORAGE_YALE(self);
|
1998
|
+
nm::dtype_t dtype = NM_DTYPE(self);
|
1999
|
+
|
2000
|
+
size_t i = FIX2INT(i_); // get the row
|
2001
|
+
size_t pos = s->ija[i];
|
2002
|
+
|
2003
|
+
// Allocate the j array and the values array
|
2004
|
+
size_t* j = NM_ALLOCA_N(size_t, len);
|
2005
|
+
void* vals = NM_ALLOCA_N(char, DTYPE_SIZES[dtype] * len);
|
2006
|
+
if (dtype == nm::RUBYOBJ){
|
2007
|
+
nm_register_values(reinterpret_cast<VALUE*>(vals), len);
|
2008
|
+
}
|
2009
|
+
|
2010
|
+
// Copy array contents
|
2011
|
+
for (size_t idx = 0; idx < len; ++idx) {
|
2012
|
+
j[idx] = FIX2INT(rb_ary_entry(jv, idx));
|
2013
|
+
rubyval_to_cval(rb_ary_entry(vv, idx), dtype, (char*)vals + idx * DTYPE_SIZES[dtype]);
|
2014
|
+
}
|
2015
|
+
|
2016
|
+
nm_yale_storage_vector_insert(s, pos, j, vals, len, false, dtype);
|
2017
|
+
nm_yale_storage_increment_ia_after(s, s->shape[0], i, len);
|
2018
|
+
s->ndnz += len;
|
2019
|
+
|
2020
|
+
if (dtype == nm::RUBYOBJ){
|
2021
|
+
nm_unregister_values(reinterpret_cast<VALUE*>(vals), len);
|
2022
|
+
}
|
2023
|
+
|
2024
|
+
NM_CONSERVATIVE(nm_unregister_value(&pos_));
|
2025
|
+
NM_CONSERVATIVE(nm_unregister_value(&vv));
|
2026
|
+
NM_CONSERVATIVE(nm_unregister_value(&jv));
|
2027
|
+
NM_CONSERVATIVE(nm_unregister_value(&i_));
|
2028
|
+
NM_CONSERVATIVE(nm_unregister_value(&self));
|
2029
|
+
|
2030
|
+
// Return the updated position
|
2031
|
+
pos += len;
|
2032
|
+
return INT2FIX(pos);
|
2033
|
+
}
|
2034
|
+
|
2035
|
+
|
2036
|
+
|
2037
|
+
|
2038
|
+
/*
|
2039
|
+
* call-seq:
|
2040
|
+
* __yale_default_value__ -> ...
|
2041
|
+
*
|
2042
|
+
* Get the default_value property from a yale matrix.
|
2043
|
+
*/
|
2044
|
+
VALUE nm_yale_default_value(VALUE self) {
|
2045
|
+
VALUE to_return = default_value(NM_STORAGE_YALE(self));
|
2046
|
+
return to_return;
|
2047
|
+
}
|
2048
|
+
|
2049
|
+
|
2050
|
+
/*
|
2051
|
+
* call-seq:
|
2052
|
+
* __yale_map_merged_stored__(right) -> Enumerator
|
2053
|
+
*
|
2054
|
+
* A map operation on two Yale matrices which only iterates across the stored indices.
|
2055
|
+
*/
|
2056
|
+
VALUE nm_yale_map_merged_stored(VALUE left, VALUE right, VALUE init) {
|
2057
|
+
NAMED_LR_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::map_merged_stored, VALUE, VALUE, VALUE, VALUE)
|
2058
|
+
return ttable[NM_DTYPE(left)][NM_DTYPE(right)](left, right, init);
|
2059
|
+
//return nm::yale_storage::map_merged_stored(left, right, init);
|
2060
|
+
}
|
2061
|
+
|
2062
|
+
|
2063
|
+
/*
|
2064
|
+
* call-seq:
|
2065
|
+
* __yale_map_stored__ -> Enumerator
|
2066
|
+
*
|
2067
|
+
* A map operation on two Yale matrices which only iterates across the stored indices.
|
2068
|
+
*/
|
2069
|
+
VALUE nm_yale_map_stored(VALUE self) {
|
2070
|
+
NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::yale_storage::map_stored, VALUE, VALUE)
|
2071
|
+
return ttable[NM_DTYPE(self)](self);
|
2072
|
+
}
|
2073
|
+
|
2074
|
+
} // end of extern "C" block
|