1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_REDUX_H
#define EIGEN_REDUX_H
namespace Eigen {
namespace internal {
// TODO
// * implement other kind of vectorization
// * factorize code
/***************************************************************************
* Part 1 : the logic deciding a strategy for vectorization and unrolling
***************************************************************************/
template<typename Func, typename Derived>
struct redux_traits
{
public:
enum {
PacketSize = packet_traits<typename Derived::Scalar>::size,
InnerMaxSize = int(Derived::IsRowMajor)
? Derived::MaxColsAtCompileTime
: Derived::MaxRowsAtCompileTime
};
enum {
MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit)
&& (functor_traits<Func>::PacketAccess),
MayLinearVectorize = MightVectorize && (int(Derived::Flags)&LinearAccessBit),
MaySliceVectorize = MightVectorize && int(InnerMaxSize)>=3*PacketSize
};
public:
enum {
Traversal = int(MayLinearVectorize) ? int(LinearVectorizedTraversal)
: int(MaySliceVectorize) ? int(SliceVectorizedTraversal)
: int(DefaultTraversal)
};
public:
enum {
Cost = ( Derived::SizeAtCompileTime == Dynamic
|| Derived::CoeffReadCost == Dynamic
|| (Derived::SizeAtCompileTime!=1 && functor_traits<Func>::Cost == Dynamic)
) ? Dynamic
: Derived::SizeAtCompileTime * Derived::CoeffReadCost
+ (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize))
};
public:
enum {
Unrolling = Cost != Dynamic && Cost <= UnrollingLimit
? CompleteUnrolling
: NoUnrolling
};
};
/***************************************************************************
* Part 2 : unrollers
***************************************************************************/
/*** no vectorization ***/
template<typename Func, typename Derived, int Start, int Length>
struct redux_novec_unroller
{
enum {
HalfLength = Length/2
};
typedef typename Derived::Scalar Scalar;
EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
{
return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
redux_novec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func));
}
};
template<typename Func, typename Derived, int Start>
struct redux_novec_unroller<Func, Derived, Start, 1>
{
enum {
outer = Start / Derived::InnerSizeAtCompileTime,
inner = Start % Derived::InnerSizeAtCompileTime
};
typedef typename Derived::Scalar Scalar;
EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
{
return mat.coeffByOuterInner(outer, inner);
}
};
// This is actually dead code and will never be called. It is required
// to prevent false warnings regarding failed inlining though
// for 0 length run() will never be called at all.
template<typename Func, typename Derived, int Start>
struct redux_novec_unroller<Func, Derived, Start, 0>
{
typedef typename Derived::Scalar Scalar;
EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
};
/*** vectorization ***/
template<typename Func, typename Derived, int Start, int Length>
struct redux_vec_unroller
{
enum {
PacketSize = packet_traits<typename Derived::Scalar>::size,
HalfLength = Length/2
};
typedef typename Derived::Scalar Scalar;
typedef typename packet_traits<Scalar>::type PacketScalar;
static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func)
{
return func.packetOp(
redux_vec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
redux_vec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func) );
}
};
template<typename Func, typename Derived, int Start>
struct redux_vec_unroller<Func, Derived, Start, 1>
{
enum {
index = Start * packet_traits<typename Derived::Scalar>::size,
outer = index / int(Derived::InnerSizeAtCompileTime),
inner = index % int(Derived::InnerSizeAtCompileTime),
alignment = (Derived::Flags & AlignedBit) ? Aligned : Unaligned
};
typedef typename Derived::Scalar Scalar;
typedef typename packet_traits<Scalar>::type PacketScalar;
static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
{
return mat.template packetByOuterInner<alignment>(outer, inner);
}
};
/***************************************************************************
* Part 3 : implementation of all cases
***************************************************************************/
template<typename Func, typename Derived,
int Traversal = redux_traits<Func, Derived>::Traversal,
int Unrolling = redux_traits<Func, Derived>::Unrolling
>
struct redux_impl;
template<typename Func, typename Derived>
struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
{
typedef typename Derived::Scalar Scalar;
typedef typename Derived::Index Index;
EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
{
eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
Scalar res;
res = mat.coeffByOuterInner(0, 0);
for(Index i = 1; i < mat.innerSize(); ++i)
res = func(res, mat.coeffByOuterInner(0, i));
for(Index i = 1; i < mat.outerSize(); ++i)
for(Index j = 0; j < mat.innerSize(); ++j)
res = func(res, mat.coeffByOuterInner(i, j));
return res;
}
};
template<typename Func, typename Derived>
struct redux_impl<Func,Derived, DefaultTraversal, CompleteUnrolling>
: public redux_novec_unroller<Func,Derived, 0, Derived::SizeAtCompileTime>
{};
template<typename Func, typename Derived>
struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
{
typedef typename Derived::Scalar Scalar;
typedef typename packet_traits<Scalar>::type PacketScalar;
typedef typename Derived::Index Index;
static Scalar run(const Derived &mat, const Func& func)
{
const Index size = mat.size();
const Index packetSize = packet_traits<Scalar>::size;
const Index alignedStart = internal::first_aligned(mat);
enum {
alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit)
? Aligned : Unaligned
};
const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
const Index alignedEnd2 = alignedStart + alignedSize2;
const Index alignedEnd = alignedStart + alignedSize;
Scalar res;
if(alignedSize)
{
PacketScalar packet_res0 = mat.template packet<alignment>(alignedStart);
if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
{
PacketScalar packet_res1 = mat.template packet<alignment>(alignedStart+packetSize);
for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
{
packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(index));
packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment>(index+packetSize));
}
packet_res0 = func.packetOp(packet_res0,packet_res1);
if(alignedEnd>alignedEnd2)
packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment>(alignedEnd2));
}
res = func.predux(packet_res0);
for(Index index = 0; index < alignedStart; ++index)
res = func(res,mat.coeff(index));
for(Index index = alignedEnd; index < size; ++index)
res = func(res,mat.coeff(index));
}
else // too small to vectorize anything.
// since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
{
res = mat.coeff(0);
for(Index index = 1; index < size; ++index)
res = func(res,mat.coeff(index));
}
return res;
}
};
template<typename Func, typename Derived>
struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
{
typedef typename Derived::Scalar Scalar;
typedef typename packet_traits<Scalar>::type PacketScalar;
typedef typename Derived::Index Index;
static Scalar run(const Derived &mat, const Func& func)
{
eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
const Index innerSize = mat.innerSize();
const Index outerSize = mat.outerSize();
enum {
packetSize = packet_traits<Scalar>::size
};
const Index packetedInnerSize = ((innerSize)/packetSize)*packetSize;
Scalar res;
if(packetedInnerSize)
{
PacketScalar packet_res = mat.template packet<Unaligned>(0,0);
for(Index j=0; j<outerSize; ++j)
for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned>(j,i));
res = func.predux(packet_res);
for(Index j=0; j<outerSize; ++j)
for(Index i=packetedInnerSize; i<innerSize; ++i)
res = func(res, mat.coeffByOuterInner(j,i));
}
else // too small to vectorize anything.
// since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
{
res = redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>::run(mat, func);
}
return res;
}
};
template<typename Func, typename Derived>
struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
{
typedef typename Derived::Scalar Scalar;
typedef typename packet_traits<Scalar>::type PacketScalar;
enum {
PacketSize = packet_traits<Scalar>::size,
Size = Derived::SizeAtCompileTime,
VectorizedSize = (Size / PacketSize) * PacketSize
};
static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
{
Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
if (VectorizedSize != Size)
res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
return res;
}
};
#ifdef EIGEN_ENABLE_EVALUATORS
// evaluator adaptor
template<typename XprType>
class redux_evaluator
{
public:
redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
typedef typename XprType::Index Index;
typedef typename XprType::Scalar Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename XprType::PacketScalar PacketScalar;
typedef typename XprType::PacketReturnType PacketReturnType;
enum {
MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
// TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
Flags = XprType::Flags & ~DirectAccessBit,
IsRowMajor = XprType::IsRowMajor,
SizeAtCompileTime = XprType::SizeAtCompileTime,
InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
CoeffReadCost = XprType::CoeffReadCost
};
Index rows() const { return m_xpr.rows(); }
Index cols() const { return m_xpr.cols(); }
Index size() const { return m_xpr.size(); }
Index innerSize() const { return m_xpr.innerSize(); }
Index outerSize() const { return m_xpr.outerSize(); }
CoeffReturnType coeff(Index row, Index col) const
{ return m_evaluator.coeff(row, col); }
CoeffReturnType coeff(Index index) const
{ return m_evaluator.coeff(index); }
template<int LoadMode>
PacketReturnType packet(Index row, Index col) const
{ return m_evaluator.template packet<LoadMode>(row, col); }
template<int LoadMode>
PacketReturnType packet(Index index) const
{ return m_evaluator.template packet<LoadMode>(index); }
CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
{ return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
template<int LoadMode>
PacketReturnType packetByOuterInner(Index outer, Index inner) const
{ return m_evaluator.template packet<LoadMode>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
protected:
typename internal::evaluator<XprType>::nestedType m_evaluator;
const XprType &m_xpr;
};
#endif
} // end namespace internal
/***************************************************************************
* Part 4 : public API
***************************************************************************/
/** \returns the result of a full redux operation on the whole matrix or vector using \a func
*
* The template parameter \a BinaryOp is the type of the functor \a func which must be
* an associative operator. Both current C++98 and C++11 functor styles are handled.
*
* \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise()
*/
template<typename Derived>
template<typename Func>
EIGEN_STRONG_INLINE typename internal::result_of<Func(typename internal::traits<Derived>::Scalar)>::type
DenseBase<Derived>::redux(const Func& func) const
{
eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
#ifdef EIGEN_TEST_EVALUATORS
typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
ThisEvaluator thisEval(derived());
return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
#else
typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
return internal::redux_impl<Func, ThisNested>
::run(derived(), func);
#endif
}
/** \returns the minimum of all coefficients of \c *this.
* \warning the result is undefined if \c *this contains NaN.
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
DenseBase<Derived>::minCoeff() const
{
return this->redux(Eigen::internal::scalar_min_op<Scalar>());
}
/** \returns the maximum of all coefficients of \c *this.
* \warning the result is undefined if \c *this contains NaN.
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
DenseBase<Derived>::maxCoeff() const
{
return this->redux(Eigen::internal::scalar_max_op<Scalar>());
}
/** \returns the sum of all coefficients of *this
*
* \sa trace(), prod(), mean()
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
DenseBase<Derived>::sum() const
{
if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
return Scalar(0);
return this->redux(Eigen::internal::scalar_sum_op<Scalar>());
}
/** \returns the mean of all coefficients of *this
*
* \sa trace(), prod(), sum()
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
DenseBase<Derived>::mean() const
{
return Scalar(this->redux(Eigen::internal::scalar_sum_op<Scalar>())) / Scalar(this->size());
}
/** \returns the product of all coefficients of *this
*
* Example: \include MatrixBase_prod.cpp
* Output: \verbinclude MatrixBase_prod.out
*
* \sa sum(), mean(), trace()
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
DenseBase<Derived>::prod() const
{
if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
return Scalar(1);
return this->redux(Eigen::internal::scalar_product_op<Scalar>());
}
/** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal.
*
* \c *this can be any matrix, not necessarily square.
*
* \sa diagonal(), sum()
*/
template<typename Derived>
EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
MatrixBase<Derived>::trace() const
{
return derived().diagonal().sum();
}
} // end namespace Eigen
#endif // EIGEN_REDUX_H
|