• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2001 Intel Corporation
5 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
6 // Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
7 //
8 // This Source Code Form is subject to the terms of the Mozilla
9 // Public License v. 2.0. If a copy of the MPL was not distributed
10 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
11 
12 // The SSE code for the 4x4 float and double matrix inverse in this file
13 // comes from the following Intel's library:
14 // http://software.intel.com/en-us/articles/optimized-matrix-library-for-use-with-the-intel-pentiumr-4-processors-sse2-instructions/
15 //
16 // Here is the respective copyright and license statement:
17 //
18 //   Copyright (c) 2001 Intel Corporation.
19 //
20 // Permition is granted to use, copy, distribute and prepare derivative works
21 // of this library for any purpose and without fee, provided, that the above
22 // copyright notice and this statement appear in all copies.
23 // Intel makes no representations about the suitability of this software for
24 // any purpose, and specifically disclaims all warranties.
25 // See LEGAL.TXT for all the legal information.
26 
27 #ifndef EIGEN_INVERSE_SSE_H
28 #define EIGEN_INVERSE_SSE_H
29 
30 namespace Eigen {
31 
32 namespace internal {
33 
34 template<typename MatrixType, typename ResultType>
35 struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType>
36 {
37   enum {
38     MatrixAlignment     = traits<MatrixType>::Alignment,
39     ResultAlignment     = traits<ResultType>::Alignment,
40     StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
41   };
42   typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
43 
44   static void run(const MatrixType& mat, ResultType& result)
45   {
46     ActualMatrixType matrix(mat);
47     EIGEN_ALIGN16 const unsigned int _Sign_PNNP[4] = { 0x00000000, 0x80000000, 0x80000000, 0x00000000 };
48 
49     // Load the full matrix into registers
50     __m128 _L1 = matrix.template packet<MatrixAlignment>( 0);
51     __m128 _L2 = matrix.template packet<MatrixAlignment>( 4);
52     __m128 _L3 = matrix.template packet<MatrixAlignment>( 8);
53     __m128 _L4 = matrix.template packet<MatrixAlignment>(12);
54 
55     // The inverse is calculated using "Divide and Conquer" technique. The
56     // original matrix is divide into four 2x2 sub-matrices. Since each
57     // register holds four matrix element, the smaller matrices are
58     // represented as a registers. Hence we get a better locality of the
59     // calculations.
60 
61     __m128 A, B, C, D; // the four sub-matrices
62     if(!StorageOrdersMatch)
63     {
64       A = _mm_unpacklo_ps(_L1, _L2);
65       B = _mm_unpacklo_ps(_L3, _L4);
66       C = _mm_unpackhi_ps(_L1, _L2);
67       D = _mm_unpackhi_ps(_L3, _L4);
68     }
69     else
70     {
71       A = _mm_movelh_ps(_L1, _L2);
72       B = _mm_movehl_ps(_L2, _L1);
73       C = _mm_movelh_ps(_L3, _L4);
74       D = _mm_movehl_ps(_L4, _L3);
75     }
76 
77     __m128 iA, iB, iC, iD,                 // partial inverse of the sub-matrices
78             DC, AB;
79     __m128 dA, dB, dC, dD;                 // determinant of the sub-matrices
80     __m128 det, d, d1, d2;
81     __m128 rd;                             // reciprocal of the determinant
82 
83     //  AB = A# * B
84     AB = _mm_mul_ps(_mm_shuffle_ps(A,A,0x0F), B);
85     AB = _mm_sub_ps(AB,_mm_mul_ps(_mm_shuffle_ps(A,A,0xA5), _mm_shuffle_ps(B,B,0x4E)));
86     //  DC = D# * C
87     DC = _mm_mul_ps(_mm_shuffle_ps(D,D,0x0F), C);
88     DC = _mm_sub_ps(DC,_mm_mul_ps(_mm_shuffle_ps(D,D,0xA5), _mm_shuffle_ps(C,C,0x4E)));
89 
90     //  dA = |A|
91     dA = _mm_mul_ps(_mm_shuffle_ps(A, A, 0x5F),A);
92     dA = _mm_sub_ss(dA, _mm_movehl_ps(dA,dA));
93     //  dB = |B|
94     dB = _mm_mul_ps(_mm_shuffle_ps(B, B, 0x5F),B);
95     dB = _mm_sub_ss(dB, _mm_movehl_ps(dB,dB));
96 
97     //  dC = |C|
98     dC = _mm_mul_ps(_mm_shuffle_ps(C, C, 0x5F),C);
99     dC = _mm_sub_ss(dC, _mm_movehl_ps(dC,dC));
100     //  dD = |D|
101     dD = _mm_mul_ps(_mm_shuffle_ps(D, D, 0x5F),D);
102     dD = _mm_sub_ss(dD, _mm_movehl_ps(dD,dD));
103 
104     //  d = trace(AB*DC) = trace(A#*B*D#*C)
105     d = _mm_mul_ps(_mm_shuffle_ps(DC,DC,0xD8),AB);
106 
107     //  iD = C*A#*B
108     iD = _mm_mul_ps(_mm_shuffle_ps(C,C,0xA0), _mm_movelh_ps(AB,AB));
109     iD = _mm_add_ps(iD,_mm_mul_ps(_mm_shuffle_ps(C,C,0xF5), _mm_movehl_ps(AB,AB)));
110     //  iA = B*D#*C
111     iA = _mm_mul_ps(_mm_shuffle_ps(B,B,0xA0), _mm_movelh_ps(DC,DC));
112     iA = _mm_add_ps(iA,_mm_mul_ps(_mm_shuffle_ps(B,B,0xF5), _mm_movehl_ps(DC,DC)));
113 
114     //  d = trace(AB*DC) = trace(A#*B*D#*C) [continue]
115     d  = _mm_add_ps(d, _mm_movehl_ps(d, d));
116     d  = _mm_add_ss(d, _mm_shuffle_ps(d, d, 1));
117     d1 = _mm_mul_ss(dA,dD);
118     d2 = _mm_mul_ss(dB,dC);
119 
120     //  iD = D*|A| - C*A#*B
121     iD = _mm_sub_ps(_mm_mul_ps(D,_mm_shuffle_ps(dA,dA,0)), iD);
122 
123     //  iA = A*|D| - B*D#*C;
124     iA = _mm_sub_ps(_mm_mul_ps(A,_mm_shuffle_ps(dD,dD,0)), iA);
125 
126     //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
127     det = _mm_sub_ss(_mm_add_ss(d1,d2),d);
128     rd  = _mm_div_ss(_mm_set_ss(1.0f), det);
129 
130 //     #ifdef ZERO_SINGULAR
131 //         rd = _mm_and_ps(_mm_cmpneq_ss(det,_mm_setzero_ps()), rd);
132 //     #endif
133 
134     //  iB = D * (A#B)# = D*B#*A
135     iB = _mm_mul_ps(D, _mm_shuffle_ps(AB,AB,0x33));
136     iB = _mm_sub_ps(iB, _mm_mul_ps(_mm_shuffle_ps(D,D,0xB1), _mm_shuffle_ps(AB,AB,0x66)));
137     //  iC = A * (D#C)# = A*C#*D
138     iC = _mm_mul_ps(A, _mm_shuffle_ps(DC,DC,0x33));
139     iC = _mm_sub_ps(iC, _mm_mul_ps(_mm_shuffle_ps(A,A,0xB1), _mm_shuffle_ps(DC,DC,0x66)));
140 
141     rd = _mm_shuffle_ps(rd,rd,0);
142     rd = _mm_xor_ps(rd, _mm_load_ps((float*)_Sign_PNNP));
143 
144     //  iB = C*|B| - D*B#*A
145     iB = _mm_sub_ps(_mm_mul_ps(C,_mm_shuffle_ps(dB,dB,0)), iB);
146 
147     //  iC = B*|C| - A*C#*D;
148     iC = _mm_sub_ps(_mm_mul_ps(B,_mm_shuffle_ps(dC,dC,0)), iC);
149 
150     //  iX = iX / det
151     iA = _mm_mul_ps(rd,iA);
152     iB = _mm_mul_ps(rd,iB);
153     iC = _mm_mul_ps(rd,iC);
154     iD = _mm_mul_ps(rd,iD);
155 
156     Index res_stride = result.outerStride();
157     float* res = result.data();
158     pstoret<float, Packet4f, ResultAlignment>(res+0,            _mm_shuffle_ps(iA,iB,0x77));
159     pstoret<float, Packet4f, ResultAlignment>(res+res_stride,   _mm_shuffle_ps(iA,iB,0x22));
160     pstoret<float, Packet4f, ResultAlignment>(res+2*res_stride, _mm_shuffle_ps(iC,iD,0x77));
161     pstoret<float, Packet4f, ResultAlignment>(res+3*res_stride, _mm_shuffle_ps(iC,iD,0x22));
162   }
163 
164 };
165 
166 template<typename MatrixType, typename ResultType>
167 struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType>
168 {
169   enum {
170     MatrixAlignment     = traits<MatrixType>::Alignment,
171     ResultAlignment     = traits<ResultType>::Alignment,
172     StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
173   };
174   typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
175 
176   static void run(const MatrixType& mat, ResultType& result)
177   {
178     ActualMatrixType matrix(mat);
179     const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
180     const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
181 
182     // The inverse is calculated using "Divide and Conquer" technique. The
183     // original matrix is divide into four 2x2 sub-matrices. Since each
184     // register of the matrix holds two elements, the smaller matrices are
185     // consisted of two registers. Hence we get a better locality of the
186     // calculations.
187 
188     // the four sub-matrices
189     __m128d A1, A2, B1, B2, C1, C2, D1, D2;
190 
191     if(StorageOrdersMatch)
192     {
193       A1 = matrix.template packet<MatrixAlignment>( 0); B1 = matrix.template packet<MatrixAlignment>( 2);
194       A2 = matrix.template packet<MatrixAlignment>( 4); B2 = matrix.template packet<MatrixAlignment>( 6);
195       C1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
196       C2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
197     }
198     else
199     {
200       __m128d tmp;
201       A1 = matrix.template packet<MatrixAlignment>( 0); C1 = matrix.template packet<MatrixAlignment>( 2);
202       A2 = matrix.template packet<MatrixAlignment>( 4); C2 = matrix.template packet<MatrixAlignment>( 6);
203       tmp = A1;
204       A1 = _mm_unpacklo_pd(A1,A2);
205       A2 = _mm_unpackhi_pd(tmp,A2);
206       tmp = C1;
207       C1 = _mm_unpacklo_pd(C1,C2);
208       C2 = _mm_unpackhi_pd(tmp,C2);
209 
210       B1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
211       B2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
212       tmp = B1;
213       B1 = _mm_unpacklo_pd(B1,B2);
214       B2 = _mm_unpackhi_pd(tmp,B2);
215       tmp = D1;
216       D1 = _mm_unpacklo_pd(D1,D2);
217       D2 = _mm_unpackhi_pd(tmp,D2);
218     }
219 
220     __m128d iA1, iA2, iB1, iB2, iC1, iC2, iD1, iD2,     // partial invese of the sub-matrices
221             DC1, DC2, AB1, AB2;
222     __m128d dA, dB, dC, dD;     // determinant of the sub-matrices
223     __m128d det, d1, d2, rd;
224 
225     //  dA = |A|
226     dA = _mm_shuffle_pd(A2, A2, 1);
227     dA = _mm_mul_pd(A1, dA);
228     dA = _mm_sub_sd(dA, _mm_shuffle_pd(dA,dA,3));
229     //  dB = |B|
230     dB = _mm_shuffle_pd(B2, B2, 1);
231     dB = _mm_mul_pd(B1, dB);
232     dB = _mm_sub_sd(dB, _mm_shuffle_pd(dB,dB,3));
233 
234     //  AB = A# * B
235     AB1 = _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,3));
236     AB2 = _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,0));
237     AB1 = _mm_sub_pd(AB1, _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,3)));
238     AB2 = _mm_sub_pd(AB2, _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,0)));
239 
240     //  dC = |C|
241     dC = _mm_shuffle_pd(C2, C2, 1);
242     dC = _mm_mul_pd(C1, dC);
243     dC = _mm_sub_sd(dC, _mm_shuffle_pd(dC,dC,3));
244     //  dD = |D|
245     dD = _mm_shuffle_pd(D2, D2, 1);
246     dD = _mm_mul_pd(D1, dD);
247     dD = _mm_sub_sd(dD, _mm_shuffle_pd(dD,dD,3));
248 
249     //  DC = D# * C
250     DC1 = _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,3));
251     DC2 = _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,0));
252     DC1 = _mm_sub_pd(DC1, _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,3)));
253     DC2 = _mm_sub_pd(DC2, _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,0)));
254 
255     //  rd = trace(AB*DC) = trace(A#*B*D#*C)
256     d1 = _mm_mul_pd(AB1, _mm_shuffle_pd(DC1, DC2, 0));
257     d2 = _mm_mul_pd(AB2, _mm_shuffle_pd(DC1, DC2, 3));
258     rd = _mm_add_pd(d1, d2);
259     rd = _mm_add_sd(rd, _mm_shuffle_pd(rd, rd,3));
260 
261     //  iD = C*A#*B
262     iD1 = _mm_mul_pd(AB1, _mm_shuffle_pd(C1,C1,0));
263     iD2 = _mm_mul_pd(AB1, _mm_shuffle_pd(C2,C2,0));
264     iD1 = _mm_add_pd(iD1, _mm_mul_pd(AB2, _mm_shuffle_pd(C1,C1,3)));
265     iD2 = _mm_add_pd(iD2, _mm_mul_pd(AB2, _mm_shuffle_pd(C2,C2,3)));
266 
267     //  iA = B*D#*C
268     iA1 = _mm_mul_pd(DC1, _mm_shuffle_pd(B1,B1,0));
269     iA2 = _mm_mul_pd(DC1, _mm_shuffle_pd(B2,B2,0));
270     iA1 = _mm_add_pd(iA1, _mm_mul_pd(DC2, _mm_shuffle_pd(B1,B1,3)));
271     iA2 = _mm_add_pd(iA2, _mm_mul_pd(DC2, _mm_shuffle_pd(B2,B2,3)));
272 
273     //  iD = D*|A| - C*A#*B
274     dA = _mm_shuffle_pd(dA,dA,0);
275     iD1 = _mm_sub_pd(_mm_mul_pd(D1, dA), iD1);
276     iD2 = _mm_sub_pd(_mm_mul_pd(D2, dA), iD2);
277 
278     //  iA = A*|D| - B*D#*C;
279     dD = _mm_shuffle_pd(dD,dD,0);
280     iA1 = _mm_sub_pd(_mm_mul_pd(A1, dD), iA1);
281     iA2 = _mm_sub_pd(_mm_mul_pd(A2, dD), iA2);
282 
283     d1 = _mm_mul_sd(dA, dD);
284     d2 = _mm_mul_sd(dB, dC);
285 
286     //  iB = D * (A#B)# = D*B#*A
287     iB1 = _mm_mul_pd(D1, _mm_shuffle_pd(AB2,AB1,1));
288     iB2 = _mm_mul_pd(D2, _mm_shuffle_pd(AB2,AB1,1));
289     iB1 = _mm_sub_pd(iB1, _mm_mul_pd(_mm_shuffle_pd(D1,D1,1), _mm_shuffle_pd(AB2,AB1,2)));
290     iB2 = _mm_sub_pd(iB2, _mm_mul_pd(_mm_shuffle_pd(D2,D2,1), _mm_shuffle_pd(AB2,AB1,2)));
291 
292     //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
293     det = _mm_add_sd(d1, d2);
294     det = _mm_sub_sd(det, rd);
295 
296     //  iC = A * (D#C)# = A*C#*D
297     iC1 = _mm_mul_pd(A1, _mm_shuffle_pd(DC2,DC1,1));
298     iC2 = _mm_mul_pd(A2, _mm_shuffle_pd(DC2,DC1,1));
299     iC1 = _mm_sub_pd(iC1, _mm_mul_pd(_mm_shuffle_pd(A1,A1,1), _mm_shuffle_pd(DC2,DC1,2)));
300     iC2 = _mm_sub_pd(iC2, _mm_mul_pd(_mm_shuffle_pd(A2,A2,1), _mm_shuffle_pd(DC2,DC1,2)));
301 
302     rd = _mm_div_sd(_mm_set_sd(1.0), det);
303 //     #ifdef ZERO_SINGULAR
304 //         rd = _mm_and_pd(_mm_cmpneq_sd(det,_mm_setzero_pd()), rd);
305 //     #endif
306     rd = _mm_shuffle_pd(rd,rd,0);
307 
308     //  iB = C*|B| - D*B#*A
309     dB = _mm_shuffle_pd(dB,dB,0);
310     iB1 = _mm_sub_pd(_mm_mul_pd(C1, dB), iB1);
311     iB2 = _mm_sub_pd(_mm_mul_pd(C2, dB), iB2);
312 
313     d1 = _mm_xor_pd(rd, _Sign_PN);
314     d2 = _mm_xor_pd(rd, _Sign_NP);
315 
316     //  iC = B*|C| - A*C#*D;
317     dC = _mm_shuffle_pd(dC,dC,0);
318     iC1 = _mm_sub_pd(_mm_mul_pd(B1, dC), iC1);
319     iC2 = _mm_sub_pd(_mm_mul_pd(B2, dC), iC2);
320 
321     Index res_stride = result.outerStride();
322     double* res = result.data();
323     pstoret<double, Packet2d, ResultAlignment>(res+0,             _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1));
324     pstoret<double, Packet2d, ResultAlignment>(res+res_stride,    _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2));
325     pstoret<double, Packet2d, ResultAlignment>(res+2,             _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1));
326     pstoret<double, Packet2d, ResultAlignment>(res+res_stride+2,  _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2));
327     pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1));
328     pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2));
329     pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1));
330     pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2));
331   }
332 };
333 
334 } // end namespace internal
335 
336 } // end namespace Eigen
337 
338 #endif // EIGEN_INVERSE_SSE_H
339