27 #ifndef EIGEN_INVERSE_SSE_H
28 #define EIGEN_INVERSE_SSE_H
34 template<
typename MatrixType,
typename ResultType>
35 struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType>
38 MatrixAlignment = traits<MatrixType>::Alignment,
39 ResultAlignment = traits<ResultType>::Alignment,
42 typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
44 static void run(
const MatrixType& mat, ResultType& result)
46 ActualMatrixType matrix(mat);
47 EIGEN_ALIGN16
const unsigned int _Sign_PNNP[4] = { 0x00000000, 0x80000000, 0x80000000, 0x00000000 };
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);
62 if(!StorageOrdersMatch)
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);
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);
77 __m128 iA, iB, iC, iD,
79 __m128 dA, dB, dC, dD;
80 __m128 det, d, d1, d2;
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)));
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)));
91 dA = _mm_mul_ps(_mm_shuffle_ps(A, A, 0x5F),A);
92 dA = _mm_sub_ss(dA, _mm_movehl_ps(dA,dA));
94 dB = _mm_mul_ps(_mm_shuffle_ps(B, B, 0x5F),B);
95 dB = _mm_sub_ss(dB, _mm_movehl_ps(dB,dB));
98 dC = _mm_mul_ps(_mm_shuffle_ps(C, C, 0x5F),C);
99 dC = _mm_sub_ss(dC, _mm_movehl_ps(dC,dC));
101 dD = _mm_mul_ps(_mm_shuffle_ps(D, D, 0x5F),D);
102 dD = _mm_sub_ss(dD, _mm_movehl_ps(dD,dD));
105 d = _mm_mul_ps(_mm_shuffle_ps(DC,DC,0xD8),AB);
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)));
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)));
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);
121 iD = _mm_sub_ps(_mm_mul_ps(D,_mm_shuffle_ps(dA,dA,0)), iD);
124 iA = _mm_sub_ps(_mm_mul_ps(A,_mm_shuffle_ps(dD,dD,0)), iA);
127 det = _mm_sub_ss(_mm_add_ss(d1,d2),d);
128 rd = _mm_div_ss(_mm_set_ss(1.0f), det);
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)));
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)));
141 rd = _mm_shuffle_ps(rd,rd,0);
142 rd = _mm_xor_ps(rd, _mm_load_ps((
float*)_Sign_PNNP));
145 iB = _mm_sub_ps(_mm_mul_ps(C,_mm_shuffle_ps(dB,dB,0)), iB);
148 iC = _mm_sub_ps(_mm_mul_ps(B,_mm_shuffle_ps(dC,dC,0)), iC);
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);
156 result.template writePacket<ResultAlignment>( 0, _mm_shuffle_ps(iA,iB,0x77));
157 result.template writePacket<ResultAlignment>( 4, _mm_shuffle_ps(iA,iB,0x22));
158 result.template writePacket<ResultAlignment>( 8, _mm_shuffle_ps(iC,iD,0x77));
159 result.template writePacket<ResultAlignment>(12, _mm_shuffle_ps(iC,iD,0x22));
164 template<
typename MatrixType,
typename ResultType>
165 struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType>
168 MatrixAlignment = traits<MatrixType>::Alignment,
169 ResultAlignment = traits<ResultType>::Alignment,
172 typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
174 static void run(
const MatrixType& mat, ResultType& result)
176 ActualMatrixType matrix(mat);
177 const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
178 const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
187 __m128d A1, A2, B1, B2, C1, C2, D1, D2;
189 if(StorageOrdersMatch)
191 A1 = matrix.template packet<MatrixAlignment>( 0); B1 = matrix.template packet<MatrixAlignment>( 2);
192 A2 = matrix.template packet<MatrixAlignment>( 4); B2 = matrix.template packet<MatrixAlignment>( 6);
193 C1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
194 C2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
199 A1 = matrix.template packet<MatrixAlignment>( 0); C1 = matrix.template packet<MatrixAlignment>( 2);
200 A2 = matrix.template packet<MatrixAlignment>( 4); C2 = matrix.template packet<MatrixAlignment>( 6);
202 A1 = _mm_unpacklo_pd(A1,A2);
203 A2 = _mm_unpackhi_pd(tmp,A2);
205 C1 = _mm_unpacklo_pd(C1,C2);
206 C2 = _mm_unpackhi_pd(tmp,C2);
208 B1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
209 B2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
211 B1 = _mm_unpacklo_pd(B1,B2);
212 B2 = _mm_unpackhi_pd(tmp,B2);
214 D1 = _mm_unpacklo_pd(D1,D2);
215 D2 = _mm_unpackhi_pd(tmp,D2);
218 __m128d iA1, iA2, iB1, iB2, iC1, iC2, iD1, iD2,
220 __m128d dA, dB, dC, dD;
221 __m128d det, d1, d2, rd;
224 dA = _mm_shuffle_pd(A2, A2, 1);
225 dA = _mm_mul_pd(A1, dA);
226 dA = _mm_sub_sd(dA, _mm_shuffle_pd(dA,dA,3));
228 dB = _mm_shuffle_pd(B2, B2, 1);
229 dB = _mm_mul_pd(B1, dB);
230 dB = _mm_sub_sd(dB, _mm_shuffle_pd(dB,dB,3));
233 AB1 = _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,3));
234 AB2 = _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,0));
235 AB1 = _mm_sub_pd(AB1, _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,3)));
236 AB2 = _mm_sub_pd(AB2, _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,0)));
239 dC = _mm_shuffle_pd(C2, C2, 1);
240 dC = _mm_mul_pd(C1, dC);
241 dC = _mm_sub_sd(dC, _mm_shuffle_pd(dC,dC,3));
243 dD = _mm_shuffle_pd(D2, D2, 1);
244 dD = _mm_mul_pd(D1, dD);
245 dD = _mm_sub_sd(dD, _mm_shuffle_pd(dD,dD,3));
248 DC1 = _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,3));
249 DC2 = _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,0));
250 DC1 = _mm_sub_pd(DC1, _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,3)));
251 DC2 = _mm_sub_pd(DC2, _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,0)));
254 d1 = _mm_mul_pd(AB1, _mm_shuffle_pd(DC1, DC2, 0));
255 d2 = _mm_mul_pd(AB2, _mm_shuffle_pd(DC1, DC2, 3));
256 rd = _mm_add_pd(d1, d2);
257 rd = _mm_add_sd(rd, _mm_shuffle_pd(rd, rd,3));
260 iD1 = _mm_mul_pd(AB1, _mm_shuffle_pd(C1,C1,0));
261 iD2 = _mm_mul_pd(AB1, _mm_shuffle_pd(C2,C2,0));
262 iD1 = _mm_add_pd(iD1, _mm_mul_pd(AB2, _mm_shuffle_pd(C1,C1,3)));
263 iD2 = _mm_add_pd(iD2, _mm_mul_pd(AB2, _mm_shuffle_pd(C2,C2,3)));
266 iA1 = _mm_mul_pd(DC1, _mm_shuffle_pd(B1,B1,0));
267 iA2 = _mm_mul_pd(DC1, _mm_shuffle_pd(B2,B2,0));
268 iA1 = _mm_add_pd(iA1, _mm_mul_pd(DC2, _mm_shuffle_pd(B1,B1,3)));
269 iA2 = _mm_add_pd(iA2, _mm_mul_pd(DC2, _mm_shuffle_pd(B2,B2,3)));
272 dA = _mm_shuffle_pd(dA,dA,0);
273 iD1 = _mm_sub_pd(_mm_mul_pd(D1, dA), iD1);
274 iD2 = _mm_sub_pd(_mm_mul_pd(D2, dA), iD2);
277 dD = _mm_shuffle_pd(dD,dD,0);
278 iA1 = _mm_sub_pd(_mm_mul_pd(A1, dD), iA1);
279 iA2 = _mm_sub_pd(_mm_mul_pd(A2, dD), iA2);
281 d1 = _mm_mul_sd(dA, dD);
282 d2 = _mm_mul_sd(dB, dC);
285 iB1 = _mm_mul_pd(D1, _mm_shuffle_pd(AB2,AB1,1));
286 iB2 = _mm_mul_pd(D2, _mm_shuffle_pd(AB2,AB1,1));
287 iB1 = _mm_sub_pd(iB1, _mm_mul_pd(_mm_shuffle_pd(D1,D1,1), _mm_shuffle_pd(AB2,AB1,2)));
288 iB2 = _mm_sub_pd(iB2, _mm_mul_pd(_mm_shuffle_pd(D2,D2,1), _mm_shuffle_pd(AB2,AB1,2)));
291 det = _mm_add_sd(d1, d2);
292 det = _mm_sub_sd(det, rd);
295 iC1 = _mm_mul_pd(A1, _mm_shuffle_pd(DC2,DC1,1));
296 iC2 = _mm_mul_pd(A2, _mm_shuffle_pd(DC2,DC1,1));
297 iC1 = _mm_sub_pd(iC1, _mm_mul_pd(_mm_shuffle_pd(A1,A1,1), _mm_shuffle_pd(DC2,DC1,2)));
298 iC2 = _mm_sub_pd(iC2, _mm_mul_pd(_mm_shuffle_pd(A2,A2,1), _mm_shuffle_pd(DC2,DC1,2)));
300 rd = _mm_div_sd(_mm_set_sd(1.0), det);
304 rd = _mm_shuffle_pd(rd,rd,0);
307 dB = _mm_shuffle_pd(dB,dB,0);
308 iB1 = _mm_sub_pd(_mm_mul_pd(C1, dB), iB1);
309 iB2 = _mm_sub_pd(_mm_mul_pd(C2, dB), iB2);
311 d1 = _mm_xor_pd(rd, _Sign_PN);
312 d2 = _mm_xor_pd(rd, _Sign_NP);
315 dC = _mm_shuffle_pd(dC,dC,0);
316 iC1 = _mm_sub_pd(_mm_mul_pd(B1, dC), iC1);
317 iC2 = _mm_sub_pd(_mm_mul_pd(B2, dC), iC2);
319 result.template writePacket<ResultAlignment>( 0, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1));
320 result.template writePacket<ResultAlignment>( 4, _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2));
321 result.template writePacket<ResultAlignment>( 2, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1));
322 result.template writePacket<ResultAlignment>( 6, _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2));
323 result.template writePacket<ResultAlignment>( 8, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1));
324 result.template writePacket<ResultAlignment>(12, _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2));
325 result.template writePacket<ResultAlignment>(10, _mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1));
326 result.template writePacket<ResultAlignment>(14, _mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2));
334 #endif // EIGEN_INVERSE_SSE_H
const unsigned int RowMajorBit
Definition: Constants.h:61
Definition: Eigen_Colamd.h:54