1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2007 Julien Pommier
5 // Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 /* The sin, cos, exp, and log functions of this file come from
12  * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
13  */
14 
15 #ifndef EIGEN_MATH_FUNCTIONS_SSE_H
16 #define EIGEN_MATH_FUNCTIONS_SSE_H
17 
18 namespace Eigen {
19 
20 namespace internal {
21 
22 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
23 Packet4f plog<Packet4f>(const Packet4f& _x)
24 {
25   Packet4f x = _x;
26   _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
27   _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
28   _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
29 
30   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
31 
32   /* the smallest non denormalized float number */
33   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
34   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000);//-1.f/0.f);
35 
36   /* natural logarithm computed for 4 simultaneous float
37     return NaN for x <= 0
38   */
39   _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
40   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
41   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
42   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
43   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
44   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
45   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
46   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
47   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
48   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
49   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
50   _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
51 
52 
53   Packet4i emm0;
54 
55   Packet4f invalid_mask = _mm_cmpnge_ps(x, _mm_setzero_ps()); // not greater equal is true if x is NaN
56   Packet4f iszero_mask = _mm_cmpeq_ps(x, _mm_setzero_ps());
57 
58   x = pmax(x, p4f_min_norm_pos);  /* cut off denormalized stuff */
59   emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23);
60 
61   /* keep only the fractional part */
62   x = _mm_and_ps(x, p4f_inv_mant_mask);
63   x = _mm_or_ps(x, p4f_half);
64 
65   emm0 = _mm_sub_epi32(emm0, p4i_0x7f);
66   Packet4f e = padd(_mm_cvtepi32_ps(emm0), p4f_1);
67 
68   /* part2:
69      if( x < SQRTHF ) {
70        e -= 1;
71        x = x + x - 1.0;
72      } else { x = x - 1.0; }
73   */
74   Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF);
75   Packet4f tmp = _mm_and_ps(x, mask);
76   x = psub(x, p4f_1);
77   e = psub(e, _mm_and_ps(p4f_1, mask));
78   x = padd(x, tmp);
79 
80   Packet4f x2 = pmul(x,x);
81   Packet4f x3 = pmul(x2,x);
82 
83   Packet4f y, y1, y2;
84   y  = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
85   y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
86   y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
87   y  = pmadd(y , x, p4f_cephes_log_p2);
88   y1 = pmadd(y1, x, p4f_cephes_log_p5);
89   y2 = pmadd(y2, x, p4f_cephes_log_p8);
90   y = pmadd(y, x3, y1);
91   y = pmadd(y, x3, y2);
92   y = pmul(y, x3);
93 
94   y1 = pmul(e, p4f_cephes_log_q1);
95   tmp = pmul(x2, p4f_half);
96   y = padd(y, y1);
97   x = psub(x, tmp);
98   y2 = pmul(e, p4f_cephes_log_q2);
99   x = padd(x, y);
100   x = padd(x, y2);
101   // negative arg will be NAN, 0 will be -INF
102   return _mm_or_ps(_mm_andnot_ps(iszero_mask, _mm_or_ps(x, invalid_mask)),
103                    _mm_and_ps(iszero_mask, p4f_minus_inf));
104 }
105 
106 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
107 Packet4f pexp<Packet4f>(const Packet4f& _x)
108 {
109   Packet4f x = _x;
110   _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
111   _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
112   _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
113 
114 
115   _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
116   _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
117 
118   _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
119   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
120   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
121 
122   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
123   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
124   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
125   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
126   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
127   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
128 
129   Packet4f tmp = _mm_setzero_ps(), fx;
130   Packet4i emm0;
131 
132   // clamp x
133   x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
134 
135   /* express exp(x) as exp(g + n*log(2)) */
136   fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
137 
138 #ifdef EIGEN_VECTORIZE_SSE4_1
139   fx = _mm_floor_ps(fx);
140 #else
141   emm0 = _mm_cvttps_epi32(fx);
142   tmp  = _mm_cvtepi32_ps(emm0);
143   /* if greater, substract 1 */
144   Packet4f mask = _mm_cmpgt_ps(tmp, fx);
145   mask = _mm_and_ps(mask, p4f_1);
146   fx = psub(tmp, mask);
147 #endif
148 
149   tmp = pmul(fx, p4f_cephes_exp_C1);
150   Packet4f z = pmul(fx, p4f_cephes_exp_C2);
151   x = psub(x, tmp);
152   x = psub(x, z);
153 
154   z = pmul(x,x);
155 
156   Packet4f y = p4f_cephes_exp_p0;
157   y = pmadd(y, x, p4f_cephes_exp_p1);
158   y = pmadd(y, x, p4f_cephes_exp_p2);
159   y = pmadd(y, x, p4f_cephes_exp_p3);
160   y = pmadd(y, x, p4f_cephes_exp_p4);
161   y = pmadd(y, x, p4f_cephes_exp_p5);
162   y = pmadd(y, z, x);
163   y = padd(y, p4f_1);
164 
165   // build 2^n
166   emm0 = _mm_cvttps_epi32(fx);
167   emm0 = _mm_add_epi32(emm0, p4i_0x7f);
168   emm0 = _mm_slli_epi32(emm0, 23);
169   return pmax(pmul(y, Packet4f(_mm_castsi128_ps(emm0))), _x);
170 }
171 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
172 Packet2d pexp<Packet2d>(const Packet2d& _x)
173 {
174   Packet2d x = _x;
175 
176   _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
177   _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
178   _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
179 
180   _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
181   _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
182 
183   _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
184 
185   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
186   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
187   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
188 
189   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
190   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
191   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
192   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
193 
194   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
195   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
196   static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0);
197 
198   Packet2d tmp = _mm_setzero_pd(), fx;
199   Packet4i emm0;
200 
201   // clamp x
202   x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
203   /* express exp(x) as exp(g + n*log(2)) */
204   fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
205 
206 #ifdef EIGEN_VECTORIZE_SSE4_1
207   fx = _mm_floor_pd(fx);
208 #else
209   emm0 = _mm_cvttpd_epi32(fx);
210   tmp  = _mm_cvtepi32_pd(emm0);
211   /* if greater, substract 1 */
212   Packet2d mask = _mm_cmpgt_pd(tmp, fx);
213   mask = _mm_and_pd(mask, p2d_1);
214   fx = psub(tmp, mask);
215 #endif
216 
217   tmp = pmul(fx, p2d_cephes_exp_C1);
218   Packet2d z = pmul(fx, p2d_cephes_exp_C2);
219   x = psub(x, tmp);
220   x = psub(x, z);
221 
222   Packet2d x2 = pmul(x,x);
223 
224   Packet2d px = p2d_cephes_exp_p0;
225   px = pmadd(px, x2, p2d_cephes_exp_p1);
226   px = pmadd(px, x2, p2d_cephes_exp_p2);
227   px = pmul (px, x);
228 
229   Packet2d qx = p2d_cephes_exp_q0;
230   qx = pmadd(qx, x2, p2d_cephes_exp_q1);
231   qx = pmadd(qx, x2, p2d_cephes_exp_q2);
232   qx = pmadd(qx, x2, p2d_cephes_exp_q3);
233 
234   x = pdiv(px,psub(qx,px));
235   x = pmadd(p2d_2,x,p2d_1);
236 
237   // build 2^n
238   emm0 = _mm_cvttpd_epi32(fx);
239   emm0 = _mm_add_epi32(emm0, p4i_1023_0);
240   emm0 = _mm_slli_epi32(emm0, 20);
241   emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3));
242   return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x);
243 }
244 
245 /* evaluation of 4 sines at onces, using SSE2 intrinsics.
246 
247    The code is the exact rewriting of the cephes sinf function.
248    Precision is excellent as long as x < 8192 (I did not bother to
249    take into account the special handling they have for greater values
250    -- it does not return garbage for arguments over 8192, though, but
251    the extra precision is missing).
252 
253    Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
254    surprising but correct result.
255 */
256 
257 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
258 Packet4f psin<Packet4f>(const Packet4f& _x)
259 {
260   Packet4f x = _x;
261   _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
262   _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
263 
264   _EIGEN_DECLARE_CONST_Packet4i(1, 1);
265   _EIGEN_DECLARE_CONST_Packet4i(not1, ~1);
266   _EIGEN_DECLARE_CONST_Packet4i(2, 2);
267   _EIGEN_DECLARE_CONST_Packet4i(4, 4);
268 
269   _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000);
270 
271   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f);
272   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
273   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
274   _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f);
275   _EIGEN_DECLARE_CONST_Packet4f(sincof_p1,  8.3321608736E-3f);
276   _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f);
277   _EIGEN_DECLARE_CONST_Packet4f(coscof_p0,  2.443315711809948E-005f);
278   _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f);
279   _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
280   _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
281 
282   Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y;
283 
284   Packet4i emm0, emm2;
285   sign_bit = x;
286   /* take the absolute value */
287   x = pabs(x);
288 
289   /* take the modulo */
290 
291   /* extract the sign bit (upper one) */
292   sign_bit = _mm_and_ps(sign_bit, p4f_sign_mask);
293 
294   /* scale by 4/Pi */
295   y = pmul(x, p4f_cephes_FOPI);
296 
297   /* store the integer part of y in mm0 */
298   emm2 = _mm_cvttps_epi32(y);
299   /* j=(j+1) & (~1) (see the cephes sources) */
300   emm2 = _mm_add_epi32(emm2, p4i_1);
301   emm2 = _mm_and_si128(emm2, p4i_not1);
302   y = _mm_cvtepi32_ps(emm2);
303   /* get the swap sign flag */
304   emm0 = _mm_and_si128(emm2, p4i_4);
305   emm0 = _mm_slli_epi32(emm0, 29);
306   /* get the polynom selection mask
307      there is one polynom for 0 <= x <= Pi/4
308      and another one for Pi/4<x<=Pi/2
309 
310      Both branches will be computed.
311   */
312   emm2 = _mm_and_si128(emm2, p4i_2);
313   emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
314 
315   Packet4f swap_sign_bit = _mm_castsi128_ps(emm0);
316   Packet4f poly_mask = _mm_castsi128_ps(emm2);
317   sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
318 
319   /* The magic pass: "Extended precision modular arithmetic"
320      x = ((x - y * DP1) - y * DP2) - y * DP3; */
321   xmm1 = pmul(y, p4f_minus_cephes_DP1);
322   xmm2 = pmul(y, p4f_minus_cephes_DP2);
323   xmm3 = pmul(y, p4f_minus_cephes_DP3);
324   x = padd(x, xmm1);
325   x = padd(x, xmm2);
326   x = padd(x, xmm3);
327 
328   /* Evaluate the first polynom  (0 <= x <= Pi/4) */
329   y = p4f_coscof_p0;
330   Packet4f z = _mm_mul_ps(x,x);
331 
332   y = pmadd(y, z, p4f_coscof_p1);
333   y = pmadd(y, z, p4f_coscof_p2);
334   y = pmul(y, z);
335   y = pmul(y, z);
336   Packet4f tmp = pmul(z, p4f_half);
337   y = psub(y, tmp);
338   y = padd(y, p4f_1);
339 
340   /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
341 
342   Packet4f y2 = p4f_sincof_p0;
343   y2 = pmadd(y2, z, p4f_sincof_p1);
344   y2 = pmadd(y2, z, p4f_sincof_p2);
345   y2 = pmul(y2, z);
346   y2 = pmul(y2, x);
347   y2 = padd(y2, x);
348 
349   /* select the correct result from the two polynoms */
350   y2 = _mm_and_ps(poly_mask, y2);
351   y = _mm_andnot_ps(poly_mask, y);
352   y = _mm_or_ps(y,y2);
353   /* update the sign */
354   return _mm_xor_ps(y, sign_bit);
355 }
356 
357 /* almost the same as psin */
358 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
359 Packet4f pcos<Packet4f>(const Packet4f& _x)
360 {
361   Packet4f x = _x;
362   _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
363   _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
364 
365   _EIGEN_DECLARE_CONST_Packet4i(1, 1);
366   _EIGEN_DECLARE_CONST_Packet4i(not1, ~1);
367   _EIGEN_DECLARE_CONST_Packet4i(2, 2);
368   _EIGEN_DECLARE_CONST_Packet4i(4, 4);
369 
370   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f);
371   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
372   _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
373   _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f);
374   _EIGEN_DECLARE_CONST_Packet4f(sincof_p1,  8.3321608736E-3f);
375   _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f);
376   _EIGEN_DECLARE_CONST_Packet4f(coscof_p0,  2.443315711809948E-005f);
377   _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f);
378   _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
379   _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
380 
381   Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
382   Packet4i emm0, emm2;
383 
384   x = pabs(x);
385 
386   /* scale by 4/Pi */
387   y = pmul(x, p4f_cephes_FOPI);
388 
389   /* get the integer part of y */
390   emm2 = _mm_cvttps_epi32(y);
391   /* j=(j+1) & (~1) (see the cephes sources) */
392   emm2 = _mm_add_epi32(emm2, p4i_1);
393   emm2 = _mm_and_si128(emm2, p4i_not1);
394   y = _mm_cvtepi32_ps(emm2);
395 
396   emm2 = _mm_sub_epi32(emm2, p4i_2);
397 
398   /* get the swap sign flag */
399   emm0 = _mm_andnot_si128(emm2, p4i_4);
400   emm0 = _mm_slli_epi32(emm0, 29);
401   /* get the polynom selection mask */
402   emm2 = _mm_and_si128(emm2, p4i_2);
403   emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
404 
405   Packet4f sign_bit = _mm_castsi128_ps(emm0);
406   Packet4f poly_mask = _mm_castsi128_ps(emm2);
407 
408   /* The magic pass: "Extended precision modular arithmetic"
409      x = ((x - y * DP1) - y * DP2) - y * DP3; */
410   xmm1 = pmul(y, p4f_minus_cephes_DP1);
411   xmm2 = pmul(y, p4f_minus_cephes_DP2);
412   xmm3 = pmul(y, p4f_minus_cephes_DP3);
413   x = padd(x, xmm1);
414   x = padd(x, xmm2);
415   x = padd(x, xmm3);
416 
417   /* Evaluate the first polynom  (0 <= x <= Pi/4) */
418   y = p4f_coscof_p0;
419   Packet4f z = pmul(x,x);
420 
421   y = pmadd(y,z,p4f_coscof_p1);
422   y = pmadd(y,z,p4f_coscof_p2);
423   y = pmul(y, z);
424   y = pmul(y, z);
425   Packet4f tmp = _mm_mul_ps(z, p4f_half);
426   y = psub(y, tmp);
427   y = padd(y, p4f_1);
428 
429   /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
430   Packet4f y2 = p4f_sincof_p0;
431   y2 = pmadd(y2, z, p4f_sincof_p1);
432   y2 = pmadd(y2, z, p4f_sincof_p2);
433   y2 = pmul(y2, z);
434   y2 = pmadd(y2, x, x);
435 
436   /* select the correct result from the two polynoms */
437   y2 = _mm_and_ps(poly_mask, y2);
438   y  = _mm_andnot_ps(poly_mask, y);
439   y  = _mm_or_ps(y,y2);
440 
441   /* update the sign */
442   return _mm_xor_ps(y, sign_bit);
443 }
444 
445 #if EIGEN_FAST_MATH
446 
447 // This is based on Quake3's fast inverse square root.
448 // For detail see here: http://www.beyond3d.com/content/articles/8/
449 // It lacks 1 (or 2 bits in some rare cases) of precision, and does not handle negative, +inf, or denormalized numbers correctly.
450 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
451 Packet4f psqrt<Packet4f>(const Packet4f& _x)
452 {
453   Packet4f half = pmul(_x, pset1<Packet4f>(.5f));
454 
455   /* select only the inverse sqrt of non-zero inputs */
456   Packet4f non_zero_mask = _mm_cmpge_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)()));
457   Packet4f x = _mm_and_ps(non_zero_mask, _mm_rsqrt_ps(_x));
458 
459   x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x))));
460   return pmul(_x,x);
461 }
462 
463 #else
464 
465 template<> EIGEN_STRONG_INLINE Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
466 
467 #endif
468 
469 template<> EIGEN_STRONG_INLINE Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); }
470 
471 } // end namespace internal
472 
473 } // end namespace Eigen
474 
475 #endif // EIGEN_MATH_FUNCTIONS_SSE_H
476