10 #ifndef EIGEN_PACKET_MATH_SSE_H 11 #define EIGEN_PACKET_MATH_SSE_H 17 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 18 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 21 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 22 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) 25 typedef __m128 Packet4f;
26 typedef __m128i Packet4i;
27 typedef __m128d Packet2d;
29 template<>
struct is_arithmetic<__m128> {
enum { value =
true }; };
30 template<>
struct is_arithmetic<__m128i> {
enum { value =
true }; };
31 template<>
struct is_arithmetic<__m128d> {
enum { value =
true }; };
33 #define vec4f_swizzle1(v,p,q,r,s) \ 34 (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p))))) 36 #define vec4i_swizzle1(v,p,q,r,s) \ 37 (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p)))) 39 #define vec2d_swizzle1(v,p,q) \ 40 (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2))))) 42 #define vec4f_swizzle2(a,b,p,q,r,s) \ 43 (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p)))) 45 #define vec4i_swizzle2(a,b,p,q,r,s) \ 46 (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p)))))) 48 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ 49 const Packet4f p4f_##NAME = pset1<Packet4f>(X) 51 #define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ 52 const Packet2d p2d_##NAME = pset1<Packet2d>(X) 54 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ 55 const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X)) 57 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ 58 const Packet4i p4i_##NAME = pset1<Packet4i>(X) 61 template<>
struct packet_traits<float> : default_packet_traits
63 typedef Packet4f type;
70 HasSin = EIGEN_FAST_MATH,
71 HasCos = EIGEN_FAST_MATH,
77 template<>
struct packet_traits<double> : default_packet_traits
79 typedef Packet2d type;
90 template<>
struct packet_traits<int> : default_packet_traits
92 typedef Packet4i type;
101 template<>
struct unpacket_traits<Packet4f> {
typedef float type;
enum {size=4}; };
102 template<>
struct unpacket_traits<Packet2d> {
typedef double type;
enum {size=2}; };
103 template<>
struct unpacket_traits<Packet4i> {
typedef int type;
enum {size=4}; };
105 #if defined(_MSC_VER) && (_MSC_VER==1500) 109 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(
const float& from) {
return _mm_set_ps(from,from,from,from); }
110 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(
const double& from) {
return _mm_set_pd(from,from); }
111 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(
const int& from) {
return _mm_set_epi32(from,from,from,from); }
113 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(
const float& from) {
return _mm_set1_ps(from); }
114 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(
const double& from) {
return _mm_set1_pd(from); }
115 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(
const int& from) {
return _mm_set1_epi32(from); }
118 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(
const float& a) {
return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
119 template<> EIGEN_STRONG_INLINE Packet2d plset<double>(
const double& a) {
return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
120 template<> EIGEN_STRONG_INLINE Packet4i plset<int>(
const int& a) {
return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
122 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_add_ps(a,b); }
123 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_add_pd(a,b); }
124 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_add_epi32(a,b); }
126 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_sub_ps(a,b); }
127 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_sub_pd(a,b); }
128 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_sub_epi32(a,b); }
130 template<> EIGEN_STRONG_INLINE Packet4f pnegate(
const Packet4f& a)
132 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
133 return _mm_xor_ps(a,mask);
135 template<> EIGEN_STRONG_INLINE Packet2d pnegate(
const Packet2d& a)
137 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
138 return _mm_xor_pd(a,mask);
140 template<> EIGEN_STRONG_INLINE Packet4i pnegate(
const Packet4i& a)
142 return psub(_mm_setr_epi32(0,0,0,0), a);
145 template<> EIGEN_STRONG_INLINE Packet4f pconj(
const Packet4f& a) {
return a; }
146 template<> EIGEN_STRONG_INLINE Packet2d pconj(
const Packet2d& a) {
return a; }
147 template<> EIGEN_STRONG_INLINE Packet4i pconj(
const Packet4i& a) {
return a; }
149 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_mul_ps(a,b); }
150 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_mul_pd(a,b); }
151 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(
const Packet4i& a,
const Packet4i& b)
153 #ifdef EIGEN_VECTORIZE_SSE4_1 154 return _mm_mullo_epi32(a,b);
157 return vec4i_swizzle1(
160 _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
161 vec4i_swizzle1(b,1,0,3,2)),
167 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_div_ps(a,b); }
168 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_div_pd(a,b); }
169 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(
const Packet4i& ,
const Packet4i& )
170 { eigen_assert(
false &&
"packet integer division are not supported by SSE");
171 return pset1<Packet4i>(0);
175 template<> EIGEN_STRONG_INLINE Packet4i pmadd(
const Packet4i& a,
const Packet4i& b,
const Packet4i& c) {
return padd(pmul(a,b), c); }
177 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_min_ps(a,b); }
178 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_min_pd(a,b); }
179 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(
const Packet4i& a,
const Packet4i& b)
181 #ifdef EIGEN_VECTORIZE_SSE4_1 182 return _mm_min_epi32(a,b);
185 Packet4i mask = _mm_cmplt_epi32(a,b);
186 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
190 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_max_ps(a,b); }
191 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_max_pd(a,b); }
192 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(
const Packet4i& a,
const Packet4i& b)
194 #ifdef EIGEN_VECTORIZE_SSE4_1 195 return _mm_max_epi32(a,b);
198 Packet4i mask = _mm_cmpgt_epi32(a,b);
199 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
203 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_and_ps(a,b); }
204 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_and_pd(a,b); }
205 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_and_si128(a,b); }
207 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_or_ps(a,b); }
208 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_or_pd(a,b); }
209 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_or_si128(a,b); }
211 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_xor_ps(a,b); }
212 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_xor_pd(a,b); }
213 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_xor_si128(a,b); }
215 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(
const Packet4f& a,
const Packet4f& b) {
return _mm_andnot_ps(a,b); }
216 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(
const Packet2d& a,
const Packet2d& b) {
return _mm_andnot_pd(a,b); }
217 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(
const Packet4i& a,
const Packet4i& b) {
return _mm_andnot_si128(a,b); }
219 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(
const float* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_ps(from); }
220 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(
const double* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_pd(from); }
221 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(
const int* from) { EIGEN_DEBUG_ALIGNED_LOAD
return _mm_load_si128(reinterpret_cast<const Packet4i*>(from)); }
223 #if defined(_MSC_VER) 224 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(
const float* from) {
225 EIGEN_DEBUG_UNALIGNED_LOAD
231 __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (
const __m64*)(from));
232 res = _mm_loadh_pi(res, (
const __m64*)(from+2));
235 return _mm_loadu_ps(from);
241 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(
const float* from)
243 EIGEN_DEBUG_UNALIGNED_LOAD
244 return _mm_loadu_ps(from);
248 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(
const double* from)
250 EIGEN_DEBUG_UNALIGNED_LOAD
251 return _mm_loadu_pd(from);
253 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(
const int* from)
255 EIGEN_DEBUG_UNALIGNED_LOAD
256 return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
260 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(
const float* from)
262 return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
264 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(
const double* from)
265 {
return pset1<Packet2d>(from[0]); }
266 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(
const int* from)
269 tmp = _mm_loadl_epi64(reinterpret_cast<const Packet4i*>(from));
270 return vec4i_swizzle1(tmp, 0, 0, 1, 1);
273 template<> EIGEN_STRONG_INLINE
void pstore<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
274 template<> EIGEN_STRONG_INLINE
void pstore<double>(
double* to,
const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
275 template<> EIGEN_STRONG_INLINE
void pstore<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<Packet4i*>(to), from); }
277 template<> EIGEN_STRONG_INLINE
void pstoreu<double>(
double* to,
const Packet2d& from) {
278 EIGEN_DEBUG_UNALIGNED_STORE
279 _mm_storel_pd((to), from);
280 _mm_storeh_pd((to+1), from);
282 template<> EIGEN_STRONG_INLINE
void pstoreu<float>(
float* to,
const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); }
283 template<> EIGEN_STRONG_INLINE
void pstoreu<int>(
int* to,
const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); }
286 template<> EIGEN_STRONG_INLINE
void pstore1<Packet4f>(
float* to,
const float& a)
288 Packet4f pa = _mm_set_ss(a);
289 pstore(to, vec4f_swizzle1(pa,0,0,0,0));
292 template<> EIGEN_STRONG_INLINE
void pstore1<Packet2d>(
double* to,
const double& a)
294 Packet2d pa = _mm_set_sd(a);
295 pstore(to, vec2d_swizzle1(pa,0,0));
298 template<> EIGEN_STRONG_INLINE
void prefetch<float>(
const float* addr) { _mm_prefetch((
const char*)(addr), _MM_HINT_T0); }
299 template<> EIGEN_STRONG_INLINE
void prefetch<double>(
const double* addr) { _mm_prefetch((
const char*)(addr), _MM_HINT_T0); }
300 template<> EIGEN_STRONG_INLINE
void prefetch<int>(
const int* addr) { _mm_prefetch((
const char*)(addr), _MM_HINT_T0); }
302 #if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER) 305 template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
return a.m128_f32[0]; }
306 template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
return a.m128d_f64[0]; }
307 template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
int x = _mm_cvtsi128_si32(a);
return x; }
308 #elif defined(_MSC_VER) && !defined(__INTEL_COMPILER) 310 template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
float x = _mm_cvtss_f32(a);
return x; }
311 template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
double x = _mm_cvtsd_f64(a);
return x; }
312 template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
int x = _mm_cvtsi128_si32(a);
return x; }
314 template<> EIGEN_STRONG_INLINE
float pfirst<Packet4f>(
const Packet4f& a) {
return _mm_cvtss_f32(a); }
315 template<> EIGEN_STRONG_INLINE
double pfirst<Packet2d>(
const Packet2d& a) {
return _mm_cvtsd_f64(a); }
316 template<> EIGEN_STRONG_INLINE
int pfirst<Packet4i>(
const Packet4i& a) {
return _mm_cvtsi128_si32(a); }
319 template<> EIGEN_STRONG_INLINE Packet4f preverse(
const Packet4f& a)
320 {
return _mm_shuffle_ps(a,a,0x1B); }
321 template<> EIGEN_STRONG_INLINE Packet2d preverse(
const Packet2d& a)
322 {
return _mm_shuffle_pd(a,a,0x1); }
323 template<> EIGEN_STRONG_INLINE Packet4i preverse(
const Packet4i& a)
324 {
return _mm_shuffle_epi32(a,0x1B); }
327 template<> EIGEN_STRONG_INLINE Packet4f pabs(
const Packet4f& a)
329 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
330 return _mm_and_ps(a,mask);
332 template<> EIGEN_STRONG_INLINE Packet2d pabs(
const Packet2d& a)
334 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
335 return _mm_and_pd(a,mask);
337 template<> EIGEN_STRONG_INLINE Packet4i pabs(
const Packet4i& a)
339 #ifdef EIGEN_VECTORIZE_SSSE3 340 return _mm_abs_epi32(a);
342 Packet4i aux = _mm_srai_epi32(a,31);
343 return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
347 EIGEN_STRONG_INLINE
void punpackp(Packet4f* vecs)
349 vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
350 vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
351 vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
352 vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
355 #ifdef EIGEN_VECTORIZE_SSE3 357 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(
const Packet4f* vecs)
359 return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
361 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(
const Packet2d* vecs)
363 return _mm_hadd_pd(vecs[0], vecs[1]);
371 template<> EIGEN_STRONG_INLINE
float predux<Packet4f>(
const Packet4f& a)
373 Packet4f tmp0 = _mm_hadd_ps(a,a);
374 return pfirst(_mm_hadd_ps(tmp0, tmp0));
377 template<> EIGEN_STRONG_INLINE
double predux<Packet2d>(
const Packet2d& a) {
return pfirst(_mm_hadd_pd(a, a)); }
387 template<> EIGEN_STRONG_INLINE
float predux<Packet4f>(
const Packet4f& a)
389 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
390 return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
392 template<> EIGEN_STRONG_INLINE
double predux<Packet2d>(
const Packet2d& a)
394 return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
397 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(
const Packet4f* vecs)
399 Packet4f tmp0, tmp1, tmp2;
400 tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
401 tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
402 tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
403 tmp0 = _mm_add_ps(tmp0, tmp1);
404 tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
405 tmp1 = _mm_add_ps(tmp1, tmp2);
406 tmp2 = _mm_movehl_ps(tmp1, tmp0);
407 tmp0 = _mm_movelh_ps(tmp0, tmp1);
408 return _mm_add_ps(tmp0, tmp2);
411 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(
const Packet2d* vecs)
413 return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
417 template<> EIGEN_STRONG_INLINE
int predux<Packet4i>(
const Packet4i& a)
419 Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
420 return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1));
423 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(
const Packet4i* vecs)
425 Packet4i tmp0, tmp1, tmp2;
426 tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
427 tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
428 tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
429 tmp0 = _mm_add_epi32(tmp0, tmp1);
430 tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
431 tmp1 = _mm_add_epi32(tmp1, tmp2);
432 tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
433 tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
434 return _mm_add_epi32(tmp0, tmp2);
440 template<> EIGEN_STRONG_INLINE
float predux_mul<Packet4f>(
const Packet4f& a)
442 Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
443 return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
445 template<> EIGEN_STRONG_INLINE
double predux_mul<Packet2d>(
const Packet2d& a)
447 return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
449 template<> EIGEN_STRONG_INLINE
int predux_mul<Packet4i>(
const Packet4i& a)
454 EIGEN_ALIGN16
int aux[4];
456 return (aux[0] * aux[1]) * (aux[2] * aux[3]);;
460 template<> EIGEN_STRONG_INLINE
float predux_min<Packet4f>(
const Packet4f& a)
462 Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
463 return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
465 template<> EIGEN_STRONG_INLINE
double predux_min<Packet2d>(
const Packet2d& a)
467 return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
469 template<> EIGEN_STRONG_INLINE
int predux_min<Packet4i>(
const Packet4i& a)
473 EIGEN_ALIGN16
int aux[4];
475 int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
476 int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
477 return aux0<aux2 ? aux0 : aux2;
481 template<> EIGEN_STRONG_INLINE
float predux_max<Packet4f>(
const Packet4f& a)
483 Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
484 return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
486 template<> EIGEN_STRONG_INLINE
double predux_max<Packet2d>(
const Packet2d& a)
488 return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
490 template<> EIGEN_STRONG_INLINE
int predux_max<Packet4i>(
const Packet4i& a)
494 EIGEN_ALIGN16
int aux[4];
496 int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
497 int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
498 return aux0>aux2 ? aux0 : aux2;
501 #if (defined __GNUC__) 516 #ifdef EIGEN_VECTORIZE_SSSE3 519 struct palign_impl<Offset,Packet4f>
521 static EIGEN_STRONG_INLINE
void run(Packet4f& first,
const Packet4f& second)
524 first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
529 struct palign_impl<Offset,Packet4i>
531 static EIGEN_STRONG_INLINE
void run(Packet4i& first,
const Packet4i& second)
534 first = _mm_alignr_epi8(second,first, Offset*4);
539 struct palign_impl<Offset,Packet2d>
541 static EIGEN_STRONG_INLINE
void run(Packet2d& first,
const Packet2d& second)
544 first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
550 struct palign_impl<Offset,Packet4f>
552 static EIGEN_STRONG_INLINE
void run(Packet4f& first,
const Packet4f& second)
556 first = _mm_move_ss(first,second);
557 first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39));
561 first = _mm_movehl_ps(first,first);
562 first = _mm_movelh_ps(first,second);
566 first = _mm_move_ss(first,second);
567 first = _mm_shuffle_ps(first,second,0x93);
573 struct palign_impl<Offset,Packet4i>
575 static EIGEN_STRONG_INLINE
void run(Packet4i& first,
const Packet4i& second)
579 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
580 first = _mm_shuffle_epi32(first,0x39);
584 first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first)));
585 first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
589 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
590 first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93));
596 struct palign_impl<Offset,Packet2d>
598 static EIGEN_STRONG_INLINE
void run(Packet2d& first,
const Packet2d& second)
602 first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first)));
603 first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second)));
613 #endif // EIGEN_PACKET_MATH_SSE_H
Definition: Eigen_Colamd.h:50